1// SPDX-License-Identifier: GPL-2.0-only
    2/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
    3 * Copyright (c) 2016 Facebook
    4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
    5 */
    6#include <uapi/linux/btf.h>
    7#include <linux/kernel.h>
    8#include <linux/types.h>
    9#include <linux/slab.h>
   10#include <linux/bpf.h>
   11#include <linux/btf.h>
   12#include <linux/bpf_verifier.h>
   13#include <linux/filter.h>
   14#include <net/netlink.h>
   15#include <linux/file.h>
   16#include <linux/vmalloc.h>
   17#include <linux/stringify.h>
   18#include <linux/bsearch.h>
   19#include <linux/sort.h>
   20#include <linux/perf_event.h>
   21#include <linux/ctype.h>
   22#include <linux/error-injection.h>
   23#include <linux/bpf_lsm.h>
   24#include <linux/btf_ids.h>
   25
   26#include "disasm.h"
   27
   28static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
   29#define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
   30	[_id] = & _name ## _verifier_ops,
   31#define BPF_MAP_TYPE(_id, _ops)
   32#define BPF_LINK_TYPE(_id, _name)
   33#include <linux/bpf_types.h>
   34#undef BPF_PROG_TYPE
   35#undef BPF_MAP_TYPE
   36#undef BPF_LINK_TYPE
   37};
   38
   39/* bpf_check() is a static code analyzer that walks eBPF program
   40 * instruction by instruction and updates register/stack state.
   41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
   42 *
   43 * The first pass is depth-first-search to check that the program is a DAG.
   44 * It rejects the following programs:
   45 * - larger than BPF_MAXINSNS insns
   46 * - if loop is present (detected via back-edge)
   47 * - unreachable insns exist (shouldn't be a forest. program = one function)
   48 * - out of bounds or malformed jumps
   49 * The second pass is all possible path descent from the 1st insn.
   50 * Since it's analyzing all paths through the program, the length of the
   51 * analysis is limited to 64k insn, which may be hit even if total number of
   52 * insn is less then 4K, but there are too many branches that change stack/regs.
   53 * Number of 'branches to be analyzed' is limited to 1k
   54 *
   55 * On entry to each instruction, each register has a type, and the instruction
   56 * changes the types of the registers depending on instruction semantics.
   57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
   58 * copied to R1.
   59 *
   60 * All registers are 64-bit.
   61 * R0 - return register
   62 * R1-R5 argument passing registers
   63 * R6-R9 callee saved registers
   64 * R10 - frame pointer read-only
   65 *
   66 * At the start of BPF program the register R1 contains a pointer to bpf_context
   67 * and has type PTR_TO_CTX.
   68 *
   69 * Verifier tracks arithmetic operations on pointers in case:
   70 *    BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
   71 *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
   72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
   73 * and 2nd arithmetic instruction is pattern matched to recognize
   74 * that it wants to construct a pointer to some element within stack.
   75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
   76 * (and -20 constant is saved for further stack bounds checking).
   77 * Meaning that this reg is a pointer to stack plus known immediate constant.
   78 *
   79 * Most of the time the registers have SCALAR_VALUE type, which
   80 * means the register has some value, but it's not a valid pointer.
   81 * (like pointer plus pointer becomes SCALAR_VALUE type)
   82 *
   83 * When verifier sees load or store instructions the type of base register
   84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
   85 * four pointer types recognized by check_mem_access() function.
   86 *
   87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
   88 * and the range of [ptr, ptr + map's value_size) is accessible.
   89 *
   90 * registers used to pass values to function calls are checked against
   91 * function argument constraints.
   92 *
   93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
   94 * It means that the register type passed to this function must be
   95 * PTR_TO_STACK and it will be used inside the function as
   96 * 'pointer to map element key'
   97 *
   98 * For example the argument constraints for bpf_map_lookup_elem():
   99 *   .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
  100 *   .arg1_type = ARG_CONST_MAP_PTR,
  101 *   .arg2_type = ARG_PTR_TO_MAP_KEY,
  102 *
  103 * ret_type says that this function returns 'pointer to map elem value or null'
  104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
  105 * 2nd argument should be a pointer to stack, which will be used inside
  106 * the helper function as a pointer to map element key.
  107 *
  108 * On the kernel side the helper function looks like:
  109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
  110 * {
  111 *    struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
  112 *    void *key = (void *) (unsigned long) r2;
  113 *    void *value;
  114 *
  115 *    here kernel can access 'key' and 'map' pointers safely, knowing that
  116 *    [key, key + map->key_size) bytes are valid and were initialized on
  117 *    the stack of eBPF program.
  118 * }
  119 *
  120 * Corresponding eBPF program may look like:
  121 *    BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),  // after this insn R2 type is FRAME_PTR
  122 *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
  123 *    BPF_LD_MAP_FD(BPF_REG_1, map_fd),      // after this insn R1 type is CONST_PTR_TO_MAP
  124 *    BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
  125 * here verifier looks at prototype of map_lookup_elem() and sees:
  126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
  127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
  128 *
  129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
  130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
  131 * and were initialized prior to this call.
  132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
  133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
  134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
  135 * returns either pointer to map value or NULL.
  136 *
  137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
  138 * insn, the register holding that pointer in the true branch changes state to
  139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
  140 * branch. See check_cond_jmp_op().
  141 *
  142 * After the call R0 is set to return type of the function and registers R1-R5
  143 * are set to NOT_INIT to indicate that they are no longer readable.
  144 *
  145 * The following reference types represent a potential reference to a kernel
  146 * resource which, after first being allocated, must be checked and freed by
  147 * the BPF program:
  148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
  149 *
  150 * When the verifier sees a helper call return a reference type, it allocates a
  151 * pointer id for the reference and stores it in the current function state.
  152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
  153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
  154 * passes through a NULL-check conditional. For the branch wherein the state is
  155 * changed to CONST_IMM, the verifier releases the reference.
  156 *
  157 * For each helper function that allocates a reference, such as
  158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
  159 * bpf_sk_release(). When a reference type passes into the release function,
  160 * the verifier also releases the reference. If any unchecked or unreleased
  161 * reference remains at the end of the program, the verifier rejects it.
  162 */
  163
  164/* verifier_state + insn_idx are pushed to stack when branch is encountered */
  165struct bpf_verifier_stack_elem {
  166	/* verifer state is 'st'
  167	 * before processing instruction 'insn_idx'
  168	 * and after processing instruction 'prev_insn_idx'
  169	 */
  170	struct bpf_verifier_state st;
  171	int insn_idx;
  172	int prev_insn_idx;
  173	struct bpf_verifier_stack_elem *next;
  174	/* length of verifier log at the time this state was pushed on stack */
  175	u32 log_pos;
  176};
  177
  178#define BPF_COMPLEXITY_LIMIT_JMP_SEQ	8192
  179#define BPF_COMPLEXITY_LIMIT_STATES	64
  180
  181#define BPF_MAP_KEY_POISON	(1ULL << 63)
  182#define BPF_MAP_KEY_SEEN	(1ULL << 62)
  183
  184#define BPF_MAP_PTR_UNPRIV	1UL
  185#define BPF_MAP_PTR_POISON	((void *)((0xeB9FUL << 1) +	\
  186					  POISON_POINTER_DELTA))
  187#define BPF_MAP_PTR(X)		((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
  188
  189static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
  190{
  191	return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
  192}
  193
  194static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
  195{
  196	return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
  197}
  198
  199static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
  200			      const struct bpf_map *map, bool unpriv)
  201{
  202	BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
  203	unpriv |= bpf_map_ptr_unpriv(aux);
  204	aux->map_ptr_state = (unsigned long)map |
  205			     (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
  206}
  207
  208static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
  209{
  210	return aux->map_key_state & BPF_MAP_KEY_POISON;
  211}
  212
  213static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
  214{
  215	return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
  216}
  217
  218static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
  219{
  220	return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
  221}
  222
  223static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
  224{
  225	bool poisoned = bpf_map_key_poisoned(aux);
  226
  227	aux->map_key_state = state | BPF_MAP_KEY_SEEN |
  228			     (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
  229}
  230
  231static bool bpf_pseudo_call(const struct bpf_insn *insn)
  232{
  233	return insn->code == (BPF_JMP | BPF_CALL) &&
  234	       insn->src_reg == BPF_PSEUDO_CALL;
  235}
  236
  237static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
  238{
  239	return insn->code == (BPF_JMP | BPF_CALL) &&
  240	       insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
  241}
  242
  243static bool bpf_pseudo_func(const struct bpf_insn *insn)
  244{
  245	return insn->code == (BPF_LD | BPF_IMM | BPF_DW) &&
  246	       insn->src_reg == BPF_PSEUDO_FUNC;
  247}
  248
  249struct bpf_call_arg_meta {
  250	struct bpf_map *map_ptr;
  251	bool raw_mode;
  252	bool pkt_access;
  253	int regno;
  254	int access_size;
  255	int mem_size;
  256	u64 msize_max_value;
  257	int ref_obj_id;
  258	int func_id;
  259	struct btf *btf;
  260	u32 btf_id;
  261	struct btf *ret_btf;
  262	u32 ret_btf_id;
  263	u32 subprogno;
  264};
  265
  266struct btf *btf_vmlinux;
  267
  268static DEFINE_MUTEX(bpf_verifier_lock);
  269
  270static const struct bpf_line_info *
  271find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
  272{
  273	const struct bpf_line_info *linfo;
  274	const struct bpf_prog *prog;
  275	u32 i, nr_linfo;
  276
  277	prog = env->prog;
  278	nr_linfo = prog->aux->nr_linfo;
  279
  280	if (!nr_linfo || insn_off >= prog->len)
  281		return NULL;
  282
  283	linfo = prog->aux->linfo;
  284	for (i = 1; i < nr_linfo; i++)
  285		if (insn_off < linfo[i].insn_off)
  286			break;
  287
  288	return &linfo[i - 1];
  289}
  290
  291void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
  292		       va_list args)
  293{
  294	unsigned int n;
  295
  296	n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
  297
  298	WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
  299		  "verifier log line truncated - local buffer too short\n");
  300
  301	n = min(log->len_total - log->len_used - 1, n);
  302	log->kbuf[n] = '\0';
  303
  304	if (log->level == BPF_LOG_KERNEL) {
  305		pr_err("BPF:%s\n", log->kbuf);
  306		return;
  307	}
  308	if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
  309		log->len_used += n;
  310	else
  311		log->ubuf = NULL;
  312}
  313
  314static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
  315{
  316	char zero = 0;
  317
  318	if (!bpf_verifier_log_needed(log))
  319		return;
  320
  321	log->len_used = new_pos;
  322	if (put_user(zero, log->ubuf + new_pos))
  323		log->ubuf = NULL;
  324}
  325
  326/* log_level controls verbosity level of eBPF verifier.
  327 * bpf_verifier_log_write() is used to dump the verification trace to the log,
  328 * so the user can figure out what's wrong with the program
  329 */
  330__printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
  331					   const char *fmt, ...)
  332{
  333	va_list args;
  334
  335	if (!bpf_verifier_log_needed(&env->log))
  336		return;
  337
  338	va_start(args, fmt);
  339	bpf_verifier_vlog(&env->log, fmt, args);
  340	va_end(args);
  341}
  342EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
  343
  344__printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
  345{
  346	struct bpf_verifier_env *env = private_data;
  347	va_list args;
  348
  349	if (!bpf_verifier_log_needed(&env->log))
  350		return;
  351
  352	va_start(args, fmt);
  353	bpf_verifier_vlog(&env->log, fmt, args);
  354	va_end(args);
  355}
  356
  357__printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
  358			    const char *fmt, ...)
  359{
  360	va_list args;
  361
  362	if (!bpf_verifier_log_needed(log))
  363		return;
  364
  365	va_start(args, fmt);
  366	bpf_verifier_vlog(log, fmt, args);
  367	va_end(args);
  368}
  369
  370static const char *ltrim(const char *s)
  371{
  372	while (isspace(*s))
  373		s++;
  374
  375	return s;
  376}
  377
  378__printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
  379					 u32 insn_off,
  380					 const char *prefix_fmt, ...)
  381{
  382	const struct bpf_line_info *linfo;
  383
  384	if (!bpf_verifier_log_needed(&env->log))
  385		return;
  386
  387	linfo = find_linfo(env, insn_off);
  388	if (!linfo || linfo == env->prev_linfo)
  389		return;
  390
  391	if (prefix_fmt) {
  392		va_list args;
  393
  394		va_start(args, prefix_fmt);
  395		bpf_verifier_vlog(&env->log, prefix_fmt, args);
  396		va_end(args);
  397	}
  398
  399	verbose(env, "%s\n",
  400		ltrim(btf_name_by_offset(env->prog->aux->btf,
  401					 linfo->line_off)));
  402
  403	env->prev_linfo = linfo;
  404}
  405
  406static void verbose_invalid_scalar(struct bpf_verifier_env *env,
  407				   struct bpf_reg_state *reg,
  408				   struct tnum *range, const char *ctx,
  409				   const char *reg_name)
  410{
  411	char tn_buf[48];
  412
  413	verbose(env, "At %s the register %s ", ctx, reg_name);
  414	if (!tnum_is_unknown(reg->var_off)) {
  415		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
  416		verbose(env, "has value %s", tn_buf);
  417	} else {
  418		verbose(env, "has unknown scalar value");
  419	}
  420	tnum_strn(tn_buf, sizeof(tn_buf), *range);
  421	verbose(env, " should have been in %s\n", tn_buf);
  422}
  423
  424static bool type_is_pkt_pointer(enum bpf_reg_type type)
  425{
  426	return type == PTR_TO_PACKET ||
  427	       type == PTR_TO_PACKET_META;
  428}
  429
  430static bool type_is_sk_pointer(enum bpf_reg_type type)
  431{
  432	return type == PTR_TO_SOCKET ||
  433		type == PTR_TO_SOCK_COMMON ||
  434		type == PTR_TO_TCP_SOCK ||
  435		type == PTR_TO_XDP_SOCK;
  436}
  437
  438static bool reg_type_not_null(enum bpf_reg_type type)
  439{
  440	return type == PTR_TO_SOCKET ||
  441		type == PTR_TO_TCP_SOCK ||
  442		type == PTR_TO_MAP_VALUE ||
  443		type == PTR_TO_MAP_KEY ||
  444		type == PTR_TO_SOCK_COMMON;
  445}
  446
  447static bool reg_type_may_be_null(enum bpf_reg_type type)
  448{
  449	return type == PTR_TO_MAP_VALUE_OR_NULL ||
  450	       type == PTR_TO_SOCKET_OR_NULL ||
  451	       type == PTR_TO_SOCK_COMMON_OR_NULL ||
  452	       type == PTR_TO_TCP_SOCK_OR_NULL ||
  453	       type == PTR_TO_BTF_ID_OR_NULL ||
  454	       type == PTR_TO_MEM_OR_NULL ||
  455	       type == PTR_TO_RDONLY_BUF_OR_NULL ||
  456	       type == PTR_TO_RDWR_BUF_OR_NULL;
  457}
  458
  459static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
  460{
  461	return reg->type == PTR_TO_MAP_VALUE &&
  462		map_value_has_spin_lock(reg->map_ptr);
  463}
  464
  465static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
  466{
  467	return type == PTR_TO_SOCKET ||
  468		type == PTR_TO_SOCKET_OR_NULL ||
  469		type == PTR_TO_TCP_SOCK ||
  470		type == PTR_TO_TCP_SOCK_OR_NULL ||
  471		type == PTR_TO_MEM ||
  472		type == PTR_TO_MEM_OR_NULL;
  473}
  474
  475static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
  476{
  477	return type == ARG_PTR_TO_SOCK_COMMON;
  478}
  479
  480static bool arg_type_may_be_null(enum bpf_arg_type type)
  481{
  482	return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
  483	       type == ARG_PTR_TO_MEM_OR_NULL ||
  484	       type == ARG_PTR_TO_CTX_OR_NULL ||
  485	       type == ARG_PTR_TO_SOCKET_OR_NULL ||
  486	       type == ARG_PTR_TO_ALLOC_MEM_OR_NULL ||
  487	       type == ARG_PTR_TO_STACK_OR_NULL;
  488}
  489
  490/* Determine whether the function releases some resources allocated by another
  491 * function call. The first reference type argument will be assumed to be
  492 * released by release_reference().
  493 */
  494static bool is_release_function(enum bpf_func_id func_id)
  495{
  496	return func_id == BPF_FUNC_sk_release ||
  497	       func_id == BPF_FUNC_ringbuf_submit ||
  498	       func_id == BPF_FUNC_ringbuf_discard;
  499}
  500
  501static bool may_be_acquire_function(enum bpf_func_id func_id)
  502{
  503	return func_id == BPF_FUNC_sk_lookup_tcp ||
  504		func_id == BPF_FUNC_sk_lookup_udp ||
  505		func_id == BPF_FUNC_skc_lookup_tcp ||
  506		func_id == BPF_FUNC_map_lookup_elem ||
  507	        func_id == BPF_FUNC_ringbuf_reserve;
  508}
  509
  510static bool is_acquire_function(enum bpf_func_id func_id,
  511				const struct bpf_map *map)
  512{
  513	enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
  514
  515	if (func_id == BPF_FUNC_sk_lookup_tcp ||
  516	    func_id == BPF_FUNC_sk_lookup_udp ||
  517	    func_id == BPF_FUNC_skc_lookup_tcp ||
  518	    func_id == BPF_FUNC_ringbuf_reserve)
  519		return true;
  520
  521	if (func_id == BPF_FUNC_map_lookup_elem &&
  522	    (map_type == BPF_MAP_TYPE_SOCKMAP ||
  523	     map_type == BPF_MAP_TYPE_SOCKHASH))
  524		return true;
  525
  526	return false;
  527}
  528
  529static bool is_ptr_cast_function(enum bpf_func_id func_id)
  530{
  531	return func_id == BPF_FUNC_tcp_sock ||
  532		func_id == BPF_FUNC_sk_fullsock ||
  533		func_id == BPF_FUNC_skc_to_tcp_sock ||
  534		func_id == BPF_FUNC_skc_to_tcp6_sock ||
  535		func_id == BPF_FUNC_skc_to_udp6_sock ||
  536		func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
  537		func_id == BPF_FUNC_skc_to_tcp_request_sock;
  538}
  539
  540static bool is_cmpxchg_insn(const struct bpf_insn *insn)
  541{
  542	return BPF_CLASS(insn->code) == BPF_STX &&
  543	       BPF_MODE(insn->code) == BPF_ATOMIC &&
  544	       insn->imm == BPF_CMPXCHG;
  545}
  546
  547/* string representation of 'enum bpf_reg_type' */
  548static const char * const reg_type_str[] = {
  549	[NOT_INIT]		= "?",
  550	[SCALAR_VALUE]		= "inv",
  551	[PTR_TO_CTX]		= "ctx",
  552	[CONST_PTR_TO_MAP]	= "map_ptr",
  553	[PTR_TO_MAP_VALUE]	= "map_value",
  554	[PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
  555	[PTR_TO_STACK]		= "fp",
  556	[PTR_TO_PACKET]		= "pkt",
  557	[PTR_TO_PACKET_META]	= "pkt_meta",
  558	[PTR_TO_PACKET_END]	= "pkt_end",
  559	[PTR_TO_FLOW_KEYS]	= "flow_keys",
  560	[PTR_TO_SOCKET]		= "sock",
  561	[PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
  562	[PTR_TO_SOCK_COMMON]	= "sock_common",
  563	[PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
  564	[PTR_TO_TCP_SOCK]	= "tcp_sock",
  565	[PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
  566	[PTR_TO_TP_BUFFER]	= "tp_buffer",
  567	[PTR_TO_XDP_SOCK]	= "xdp_sock",
  568	[PTR_TO_BTF_ID]		= "ptr_",
  569	[PTR_TO_BTF_ID_OR_NULL]	= "ptr_or_null_",
  570	[PTR_TO_PERCPU_BTF_ID]	= "percpu_ptr_",
  571	[PTR_TO_MEM]		= "mem",
  572	[PTR_TO_MEM_OR_NULL]	= "mem_or_null",
  573	[PTR_TO_RDONLY_BUF]	= "rdonly_buf",
  574	[PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
  575	[PTR_TO_RDWR_BUF]	= "rdwr_buf",
  576	[PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
  577	[PTR_TO_FUNC]		= "func",
  578	[PTR_TO_MAP_KEY]	= "map_key",
  579};
  580
  581static char slot_type_char[] = {
  582	[STACK_INVALID]	= '?',
  583	[STACK_SPILL]	= 'r',
  584	[STACK_MISC]	= 'm',
  585	[STACK_ZERO]	= '0',
  586};
  587
  588static void print_liveness(struct bpf_verifier_env *env,
  589			   enum bpf_reg_liveness live)
  590{
  591	if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
  592	    verbose(env, "_");
  593	if (live & REG_LIVE_READ)
  594		verbose(env, "r");
  595	if (live & REG_LIVE_WRITTEN)
  596		verbose(env, "w");
  597	if (live & REG_LIVE_DONE)
  598		verbose(env, "D");
  599}
  600
  601static struct bpf_func_state *func(struct bpf_verifier_env *env,
  602				   const struct bpf_reg_state *reg)
  603{
  604	struct bpf_verifier_state *cur = env->cur_state;
  605
  606	return cur->frame[reg->frameno];
  607}
  608
  609static const char *kernel_type_name(const struct btf* btf, u32 id)
  610{
  611	return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
  612}
  613
  614static void print_verifier_state(struct bpf_verifier_env *env,
  615				 const struct bpf_func_state *state)
  616{
  617	const struct bpf_reg_state *reg;
  618	enum bpf_reg_type t;
  619	int i;
  620
  621	if (state->frameno)
  622		verbose(env, " frame%d:", state->frameno);
  623	for (i = 0; i < MAX_BPF_REG; i++) {
  624		reg = &state->regs[i];
  625		t = reg->type;
  626		if (t == NOT_INIT)
  627			continue;
  628		verbose(env, " R%d", i);
  629		print_liveness(env, reg->live);
  630		verbose(env, "=%s", reg_type_str[t]);
  631		if (t == SCALAR_VALUE && reg->precise)
  632			verbose(env, "P");
  633		if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
  634		    tnum_is_const(reg->var_off)) {
  635			/* reg->off should be 0 for SCALAR_VALUE */
  636			verbose(env, "%lld", reg->var_off.value + reg->off);
  637		} else {
  638			if (t == PTR_TO_BTF_ID ||
  639			    t == PTR_TO_BTF_ID_OR_NULL ||
  640			    t == PTR_TO_PERCPU_BTF_ID)
  641				verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
  642			verbose(env, "(id=%d", reg->id);
  643			if (reg_type_may_be_refcounted_or_null(t))
  644				verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
  645			if (t != SCALAR_VALUE)
  646				verbose(env, ",off=%d", reg->off);
  647			if (type_is_pkt_pointer(t))
  648				verbose(env, ",r=%d", reg->range);
  649			else if (t == CONST_PTR_TO_MAP ||
  650				 t == PTR_TO_MAP_KEY ||
  651				 t == PTR_TO_MAP_VALUE ||
  652				 t == PTR_TO_MAP_VALUE_OR_NULL)
  653				verbose(env, ",ks=%d,vs=%d",
  654					reg->map_ptr->key_size,
  655					reg->map_ptr->value_size);
  656			if (tnum_is_const(reg->var_off)) {
  657				/* Typically an immediate SCALAR_VALUE, but
  658				 * could be a pointer whose offset is too big
  659				 * for reg->off
  660				 */
  661				verbose(env, ",imm=%llx", reg->var_off.value);
  662			} else {
  663				if (reg->smin_value != reg->umin_value &&
  664				    reg->smin_value != S64_MIN)
  665					verbose(env, ",smin_value=%lld",
  666						(long long)reg->smin_value);
  667				if (reg->smax_value != reg->umax_value &&
  668				    reg->smax_value != S64_MAX)
  669					verbose(env, ",smax_value=%lld",
  670						(long long)reg->smax_value);
  671				if (reg->umin_value != 0)
  672					verbose(env, ",umin_value=%llu",
  673						(unsigned long long)reg->umin_value);
  674				if (reg->umax_value != U64_MAX)
  675					verbose(env, ",umax_value=%llu",
  676						(unsigned long long)reg->umax_value);
  677				if (!tnum_is_unknown(reg->var_off)) {
  678					char tn_buf[48];
  679
  680					tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
  681					verbose(env, ",var_off=%s", tn_buf);
  682				}
  683				if (reg->s32_min_value != reg->smin_value &&
  684				    reg->s32_min_value != S32_MIN)
  685					verbose(env, ",s32_min_value=%d",
  686						(int)(reg->s32_min_value));
  687				if (reg->s32_max_value != reg->smax_value &&
  688				    reg->s32_max_value != S32_MAX)
  689					verbose(env, ",s32_max_value=%d",
  690						(int)(reg->s32_max_value));
  691				if (reg->u32_min_value != reg->umin_value &&
  692				    reg->u32_min_value != U32_MIN)
  693					verbose(env, ",u32_min_value=%d",
  694						(int)(reg->u32_min_value));
  695				if (reg->u32_max_value != reg->umax_value &&
  696				    reg->u32_max_value != U32_MAX)
  697					verbose(env, ",u32_max_value=%d",
  698						(int)(reg->u32_max_value));
  699			}
  700			verbose(env, ")");
  701		}
  702	}
  703	for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
  704		char types_buf[BPF_REG_SIZE + 1];
  705		bool valid = false;
  706		int j;
  707
  708		for (j = 0; j < BPF_REG_SIZE; j++) {
  709			if (state->stack[i].slot_type[j] != STACK_INVALID)
  710				valid = true;
  711			types_buf[j] = slot_type_char[
  712					state->stack[i].slot_type[j]];
  713		}
  714		types_buf[BPF_REG_SIZE] = 0;
  715		if (!valid)
  716			continue;
  717		verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
  718		print_liveness(env, state->stack[i].spilled_ptr.live);
  719		if (state->stack[i].slot_type[0] == STACK_SPILL) {
  720			reg = &state->stack[i].spilled_ptr;
  721			t = reg->type;
  722			verbose(env, "=%s", reg_type_str[t]);
  723			if (t == SCALAR_VALUE && reg->precise)
  724				verbose(env, "P");
  725			if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
  726				verbose(env, "%lld", reg->var_off.value + reg->off);
  727		} else {
  728			verbose(env, "=%s", types_buf);
  729		}
  730	}
  731	if (state->acquired_refs && state->refs[0].id) {
  732		verbose(env, " refs=%d", state->refs[0].id);
  733		for (i = 1; i < state->acquired_refs; i++)
  734			if (state->refs[i].id)
  735				verbose(env, ",%d", state->refs[i].id);
  736	}
  737	verbose(env, "\n");
  738}
  739
  740/* copy array src of length n * size bytes to dst. dst is reallocated if it's too
  741 * small to hold src. This is different from krealloc since we don't want to preserve
  742 * the contents of dst.
  743 *
  744 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
  745 * not be allocated.
  746 */
  747static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
  748{
  749	size_t bytes;
  750
  751	if (ZERO_OR_NULL_PTR(src))
  752		goto out;
  753
  754	if (unlikely(check_mul_overflow(n, size, &bytes)))
  755		return NULL;
  756
  757	if (ksize(dst) < bytes) {
  758		kfree(dst);
  759		dst = kmalloc_track_caller(bytes, flags);
  760		if (!dst)
  761			return NULL;
  762	}
  763
  764	memcpy(dst, src, bytes);
  765out:
  766	return dst ? dst : ZERO_SIZE_PTR;
  767}
  768
  769/* resize an array from old_n items to new_n items. the array is reallocated if it's too
  770 * small to hold new_n items. new items are zeroed out if the array grows.
  771 *
  772 * Contrary to krealloc_array, does not free arr if new_n is zero.
  773 */
  774static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
  775{
  776	if (!new_n || old_n == new_n)
  777		goto out;
  778
  779	arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
  780	if (!arr)
  781		return NULL;
  782
  783	if (new_n > old_n)
  784		memset(arr + old_n * size, 0, (new_n - old_n) * size);
  785
  786out:
  787	return arr ? arr : ZERO_SIZE_PTR;
  788}
  789
  790static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
  791{
  792	dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
  793			       sizeof(struct bpf_reference_state), GFP_KERNEL);
  794	if (!dst->refs)
  795		return -ENOMEM;
  796
  797	dst->acquired_refs = src->acquired_refs;
  798	return 0;
  799}
  800
  801static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
  802{
  803	size_t n = src->allocated_stack / BPF_REG_SIZE;
  804
  805	dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
  806				GFP_KERNEL);
  807	if (!dst->stack)
  808		return -ENOMEM;
  809
  810	dst->allocated_stack = src->allocated_stack;
  811	return 0;
  812}
  813
  814static int resize_reference_state(struct bpf_func_state *state, size_t n)
  815{
  816	state->refs = realloc_array(state->refs, state->acquired_refs, n,
  817				    sizeof(struct bpf_reference_state));
  818	if (!state->refs)
  819		return -ENOMEM;
  820
  821	state->acquired_refs = n;
  822	return 0;
  823}
  824
  825static int grow_stack_state(struct bpf_func_state *state, int size)
  826{
  827	size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
  828
  829	if (old_n >= n)
  830		return 0;
  831
  832	state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
  833	if (!state->stack)
  834		return -ENOMEM;
  835
  836	state->allocated_stack = size;
  837	return 0;
  838}
  839
  840/* Acquire a pointer id from the env and update the state->refs to include
  841 * this new pointer reference.
  842 * On success, returns a valid pointer id to associate with the register
  843 * On failure, returns a negative errno.
  844 */
  845static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
  846{
  847	struct bpf_func_state *state = cur_func(env);
  848	int new_ofs = state->acquired_refs;
  849	int id, err;
  850
  851	err = resize_reference_state(state, state->acquired_refs + 1);
  852	if (err)
  853		return err;
  854	id = ++env->id_gen;
  855	state->refs[new_ofs].id = id;
  856	state->refs[new_ofs].insn_idx = insn_idx;
  857
  858	return id;
  859}
  860
  861/* release function corresponding to acquire_reference_state(). Idempotent. */
  862static int release_reference_state(struct bpf_func_state *state, int ptr_id)
  863{
  864	int i, last_idx;
  865
  866	last_idx = state->acquired_refs - 1;
  867	for (i = 0; i < state->acquired_refs; i++) {
  868		if (state->refs[i].id == ptr_id) {
  869			if (last_idx && i != last_idx)
  870				memcpy(&state->refs[i], &state->refs[last_idx],
  871				       sizeof(*state->refs));
  872			memset(&state->refs[last_idx], 0, sizeof(*state->refs));
  873			state->acquired_refs--;
  874			return 0;
  875		}
  876	}
  877	return -EINVAL;
  878}
  879
  880static void free_func_state(struct bpf_func_state *state)
  881{
  882	if (!state)
  883		return;
  884	kfree(state->refs);
  885	kfree(state->stack);
  886	kfree(state);
  887}
  888
  889static void clear_jmp_history(struct bpf_verifier_state *state)
  890{
  891	kfree(state->jmp_history);
  892	state->jmp_history = NULL;
  893	state->jmp_history_cnt = 0;
  894}
  895
  896static void free_verifier_state(struct bpf_verifier_state *state,
  897				bool free_self)
  898{
  899	int i;
  900
  901	for (i = 0; i <= state->curframe; i++) {
  902		free_func_state(state->frame[i]);
  903		state->frame[i] = NULL;
  904	}
  905	clear_jmp_history(state);
  906	if (free_self)
  907		kfree(state);
  908}
  909
  910/* copy verifier state from src to dst growing dst stack space
  911 * when necessary to accommodate larger src stack
  912 */
  913static int copy_func_state(struct bpf_func_state *dst,
  914			   const struct bpf_func_state *src)
  915{
  916	int err;
  917
  918	memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
  919	err = copy_reference_state(dst, src);
  920	if (err)
  921		return err;
  922	return copy_stack_state(dst, src);
  923}
  924
  925static int copy_verifier_state(struct bpf_verifier_state *dst_state,
  926			       const struct bpf_verifier_state *src)
  927{
  928	struct bpf_func_state *dst;
  929	int i, err;
  930
  931	dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
  932					    src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
  933					    GFP_USER);
  934	if (!dst_state->jmp_history)
  935		return -ENOMEM;
  936	dst_state->jmp_history_cnt = src->jmp_history_cnt;
  937
  938	/* if dst has more stack frames then src frame, free them */
  939	for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
  940		free_func_state(dst_state->frame[i]);
  941		dst_state->frame[i] = NULL;
  942	}
  943	dst_state->speculative = src->speculative;
  944	dst_state->curframe = src->curframe;
  945	dst_state->active_spin_lock = src->active_spin_lock;
  946	dst_state->branches = src->branches;
  947	dst_state->parent = src->parent;
  948	dst_state->first_insn_idx = src->first_insn_idx;
  949	dst_state->last_insn_idx = src->last_insn_idx;
  950	for (i = 0; i <= src->curframe; i++) {
  951		dst = dst_state->frame[i];
  952		if (!dst) {
  953			dst = kzalloc(sizeof(*dst), GFP_KERNEL);
  954			if (!dst)
  955				return -ENOMEM;
  956			dst_state->frame[i] = dst;
  957		}
  958		err = copy_func_state(dst, src->frame[i]);
  959		if (err)
  960			return err;
  961	}
  962	return 0;
  963}
  964
  965static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
  966{
  967	while (st) {
  968		u32 br = --st->branches;
  969
  970		/* WARN_ON(br > 1) technically makes sense here,
  971		 * but see comment in push_stack(), hence:
  972		 */
  973		WARN_ONCE((int)br < 0,
  974			  "BUG update_branch_counts:branches_to_explore=%d\n",
  975			  br);
  976		if (br)
  977			break;
  978		st = st->parent;
  979	}
  980}
  981
  982static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
  983		     int *insn_idx, bool pop_log)
  984{
  985	struct bpf_verifier_state *cur = env->cur_state;
  986	struct bpf_verifier_stack_elem *elem, *head = env->head;
  987	int err;
  988
  989	if (env->head == NULL)
  990		return -ENOENT;
  991
  992	if (cur) {
  993		err = copy_verifier_state(cur, &head->st);
  994		if (err)
  995			return err;
  996	}
  997	if (pop_log)
  998		bpf_vlog_reset(&env->log, head->log_pos);
  999	if (insn_idx)
 1000		*insn_idx = head->insn_idx;
 1001	if (prev_insn_idx)
 1002		*prev_insn_idx = head->prev_insn_idx;
 1003	elem = head->next;
 1004	free_verifier_state(&head->st, false);
 1005	kfree(head);
 1006	env->head = elem;
 1007	env->stack_size--;
 1008	return 0;
 1009}
 1010
 1011static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
 1012					     int insn_idx, int prev_insn_idx,
 1013					     bool speculative)
 1014{
 1015	struct bpf_verifier_state *cur = env->cur_state;
 1016	struct bpf_verifier_stack_elem *elem;
 1017	int err;
 1018
 1019	elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
 1020	if (!elem)
 1021		goto err;
 1022
 1023	elem->insn_idx = insn_idx;
 1024	elem->prev_insn_idx = prev_insn_idx;
 1025	elem->next = env->head;
 1026	elem->log_pos = env->log.len_used;
 1027	env->head = elem;
 1028	env->stack_size++;
 1029	err = copy_verifier_state(&elem->st, cur);
 1030	if (err)
 1031		goto err;
 1032	elem->st.speculative |= speculative;
 1033	if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
 1034		verbose(env, "The sequence of %d jumps is too complex.\n",
 1035			env->stack_size);
 1036		goto err;
 1037	}
 1038	if (elem->st.parent) {
 1039		++elem->st.parent->branches;
 1040		/* WARN_ON(branches > 2) technically makes sense here,
 1041		 * but
 1042		 * 1. speculative states will bump 'branches' for non-branch
 1043		 * instructions
 1044		 * 2. is_state_visited() heuristics may decide not to create
 1045		 * a new state for a sequence of branches and all such current
 1046		 * and cloned states will be pointing to a single parent state
 1047		 * which might have large 'branches' count.
 1048		 */
 1049	}
 1050	return &elem->st;
 1051err:
 1052	free_verifier_state(env->cur_state, true);
 1053	env->cur_state = NULL;
 1054	/* pop all elements and return */
 1055	while (!pop_stack(env, NULL, NULL, false));
 1056	return NULL;
 1057}
 1058
 1059#define CALLER_SAVED_REGS 6
 1060static const int caller_saved[CALLER_SAVED_REGS] = {
 1061	BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
 1062};
 1063
 1064static void __mark_reg_not_init(const struct bpf_verifier_env *env,
 1065				struct bpf_reg_state *reg);
 1066
 1067/* This helper doesn't clear reg->id */
 1068static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
 1069{
 1070	reg->var_off = tnum_const(imm);
 1071	reg->smin_value = (s64)imm;
 1072	reg->smax_value = (s64)imm;
 1073	reg->umin_value = imm;
 1074	reg->umax_value = imm;
 1075
 1076	reg->s32_min_value = (s32)imm;
 1077	reg->s32_max_value = (s32)imm;
 1078	reg->u32_min_value = (u32)imm;
 1079	reg->u32_max_value = (u32)imm;
 1080}
 1081
 1082/* Mark the unknown part of a register (variable offset or scalar value) as
 1083 * known to have the value @imm.
 1084 */
 1085static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
 1086{
 1087	/* Clear id, off, and union(map_ptr, range) */
 1088	memset(((u8 *)reg) + sizeof(reg->type), 0,
 1089	       offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
 1090	___mark_reg_known(reg, imm);
 1091}
 1092
 1093static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
 1094{
 1095	reg->var_off = tnum_const_subreg(reg->var_off, imm);
 1096	reg->s32_min_value = (s32)imm;
 1097	reg->s32_max_value = (s32)imm;
 1098	reg->u32_min_value = (u32)imm;
 1099	reg->u32_max_value = (u32)imm;
 1100}
 1101
 1102/* Mark the 'variable offset' part of a register as zero.  This should be
 1103 * used only on registers holding a pointer type.
 1104 */
 1105static void __mark_reg_known_zero(struct bpf_reg_state *reg)
 1106{
 1107	__mark_reg_known(reg, 0);
 1108}
 1109
 1110static void __mark_reg_const_zero(struct bpf_reg_state *reg)
 1111{
 1112	__mark_reg_known(reg, 0);
 1113	reg->type = SCALAR_VALUE;
 1114}
 1115
 1116static void mark_reg_known_zero(struct bpf_verifier_env *env,
 1117				struct bpf_reg_state *regs, u32 regno)
 1118{
 1119	if (WARN_ON(regno >= MAX_BPF_REG)) {
 1120		verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
 1121		/* Something bad happened, let's kill all regs */
 1122		for (regno = 0; regno < MAX_BPF_REG; regno++)
 1123			__mark_reg_not_init(env, regs + regno);
 1124		return;
 1125	}
 1126	__mark_reg_known_zero(regs + regno);
 1127}
 1128
 1129static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
 1130{
 1131	switch (reg->type) {
 1132	case PTR_TO_MAP_VALUE_OR_NULL: {
 1133		const struct bpf_map *map = reg->map_ptr;
 1134
 1135		if (map->inner_map_meta) {
 1136			reg->type = CONST_PTR_TO_MAP;
 1137			reg->map_ptr = map->inner_map_meta;
 1138		} else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
 1139			reg->type = PTR_TO_XDP_SOCK;
 1140		} else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
 1141			   map->map_type == BPF_MAP_TYPE_SOCKHASH) {
 1142			reg->type = PTR_TO_SOCKET;
 1143		} else {
 1144			reg->type = PTR_TO_MAP_VALUE;
 1145		}
 1146		break;
 1147	}
 1148	case PTR_TO_SOCKET_OR_NULL:
 1149		reg->type = PTR_TO_SOCKET;
 1150		break;
 1151	case PTR_TO_SOCK_COMMON_OR_NULL:
 1152		reg->type = PTR_TO_SOCK_COMMON;
 1153		break;
 1154	case PTR_TO_TCP_SOCK_OR_NULL:
 1155		reg->type = PTR_TO_TCP_SOCK;
 1156		break;
 1157	case PTR_TO_BTF_ID_OR_NULL:
 1158		reg->type = PTR_TO_BTF_ID;
 1159		break;
 1160	case PTR_TO_MEM_OR_NULL:
 1161		reg->type = PTR_TO_MEM;
 1162		break;
 1163	case PTR_TO_RDONLY_BUF_OR_NULL:
 1164		reg->type = PTR_TO_RDONLY_BUF;
 1165		break;
 1166	case PTR_TO_RDWR_BUF_OR_NULL:
 1167		reg->type = PTR_TO_RDWR_BUF;
 1168		break;
 1169	default:
 1170		WARN_ONCE(1, "unknown nullable register type");
 1171	}
 1172}
 1173
 1174static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
 1175{
 1176	return type_is_pkt_pointer(reg->type);
 1177}
 1178
 1179static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
 1180{
 1181	return reg_is_pkt_pointer(reg) ||
 1182	       reg->type == PTR_TO_PACKET_END;
 1183}
 1184
 1185/* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
 1186static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
 1187				    enum bpf_reg_type which)
 1188{
 1189	/* The register can already have a range from prior markings.
 1190	 * This is fine as long as it hasn't been advanced from its
 1191	 * origin.
 1192	 */
 1193	return reg->type == which &&
 1194	       reg->id == 0 &&
 1195	       reg->off == 0 &&
 1196	       tnum_equals_const(reg->var_off, 0);
 1197}
 1198
 1199/* Reset the min/max bounds of a register */
 1200static void __mark_reg_unbounded(struct bpf_reg_state *reg)
 1201{
 1202	reg->smin_value = S64_MIN;
 1203	reg->smax_value = S64_MAX;
 1204	reg->umin_value = 0;
 1205	reg->umax_value = U64_MAX;
 1206
 1207	reg->s32_min_value = S32_MIN;
 1208	reg->s32_max_value = S32_MAX;
 1209	reg->u32_min_value = 0;
 1210	reg->u32_max_value = U32_MAX;
 1211}
 1212
 1213static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
 1214{
 1215	reg->smin_value = S64_MIN;
 1216	reg->smax_value = S64_MAX;
 1217	reg->umin_value = 0;
 1218	reg->umax_value = U64_MAX;
 1219}
 1220
 1221static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
 1222{
 1223	reg->s32_min_value = S32_MIN;
 1224	reg->s32_max_value = S32_MAX;
 1225	reg->u32_min_value = 0;
 1226	reg->u32_max_value = U32_MAX;
 1227}
 1228
 1229static void __update_reg32_bounds(struct bpf_reg_state *reg)
 1230{
 1231	struct tnum var32_off = tnum_subreg(reg->var_off);
 1232
 1233	/* min signed is max(sign bit) | min(other bits) */
 1234	reg->s32_min_value = max_t(s32, reg->s32_min_value,
 1235			var32_off.value | (var32_off.mask & S32_MIN));
 1236	/* max signed is min(sign bit) | max(other bits) */
 1237	reg->s32_max_value = min_t(s32, reg->s32_max_value,
 1238			var32_off.value | (var32_off.mask & S32_MAX));
 1239	reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
 1240	reg->u32_max_value = min(reg->u32_max_value,
 1241				 (u32)(var32_off.value | var32_off.mask));
 1242}
 1243
 1244static void __update_reg64_bounds(struct bpf_reg_state *reg)
 1245{
 1246	/* min signed is max(sign bit) | min(other bits) */
 1247	reg->smin_value = max_t(s64, reg->smin_value,
 1248				reg->var_off.value | (reg->var_off.mask & S64_MIN));
 1249	/* max signed is min(sign bit) | max(other bits) */
 1250	reg->smax_value = min_t(s64, reg->smax_value,
 1251				reg->var_off.value | (reg->var_off.mask & S64_MAX));
 1252	reg->umin_value = max(reg->umin_value, reg->var_off.value);
 1253	reg->umax_value = min(reg->umax_value,
 1254			      reg->var_off.value | reg->var_off.mask);
 1255}
 1256
 1257static void __update_reg_bounds(struct bpf_reg_state *reg)
 1258{
 1259	__update_reg32_bounds(reg);
 1260	__update_reg64_bounds(reg);
 1261}
 1262
 1263/* Uses signed min/max values to inform unsigned, and vice-versa */
 1264static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
 1265{
 1266	/* Learn sign from signed bounds.
 1267	 * If we cannot cross the sign boundary, then signed and unsigned bounds
 1268	 * are the same, so combine.  This works even in the negative case, e.g.
 1269	 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
 1270	 */
 1271	if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
 1272		reg->s32_min_value = reg->u32_min_value =
 1273			max_t(u32, reg->s32_min_value, reg->u32_min_value);
 1274		reg->s32_max_value = reg->u32_max_value =
 1275			min_t(u32, reg->s32_max_value, reg->u32_max_value);
 1276		return;
 1277	}
 1278	/* Learn sign from unsigned bounds.  Signed bounds cross the sign
 1279	 * boundary, so we must be careful.
 1280	 */
 1281	if ((s32)reg->u32_max_value >= 0) {
 1282		/* Positive.  We can't learn anything from the smin, but smax
 1283		 * is positive, hence safe.
 1284		 */
 1285		reg->s32_min_value = reg->u32_min_value;
 1286		reg->s32_max_value = reg->u32_max_value =
 1287			min_t(u32, reg->s32_max_value, reg->u32_max_value);
 1288	} else if ((s32)reg->u32_min_value < 0) {
 1289		/* Negative.  We can't learn anything from the smax, but smin
 1290		 * is negative, hence safe.
 1291		 */
 1292		reg->s32_min_value = reg->u32_min_value =
 1293			max_t(u32, reg->s32_min_value, reg->u32_min_value);
 1294		reg->s32_max_value = reg->u32_max_value;
 1295	}
 1296}
 1297
 1298static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
 1299{
 1300	/* Learn sign from signed bounds.
 1301	 * If we cannot cross the sign boundary, then signed and unsigned bounds
 1302	 * are the same, so combine.  This works even in the negative case, e.g.
 1303	 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
 1304	 */
 1305	if (reg->smin_value >= 0 || reg->smax_value < 0) {
 1306		reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
 1307							  reg->umin_value);
 1308		reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
 1309							  reg->umax_value);
 1310		return;
 1311	}
 1312	/* Learn sign from unsigned bounds.  Signed bounds cross the sign
 1313	 * boundary, so we must be careful.
 1314	 */
 1315	if ((s64)reg->umax_value >= 0) {
 1316		/* Positive.  We can't learn anything from the smin, but smax
 1317		 * is positive, hence safe.
 1318		 */
 1319		reg->smin_value = reg->umin_value;
 1320		reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
 1321							  reg->umax_value);
 1322	} else if ((s64)reg->umin_value < 0) {
 1323		/* Negative.  We can't learn anything from the smax, but smin
 1324		 * is negative, hence safe.
 1325		 */
 1326		reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
 1327							  reg->umin_value);
 1328		reg->smax_value = reg->umax_value;
 1329	}
 1330}
 1331
 1332static void __reg_deduce_bounds(struct bpf_reg_state *reg)
 1333{
 1334	__reg32_deduce_bounds(reg);
 1335	__reg64_deduce_bounds(reg);
 1336}
 1337
 1338/* Attempts to improve var_off based on unsigned min/max information */
 1339static void __reg_bound_offset(struct bpf_reg_state *reg)
 1340{
 1341	struct tnum var64_off = tnum_intersect(reg->var_off,
 1342					       tnum_range(reg->umin_value,
 1343							  reg->umax_value));
 1344	struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
 1345						tnum_range(reg->u32_min_value,
 1346							   reg->u32_max_value));
 1347
 1348	reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
 1349}
 1350
 1351static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
 1352{
 1353	reg->umin_value = reg->u32_min_value;
 1354	reg->umax_value = reg->u32_max_value;
 1355	/* Attempt to pull 32-bit signed bounds into 64-bit bounds
 1356	 * but must be positive otherwise set to worse case bounds
 1357	 * and refine later from tnum.
 1358	 */
 1359	if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
 1360		reg->smax_value = reg->s32_max_value;
 1361	else
 1362		reg->smax_value = U32_MAX;
 1363	if (reg->s32_min_value >= 0)
 1364		reg->smin_value = reg->s32_min_value;
 1365	else
 1366		reg->smin_value = 0;
 1367}
 1368
 1369static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
 1370{
 1371	/* special case when 64-bit register has upper 32-bit register
 1372	 * zeroed. Typically happens after zext or <<32, >>32 sequence
 1373	 * allowing us to use 32-bit bounds directly,
 1374	 */
 1375	if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
 1376		__reg_assign_32_into_64(reg);
 1377	} else {
 1378		/* Otherwise the best we can do is push lower 32bit known and
 1379		 * unknown bits into register (var_off set from jmp logic)
 1380		 * then learn as much as possible from the 64-bit tnum
 1381		 * known and unknown bits. The previous smin/smax bounds are
 1382		 * invalid here because of jmp32 compare so mark them unknown
 1383		 * so they do not impact tnum bounds calculation.
 1384		 */
 1385		__mark_reg64_unbounded(reg);
 1386		__update_reg_bounds(reg);
 1387	}
 1388
 1389	/* Intersecting with the old var_off might have improved our bounds
 1390	 * slightly.  e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
 1391	 * then new var_off is (0; 0x7f...fc) which improves our umax.
 1392	 */
 1393	__reg_deduce_bounds(reg);
 1394	__reg_bound_offset(reg);
 1395	__update_reg_bounds(reg);
 1396}
 1397
 1398static bool __reg64_bound_s32(s64 a)
 1399{
 1400	return a > S32_MIN && a < S32_MAX;
 1401}
 1402
 1403static bool __reg64_bound_u32(u64 a)
 1404{
 1405	return a > U32_MIN && a < U32_MAX;
 1406}
 1407
 1408static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
 1409{
 1410	__mark_reg32_unbounded(reg);
 1411
 1412	if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
 1413		reg->s32_min_value = (s32)reg->smin_value;
 1414		reg->s32_max_value = (s32)reg->smax_value;
 1415	}
 1416	if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
 1417		reg->u32_min_value = (u32)reg->umin_value;
 1418		reg->u32_max_value = (u32)reg->umax_value;
 1419	}
 1420
 1421	/* Intersecting with the old var_off might have improved our bounds
 1422	 * slightly.  e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
 1423	 * then new var_off is (0; 0x7f...fc) which improves our umax.
 1424	 */
 1425	__reg_deduce_bounds(reg);
 1426	__reg_bound_offset(reg);
 1427	__update_reg_bounds(reg);
 1428}
 1429
 1430/* Mark a register as having a completely unknown (scalar) value. */
 1431static void __mark_reg_unknown(const struct bpf_verifier_env *env,
 1432			       struct bpf_reg_state *reg)
 1433{
 1434	/*
 1435	 * Clear type, id, off, and union(map_ptr, range) and
 1436	 * padding between 'type' and union
 1437	 */
 1438	memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
 1439	reg->type = SCALAR_VALUE;
 1440	reg->var_off = tnum_unknown;
 1441	reg->frameno = 0;
 1442	reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
 1443	__mark_reg_unbounded(reg);
 1444}
 1445
 1446static void mark_reg_unknown(struct bpf_verifier_env *env,
 1447			     struct bpf_reg_state *regs, u32 regno)
 1448{
 1449	if (WARN_ON(regno >= MAX_BPF_REG)) {
 1450		verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
 1451		/* Something bad happened, let's kill all regs except FP */
 1452		for (regno = 0; regno < BPF_REG_FP; regno++)
 1453			__mark_reg_not_init(env, regs + regno);
 1454		return;
 1455	}
 1456	__mark_reg_unknown(env, regs + regno);
 1457}
 1458
 1459static void __mark_reg_not_init(const struct bpf_verifier_env *env,
 1460				struct bpf_reg_state *reg)
 1461{
 1462	__mark_reg_unknown(env, reg);
 1463	reg->type = NOT_INIT;
 1464}
 1465
 1466static void mark_reg_not_init(struct bpf_verifier_env *env,
 1467			      struct bpf_reg_state *regs, u32 regno)
 1468{
 1469	if (WARN_ON(regno >= MAX_BPF_REG)) {
 1470		verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
 1471		/* Something bad happened, let's kill all regs except FP */
 1472		for (regno = 0; regno < BPF_REG_FP; regno++)
 1473			__mark_reg_not_init(env, regs + regno);
 1474		return;
 1475	}
 1476	__mark_reg_not_init(env, regs + regno);
 1477}
 1478
 1479static void mark_btf_ld_reg(struct bpf_verifier_env *env,
 1480			    struct bpf_reg_state *regs, u32 regno,
 1481			    enum bpf_reg_type reg_type,
 1482			    struct btf *btf, u32 btf_id)
 1483{
 1484	if (reg_type == SCALAR_VALUE) {
 1485		mark_reg_unknown(env, regs, regno);
 1486		return;
 1487	}
 1488	mark_reg_known_zero(env, regs, regno);
 1489	regs[regno].type = PTR_TO_BTF_ID;
 1490	regs[regno].btf = btf;
 1491	regs[regno].btf_id = btf_id;
 1492}
 1493
 1494#define DEF_NOT_SUBREG	(0)
 1495static void init_reg_state(struct bpf_verifier_env *env,
 1496			   struct bpf_func_state *state)
 1497{
 1498	struct bpf_reg_state *regs = state->regs;
 1499	int i;
 1500
 1501	for (i = 0; i < MAX_BPF_REG; i++) {
 1502		mark_reg_not_init(env, regs, i);
 1503		regs[i].live = REG_LIVE_NONE;
 1504		regs[i].parent = NULL;
 1505		regs[i].subreg_def = DEF_NOT_SUBREG;
 1506	}
 1507
 1508	/* frame pointer */
 1509	regs[BPF_REG_FP].type = PTR_TO_STACK;
 1510	mark_reg_known_zero(env, regs, BPF_REG_FP);
 1511	regs[BPF_REG_FP].frameno = state->frameno;
 1512}
 1513
 1514#define BPF_MAIN_FUNC (-1)
 1515static void init_func_state(struct bpf_verifier_env *env,
 1516			    struct bpf_func_state *state,
 1517			    int callsite, int frameno, int subprogno)
 1518{
 1519	state->callsite = callsite;
 1520	state->frameno = frameno;
 1521	state->subprogno = subprogno;
 1522	init_reg_state(env, state);
 1523}
 1524
 1525enum reg_arg_type {
 1526	SRC_OP,		/* register is used as source operand */
 1527	DST_OP,		/* register is used as destination operand */
 1528	DST_OP_NO_MARK	/* same as above, check only, don't mark */
 1529};
 1530
 1531static int cmp_subprogs(const void *a, const void *b)
 1532{
 1533	return ((struct bpf_subprog_info *)a)->start -
 1534	       ((struct bpf_subprog_info *)b)->start;
 1535}
 1536
 1537static int find_subprog(struct bpf_verifier_env *env, int off)
 1538{
 1539	struct bpf_subprog_info *p;
 1540
 1541	p = bsearch(&off, env->subprog_info, env->subprog_cnt,
 1542		    sizeof(env->subprog_info[0]), cmp_subprogs);
 1543	if (!p)
 1544		return -ENOENT;
 1545	return p - env->subprog_info;
 1546
 1547}
 1548
 1549static int add_subprog(struct bpf_verifier_env *env, int off)
 1550{
 1551	int insn_cnt = env->prog->len;
 1552	int ret;
 1553
 1554	if (off >= insn_cnt || off < 0) {
 1555		verbose(env, "call to invalid destination\n");
 1556		return -EINVAL;
 1557	}
 1558	ret = find_subprog(env, off);
 1559	if (ret >= 0)
 1560		return ret;
 1561	if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
 1562		verbose(env, "too many subprograms\n");
 1563		return -E2BIG;
 1564	}
 1565	/* determine subprog starts. The end is one before the next starts */
 1566	env->subprog_info[env->subprog_cnt++].start = off;
 1567	sort(env->subprog_info, env->subprog_cnt,
 1568	     sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
 1569	return env->subprog_cnt - 1;
 1570}
 1571
 1572struct bpf_kfunc_desc {
 1573	struct btf_func_model func_model;
 1574	u32 func_id;
 1575	s32 imm;
 1576};
 1577
 1578#define MAX_KFUNC_DESCS 256
 1579struct bpf_kfunc_desc_tab {
 1580	struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
 1581	u32 nr_descs;
 1582};
 1583
 1584static int kfunc_desc_cmp_by_id(const void *a, const void *b)
 1585{
 1586	const struct bpf_kfunc_desc *d0 = a;
 1587	const struct bpf_kfunc_desc *d1 = b;
 1588
 1589	/* func_id is not greater than BTF_MAX_TYPE */
 1590	return d0->func_id - d1->func_id;
 1591}
 1592
 1593static const struct bpf_kfunc_desc *
 1594find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
 1595{
 1596	struct bpf_kfunc_desc desc = {
 1597		.func_id = func_id,
 1598	};
 1599	struct bpf_kfunc_desc_tab *tab;
 1600
 1601	tab = prog->aux->kfunc_tab;
 1602	return bsearch(&desc, tab->descs, tab->nr_descs,
 1603		       sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
 1604}
 1605
 1606static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
 1607{
 1608	const struct btf_type *func, *func_proto;
 1609	struct bpf_kfunc_desc_tab *tab;
 1610	struct bpf_prog_aux *prog_aux;
 1611	struct bpf_kfunc_desc *desc;
 1612	const char *func_name;
 1613	unsigned long addr;
 1614	int err;
 1615
 1616	prog_aux = env->prog->aux;
 1617	tab = prog_aux->kfunc_tab;
 1618	if (!tab) {
 1619		if (!btf_vmlinux) {
 1620			verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
 1621			return -ENOTSUPP;
 1622		}
 1623
 1624		if (!env->prog->jit_requested) {
 1625			verbose(env, "JIT is required for calling kernel function\n");
 1626			return -ENOTSUPP;
 1627		}
 1628
 1629		if (!bpf_jit_supports_kfunc_call()) {
 1630			verbose(env, "JIT does not support calling kernel function\n");
 1631			return -ENOTSUPP;
 1632		}
 1633
 1634		if (!env->prog->gpl_compatible) {
 1635			verbose(env, "cannot call kernel function from non-GPL compatible program\n");
 1636			return -EINVAL;
 1637		}
 1638
 1639		tab = kzalloc(sizeof(*tab), GFP_KERNEL);
 1640		if (!tab)
 1641			return -ENOMEM;
 1642		prog_aux->kfunc_tab = tab;
 1643	}
 1644
 1645	if (find_kfunc_desc(env->prog, func_id))
 1646		return 0;
 1647
 1648	if (tab->nr_descs == MAX_KFUNC_DESCS) {
 1649		verbose(env, "too many different kernel function calls\n");
 1650		return -E2BIG;
 1651	}
 1652
 1653	func = btf_type_by_id(btf_vmlinux, func_id);
 1654	if (!func || !btf_type_is_func(func)) {
 1655		verbose(env, "kernel btf_id %u is not a function\n",
 1656			func_id);
 1657		return -EINVAL;
 1658	}
 1659	func_proto = btf_type_by_id(btf_vmlinux, func->type);
 1660	if (!func_proto || !btf_type_is_func_proto(func_proto)) {
 1661		verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
 1662			func_id);
 1663		return -EINVAL;
 1664	}
 1665
 1666	func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
 1667	addr = kallsyms_lookup_name(func_name);
 1668	if (!addr) {
 1669		verbose(env, "cannot find address for kernel function %s\n",
 1670			func_name);
 1671		return -EINVAL;
 1672	}
 1673
 1674	desc = &tab->descs[tab->nr_descs++];
 1675	desc->func_id = func_id;
 1676	desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
 1677	err = btf_distill_func_proto(&env->log, btf_vmlinux,
 1678				     func_proto, func_name,
 1679				     &desc->func_model);
 1680	if (!err)
 1681		sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
 1682		     kfunc_desc_cmp_by_id, NULL);
 1683	return err;
 1684}
 1685
 1686static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
 1687{
 1688	const struct bpf_kfunc_desc *d0 = a;
 1689	const struct bpf_kfunc_desc *d1 = b;
 1690
 1691	if (d0->imm > d1->imm)
 1692		return 1;
 1693	else if (d0->imm < d1->imm)
 1694		return -1;
 1695	return 0;
 1696}
 1697
 1698static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
 1699{
 1700	struct bpf_kfunc_desc_tab *tab;
 1701
 1702	tab = prog->aux->kfunc_tab;
 1703	if (!tab)
 1704		return;
 1705
 1706	sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
 1707	     kfunc_desc_cmp_by_imm, NULL);
 1708}
 1709
 1710bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
 1711{
 1712	return !!prog->aux->kfunc_tab;
 1713}
 1714
 1715const struct btf_func_model *
 1716bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
 1717			 const struct bpf_insn *insn)
 1718{
 1719	const struct bpf_kfunc_desc desc = {
 1720		.imm = insn->imm,
 1721	};
 1722	const struct bpf_kfunc_desc *res;
 1723	struct bpf_kfunc_desc_tab *tab;
 1724
 1725	tab = prog->aux->kfunc_tab;
 1726	res = bsearch(&desc, tab->descs, tab->nr_descs,
 1727		      sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
 1728
 1729	return res ? &res->func_model : NULL;
 1730}
 1731
 1732static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
 1733{
 1734	struct bpf_subprog_info *subprog = env->subprog_info;
 1735	struct bpf_insn *insn = env->prog->insnsi;
 1736	int i, ret, insn_cnt = env->prog->len;
 1737
 1738	/* Add entry function. */
 1739	ret = add_subprog(env, 0);
 1740	if (ret)
 1741		return ret;
 1742
 1743	for (i = 0; i < insn_cnt; i++, insn++) {
 1744		if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
 1745		    !bpf_pseudo_kfunc_call(insn))
 1746			continue;
 1747
 1748		if (!env->bpf_capable) {
 1749			verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
 1750			return -EPERM;
 1751		}
 1752
 1753		if (bpf_pseudo_func(insn)) {
 1754			ret = add_subprog(env, i + insn->imm + 1);
 1755			if (ret >= 0)
 1756				/* remember subprog */
 1757				insn[1].imm = ret;
 1758		} else if (bpf_pseudo_call(insn)) {
 1759			ret = add_subprog(env, i + insn->imm + 1);
 1760		} else {
 1761			ret = add_kfunc_call(env, insn->imm);
 1762		}
 1763
 1764		if (ret < 0)
 1765			return ret;
 1766	}
 1767
 1768	/* Add a fake 'exit' subprog which could simplify subprog iteration
 1769	 * logic. 'subprog_cnt' should not be increased.
 1770	 */
 1771	subprog[env->subprog_cnt].start = insn_cnt;
 1772
 1773	if (env->log.level & BPF_LOG_LEVEL2)
 1774		for (i = 0; i < env->subprog_cnt; i++)
 1775			verbose(env, "func#%d @%d\n", i, subprog[i].start);
 1776
 1777	return 0;
 1778}
 1779
 1780static int check_subprogs(struct bpf_verifier_env *env)
 1781{
 1782	int i, subprog_start, subprog_end, off, cur_subprog = 0;
 1783	struct bpf_subprog_info *subprog = env->subprog_info;
 1784	struct bpf_insn *insn = env->prog->insnsi;
 1785	int insn_cnt = env->prog->len;
 1786
 1787	/* now check that all jumps are within the same subprog */
 1788	subprog_start = subprog[cur_subprog].start;
 1789	subprog_end = subprog[cur_subprog + 1].start;
 1790	for (i = 0; i < insn_cnt; i++) {
 1791		u8 code = insn[i].code;
 1792
 1793		if (code == (BPF_JMP | BPF_CALL) &&
 1794		    insn[i].imm == BPF_FUNC_tail_call &&
 1795		    insn[i].src_reg != BPF_PSEUDO_CALL)
 1796			subprog[cur_subprog].has_tail_call = true;
 1797		if (BPF_CLASS(code) == BPF_LD &&
 1798		    (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
 1799			subprog[cur_subprog].has_ld_abs = true;
 1800		if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
 1801			goto next;
 1802		if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
 1803			goto next;
 1804		off = i + insn[i].off + 1;
 1805		if (off < subprog_start || off >= subprog_end) {
 1806			verbose(env, "jump out of range from insn %d to %d\n", i, off);
 1807			return -EINVAL;
 1808		}
 1809next:
 1810		if (i == subprog_end - 1) {
 1811			/* to avoid fall-through from one subprog into another
 1812			 * the last insn of the subprog should be either exit
 1813			 * or unconditional jump back
 1814			 */
 1815			if (code != (BPF_JMP | BPF_EXIT) &&
 1816			    code != (BPF_JMP | BPF_JA)) {
 1817				verbose(env, "last insn is not an exit or jmp\n");
 1818				return -EINVAL;
 1819			}
 1820			subprog_start = subprog_end;
 1821			cur_subprog++;
 1822			if (cur_subprog < env->subprog_cnt)
 1823				subprog_end = subprog[cur_subprog + 1].start;
 1824		}
 1825	}
 1826	return 0;
 1827}
 1828
 1829/* Parentage chain of this register (or stack slot) should take care of all
 1830 * issues like callee-saved registers, stack slot allocation time, etc.
 1831 */
 1832static int mark_reg_read(struct bpf_verifier_env *env,
 1833			 const struct bpf_reg_state *state,
 1834			 struct bpf_reg_state *parent, u8 flag)
 1835{
 1836	bool writes = parent == state->parent; /* Observe write marks */
 1837	int cnt = 0;
 1838
 1839	while (parent) {
 1840		/* if read wasn't screened by an earlier write ... */
 1841		if (writes && state->live & REG_LIVE_WRITTEN)
 1842			break;
 1843		if (parent->live & REG_LIVE_DONE) {
 1844			verbose(env, "verifier BUG type %s var_off %lld off %d\n",
 1845				reg_type_str[parent->type],
 1846				parent->var_off.value, parent->off);
 1847			return -EFAULT;
 1848		}
 1849		/* The first condition is more likely to be true than the
 1850		 * second, checked it first.
 1851		 */
 1852		if ((parent->live & REG_LIVE_READ) == flag ||
 1853		    parent->live & REG_LIVE_READ64)
 1854			/* The parentage chain never changes and
 1855			 * this parent was already marked as LIVE_READ.
 1856			 * There is no need to keep walking the chain again and
 1857			 * keep re-marking all parents as LIVE_READ.
 1858			 * This case happens when the same register is read
 1859			 * multiple times without writes into it in-between.
 1860			 * Also, if parent has the stronger REG_LIVE_READ64 set,
 1861			 * then no need to set the weak REG_LIVE_READ32.
 1862			 */
 1863			break;
 1864		/* ... then we depend on parent's value */
 1865		parent->live |= flag;
 1866		/* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
 1867		if (flag == REG_LIVE_READ64)
 1868			parent->live &= ~REG_LIVE_READ32;
 1869		state = parent;
 1870		parent = state->parent;
 1871		writes = true;
 1872		cnt++;
 1873	}
 1874
 1875	if (env->longest_mark_read_walk < cnt)
 1876		env->longest_mark_read_walk = cnt;
 1877	return 0;
 1878}
 1879
 1880/* This function is supposed to be used by the following 32-bit optimization
 1881 * code only. It returns TRUE if the source or destination register operates
 1882 * on 64-bit, otherwise return FALSE.
 1883 */
 1884static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
 1885		     u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
 1886{
 1887	u8 code, class, op;
 1888
 1889	code = insn->code;
 1890	class = BPF_CLASS(code);
 1891	op = BPF_OP(code);
 1892	if (class == BPF_JMP) {
 1893		/* BPF_EXIT for "main" will reach here. Return TRUE
 1894		 * conservatively.
 1895		 */
 1896		if (op == BPF_EXIT)
 1897			return true;
 1898		if (op == BPF_CALL) {
 1899			/* BPF to BPF call will reach here because of marking
 1900			 * caller saved clobber with DST_OP_NO_MARK for which we
 1901			 * don't care the register def because they are anyway
 1902			 * marked as NOT_INIT already.
 1903			 */
 1904			if (insn->src_reg == BPF_PSEUDO_CALL)
 1905				return false;
 1906			/* Helper call will reach here because of arg type
 1907			 * check, conservatively return TRUE.
 1908			 */
 1909			if (t == SRC_OP)
 1910				return true;
 1911
 1912			return false;
 1913		}
 1914	}
 1915
 1916	if (class == BPF_ALU64 || class == BPF_JMP ||
 1917	    /* BPF_END always use BPF_ALU class. */
 1918	    (class == BPF_ALU && op == BPF_END && insn->imm == 64))
 1919		return true;
 1920
 1921	if (class == BPF_ALU || class == BPF_JMP32)
 1922		return false;
 1923
 1924	if (class == BPF_LDX) {
 1925		if (t != SRC_OP)
 1926			return BPF_SIZE(code) == BPF_DW;
 1927		/* LDX source must be ptr. */
 1928		return true;
 1929	}
 1930
 1931	if (class == BPF_STX) {
 1932		/* BPF_STX (including atomic variants) has multiple source
 1933		 * operands, one of which is a ptr. Check whether the caller is
 1934		 * asking about it.
 1935		 */
 1936		if (t == SRC_OP && reg->type != SCALAR_VALUE)
 1937			return true;
 1938		return BPF_SIZE(code) == BPF_DW;
 1939	}
 1940
 1941	if (class == BPF_LD) {
 1942		u8 mode = BPF_MODE(code);
 1943
 1944		/* LD_IMM64 */
 1945		if (mode == BPF_IMM)
 1946			return true;
 1947
 1948		/* Both LD_IND and LD_ABS return 32-bit data. */
 1949		if (t != SRC_OP)
 1950			return  false;
 1951
 1952		/* Implicit ctx ptr. */
 1953		if (regno == BPF_REG_6)
 1954			return true;
 1955
 1956		/* Explicit source could be any width. */
 1957		return true;
 1958	}
 1959
 1960	if (class == BPF_ST)
 1961		/* The only source register for BPF_ST is a ptr. */
 1962		return true;
 1963
 1964	/* Conservatively return true at default. */
 1965	return true;
 1966}
 1967
 1968/* Return the regno defined by the insn, or -1. */
 1969static int insn_def_regno(const struct bpf_insn *insn)
 1970{
 1971	switch (BPF_CLASS(insn->code)) {
 1972	case BPF_JMP:
 1973	case BPF_JMP32:
 1974	case BPF_ST:
 1975		return -1;
 1976	case BPF_STX:
 1977		if (BPF_MODE(insn->code) == BPF_ATOMIC &&
 1978		    (insn->imm & BPF_FETCH)) {
 1979			if (insn->imm == BPF_CMPXCHG)
 1980				return BPF_REG_0;
 1981			else
 1982				return insn->src_reg;
 1983		} else {
 1984			return -1;
 1985		}
 1986	default:
 1987		return insn->dst_reg;
 1988	}
 1989}
 1990
 1991/* Return TRUE if INSN has defined any 32-bit value explicitly. */
 1992static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
 1993{
 1994	int dst_reg = insn_def_regno(insn);
 1995
 1996	if (dst_reg == -1)
 1997		return false;
 1998
 1999	return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
 2000}
 2001
 2002static void mark_insn_zext(struct bpf_verifier_env *env,
 2003			   struct bpf_reg_state *reg)
 2004{
 2005	s32 def_idx = reg->subreg_def;
 2006
 2007	if (def_idx == DEF_NOT_SUBREG)
 2008		return;
 2009
 2010	env->insn_aux_data[def_idx - 1].zext_dst = true;
 2011	/* The dst will be zero extended, so won't be sub-register anymore. */
 2012	reg->subreg_def = DEF_NOT_SUBREG;
 2013}
 2014
 2015static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
 2016			 enum reg_arg_type t)
 2017{
 2018	struct bpf_verifier_state *vstate = env->cur_state;
 2019	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 2020	struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
 2021	struct bpf_reg_state *reg, *regs = state->regs;
 2022	bool rw64;
 2023
 2024	if (regno >= MAX_BPF_REG) {
 2025		verbose(env, "R%d is invalid\n", regno);
 2026		return -EINVAL;
 2027	}
 2028
 2029	reg = &regs[regno];
 2030	rw64 = is_reg64(env, insn, regno, reg, t);
 2031	if (t == SRC_OP) {
 2032		/* check whether register used as source operand can be read */
 2033		if (reg->type == NOT_INIT) {
 2034			verbose(env, "R%d !read_ok\n", regno);
 2035			return -EACCES;
 2036		}
 2037		/* We don't need to worry about FP liveness because it's read-only */
 2038		if (regno == BPF_REG_FP)
 2039			return 0;
 2040
 2041		if (rw64)
 2042			mark_insn_zext(env, reg);
 2043
 2044		return mark_reg_read(env, reg, reg->parent,
 2045				     rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
 2046	} else {
 2047		/* check whether register used as dest operand can be written to */
 2048		if (regno == BPF_REG_FP) {
 2049			verbose(env, "frame pointer is read only\n");
 2050			return -EACCES;
 2051		}
 2052		reg->live |= REG_LIVE_WRITTEN;
 2053		reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
 2054		if (t == DST_OP)
 2055			mark_reg_unknown(env, regs, regno);
 2056	}
 2057	return 0;
 2058}
 2059
 2060/* for any branch, call, exit record the history of jmps in the given state */
 2061static int push_jmp_history(struct bpf_verifier_env *env,
 2062			    struct bpf_verifier_state *cur)
 2063{
 2064	u32 cnt = cur->jmp_history_cnt;
 2065	struct bpf_idx_pair *p;
 2066
 2067	cnt++;
 2068	p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
 2069	if (!p)
 2070		return -ENOMEM;
 2071	p[cnt - 1].idx = env->insn_idx;
 2072	p[cnt - 1].prev_idx = env->prev_insn_idx;
 2073	cur->jmp_history = p;
 2074	cur->jmp_history_cnt = cnt;
 2075	return 0;
 2076}
 2077
 2078/* Backtrack one insn at a time. If idx is not at the top of recorded
 2079 * history then previous instruction came from straight line execution.
 2080 */
 2081static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
 2082			     u32 *history)
 2083{
 2084	u32 cnt = *history;
 2085
 2086	if (cnt && st->jmp_history[cnt - 1].idx == i) {
 2087		i = st->jmp_history[cnt - 1].prev_idx;
 2088		(*history)--;
 2089	} else {
 2090		i--;
 2091	}
 2092	return i;
 2093}
 2094
 2095static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
 2096{
 2097	const struct btf_type *func;
 2098
 2099	if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
 2100		return NULL;
 2101
 2102	func = btf_type_by_id(btf_vmlinux, insn->imm);
 2103	return btf_name_by_offset(btf_vmlinux, func->name_off);
 2104}
 2105
 2106/* For given verifier state backtrack_insn() is called from the last insn to
 2107 * the first insn. Its purpose is to compute a bitmask of registers and
 2108 * stack slots that needs precision in the parent verifier state.
 2109 */
 2110static int backtrack_insn(struct bpf_verifier_env *env, int idx,
 2111			  u32 *reg_mask, u64 *stack_mask)
 2112{
 2113	const struct bpf_insn_cbs cbs = {
 2114		.cb_call	= disasm_kfunc_name,
 2115		.cb_print	= verbose,
 2116		.private_data	= env,
 2117	};
 2118	struct bpf_insn *insn = env->prog->insnsi + idx;
 2119	u8 class = BPF_CLASS(insn->code);
 2120	u8 opcode = BPF_OP(insn->code);
 2121	u8 mode = BPF_MODE(insn->code);
 2122	u32 dreg = 1u << insn->dst_reg;
 2123	u32 sreg = 1u << insn->src_reg;
 2124	u32 spi;
 2125
 2126	if (insn->code == 0)
 2127		return 0;
 2128	if (env->log.level & BPF_LOG_LEVEL) {
 2129		verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
 2130		verbose(env, "%d: ", idx);
 2131		print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
 2132	}
 2133
 2134	if (class == BPF_ALU || class == BPF_ALU64) {
 2135		if (!(*reg_mask & dreg))
 2136			return 0;
 2137		if (opcode == BPF_MOV) {
 2138			if (BPF_SRC(insn->code) == BPF_X) {
 2139				/* dreg = sreg
 2140				 * dreg needs precision after this insn
 2141				 * sreg needs precision before this insn
 2142				 */
 2143				*reg_mask &= ~dreg;
 2144				*reg_mask |= sreg;
 2145			} else {
 2146				/* dreg = K
 2147				 * dreg needs precision after this insn.
 2148				 * Corresponding register is already marked
 2149				 * as precise=true in this verifier state.
 2150				 * No further markings in parent are necessary
 2151				 */
 2152				*reg_mask &= ~dreg;
 2153			}
 2154		} else {
 2155			if (BPF_SRC(insn->code) == BPF_X) {
 2156				/* dreg += sreg
 2157				 * both dreg and sreg need precision
 2158				 * before this insn
 2159				 */
 2160				*reg_mask |= sreg;
 2161			} /* else dreg += K
 2162			   * dreg still needs precision before this insn
 2163			   */
 2164		}
 2165	} else if (class == BPF_LDX) {
 2166		if (!(*reg_mask & dreg))
 2167			return 0;
 2168		*reg_mask &= ~dreg;
 2169
 2170		/* scalars can only be spilled into stack w/o losing precision.
 2171		 * Load from any other memory can be zero extended.
 2172		 * The desire to keep that precision is already indicated
 2173		 * by 'precise' mark in corresponding register of this state.
 2174		 * No further tracking necessary.
 2175		 */
 2176		if (insn->src_reg != BPF_REG_FP)
 2177			return 0;
 2178		if (BPF_SIZE(insn->code) != BPF_DW)
 2179			return 0;
 2180
 2181		/* dreg = *(u64 *)[fp - off] was a fill from the stack.
 2182		 * that [fp - off] slot contains scalar that needs to be
 2183		 * tracked with precision
 2184		 */
 2185		spi = (-insn->off - 1) / BPF_REG_SIZE;
 2186		if (spi >= 64) {
 2187			verbose(env, "BUG spi %d\n", spi);
 2188			WARN_ONCE(1, "verifier backtracking bug");
 2189			return -EFAULT;
 2190		}
 2191		*stack_mask |= 1ull << spi;
 2192	} else if (class == BPF_STX || class == BPF_ST) {
 2193		if (*reg_mask & dreg)
 2194			/* stx & st shouldn't be using _scalar_ dst_reg
 2195			 * to access memory. It means backtracking
 2196			 * encountered a case of pointer subtraction.
 2197			 */
 2198			return -ENOTSUPP;
 2199		/* scalars can only be spilled into stack */
 2200		if (insn->dst_reg != BPF_REG_FP)
 2201			return 0;
 2202		if (BPF_SIZE(insn->code) != BPF_DW)
 2203			return 0;
 2204		spi = (-insn->off - 1) / BPF_REG_SIZE;
 2205		if (spi >= 64) {
 2206			verbose(env, "BUG spi %d\n", spi);
 2207			WARN_ONCE(1, "verifier backtracking bug");
 2208			return -EFAULT;
 2209		}
 2210		if (!(*stack_mask & (1ull << spi)))
 2211			return 0;
 2212		*stack_mask &= ~(1ull << spi);
 2213		if (class == BPF_STX)
 2214			*reg_mask |= sreg;
 2215	} else if (class == BPF_JMP || class == BPF_JMP32) {
 2216		if (opcode == BPF_CALL) {
 2217			if (insn->src_reg == BPF_PSEUDO_CALL)
 2218				return -ENOTSUPP;
 2219			/* regular helper call sets R0 */
 2220			*reg_mask &= ~1;
 2221			if (*reg_mask & 0x3f) {
 2222				/* if backtracing was looking for registers R1-R5
 2223				 * they should have been found already.
 2224				 */
 2225				verbose(env, "BUG regs %x\n", *reg_mask);
 2226				WARN_ONCE(1, "verifier backtracking bug");
 2227				return -EFAULT;
 2228			}
 2229		} else if (opcode == BPF_EXIT) {
 2230			return -ENOTSUPP;
 2231		}
 2232	} else if (class == BPF_LD) {
 2233		if (!(*reg_mask & dreg))
 2234			return 0;
 2235		*reg_mask &= ~dreg;
 2236		/* It's ld_imm64 or ld_abs or ld_ind.
 2237		 * For ld_imm64 no further tracking of precision
 2238		 * into parent is necessary
 2239		 */
 2240		if (mode == BPF_IND || mode == BPF_ABS)
 2241			/* to be analyzed */
 2242			return -ENOTSUPP;
 2243	}
 2244	return 0;
 2245}
 2246
 2247/* the scalar precision tracking algorithm:
 2248 * . at the start all registers have precise=false.
 2249 * . scalar ranges are tracked as normal through alu and jmp insns.
 2250 * . once precise value of the scalar register is used in:
 2251 *   .  ptr + scalar alu
 2252 *   . if (scalar cond K|scalar)
 2253 *   .  helper_call(.., scalar, ...) where ARG_CONST is expected
 2254 *   backtrack through the verifier states and mark all registers and
 2255 *   stack slots with spilled constants that these scalar regisers
 2256 *   should be precise.
 2257 * . during state pruning two registers (or spilled stack slots)
 2258 *   are equivalent if both are not precise.
 2259 *
 2260 * Note the verifier cannot simply walk register parentage chain,
 2261 * since many different registers and stack slots could have been
 2262 * used to compute single precise scalar.
 2263 *
 2264 * The approach of starting with precise=true for all registers and then
 2265 * backtrack to mark a register as not precise when the verifier detects
 2266 * that program doesn't care about specific value (e.g., when helper
 2267 * takes register as ARG_ANYTHING parameter) is not safe.
 2268 *
 2269 * It's ok to walk single parentage chain of the verifier states.
 2270 * It's possible that this backtracking will go all the way till 1st insn.
 2271 * All other branches will be explored for needing precision later.
 2272 *
 2273 * The backtracking needs to deal with cases like:
 2274 *   R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0)
 2275 * r9 -= r8
 2276 * r5 = r9
 2277 * if r5 > 0x79f goto pc+7
 2278 *    R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
 2279 * r5 += 1
 2280 * ...
 2281 * call bpf_perf_event_output#25
 2282 *   where .arg5_type = ARG_CONST_SIZE_OR_ZERO
 2283 *
 2284 * and this case:
 2285 * r6 = 1
 2286 * call foo // uses callee's r6 inside to compute r0
 2287 * r0 += r6
 2288 * if r0 == 0 goto
 2289 *
 2290 * to track above reg_mask/stack_mask needs to be independent for each frame.
 2291 *
 2292 * Also if parent's curframe > frame where backtracking started,
 2293 * the verifier need to mark registers in both frames, otherwise callees
 2294 * may incorrectly prune callers. This is similar to
 2295 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
 2296 *
 2297 * For now backtracking falls back into conservative marking.
 2298 */
 2299static void mark_all_scalars_precise(struct bpf_verifier_env *env,
 2300				     struct bpf_verifier_state *st)
 2301{
 2302	struct bpf_func_state *func;
 2303	struct bpf_reg_state *reg;
 2304	int i, j;
 2305
 2306	/* big hammer: mark all scalars precise in this path.
 2307	 * pop_stack may still get !precise scalars.
 2308	 */
 2309	for (; st; st = st->parent)
 2310		for (i = 0; i <= st->curframe; i++) {
 2311			func = st->frame[i];
 2312			for (j = 0; j < BPF_REG_FP; j++) {
 2313				reg = &func->regs[j];
 2314				if (reg->type != SCALAR_VALUE)
 2315					continue;
 2316				reg->precise = true;
 2317			}
 2318			for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
 2319				if (func->stack[j].slot_type[0] != STACK_SPILL)
 2320					continue;
 2321				reg = &func->stack[j].spilled_ptr;
 2322				if (reg->type != SCALAR_VALUE)
 2323					continue;
 2324				reg->precise = true;
 2325			}
 2326		}
 2327}
 2328
 2329static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
 2330				  int spi)
 2331{
 2332	struct bpf_verifier_state *st = env->cur_state;
 2333	int first_idx = st->first_insn_idx;
 2334	int last_idx = env->insn_idx;
 2335	struct bpf_func_state *func;
 2336	struct bpf_reg_state *reg;
 2337	u32 reg_mask = regno >= 0 ? 1u << regno : 0;
 2338	u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
 2339	bool skip_first = true;
 2340	bool new_marks = false;
 2341	int i, err;
 2342
 2343	if (!env->bpf_capable)
 2344		return 0;
 2345
 2346	func = st->frame[st->curframe];
 2347	if (regno >= 0) {
 2348		reg = &func->regs[regno];
 2349		if (reg->type != SCALAR_VALUE) {
 2350			WARN_ONCE(1, "backtracing misuse");
 2351			return -EFAULT;
 2352		}
 2353		if (!reg->precise)
 2354			new_marks = true;
 2355		else
 2356			reg_mask = 0;
 2357		reg->precise = true;
 2358	}
 2359
 2360	while (spi >= 0) {
 2361		if (func->stack[spi].slot_type[0] != STACK_SPILL) {
 2362			stack_mask = 0;
 2363			break;
 2364		}
 2365		reg = &func->stack[spi].spilled_ptr;
 2366		if (reg->type != SCALAR_VALUE) {
 2367			stack_mask = 0;
 2368			break;
 2369		}
 2370		if (!reg->precise)
 2371			new_marks = true;
 2372		else
 2373			stack_mask = 0;
 2374		reg->precise = true;
 2375		break;
 2376	}
 2377
 2378	if (!new_marks)
 2379		return 0;
 2380	if (!reg_mask && !stack_mask)
 2381		return 0;
 2382	for (;;) {
 2383		DECLARE_BITMAP(mask, 64);
 2384		u32 history = st->jmp_history_cnt;
 2385
 2386		if (env->log.level & BPF_LOG_LEVEL)
 2387			verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
 2388		for (i = last_idx;;) {
 2389			if (skip_first) {
 2390				err = 0;
 2391				skip_first = false;
 2392			} else {
 2393				err = backtrack_insn(env, i, &reg_mask, &stack_mask);
 2394			}
 2395			if (err == -ENOTSUPP) {
 2396				mark_all_scalars_precise(env, st);
 2397				return 0;
 2398			} else if (err) {
 2399				return err;
 2400			}
 2401			if (!reg_mask && !stack_mask)
 2402				/* Found assignment(s) into tracked register in this state.
 2403				 * Since this state is already marked, just return.
 2404				 * Nothing to be tracked further in the parent state.
 2405				 */
 2406				return 0;
 2407			if (i == first_idx)
 2408				break;
 2409			i = get_prev_insn_idx(st, i, &history);
 2410			if (i >= env->prog->len) {
 2411				/* This can happen if backtracking reached insn 0
 2412				 * and there are still reg_mask or stack_mask
 2413				 * to backtrack.
 2414				 * It means the backtracking missed the spot where
 2415				 * particular register was initialized with a constant.
 2416				 */
 2417				verbose(env, "BUG backtracking idx %d\n", i);
 2418				WARN_ONCE(1, "verifier backtracking bug");
 2419				return -EFAULT;
 2420			}
 2421		}
 2422		st = st->parent;
 2423		if (!st)
 2424			break;
 2425
 2426		new_marks = false;
 2427		func = st->frame[st->curframe];
 2428		bitmap_from_u64(mask, reg_mask);
 2429		for_each_set_bit(i, mask, 32) {
 2430			reg = &func->regs[i];
 2431			if (reg->type != SCALAR_VALUE) {
 2432				reg_mask &= ~(1u << i);
 2433				continue;
 2434			}
 2435			if (!reg->precise)
 2436				new_marks = true;
 2437			reg->precise = true;
 2438		}
 2439
 2440		bitmap_from_u64(mask, stack_mask);
 2441		for_each_set_bit(i, mask, 64) {
 2442			if (i >= func->allocated_stack / BPF_REG_SIZE) {
 2443				/* the sequence of instructions:
 2444				 * 2: (bf) r3 = r10
 2445				 * 3: (7b) *(u64 *)(r3 -8) = r0
 2446				 * 4: (79) r4 = *(u64 *)(r10 -8)
 2447				 * doesn't contain jmps. It's backtracked
 2448				 * as a single block.
 2449				 * During backtracking insn 3 is not recognized as
 2450				 * stack access, so at the end of backtracking
 2451				 * stack slot fp-8 is still marked in stack_mask.
 2452				 * However the parent state may not have accessed
 2453				 * fp-8 and it's "unallocated" stack space.
 2454				 * In such case fallback to conservative.
 2455				 */
 2456				mark_all_scalars_precise(env, st);
 2457				return 0;
 2458			}
 2459
 2460			if (func->stack[i].slot_type[0] != STACK_SPILL) {
 2461				stack_mask &= ~(1ull << i);
 2462				continue;
 2463			}
 2464			reg = &func->stack[i].spilled_ptr;
 2465			if (reg->type != SCALAR_VALUE) {
 2466				stack_mask &= ~(1ull << i);
 2467				continue;
 2468			}
 2469			if (!reg->precise)
 2470				new_marks = true;
 2471			reg->precise = true;
 2472		}
 2473		if (env->log.level & BPF_LOG_LEVEL) {
 2474			print_verifier_state(env, func);
 2475			verbose(env, "parent %s regs=%x stack=%llx marks\n",
 2476				new_marks ? "didn't have" : "already had",
 2477				reg_mask, stack_mask);
 2478		}
 2479
 2480		if (!reg_mask && !stack_mask)
 2481			break;
 2482		if (!new_marks)
 2483			break;
 2484
 2485		last_idx = st->last_insn_idx;
 2486		first_idx = st->first_insn_idx;
 2487	}
 2488	return 0;
 2489}
 2490
 2491static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
 2492{
 2493	return __mark_chain_precision(env, regno, -1);
 2494}
 2495
 2496static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
 2497{
 2498	return __mark_chain_precision(env, -1, spi);
 2499}
 2500
 2501static bool is_spillable_regtype(enum bpf_reg_type type)
 2502{
 2503	switch (type) {
 2504	case PTR_TO_MAP_VALUE:
 2505	case PTR_TO_MAP_VALUE_OR_NULL:
 2506	case PTR_TO_STACK:
 2507	case PTR_TO_CTX:
 2508	case PTR_TO_PACKET:
 2509	case PTR_TO_PACKET_META:
 2510	case PTR_TO_PACKET_END:
 2511	case PTR_TO_FLOW_KEYS:
 2512	case CONST_PTR_TO_MAP:
 2513	case PTR_TO_SOCKET:
 2514	case PTR_TO_SOCKET_OR_NULL:
 2515	case PTR_TO_SOCK_COMMON:
 2516	case PTR_TO_SOCK_COMMON_OR_NULL:
 2517	case PTR_TO_TCP_SOCK:
 2518	case PTR_TO_TCP_SOCK_OR_NULL:
 2519	case PTR_TO_XDP_SOCK:
 2520	case PTR_TO_BTF_ID:
 2521	case PTR_TO_BTF_ID_OR_NULL:
 2522	case PTR_TO_RDONLY_BUF:
 2523	case PTR_TO_RDONLY_BUF_OR_NULL:
 2524	case PTR_TO_RDWR_BUF:
 2525	case PTR_TO_RDWR_BUF_OR_NULL:
 2526	case PTR_TO_PERCPU_BTF_ID:
 2527	case PTR_TO_MEM:
 2528	case PTR_TO_MEM_OR_NULL:
 2529	case PTR_TO_FUNC:
 2530	case PTR_TO_MAP_KEY:
 2531		return true;
 2532	default:
 2533		return false;
 2534	}
 2535}
 2536
 2537/* Does this register contain a constant zero? */
 2538static bool register_is_null(struct bpf_reg_state *reg)
 2539{
 2540	return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
 2541}
 2542
 2543static bool register_is_const(struct bpf_reg_state *reg)
 2544{
 2545	return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
 2546}
 2547
 2548static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
 2549{
 2550	return tnum_is_unknown(reg->var_off) &&
 2551	       reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
 2552	       reg->umin_value == 0 && reg->umax_value == U64_MAX &&
 2553	       reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
 2554	       reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
 2555}
 2556
 2557static bool register_is_bounded(struct bpf_reg_state *reg)
 2558{
 2559	return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
 2560}
 2561
 2562static bool __is_pointer_value(bool allow_ptr_leaks,
 2563			       const struct bpf_reg_state *reg)
 2564{
 2565	if (allow_ptr_leaks)
 2566		return false;
 2567
 2568	return reg->type != SCALAR_VALUE;
 2569}
 2570
 2571static void save_register_state(struct bpf_func_state *state,
 2572				int spi, struct bpf_reg_state *reg)
 2573{
 2574	int i;
 2575
 2576	state->stack[spi].spilled_ptr = *reg;
 2577	state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
 2578
 2579	for (i = 0; i < BPF_REG_SIZE; i++)
 2580		state->stack[spi].slot_type[i] = STACK_SPILL;
 2581}
 2582
 2583/* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
 2584 * stack boundary and alignment are checked in check_mem_access()
 2585 */
 2586static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
 2587				       /* stack frame we're writing to */
 2588				       struct bpf_func_state *state,
 2589				       int off, int size, int value_regno,
 2590				       int insn_idx)
 2591{
 2592	struct bpf_func_state *cur; /* state of the current function */
 2593	int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
 2594	u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
 2595	struct bpf_reg_state *reg = NULL;
 2596
 2597	err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
 2598	if (err)
 2599		return err;
 2600	/* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
 2601	 * so it's aligned access and [off, off + size) are within stack limits
 2602	 */
 2603	if (!env->allow_ptr_leaks &&
 2604	    state->stack[spi].slot_type[0] == STACK_SPILL &&
 2605	    size != BPF_REG_SIZE) {
 2606		verbose(env, "attempt to corrupt spilled pointer on stack\n");
 2607		return -EACCES;
 2608	}
 2609
 2610	cur = env->cur_state->frame[env->cur_state->curframe];
 2611	if (value_regno >= 0)
 2612		reg = &cur->regs[value_regno];
 2613	if (!env->bypass_spec_v4) {
 2614		bool sanitize = reg && is_spillable_regtype(reg->type);
 2615
 2616		for (i = 0; i < size; i++) {
 2617			if (state->stack[spi].slot_type[i] == STACK_INVALID) {
 2618				sanitize = true;
 2619				break;
 2620			}
 2621		}
 2622
 2623		if (sanitize)
 2624			env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
 2625	}
 2626
 2627	if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
 2628	    !register_is_null(reg) && env->bpf_capable) {
 2629		if (dst_reg != BPF_REG_FP) {
 2630			/* The backtracking logic can only recognize explicit
 2631			 * stack slot address like [fp - 8]. Other spill of
 2632			 * scalar via different register has to be conservative.
 2633			 * Backtrack from here and mark all registers as precise
 2634			 * that contributed into 'reg' being a constant.
 2635			 */
 2636			err = mark_chain_precision(env, value_regno);
 2637			if (err)
 2638				return err;
 2639		}
 2640		save_register_state(state, spi, reg);
 2641	} else if (reg && is_spillable_regtype(reg->type)) {
 2642		/* register containing pointer is being spilled into stack */
 2643		if (size != BPF_REG_SIZE) {
 2644			verbose_linfo(env, insn_idx, "; ");
 2645			verbose(env, "invalid size of register spill\n");
 2646			return -EACCES;
 2647		}
 2648		if (state != cur && reg->type == PTR_TO_STACK) {
 2649			verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
 2650			return -EINVAL;
 2651		}
 2652		save_register_state(state, spi, reg);
 2653	} else {
 2654		u8 type = STACK_MISC;
 2655
 2656		/* regular write of data into stack destroys any spilled ptr */
 2657		state->stack[spi].spilled_ptr.type = NOT_INIT;
 2658		/* Mark slots as STACK_MISC if they belonged to spilled ptr. */
 2659		if (state->stack[spi].slot_type[0] == STACK_SPILL)
 2660			for (i = 0; i < BPF_REG_SIZE; i++)
 2661				state->stack[spi].slot_type[i] = STACK_MISC;
 2662
 2663		/* only mark the slot as written if all 8 bytes were written
 2664		 * otherwise read propagation may incorrectly stop too soon
 2665		 * when stack slots are partially written.
 2666		 * This heuristic means that read propagation will be
 2667		 * conservative, since it will add reg_live_read marks
 2668		 * to stack slots all the way to first state when programs
 2669		 * writes+reads less than 8 bytes
 2670		 */
 2671		if (size == BPF_REG_SIZE)
 2672			state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
 2673
 2674		/* when we zero initialize stack slots mark them as such */
 2675		if (reg && register_is_null(reg)) {
 2676			/* backtracking doesn't work for STACK_ZERO yet. */
 2677			err = mark_chain_precision(env, value_regno);
 2678			if (err)
 2679				return err;
 2680			type = STACK_ZERO;
 2681		}
 2682
 2683		/* Mark slots affected by this stack write. */
 2684		for (i = 0; i < size; i++)
 2685			state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
 2686				type;
 2687	}
 2688	return 0;
 2689}
 2690
 2691/* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
 2692 * known to contain a variable offset.
 2693 * This function checks whether the write is permitted and conservatively
 2694 * tracks the effects of the write, considering that each stack slot in the
 2695 * dynamic range is potentially written to.
 2696 *
 2697 * 'off' includes 'regno->off'.
 2698 * 'value_regno' can be -1, meaning that an unknown value is being written to
 2699 * the stack.
 2700 *
 2701 * Spilled pointers in range are not marked as written because we don't know
 2702 * what's going to be actually written. This means that read propagation for
 2703 * future reads cannot be terminated by this write.
 2704 *
 2705 * For privileged programs, uninitialized stack slots are considered
 2706 * initialized by this write (even though we don't know exactly what offsets
 2707 * are going to be written to). The idea is that we don't want the verifier to
 2708 * reject future reads that access slots written to through variable offsets.
 2709 */
 2710static int check_stack_write_var_off(struct bpf_verifier_env *env,
 2711				     /* func where register points to */
 2712				     struct bpf_func_state *state,
 2713				     int ptr_regno, int off, int size,
 2714				     int value_regno, int insn_idx)
 2715{
 2716	struct bpf_func_state *cur; /* state of the current function */
 2717	int min_off, max_off;
 2718	int i, err;
 2719	struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
 2720	bool writing_zero = false;
 2721	/* set if the fact that we're writing a zero is used to let any
 2722	 * stack slots remain STACK_ZERO
 2723	 */
 2724	bool zero_used = false;
 2725
 2726	cur = env->cur_state->frame[env->cur_state->curframe];
 2727	ptr_reg = &cur->regs[ptr_regno];
 2728	min_off = ptr_reg->smin_value + off;
 2729	max_off = ptr_reg->smax_value + off + size;
 2730	if (value_regno >= 0)
 2731		value_reg = &cur->regs[value_regno];
 2732	if (value_reg && register_is_null(value_reg))
 2733		writing_zero = true;
 2734
 2735	err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
 2736	if (err)
 2737		return err;
 2738
 2739
 2740	/* Variable offset writes destroy any spilled pointers in range. */
 2741	for (i = min_off; i < max_off; i++) {
 2742		u8 new_type, *stype;
 2743		int slot, spi;
 2744
 2745		slot = -i - 1;
 2746		spi = slot / BPF_REG_SIZE;
 2747		stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
 2748
 2749		if (!env->allow_ptr_leaks
 2750				&& *stype != NOT_INIT
 2751				&& *stype != SCALAR_VALUE) {
 2752			/* Reject the write if there's are spilled pointers in
 2753			 * range. If we didn't reject here, the ptr status
 2754			 * would be erased below (even though not all slots are
 2755			 * actually overwritten), possibly opening the door to
 2756			 * leaks.
 2757			 */
 2758			verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
 2759				insn_idx, i);
 2760			return -EINVAL;
 2761		}
 2762
 2763		/* Erase all spilled pointers. */
 2764		state->stack[spi].spilled_ptr.type = NOT_INIT;
 2765
 2766		/* Update the slot type. */
 2767		new_type = STACK_MISC;
 2768		if (writing_zero && *stype == STACK_ZERO) {
 2769			new_type = STACK_ZERO;
 2770			zero_used = true;
 2771		}
 2772		/* If the slot is STACK_INVALID, we check whether it's OK to
 2773		 * pretend that it will be initialized by this write. The slot
 2774		 * might not actually be written to, and so if we mark it as
 2775		 * initialized future reads might leak uninitialized memory.
 2776		 * For privileged programs, we will accept such reads to slots
 2777		 * that may or may not be written because, if we're reject
 2778		 * them, the error would be too confusing.
 2779		 */
 2780		if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
 2781			verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
 2782					insn_idx, i);
 2783			return -EINVAL;
 2784		}
 2785		*stype = new_type;
 2786	}
 2787	if (zero_used) {
 2788		/* backtracking doesn't work for STACK_ZERO yet. */
 2789		err = mark_chain_precision(env, value_regno);
 2790		if (err)
 2791			return err;
 2792	}
 2793	return 0;
 2794}
 2795
 2796/* When register 'dst_regno' is assigned some values from stack[min_off,
 2797 * max_off), we set the register's type according to the types of the
 2798 * respective stack slots. If all the stack values are known to be zeros, then
 2799 * so is the destination reg. Otherwise, the register is considered to be
 2800 * SCALAR. This function does not deal with register filling; the caller must
 2801 * ensure that all spilled registers in the stack range have been marked as
 2802 * read.
 2803 */
 2804static void mark_reg_stack_read(struct bpf_verifier_env *env,
 2805				/* func where src register points to */
 2806				struct bpf_func_state *ptr_state,
 2807				int min_off, int max_off, int dst_regno)
 2808{
 2809	struct bpf_verifier_state *vstate = env->cur_state;
 2810	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 2811	int i, slot, spi;
 2812	u8 *stype;
 2813	int zeros = 0;
 2814
 2815	for (i = min_off; i < max_off; i++) {
 2816		slot = -i - 1;
 2817		spi = slot / BPF_REG_SIZE;
 2818		stype = ptr_state->stack[spi].slot_type;
 2819		if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
 2820			break;
 2821		zeros++;
 2822	}
 2823	if (zeros == max_off - min_off) {
 2824		/* any access_size read into register is zero extended,
 2825		 * so the whole register == const_zero
 2826		 */
 2827		__mark_reg_const_zero(&state->regs[dst_regno]);
 2828		/* backtracking doesn't support STACK_ZERO yet,
 2829		 * so mark it precise here, so that later
 2830		 * backtracking can stop here.
 2831		 * Backtracking may not need this if this register
 2832		 * doesn't participate in pointer adjustment.
 2833		 * Forward propagation of precise flag is not
 2834		 * necessary either. This mark is only to stop
 2835		 * backtracking. Any register that contributed
 2836		 * to const 0 was marked precise before spill.
 2837		 */
 2838		state->regs[dst_regno].precise = true;
 2839	} else {
 2840		/* have read misc data from the stack */
 2841		mark_reg_unknown(env, state->regs, dst_regno);
 2842	}
 2843	state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
 2844}
 2845
 2846/* Read the stack at 'off' and put the results into the register indicated by
 2847 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
 2848 * spilled reg.
 2849 *
 2850 * 'dst_regno' can be -1, meaning that the read value is not going to a
 2851 * register.
 2852 *
 2853 * The access is assumed to be within the current stack bounds.
 2854 */
 2855static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
 2856				      /* func where src register points to */
 2857				      struct bpf_func_state *reg_state,
 2858				      int off, int size, int dst_regno)
 2859{
 2860	struct bpf_verifier_state *vstate = env->cur_state;
 2861	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 2862	int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
 2863	struct bpf_reg_state *reg;
 2864	u8 *stype;
 2865
 2866	stype = reg_state->stack[spi].slot_type;
 2867	reg = &reg_state->stack[spi].spilled_ptr;
 2868
 2869	if (stype[0] == STACK_SPILL) {
 2870		if (size != BPF_REG_SIZE) {
 2871			if (reg->type != SCALAR_VALUE) {
 2872				verbose_linfo(env, env->insn_idx, "; ");
 2873				verbose(env, "invalid size of register fill\n");
 2874				return -EACCES;
 2875			}
 2876			if (dst_regno >= 0) {
 2877				mark_reg_unknown(env, state->regs, dst_regno);
 2878				state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
 2879			}
 2880			mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
 2881			return 0;
 2882		}
 2883		for (i = 1; i < BPF_REG_SIZE; i++) {
 2884			if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
 2885				verbose(env, "corrupted spill memory\n");
 2886				return -EACCES;
 2887			}
 2888		}
 2889
 2890		if (dst_regno >= 0) {
 2891			/* restore register state from stack */
 2892			state->regs[dst_regno] = *reg;
 2893			/* mark reg as written since spilled pointer state likely
 2894			 * has its liveness marks cleared by is_state_visited()
 2895			 * which resets stack/reg liveness for state transitions
 2896			 */
 2897			state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
 2898		} else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
 2899			/* If dst_regno==-1, the caller is asking us whether
 2900			 * it is acceptable to use this value as a SCALAR_VALUE
 2901			 * (e.g. for XADD).
 2902			 * We must not allow unprivileged callers to do that
 2903			 * with spilled pointers.
 2904			 */
 2905			verbose(env, "leaking pointer from stack off %d\n",
 2906				off);
 2907			return -EACCES;
 2908		}
 2909		mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
 2910	} else {
 2911		u8 type;
 2912
 2913		for (i = 0; i < size; i++) {
 2914			type = stype[(slot - i) % BPF_REG_SIZE];
 2915			if (type == STACK_MISC)
 2916				continue;
 2917			if (type == STACK_ZERO)
 2918				continue;
 2919			verbose(env, "invalid read from stack off %d+%d size %d\n",
 2920				off, i, size);
 2921			return -EACCES;
 2922		}
 2923		mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
 2924		if (dst_regno >= 0)
 2925			mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
 2926	}
 2927	return 0;
 2928}
 2929
 2930enum stack_access_src {
 2931	ACCESS_DIRECT = 1,  /* the access is performed by an instruction */
 2932	ACCESS_HELPER = 2,  /* the access is performed by a helper */
 2933};
 2934
 2935static int check_stack_range_initialized(struct bpf_verifier_env *env,
 2936					 int regno, int off, int access_size,
 2937					 bool zero_size_allowed,
 2938					 enum stack_access_src type,
 2939					 struct bpf_call_arg_meta *meta);
 2940
 2941static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
 2942{
 2943	return cur_regs(env) + regno;
 2944}
 2945
 2946/* Read the stack at 'ptr_regno + off' and put the result into the register
 2947 * 'dst_regno'.
 2948 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
 2949 * but not its variable offset.
 2950 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
 2951 *
 2952 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
 2953 * filling registers (i.e. reads of spilled register cannot be detected when
 2954 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
 2955 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
 2956 * offset; for a fixed offset check_stack_read_fixed_off should be used
 2957 * instead.
 2958 */
 2959static int check_stack_read_var_off(struct bpf_verifier_env *env,
 2960				    int ptr_regno, int off, int size, int dst_regno)
 2961{
 2962	/* The state of the source register. */
 2963	struct bpf_reg_state *reg = reg_state(env, ptr_regno);
 2964	struct bpf_func_state *ptr_state = func(env, reg);
 2965	int err;
 2966	int min_off, max_off;
 2967
 2968	/* Note that we pass a NULL meta, so raw access will not be permitted.
 2969	 */
 2970	err = check_stack_range_initialized(env, ptr_regno, off, size,
 2971					    false, ACCESS_DIRECT, NULL);
 2972	if (err)
 2973		return err;
 2974
 2975	min_off = reg->smin_value + off;
 2976	max_off = reg->smax_value + off;
 2977	mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
 2978	return 0;
 2979}
 2980
 2981/* check_stack_read dispatches to check_stack_read_fixed_off or
 2982 * check_stack_read_var_off.
 2983 *
 2984 * The caller must ensure that the offset falls within the allocated stack
 2985 * bounds.
 2986 *
 2987 * 'dst_regno' is a register which will receive the value from the stack. It
 2988 * can be -1, meaning that the read value is not going to a register.
 2989 */
 2990static int check_stack_read(struct bpf_verifier_env *env,
 2991			    int ptr_regno, int off, int size,
 2992			    int dst_regno)
 2993{
 2994	struct bpf_reg_state *reg = reg_state(env, ptr_regno);
 2995	struct bpf_func_state *state = func(env, reg);
 2996	int err;
 2997	/* Some accesses are only permitted with a static offset. */
 2998	bool var_off = !tnum_is_const(reg->var_off);
 2999
 3000	/* The offset is required to be static when reads don't go to a
 3001	 * register, in order to not leak pointers (see
 3002	 * check_stack_read_fixed_off).
 3003	 */
 3004	if (dst_regno < 0 && var_off) {
 3005		char tn_buf[48];
 3006
 3007		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 3008		verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
 3009			tn_buf, off, size);
 3010		return -EACCES;
 3011	}
 3012	/* Variable offset is prohibited for unprivileged mode for simplicity
 3013	 * since it requires corresponding support in Spectre masking for stack
 3014	 * ALU. See also retrieve_ptr_limit().
 3015	 */
 3016	if (!env->bypass_spec_v1 && var_off) {
 3017		char tn_buf[48];
 3018
 3019		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 3020		verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
 3021				ptr_regno, tn_buf);
 3022		return -EACCES;
 3023	}
 3024
 3025	if (!var_off) {
 3026		off += reg->var_off.value;
 3027		err = check_stack_read_fixed_off(env, state, off, size,
 3028						 dst_regno);
 3029	} else {
 3030		/* Variable offset stack reads need more conservative handling
 3031		 * than fixed offset ones. Note that dst_regno >= 0 on this
 3032		 * branch.
 3033		 */
 3034		err = check_stack_read_var_off(env, ptr_regno, off, size,
 3035					       dst_regno);
 3036	}
 3037	return err;
 3038}
 3039
 3040
 3041/* check_stack_write dispatches to check_stack_write_fixed_off or
 3042 * check_stack_write_var_off.
 3043 *
 3044 * 'ptr_regno' is the register used as a pointer into the stack.
 3045 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
 3046 * 'value_regno' is the register whose value we're writing to the stack. It can
 3047 * be -1, meaning that we're not writing from a register.
 3048 *
 3049 * The caller must ensure that the offset falls within the maximum stack size.
 3050 */
 3051static int check_stack_write(struct bpf_verifier_env *env,
 3052			     int ptr_regno, int off, int size,
 3053			     int value_regno, int insn_idx)
 3054{
 3055	struct bpf_reg_state *reg = reg_state(env, ptr_regno);
 3056	struct bpf_func_state *state = func(env, reg);
 3057	int err;
 3058
 3059	if (tnum_is_const(reg->var_off)) {
 3060		off += reg->var_off.value;
 3061		err = check_stack_write_fixed_off(env, state, off, size,
 3062						  value_regno, insn_idx);
 3063	} else {
 3064		/* Variable offset stack reads need more conservative handling
 3065		 * than fixed offset ones.
 3066		 */
 3067		err = check_stack_write_var_off(env, state,
 3068						ptr_regno, off, size,
 3069						value_regno, insn_idx);
 3070	}
 3071	return err;
 3072}
 3073
 3074static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
 3075				 int off, int size, enum bpf_access_type type)
 3076{
 3077	struct bpf_reg_state *regs = cur_regs(env);
 3078	struct bpf_map *map = regs[regno].map_ptr;
 3079	u32 cap = bpf_map_flags_to_cap(map);
 3080
 3081	if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
 3082		verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
 3083			map->value_size, off, size);
 3084		return -EACCES;
 3085	}
 3086
 3087	if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
 3088		verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
 3089			map->value_size, off, size);
 3090		return -EACCES;
 3091	}
 3092
 3093	return 0;
 3094}
 3095
 3096/* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
 3097static int __check_mem_access(struct bpf_verifier_env *env, int regno,
 3098			      int off, int size, u32 mem_size,
 3099			      bool zero_size_allowed)
 3100{
 3101	bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
 3102	struct bpf_reg_state *reg;
 3103
 3104	if (off >= 0 && size_ok && (u64)off + size <= mem_size)
 3105		return 0;
 3106
 3107	reg = &cur_regs(env)[regno];
 3108	switch (reg->type) {
 3109	case PTR_TO_MAP_KEY:
 3110		verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
 3111			mem_size, off, size);
 3112		break;
 3113	case PTR_TO_MAP_VALUE:
 3114		verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
 3115			mem_size, off, size);
 3116		break;
 3117	case PTR_TO_PACKET:
 3118	case PTR_TO_PACKET_META:
 3119	case PTR_TO_PACKET_END:
 3120		verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
 3121			off, size, regno, reg->id, off, mem_size);
 3122		break;
 3123	case PTR_TO_MEM:
 3124	default:
 3125		verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
 3126			mem_size, off, size);
 3127	}
 3128
 3129	return -EACCES;
 3130}
 3131
 3132/* check read/write into a memory region with possible variable offset */
 3133static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
 3134				   int off, int size, u32 mem_size,
 3135				   bool zero_size_allowed)
 3136{
 3137	struct bpf_verifier_state *vstate = env->cur_state;
 3138	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 3139	struct bpf_reg_state *reg = &state->regs[regno];
 3140	int err;
 3141
 3142	/* We may have adjusted the register pointing to memory region, so we
 3143	 * need to try adding each of min_value and max_value to off
 3144	 * to make sure our theoretical access will be safe.
 3145	 */
 3146	if (env->log.level & BPF_LOG_LEVEL)
 3147		print_verifier_state(env, state);
 3148
 3149	/* The minimum value is only important with signed
 3150	 * comparisons where we can't assume the floor of a
 3151	 * value is 0.  If we are using signed variables for our
 3152	 * index'es we need to make sure that whatever we use
 3153	 * will have a set floor within our range.
 3154	 */
 3155	if (reg->smin_value < 0 &&
 3156	    (reg->smin_value == S64_MIN ||
 3157	     (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
 3158	      reg->smin_value + off < 0)) {
 3159		verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
 3160			regno);
 3161		return -EACCES;
 3162	}
 3163	err = __check_mem_access(env, regno, reg->smin_value + off, size,
 3164				 mem_size, zero_size_allowed);
 3165	if (err) {
 3166		verbose(env, "R%d min value is outside of the allowed memory range\n",
 3167			regno);
 3168		return err;
 3169	}
 3170
 3171	/* If we haven't set a max value then we need to bail since we can't be
 3172	 * sure we won't do bad things.
 3173	 * If reg->umax_value + off could overflow, treat that as unbounded too.
 3174	 */
 3175	if (reg->umax_value >= BPF_MAX_VAR_OFF) {
 3176		verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
 3177			regno);
 3178		return -EACCES;
 3179	}
 3180	err = __check_mem_access(env, regno, reg->umax_value + off, size,
 3181				 mem_size, zero_size_allowed);
 3182	if (err) {
 3183		verbose(env, "R%d max value is outside of the allowed memory range\n",
 3184			regno);
 3185		return err;
 3186	}
 3187
 3188	return 0;
 3189}
 3190
 3191/* check read/write into a map element with possible variable offset */
 3192static int check_map_access(struct bpf_verifier_env *env, u32 regno,
 3193			    int off, int size, bool zero_size_allowed)
 3194{
 3195	struct bpf_verifier_state *vstate = env->cur_state;
 3196	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 3197	struct bpf_reg_state *reg = &state->regs[regno];
 3198	struct bpf_map *map = reg->map_ptr;
 3199	int err;
 3200
 3201	err = check_mem_region_access(env, regno, off, size, map->value_size,
 3202				      zero_size_allowed);
 3203	if (err)
 3204		return err;
 3205
 3206	if (map_value_has_spin_lock(map)) {
 3207		u32 lock = map->spin_lock_off;
 3208
 3209		/* if any part of struct bpf_spin_lock can be touched by
 3210		 * load/store reject this program.
 3211		 * To check that [x1, x2) overlaps with [y1, y2)
 3212		 * it is sufficient to check x1 < y2 && y1 < x2.
 3213		 */
 3214		if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
 3215		     lock < reg->umax_value + off + size) {
 3216			verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
 3217			return -EACCES;
 3218		}
 3219	}
 3220	return err;
 3221}
 3222
 3223#define MAX_PACKET_OFF 0xffff
 3224
 3225static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
 3226{
 3227	return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
 3228}
 3229
 3230static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
 3231				       const struct bpf_call_arg_meta *meta,
 3232				       enum bpf_access_type t)
 3233{
 3234	enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
 3235
 3236	switch (prog_type) {
 3237	/* Program types only with direct read access go here! */
 3238	case BPF_PROG_TYPE_LWT_IN:
 3239	case BPF_PROG_TYPE_LWT_OUT:
 3240	case BPF_PROG_TYPE_LWT_SEG6LOCAL:
 3241	case BPF_PROG_TYPE_SK_REUSEPORT:
 3242	case BPF_PROG_TYPE_FLOW_DISSECTOR:
 3243	case BPF_PROG_TYPE_CGROUP_SKB:
 3244		if (t == BPF_WRITE)
 3245			return false;
 3246		fallthrough;
 3247
 3248	/* Program types with direct read + write access go here! */
 3249	case BPF_PROG_TYPE_SCHED_CLS:
 3250	case BPF_PROG_TYPE_SCHED_ACT:
 3251	case BPF_PROG_TYPE_XDP:
 3252	case BPF_PROG_TYPE_LWT_XMIT:
 3253	case BPF_PROG_TYPE_SK_SKB:
 3254	case BPF_PROG_TYPE_SK_MSG:
 3255		if (meta)
 3256			return meta->pkt_access;
 3257
 3258		env->seen_direct_write = true;
 3259		return true;
 3260
 3261	case BPF_PROG_TYPE_CGROUP_SOCKOPT:
 3262		if (t == BPF_WRITE)
 3263			env->seen_direct_write = true;
 3264
 3265		return true;
 3266
 3267	default:
 3268		return false;
 3269	}
 3270}
 3271
 3272static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
 3273			       int size, bool zero_size_allowed)
 3274{
 3275	struct bpf_reg_state *regs = cur_regs(env);
 3276	struct bpf_reg_state *reg = &regs[regno];
 3277	int err;
 3278
 3279	/* We may have added a variable offset to the packet pointer; but any
 3280	 * reg->range we have comes after that.  We are only checking the fixed
 3281	 * offset.
 3282	 */
 3283
 3284	/* We don't allow negative numbers, because we aren't tracking enough
 3285	 * detail to prove they're safe.
 3286	 */
 3287	if (reg->smin_value < 0) {
 3288		verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
 3289			regno);
 3290		return -EACCES;
 3291	}
 3292
 3293	err = reg->range < 0 ? -EINVAL :
 3294	      __check_mem_access(env, regno, off, size, reg->range,
 3295				 zero_size_allowed);
 3296	if (err) {
 3297		verbose(env, "R%d offset is outside of the packet\n", regno);
 3298		return err;
 3299	}
 3300
 3301	/* __check_mem_access has made sure "off + size - 1" is within u16.
 3302	 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
 3303	 * otherwise find_good_pkt_pointers would have refused to set range info
 3304	 * that __check_mem_access would have rejected this pkt access.
 3305	 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
 3306	 */
 3307	env->prog->aux->max_pkt_offset =
 3308		max_t(u32, env->prog->aux->max_pkt_offset,
 3309		      off + reg->umax_value + size - 1);
 3310
 3311	return err;
 3312}
 3313
 3314/* check access to 'struct bpf_context' fields.  Supports fixed offsets only */
 3315static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
 3316			    enum bpf_access_type t, enum bpf_reg_type *reg_type,
 3317			    struct btf **btf, u32 *btf_id)
 3318{
 3319	struct bpf_insn_access_aux info = {
 3320		.reg_type = *reg_type,
 3321		.log = &env->log,
 3322	};
 3323
 3324	if (env->ops->is_valid_access &&
 3325	    env->ops->is_valid_access(off, size, t, env->prog, &info)) {
 3326		/* A non zero info.ctx_field_size indicates that this field is a
 3327		 * candidate for later verifier transformation to load the whole
 3328		 * field and then apply a mask when accessed with a narrower
 3329		 * access than actual ctx access size. A zero info.ctx_field_size
 3330		 * will only allow for whole field access and rejects any other
 3331		 * type of narrower access.
 3332		 */
 3333		*reg_type = info.reg_type;
 3334
 3335		if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
 3336			*btf = info.btf;
 3337			*btf_id = info.btf_id;
 3338		} else {
 3339			env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
 3340		}
 3341		/* remember the offset of last byte accessed in ctx */
 3342		if (env->prog->aux->max_ctx_offset < off + size)
 3343			env->prog->aux->max_ctx_offset = off + size;
 3344		return 0;
 3345	}
 3346
 3347	verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
 3348	return -EACCES;
 3349}
 3350
 3351static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
 3352				  int size)
 3353{
 3354	if (size < 0 || off < 0 ||
 3355	    (u64)off + size > sizeof(struct bpf_flow_keys)) {
 3356		verbose(env, "invalid access to flow keys off=%d size=%d\n",
 3357			off, size);
 3358		return -EACCES;
 3359	}
 3360	return 0;
 3361}
 3362
 3363static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
 3364			     u32 regno, int off, int size,
 3365			     enum bpf_access_type t)
 3366{
 3367	struct bpf_reg_state *regs = cur_regs(env);
 3368	struct bpf_reg_state *reg = &regs[regno];
 3369	struct bpf_insn_access_aux info = {};
 3370	bool valid;
 3371
 3372	if (reg->smin_value < 0) {
 3373		verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
 3374			regno);
 3375		return -EACCES;
 3376	}
 3377
 3378	switch (reg->type) {
 3379	case PTR_TO_SOCK_COMMON:
 3380		valid = bpf_sock_common_is_valid_access(off, size, t, &info);
 3381		break;
 3382	case PTR_TO_SOCKET:
 3383		valid = bpf_sock_is_valid_access(off, size, t, &info);
 3384		break;
 3385	case PTR_TO_TCP_SOCK:
 3386		valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
 3387		break;
 3388	case PTR_TO_XDP_SOCK:
 3389		valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
 3390		break;
 3391	default:
 3392		valid = false;
 3393	}
 3394
 3395
 3396	if (valid) {
 3397		env->insn_aux_data[insn_idx].ctx_field_size =
 3398			info.ctx_field_size;
 3399		return 0;
 3400	}
 3401
 3402	verbose(env, "R%d invalid %s access off=%d size=%d\n",
 3403		regno, reg_type_str[reg->type], off, size);
 3404
 3405	return -EACCES;
 3406}
 3407
 3408static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
 3409{
 3410	return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
 3411}
 3412
 3413static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
 3414{
 3415	const struct bpf_reg_state *reg = reg_state(env, regno);
 3416
 3417	return reg->type == PTR_TO_CTX;
 3418}
 3419
 3420static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
 3421{
 3422	const struct bpf_reg_state *reg = reg_state(env, regno);
 3423
 3424	return type_is_sk_pointer(reg->type);
 3425}
 3426
 3427static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
 3428{
 3429	const struct bpf_reg_state *reg = reg_state(env, regno);
 3430
 3431	return type_is_pkt_pointer(reg->type);
 3432}
 3433
 3434static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
 3435{
 3436	const struct bpf_reg_state *reg = reg_state(env, regno);
 3437
 3438	/* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
 3439	return reg->type == PTR_TO_FLOW_KEYS;
 3440}
 3441
 3442static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
 3443				   const struct bpf_reg_state *reg,
 3444				   int off, int size, bool strict)
 3445{
 3446	struct tnum reg_off;
 3447	int ip_align;
 3448
 3449	/* Byte size accesses are always allowed. */
 3450	if (!strict || size == 1)
 3451		return 0;
 3452
 3453	/* For platforms that do not have a Kconfig enabling
 3454	 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
 3455	 * NET_IP_ALIGN is universally set to '2'.  And on platforms
 3456	 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
 3457	 * to this code only in strict mode where we want to emulate
 3458	 * the NET_IP_ALIGN==2 checking.  Therefore use an
 3459	 * unconditional IP align value of '2'.
 3460	 */
 3461	ip_align = 2;
 3462
 3463	reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
 3464	if (!tnum_is_aligned(reg_off, size)) {
 3465		char tn_buf[48];
 3466
 3467		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 3468		verbose(env,
 3469			"misaligned packet access off %d+%s+%d+%d size %d\n",
 3470			ip_align, tn_buf, reg->off, off, size);
 3471		return -EACCES;
 3472	}
 3473
 3474	return 0;
 3475}
 3476
 3477static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
 3478				       const struct bpf_reg_state *reg,
 3479				       const char *pointer_desc,
 3480				       int off, int size, bool strict)
 3481{
 3482	struct tnum reg_off;
 3483
 3484	/* Byte size accesses are always allowed. */
 3485	if (!strict || size == 1)
 3486		return 0;
 3487
 3488	reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
 3489	if (!tnum_is_aligned(reg_off, size)) {
 3490		char tn_buf[48];
 3491
 3492		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 3493		verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
 3494			pointer_desc, tn_buf, reg->off, off, size);
 3495		return -EACCES;
 3496	}
 3497
 3498	return 0;
 3499}
 3500
 3501static int check_ptr_alignment(struct bpf_verifier_env *env,
 3502			       const struct bpf_reg_state *reg, int off,
 3503			       int size, bool strict_alignment_once)
 3504{
 3505	bool strict = env->strict_alignment || strict_alignment_once;
 3506	const char *pointer_desc = "";
 3507
 3508	switch (reg->type) {
 3509	case PTR_TO_PACKET:
 3510	case PTR_TO_PACKET_META:
 3511		/* Special case, because of NET_IP_ALIGN. Given metadata sits
 3512		 * right in front, treat it the very same way.
 3513		 */
 3514		return check_pkt_ptr_alignment(env, reg, off, size, strict);
 3515	case PTR_TO_FLOW_KEYS:
 3516		pointer_desc = "flow keys ";
 3517		break;
 3518	case PTR_TO_MAP_KEY:
 3519		pointer_desc = "key ";
 3520		break;
 3521	case PTR_TO_MAP_VALUE:
 3522		pointer_desc = "value ";
 3523		break;
 3524	case PTR_TO_CTX:
 3525		pointer_desc = "context ";
 3526		break;
 3527	case PTR_TO_STACK:
 3528		pointer_desc = "stack ";
 3529		/* The stack spill tracking logic in check_stack_write_fixed_off()
 3530		 * and check_stack_read_fixed_off() relies on stack accesses being
 3531		 * aligned.
 3532		 */
 3533		strict = true;
 3534		break;
 3535	case PTR_TO_SOCKET:
 3536		pointer_desc = "sock ";
 3537		break;
 3538	case PTR_TO_SOCK_COMMON:
 3539		pointer_desc = "sock_common ";
 3540		break;
 3541	case PTR_TO_TCP_SOCK:
 3542		pointer_desc = "tcp_sock ";
 3543		break;
 3544	case PTR_TO_XDP_SOCK:
 3545		pointer_desc = "xdp_sock ";
 3546		break;
 3547	default:
 3548		break;
 3549	}
 3550	return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
 3551					   strict);
 3552}
 3553
 3554static int update_stack_depth(struct bpf_verifier_env *env,
 3555			      const struct bpf_func_state *func,
 3556			      int off)
 3557{
 3558	u16 stack = env->subprog_info[func->subprogno].stack_depth;
 3559
 3560	if (stack >= -off)
 3561		return 0;
 3562
 3563	/* update known max for given subprogram */
 3564	env->subprog_info[func->subprogno].stack_depth = -off;
 3565	return 0;
 3566}
 3567
 3568/* starting from main bpf function walk all instructions of the function
 3569 * and recursively walk all callees that given function can call.
 3570 * Ignore jump and exit insns.
 3571 * Since recursion is prevented by check_cfg() this algorithm
 3572 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
 3573 */
 3574static int check_max_stack_depth(struct bpf_verifier_env *env)
 3575{
 3576	int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
 3577	struct bpf_subprog_info *subprog = env->subprog_info;
 3578	struct bpf_insn *insn = env->prog->insnsi;
 3579	bool tail_call_reachable = false;
 3580	int ret_insn[MAX_CALL_FRAMES];
 3581	int ret_prog[MAX_CALL_FRAMES];
 3582	int j;
 3583
 3584process_func:
 3585	/* protect against potential stack overflow that might happen when
 3586	 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
 3587	 * depth for such case down to 256 so that the worst case scenario
 3588	 * would result in 8k stack size (32 which is tailcall limit * 256 =
 3589	 * 8k).
 3590	 *
 3591	 * To get the idea what might happen, see an example:
 3592	 * func1 -> sub rsp, 128
 3593	 *  subfunc1 -> sub rsp, 256
 3594	 *  tailcall1 -> add rsp, 256
 3595	 *   func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
 3596	 *   subfunc2 -> sub rsp, 64
 3597	 *   subfunc22 -> sub rsp, 128
 3598	 *   tailcall2 -> add rsp, 128
 3599	 *    func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
 3600	 *
 3601	 * tailcall will unwind the current stack frame but it will not get rid
 3602	 * of caller's stack as shown on the example above.
 3603	 */
 3604	if (idx && subprog[idx].has_tail_call && depth >= 256) {
 3605		verbose(env,
 3606			"tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
 3607			depth);
 3608		return -EACCES;
 3609	}
 3610	/* round up to 32-bytes, since this is granularity
 3611	 * of interpreter stack size
 3612	 */
 3613	depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
 3614	if (depth > MAX_BPF_STACK) {
 3615		verbose(env, "combined stack size of %d calls is %d. Too large\n",
 3616			frame + 1, depth);
 3617		return -EACCES;
 3618	}
 3619continue_func:
 3620	subprog_end = subprog[idx + 1].start;
 3621	for (; i < subprog_end; i++) {
 3622		if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
 3623			continue;
 3624		/* remember insn and function to return to */
 3625		ret_insn[frame] = i + 1;
 3626		ret_prog[frame] = idx;
 3627
 3628		/* find the callee */
 3629		i = i + insn[i].imm + 1;
 3630		idx = find_subprog(env, i);
 3631		if (idx < 0) {
 3632			WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
 3633				  i);
 3634			return -EFAULT;
 3635		}
 3636
 3637		if (subprog[idx].has_tail_call)
 3638			tail_call_reachable = true;
 3639
 3640		frame++;
 3641		if (frame >= MAX_CALL_FRAMES) {
 3642			verbose(env, "the call stack of %d frames is too deep !\n",
 3643				frame);
 3644			return -E2BIG;
 3645		}
 3646		goto process_func;
 3647	}
 3648	/* if tail call got detected across bpf2bpf calls then mark each of the
 3649	 * currently present subprog frames as tail call reachable subprogs;
 3650	 * this info will be utilized by JIT so that we will be preserving the
 3651	 * tail call counter throughout bpf2bpf calls combined with tailcalls
 3652	 */
 3653	if (tail_call_reachable)
 3654		for (j = 0; j < frame; j++)
 3655			subprog[ret_prog[j]].tail_call_reachable = true;
 3656	if (subprog[0].tail_call_reachable)
 3657		env->prog->aux->tail_call_reachable = true;
 3658
 3659	/* end of for() loop means the last insn of the 'subprog'
 3660	 * was reached. Doesn't matter whether it was JA or EXIT
 3661	 */
 3662	if (frame == 0)
 3663		return 0;
 3664	depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
 3665	frame--;
 3666	i = ret_insn[frame];
 3667	idx = ret_prog[frame];
 3668	goto continue_func;
 3669}
 3670
 3671#ifndef CONFIG_BPF_JIT_ALWAYS_ON
 3672static int get_callee_stack_depth(struct bpf_verifier_env *env,
 3673				  const struct bpf_insn *insn, int idx)
 3674{
 3675	int start = idx + insn->imm + 1, subprog;
 3676
 3677	subprog = find_subprog(env, start);
 3678	if (subprog < 0) {
 3679		WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
 3680			  start);
 3681		return -EFAULT;
 3682	}
 3683	return env->subprog_info[subprog].stack_depth;
 3684}
 3685#endif
 3686
 3687int check_ctx_reg(struct bpf_verifier_env *env,
 3688		  const struct bpf_reg_state *reg, int regno)
 3689{
 3690	/* Access to ctx or passing it to a helper is only allowed in
 3691	 * its original, unmodified form.
 3692	 */
 3693
 3694	if (reg->off) {
 3695		verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
 3696			regno, reg->off);
 3697		return -EACCES;
 3698	}
 3699
 3700	if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
 3701		char tn_buf[48];
 3702
 3703		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 3704		verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
 3705		return -EACCES;
 3706	}
 3707
 3708	return 0;
 3709}
 3710
 3711static int __check_buffer_access(struct bpf_verifier_env *env,
 3712				 const char *buf_info,
 3713				 const struct bpf_reg_state *reg,
 3714				 int regno, int off, int size)
 3715{
 3716	if (off < 0) {
 3717		verbose(env,
 3718			"R%d invalid %s buffer access: off=%d, size=%d\n",
 3719			regno, buf_info, off, size);
 3720		return -EACCES;
 3721	}
 3722	if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
 3723		char tn_buf[48];
 3724
 3725		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 3726		verbose(env,
 3727			"R%d invalid variable buffer offset: off=%d, var_off=%s\n",
 3728			regno, off, tn_buf);
 3729		return -EACCES;
 3730	}
 3731
 3732	return 0;
 3733}
 3734
 3735static int check_tp_buffer_access(struct bpf_verifier_env *env,
 3736				  const struct bpf_reg_state *reg,
 3737				  int regno, int off, int size)
 3738{
 3739	int err;
 3740
 3741	err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
 3742	if (err)
 3743		return err;
 3744
 3745	if (off + size > env->prog->aux->max_tp_access)
 3746		env->prog->aux->max_tp_access = off + size;
 3747
 3748	return 0;
 3749}
 3750
 3751static int check_buffer_access(struct bpf_verifier_env *env,
 3752			       const struct bpf_reg_state *reg,
 3753			       int regno, int off, int size,
 3754			       bool zero_size_allowed,
 3755			       const char *buf_info,
 3756			       u32 *max_access)
 3757{
 3758	int err;
 3759
 3760	err = __check_buffer_access(env, buf_info, reg, regno, off, size);
 3761	if (err)
 3762		return err;
 3763
 3764	if (off + size > *max_access)
 3765		*max_access = off + size;
 3766
 3767	return 0;
 3768}
 3769
 3770/* BPF architecture zero extends alu32 ops into 64-bit registesr */
 3771static void zext_32_to_64(struct bpf_reg_state *reg)
 3772{
 3773	reg->var_off = tnum_subreg(reg->var_off);
 3774	__reg_assign_32_into_64(reg);
 3775}
 3776
 3777/* truncate register to smaller size (in bytes)
 3778 * must be called with size < BPF_REG_SIZE
 3779 */
 3780static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
 3781{
 3782	u64 mask;
 3783
 3784	/* clear high bits in bit representation */
 3785	reg->var_off = tnum_cast(reg->var_off, size);
 3786
 3787	/* fix arithmetic bounds */
 3788	mask = ((u64)1 << (size * 8)) - 1;
 3789	if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
 3790		reg->umin_value &= mask;
 3791		reg->umax_value &= mask;
 3792	} else {
 3793		reg->umin_value = 0;
 3794		reg->umax_value = mask;
 3795	}
 3796	reg->smin_value = reg->umin_value;
 3797	reg->smax_value = reg->umax_value;
 3798
 3799	/* If size is smaller than 32bit register the 32bit register
 3800	 * values are also truncated so we push 64-bit bounds into
 3801	 * 32-bit bounds. Above were truncated < 32-bits already.
 3802	 */
 3803	if (size >= 4)
 3804		return;
 3805	__reg_combine_64_into_32(reg);
 3806}
 3807
 3808static bool bpf_map_is_rdonly(const struct bpf_map *map)
 3809{
 3810	return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
 3811}
 3812
 3813static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
 3814{
 3815	void *ptr;
 3816	u64 addr;
 3817	int err;
 3818
 3819	err = map->ops->map_direct_value_addr(map, &addr, off);
 3820	if (err)
 3821		return err;
 3822	ptr = (void *)(long)addr + off;
 3823
 3824	switch (size) {
 3825	case sizeof(u8):
 3826		*val = (u64)*(u8 *)ptr;
 3827		break;
 3828	case sizeof(u16):
 3829		*val = (u64)*(u16 *)ptr;
 3830		break;
 3831	case sizeof(u32):
 3832		*val = (u64)*(u32 *)ptr;
 3833		break;
 3834	case sizeof(u64):
 3835		*val = *(u64 *)ptr;
 3836		break;
 3837	default:
 3838		return -EINVAL;
 3839	}
 3840	return 0;
 3841}
 3842
 3843static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
 3844				   struct bpf_reg_state *regs,
 3845				   int regno, int off, int size,
 3846				   enum bpf_access_type atype,
 3847				   int value_regno)
 3848{
 3849	struct bpf_reg_state *reg = regs + regno;
 3850	const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
 3851	const char *tname = btf_name_by_offset(reg->btf, t->name_off);
 3852	u32 btf_id;
 3853	int ret;
 3854
 3855	if (off < 0) {
 3856		verbose(env,
 3857			"R%d is ptr_%s invalid negative access: off=%d\n",
 3858			regno, tname, off);
 3859		return -EACCES;
 3860	}
 3861	if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
 3862		char tn_buf[48];
 3863
 3864		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 3865		verbose(env,
 3866			"R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
 3867			regno, tname, off, tn_buf);
 3868		return -EACCES;
 3869	}
 3870
 3871	if (env->ops->btf_struct_access) {
 3872		ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
 3873						  off, size, atype, &btf_id);
 3874	} else {
 3875		if (atype != BPF_READ) {
 3876			verbose(env, "only read is supported\n");
 3877			return -EACCES;
 3878		}
 3879
 3880		ret = btf_struct_access(&env->log, reg->btf, t, off, size,
 3881					atype, &btf_id);
 3882	}
 3883
 3884	if (ret < 0)
 3885		return ret;
 3886
 3887	if (atype == BPF_READ && value_regno >= 0)
 3888		mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
 3889
 3890	return 0;
 3891}
 3892
 3893static int check_ptr_to_map_access(struct bpf_verifier_env *env,
 3894				   struct bpf_reg_state *regs,
 3895				   int regno, int off, int size,
 3896				   enum bpf_access_type atype,
 3897				   int value_regno)
 3898{
 3899	struct bpf_reg_state *reg = regs + regno;
 3900	struct bpf_map *map = reg->map_ptr;
 3901	const struct btf_type *t;
 3902	const char *tname;
 3903	u32 btf_id;
 3904	int ret;
 3905
 3906	if (!btf_vmlinux) {
 3907		verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
 3908		return -ENOTSUPP;
 3909	}
 3910
 3911	if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
 3912		verbose(env, "map_ptr access not supported for map type %d\n",
 3913			map->map_type);
 3914		return -ENOTSUPP;
 3915	}
 3916
 3917	t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
 3918	tname = btf_name_by_offset(btf_vmlinux, t->name_off);
 3919
 3920	if (!env->allow_ptr_to_map_access) {
 3921		verbose(env,
 3922			"%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
 3923			tname);
 3924		return -EPERM;
 3925	}
 3926
 3927	if (off < 0) {
 3928		verbose(env, "R%d is %s invalid negative access: off=%d\n",
 3929			regno, tname, off);
 3930		return -EACCES;
 3931	}
 3932
 3933	if (atype != BPF_READ) {
 3934		verbose(env, "only read from %s is supported\n", tname);
 3935		return -EACCES;
 3936	}
 3937
 3938	ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
 3939	if (ret < 0)
 3940		return ret;
 3941
 3942	if (value_regno >= 0)
 3943		mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
 3944
 3945	return 0;
 3946}
 3947
 3948/* Check that the stack access at the given offset is within bounds. The
 3949 * maximum valid offset is -1.
 3950 *
 3951 * The minimum valid offset is -MAX_BPF_STACK for writes, and
 3952 * -state->allocated_stack for reads.
 3953 */
 3954static int check_stack_slot_within_bounds(int off,
 3955					  struct bpf_func_state *state,
 3956					  enum bpf_access_type t)
 3957{
 3958	int min_valid_off;
 3959
 3960	if (t == BPF_WRITE)
 3961		min_valid_off = -MAX_BPF_STACK;
 3962	else
 3963		min_valid_off = -state->allocated_stack;
 3964
 3965	if (off < min_valid_off || off > -1)
 3966		return -EACCES;
 3967	return 0;
 3968}
 3969
 3970/* Check that the stack access at 'regno + off' falls within the maximum stack
 3971 * bounds.
 3972 *
 3973 * 'off' includes `regno->offset`, but not its dynamic part (if any).
 3974 */
 3975static int check_stack_access_within_bounds(
 3976		struct bpf_verifier_env *env,
 3977		int regno, int off, int access_size,
 3978		enum stack_access_src src, enum bpf_access_type type)
 3979{
 3980	struct bpf_reg_state *regs = cur_regs(env);
 3981	struct bpf_reg_state *reg = regs + regno;
 3982	struct bpf_func_state *state = func(env, reg);
 3983	int min_off, max_off;
 3984	int err;
 3985	char *err_extra;
 3986
 3987	if (src == ACCESS_HELPER)
 3988		/* We don't know if helpers are reading or writing (or both). */
 3989		err_extra = " indirect access to";
 3990	else if (type == BPF_READ)
 3991		err_extra = " read from";
 3992	else
 3993		err_extra = " write to";
 3994
 3995	if (tnum_is_const(reg->var_off)) {
 3996		min_off = reg->var_off.value + off;
 3997		if (access_size > 0)
 3998			max_off = min_off + access_size - 1;
 3999		else
 4000			max_off = min_off;
 4001	} else {
 4002		if (reg->smax_value >= BPF_MAX_VAR_OFF ||
 4003		    reg->smin_value <= -BPF_MAX_VAR_OFF) {
 4004			verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
 4005				err_extra, regno);
 4006			return -EACCES;
 4007		}
 4008		min_off = reg->smin_value + off;
 4009		if (access_size > 0)
 4010			max_off = reg->smax_value + off + access_size - 1;
 4011		else
 4012			max_off = min_off;
 4013	}
 4014
 4015	err = check_stack_slot_within_bounds(min_off, state, type);
 4016	if (!err)
 4017		err = check_stack_slot_within_bounds(max_off, state, type);
 4018
 4019	if (err) {
 4020		if (tnum_is_const(reg->var_off)) {
 4021			verbose(env, "invalid%s stack R%d off=%d size=%d\n",
 4022				err_extra, regno, off, access_size);
 4023		} else {
 4024			char tn_buf[48];
 4025
 4026			tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 4027			verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
 4028				err_extra, regno, tn_buf, access_size);
 4029		}
 4030	}
 4031	return err;
 4032}
 4033
 4034/* check whether memory at (regno + off) is accessible for t = (read | write)
 4035 * if t==write, value_regno is a register which value is stored into memory
 4036 * if t==read, value_regno is a register which will receive the value from memory
 4037 * if t==write && value_regno==-1, some unknown value is stored into memory
 4038 * if t==read && value_regno==-1, don't care what we read from memory
 4039 */
 4040static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
 4041			    int off, int bpf_size, enum bpf_access_type t,
 4042			    int value_regno, bool strict_alignment_once)
 4043{
 4044	struct bpf_reg_state *regs = cur_regs(env);
 4045	struct bpf_reg_state *reg = regs + regno;
 4046	struct bpf_func_state *state;
 4047	int size, err = 0;
 4048
 4049	size = bpf_size_to_bytes(bpf_size);
 4050	if (size < 0)
 4051		return size;
 4052
 4053	/* alignment checks will add in reg->off themselves */
 4054	err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
 4055	if (err)
 4056		return err;
 4057
 4058	/* for access checks, reg->off is just part of off */
 4059	off += reg->off;
 4060
 4061	if (reg->type == PTR_TO_MAP_KEY) {
 4062		if (t == BPF_WRITE) {
 4063			verbose(env, "write to change key R%d not allowed\n", regno);
 4064			return -EACCES;
 4065		}
 4066
 4067		err = check_mem_region_access(env, regno, off, size,
 4068					      reg->map_ptr->key_size, false);
 4069		if (err)
 4070			return err;
 4071		if (value_regno >= 0)
 4072			mark_reg_unknown(env, regs, value_regno);
 4073	} else if (reg->type == PTR_TO_MAP_VALUE) {
 4074		if (t == BPF_WRITE && value_regno >= 0 &&
 4075		    is_pointer_value(env, value_regno)) {
 4076			verbose(env, "R%d leaks addr into map\n", value_regno);
 4077			return -EACCES;
 4078		}
 4079		err = check_map_access_type(env, regno, off, size, t);
 4080		if (err)
 4081			return err;
 4082		err = check_map_access(env, regno, off, size, false);
 4083		if (!err && t == BPF_READ && value_regno >= 0) {
 4084			struct bpf_map *map = reg->map_ptr;
 4085
 4086			/* if map is read-only, track its contents as scalars */
 4087			if (tnum_is_const(reg->var_off) &&
 4088			    bpf_map_is_rdonly(map) &&
 4089			    map->ops->map_direct_value_addr) {
 4090				int map_off = off + reg->var_off.value;
 4091				u64 val = 0;
 4092
 4093				err = bpf_map_direct_read(map, map_off, size,
 4094							  &val);
 4095				if (err)
 4096					return err;
 4097
 4098				regs[value_regno].type = SCALAR_VALUE;
 4099				__mark_reg_known(&regs[value_regno], val);
 4100			} else {
 4101				mark_reg_unknown(env, regs, value_regno);
 4102			}
 4103		}
 4104	} else if (reg->type == PTR_TO_MEM) {
 4105		if (t == BPF_WRITE && value_regno >= 0 &&
 4106		    is_pointer_value(env, value_regno)) {
 4107			verbose(env, "R%d leaks addr into mem\n", value_regno);
 4108			return -EACCES;
 4109		}
 4110		err = check_mem_region_access(env, regno, off, size,
 4111					      reg->mem_size, false);
 4112		if (!err && t == BPF_READ && value_regno >= 0)
 4113			mark_reg_unknown(env, regs, value_regno);
 4114	} else if (reg->type == PTR_TO_CTX) {
 4115		enum bpf_reg_type reg_type = SCALAR_VALUE;
 4116		struct btf *btf = NULL;
 4117		u32 btf_id = 0;
 4118
 4119		if (t == BPF_WRITE && value_regno >= 0 &&
 4120		    is_pointer_value(env, value_regno)) {
 4121			verbose(env, "R%d leaks addr into ctx\n", value_regno);
 4122			return -EACCES;
 4123		}
 4124
 4125		err = check_ctx_reg(env, reg, regno);
 4126		if (err < 0)
 4127			return err;
 4128
 4129		err = check_ctx_access(env, insn_idx, off, size, t, &reg_type, &btf, &btf_id);
 4130		if (err)
 4131			verbose_linfo(env, insn_idx, "; ");
 4132		if (!err && t == BPF_READ && value_regno >= 0) {
 4133			/* ctx access returns either a scalar, or a
 4134			 * PTR_TO_PACKET[_META,_END]. In the latter
 4135			 * case, we know the offset is zero.
 4136			 */
 4137			if (reg_type == SCALAR_VALUE) {
 4138				mark_reg_unknown(env, regs, value_regno);
 4139			} else {
 4140				mark_reg_known_zero(env, regs,
 4141						    value_regno);
 4142				if (reg_type_may_be_null(reg_type))
 4143					regs[value_regno].id = ++env->id_gen;
 4144				/* A load of ctx field could have different
 4145				 * actual load size with the one encoded in the
 4146				 * insn. When the dst is PTR, it is for sure not
 4147				 * a sub-register.
 4148				 */
 4149				regs[value_regno].subreg_def = DEF_NOT_SUBREG;
 4150				if (reg_type == PTR_TO_BTF_ID ||
 4151				    reg_type == PTR_TO_BTF_ID_OR_NULL) {
 4152					regs[value_regno].btf = btf;
 4153					regs[value_regno].btf_id = btf_id;
 4154				}
 4155			}
 4156			regs[value_regno].type = reg_type;
 4157		}
 4158
 4159	} else if (reg->type == PTR_TO_STACK) {
 4160		/* Basic bounds checks. */
 4161		err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
 4162		if (err)
 4163			return err;
 4164
 4165		state = func(env, reg);
 4166		err = update_stack_depth(env, state, off);
 4167		if (err)
 4168			return err;
 4169
 4170		if (t == BPF_READ)
 4171			err = check_stack_read(env, regno, off, size,
 4172					       value_regno);
 4173		else
 4174			err = check_stack_write(env, regno, off, size,
 4175						value_regno, insn_idx);
 4176	} else if (reg_is_pkt_pointer(reg)) {
 4177		if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
 4178			verbose(env, "cannot write into packet\n");
 4179			return -EACCES;
 4180		}
 4181		if (t == BPF_WRITE && value_regno >= 0 &&
 4182		    is_pointer_value(env, value_regno)) {
 4183			verbose(env, "R%d leaks addr into packet\n",
 4184				value_regno);
 4185			return -EACCES;
 4186		}
 4187		err = check_packet_access(env, regno, off, size, false);
 4188		if (!err && t == BPF_READ && value_regno >= 0)
 4189			mark_reg_unknown(env, regs, value_regno);
 4190	} else if (reg->type == PTR_TO_FLOW_KEYS) {
 4191		if (t == BPF_WRITE && value_regno >= 0 &&
 4192		    is_pointer_value(env, value_regno)) {
 4193			verbose(env, "R%d leaks addr into flow keys\n",
 4194				value_regno);
 4195			return -EACCES;
 4196		}
 4197
 4198		err = check_flow_keys_access(env, off, size);
 4199		if (!err && t == BPF_READ && value_regno >= 0)
 4200			mark_reg_unknown(env, regs, value_regno);
 4201	} else if (type_is_sk_pointer(reg->type)) {
 4202		if (t == BPF_WRITE) {
 4203			verbose(env, "R%d cannot write into %s\n",
 4204				regno, reg_type_str[reg->type]);
 4205			return -EACCES;
 4206		}
 4207		err = check_sock_access(env, insn_idx, regno, off, size, t);
 4208		if (!err && value_regno >= 0)
 4209			mark_reg_unknown(env, regs, value_regno);
 4210	} else if (reg->type == PTR_TO_TP_BUFFER) {
 4211		err = check_tp_buffer_access(env, reg, regno, off, size);
 4212		if (!err && t == BPF_READ && value_regno >= 0)
 4213			mark_reg_unknown(env, regs, value_regno);
 4214	} else if (reg->type == PTR_TO_BTF_ID) {
 4215		err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
 4216					      value_regno);
 4217	} else if (reg->type == CONST_PTR_TO_MAP) {
 4218		err = check_ptr_to_map_access(env, regs, regno, off, size, t,
 4219					      value_regno);
 4220	} else if (reg->type == PTR_TO_RDONLY_BUF) {
 4221		if (t == BPF_WRITE) {
 4222			verbose(env, "R%d cannot write into %s\n",
 4223				regno, reg_type_str[reg->type]);
 4224			return -EACCES;
 4225		}
 4226		err = check_buffer_access(env, reg, regno, off, size, false,
 4227					  "rdonly",
 4228					  &env->prog->aux->max_rdonly_access);
 4229		if (!err && value_regno >= 0)
 4230			mark_reg_unknown(env, regs, value_regno);
 4231	} else if (reg->type == PTR_TO_RDWR_BUF) {
 4232		err = check_buffer_access(env, reg, regno, off, size, false,
 4233					  "rdwr",
 4234					  &env->prog->aux->max_rdwr_access);
 4235		if (!err && t == BPF_READ && value_regno >= 0)
 4236			mark_reg_unknown(env, regs, value_regno);
 4237	} else {
 4238		verbose(env, "R%d invalid mem access '%s'\n", regno,
 4239			reg_type_str[reg->type]);
 4240		return -EACCES;
 4241	}
 4242
 4243	if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
 4244	    regs[value_regno].type == SCALAR_VALUE) {
 4245		/* b/h/w load zero-extends, mark upper bits as known 0 */
 4246		coerce_reg_to_size(&regs[value_regno], size);
 4247	}
 4248	return err;
 4249}
 4250
 4251static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
 4252{
 4253	int load_reg;
 4254	int err;
 4255
 4256	switch (insn->imm) {
 4257	case BPF_ADD:
 4258	case BPF_ADD | BPF_FETCH:
 4259	case BPF_AND:
 4260	case BPF_AND | BPF_FETCH:
 4261	case BPF_OR:
 4262	case BPF_OR | BPF_FETCH:
 4263	case BPF_XOR:
 4264	case BPF_XOR | BPF_FETCH:
 4265	case BPF_XCHG:
 4266	case BPF_CMPXCHG:
 4267		break;
 4268	default:
 4269		verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
 4270		return -EINVAL;
 4271	}
 4272
 4273	if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
 4274		verbose(env, "invalid atomic operand size\n");
 4275		return -EINVAL;
 4276	}
 4277
 4278	/* check src1 operand */
 4279	err = check_reg_arg(env, insn->src_reg, SRC_OP);
 4280	if (err)
 4281		return err;
 4282
 4283	/* check src2 operand */
 4284	err = check_reg_arg(env, insn->dst_reg, SRC_OP);
 4285	if (err)
 4286		return err;
 4287
 4288	if (insn->imm == BPF_CMPXCHG) {
 4289		/* Check comparison of R0 with memory location */
 4290		err = check_reg_arg(env, BPF_REG_0, SRC_OP);
 4291		if (err)
 4292			return err;
 4293	}
 4294
 4295	if (is_pointer_value(env, insn->src_reg)) {
 4296		verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
 4297		return -EACCES;
 4298	}
 4299
 4300	if (is_ctx_reg(env, insn->dst_reg) ||
 4301	    is_pkt_reg(env, insn->dst_reg) ||
 4302	    is_flow_key_reg(env, insn->dst_reg) ||
 4303	    is_sk_reg(env, insn->dst_reg)) {
 4304		verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
 4305			insn->dst_reg,
 4306			reg_type_str[reg_state(env, insn->dst_reg)->type]);
 4307		return -EACCES;
 4308	}
 4309
 4310	if (insn->imm & BPF_FETCH) {
 4311		if (insn->imm == BPF_CMPXCHG)
 4312			load_reg = BPF_REG_0;
 4313		else
 4314			load_reg = insn->src_reg;
 4315
 4316		/* check and record load of old value */
 4317		err = check_reg_arg(env, load_reg, DST_OP);
 4318		if (err)
 4319			return err;
 4320	} else {
 4321		/* This instruction accesses a memory location but doesn't
 4322		 * actually load it into a register.
 4323		 */
 4324		load_reg = -1;
 4325	}
 4326
 4327	/* check whether we can read the memory */
 4328	err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
 4329			       BPF_SIZE(insn->code), BPF_READ, load_reg, true);
 4330	if (err)
 4331		return err;
 4332
 4333	/* check whether we can write into the same memory */
 4334	err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
 4335			       BPF_SIZE(insn->code), BPF_WRITE, -1, true);
 4336	if (err)
 4337		return err;
 4338
 4339	return 0;
 4340}
 4341
 4342/* When register 'regno' is used to read the stack (either directly or through
 4343 * a helper function) make sure that it's within stack boundary and, depending
 4344 * on the access type, that all elements of the stack are initialized.
 4345 *
 4346 * 'off' includes 'regno->off', but not its dynamic part (if any).
 4347 *
 4348 * All registers that have been spilled on the stack in the slots within the
 4349 * read offsets are marked as read.
 4350 */
 4351static int check_stack_range_initialized(
 4352		struct bpf_verifier_env *env, int regno, int off,
 4353		int access_size, bool zero_size_allowed,
 4354		enum stack_access_src type, struct bpf_call_arg_meta *meta)
 4355{
 4356	struct bpf_reg_state *reg = reg_state(env, regno);
 4357	struct bpf_func_state *state = func(env, reg);
 4358	int err, min_off, max_off, i, j, slot, spi;
 4359	char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
 4360	enum bpf_access_type bounds_check_type;
 4361	/* Some accesses can write anything into the stack, others are
 4362	 * read-only.
 4363	 */
 4364	bool clobber = false;
 4365
 4366	if (access_size == 0 && !zero_size_allowed) {
 4367		verbose(env, "invalid zero-sized read\n");
 4368		return -EACCES;
 4369	}
 4370
 4371	if (type == ACCESS_HELPER) {
 4372		/* The bounds checks for writes are more permissive than for
 4373		 * reads. However, if raw_mode is not set, we'll do extra
 4374		 * checks below.
 4375		 */
 4376		bounds_check_type = BPF_WRITE;
 4377		clobber = true;
 4378	} else {
 4379		bounds_check_type = BPF_READ;
 4380	}
 4381	err = check_stack_access_within_bounds(env, regno, off, access_size,
 4382					       type, bounds_check_type);
 4383	if (err)
 4384		return err;
 4385
 4386
 4387	if (tnum_is_const(reg->var_off)) {
 4388		min_off = max_off = reg->var_off.value + off;
 4389	} else {
 4390		/* Variable offset is prohibited for unprivileged mode for
 4391		 * simplicity since it requires corresponding support in
 4392		 * Spectre masking for stack ALU.
 4393		 * See also retrieve_ptr_limit().
 4394		 */
 4395		if (!env->bypass_spec_v1) {
 4396			char tn_buf[48];
 4397
 4398			tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 4399			verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
 4400				regno, err_extra, tn_buf);
 4401			return -EACCES;
 4402		}
 4403		/* Only initialized buffer on stack is allowed to be accessed
 4404		 * with variable offset. With uninitialized buffer it's hard to
 4405		 * guarantee that whole memory is marked as initialized on
 4406		 * helper return since specific bounds are unknown what may
 4407		 * cause uninitialized stack leaking.
 4408		 */
 4409		if (meta && meta->raw_mode)
 4410			meta = NULL;
 4411
 4412		min_off = reg->smin_value + off;
 4413		max_off = reg->smax_value + off;
 4414	}
 4415
 4416	if (meta && meta->raw_mode) {
 4417		meta->access_size = access_size;
 4418		meta->regno = regno;
 4419		return 0;
 4420	}
 4421
 4422	for (i = min_off; i < max_off + access_size; i++) {
 4423		u8 *stype;
 4424
 4425		slot = -i - 1;
 4426		spi = slot / BPF_REG_SIZE;
 4427		if (state->allocated_stack <= slot)
 4428			goto err;
 4429		stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
 4430		if (*stype == STACK_MISC)
 4431			goto mark;
 4432		if (*stype == STACK_ZERO) {
 4433			if (clobber) {
 4434				/* helper can write anything into the stack */
 4435				*stype = STACK_MISC;
 4436			}
 4437			goto mark;
 4438		}
 4439
 4440		if (state->stack[spi].slot_type[0] == STACK_SPILL &&
 4441		    state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
 4442			goto mark;
 4443
 4444		if (state->stack[spi].slot_type[0] == STACK_SPILL &&
 4445		    (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
 4446		     env->allow_ptr_leaks)) {
 4447			if (clobber) {
 4448				__mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
 4449				for (j = 0; j < BPF_REG_SIZE; j++)
 4450					state->stack[spi].slot_type[j] = STACK_MISC;
 4451			}
 4452			goto mark;
 4453		}
 4454
 4455err:
 4456		if (tnum_is_const(reg->var_off)) {
 4457			verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
 4458				err_extra, regno, min_off, i - min_off, access_size);
 4459		} else {
 4460			char tn_buf[48];
 4461
 4462			tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 4463			verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
 4464				err_extra, regno, tn_buf, i - min_off, access_size);
 4465		}
 4466		return -EACCES;
 4467mark:
 4468		/* reading any byte out of 8-byte 'spill_slot' will cause
 4469		 * the whole slot to be marked as 'read'
 4470		 */
 4471		mark_reg_read(env, &state->stack[spi].spilled_ptr,
 4472			      state->stack[spi].spilled_ptr.parent,
 4473			      REG_LIVE_READ64);
 4474	}
 4475	return update_stack_depth(env, state, min_off);
 4476}
 4477
 4478static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
 4479				   int access_size, bool zero_size_allowed,
 4480				   struct bpf_call_arg_meta *meta)
 4481{
 4482	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
 4483
 4484	switch (reg->type) {
 4485	case PTR_TO_PACKET:
 4486	case PTR_TO_PACKET_META:
 4487		return check_packet_access(env, regno, reg->off, access_size,
 4488					   zero_size_allowed);
 4489	case PTR_TO_MAP_KEY:
 4490		return check_mem_region_access(env, regno, reg->off, access_size,
 4491					       reg->map_ptr->key_size, false);
 4492	case PTR_TO_MAP_VALUE:
 4493		if (check_map_access_type(env, regno, reg->off, access_size,
 4494					  meta && meta->raw_mode ? BPF_WRITE :
 4495					  BPF_READ))
 4496			return -EACCES;
 4497		return check_map_access(env, regno, reg->off, access_size,
 4498					zero_size_allowed);
 4499	case PTR_TO_MEM:
 4500		return check_mem_region_access(env, regno, reg->off,
 4501					       access_size, reg->mem_size,
 4502					       zero_size_allowed);
 4503	case PTR_TO_RDONLY_BUF:
 4504		if (meta && meta->raw_mode)
 4505			return -EACCES;
 4506		return check_buffer_access(env, reg, regno, reg->off,
 4507					   access_size, zero_size_allowed,
 4508					   "rdonly",
 4509					   &env->prog->aux->max_rdonly_access);
 4510	case PTR_TO_RDWR_BUF:
 4511		return check_buffer_access(env, reg, regno, reg->off,
 4512					   access_size, zero_size_allowed,
 4513					   "rdwr",
 4514					   &env->prog->aux->max_rdwr_access);
 4515	case PTR_TO_STACK:
 4516		return check_stack_range_initialized(
 4517				env,
 4518				regno, reg->off, access_size,
 4519				zero_size_allowed, ACCESS_HELPER, meta);
 4520	default: /* scalar_value or invalid ptr */
 4521		/* Allow zero-byte read from NULL, regardless of pointer type */
 4522		if (zero_size_allowed && access_size == 0 &&
 4523		    register_is_null(reg))
 4524			return 0;
 4525
 4526		verbose(env, "R%d type=%s expected=%s\n", regno,
 4527			reg_type_str[reg->type],
 4528			reg_type_str[PTR_TO_STACK]);
 4529		return -EACCES;
 4530	}
 4531}
 4532
 4533int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
 4534		   u32 regno, u32 mem_size)
 4535{
 4536	if (register_is_null(reg))
 4537		return 0;
 4538
 4539	if (reg_type_may_be_null(reg->type)) {
 4540		/* Assuming that the register contains a value check if the memory
 4541		 * access is safe. Temporarily save and restore the register's state as
 4542		 * the conversion shouldn't be visible to a caller.
 4543		 */
 4544		const struct bpf_reg_state saved_reg = *reg;
 4545		int rv;
 4546
 4547		mark_ptr_not_null_reg(reg);
 4548		rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
 4549		*reg = saved_reg;
 4550		return rv;
 4551	}
 4552
 4553	return check_helper_mem_access(env, regno, mem_size, true, NULL);
 4554}
 4555
 4556/* Implementation details:
 4557 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
 4558 * Two bpf_map_lookups (even with the same key) will have different reg->id.
 4559 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
 4560 * value_or_null->value transition, since the verifier only cares about
 4561 * the range of access to valid map value pointer and doesn't care about actual
 4562 * address of the map element.
 4563 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
 4564 * reg->id > 0 after value_or_null->value transition. By doing so
 4565 * two bpf_map_lookups will be considered two different pointers that
 4566 * point to different bpf_spin_locks.
 4567 * The verifier allows taking only one bpf_spin_lock at a time to avoid
 4568 * dead-locks.
 4569 * Since only one bpf_spin_lock is allowed the checks are simpler than
 4570 * reg_is_refcounted() logic. The verifier needs to remember only
 4571 * one spin_lock instead of array of acquired_refs.
 4572 * cur_state->active_spin_lock remembers which map value element got locked
 4573 * and clears it after bpf_spin_unlock.
 4574 */
 4575static int process_spin_lock(struct bpf_verifier_env *env, int regno,
 4576			     bool is_lock)
 4577{
 4578	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
 4579	struct bpf_verifier_state *cur = env->cur_state;
 4580	bool is_const = tnum_is_const(reg->var_off);
 4581	struct bpf_map *map = reg->map_ptr;
 4582	u64 val = reg->var_off.value;
 4583
 4584	if (!is_const) {
 4585		verbose(env,
 4586			"R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
 4587			regno);
 4588		return -EINVAL;
 4589	}
 4590	if (!map->btf) {
 4591		verbose(env,
 4592			"map '%s' has to have BTF in order to use bpf_spin_lock\n",
 4593			map->name);
 4594		return -EINVAL;
 4595	}
 4596	if (!map_value_has_spin_lock(map)) {
 4597		if (map->spin_lock_off == -E2BIG)
 4598			verbose(env,
 4599				"map '%s' has more than one 'struct bpf_spin_lock'\n",
 4600				map->name);
 4601		else if (map->spin_lock_off == -ENOENT)
 4602			verbose(env,
 4603				"map '%s' doesn't have 'struct bpf_spin_lock'\n",
 4604				map->name);
 4605		else
 4606			verbose(env,
 4607				"map '%s' is not a struct type or bpf_spin_lock is mangled\n",
 4608				map->name);
 4609		return -EINVAL;
 4610	}
 4611	if (map->spin_lock_off != val + reg->off) {
 4612		verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
 4613			val + reg->off);
 4614		return -EINVAL;
 4615	}
 4616	if (is_lock) {
 4617		if (cur->active_spin_lock) {
 4618			verbose(env,
 4619				"Locking two bpf_spin_locks are not allowed\n");
 4620			return -EINVAL;
 4621		}
 4622		cur->active_spin_lock = reg->id;
 4623	} else {
 4624		if (!cur->active_spin_lock) {
 4625			verbose(env, "bpf_spin_unlock without taking a lock\n");
 4626			return -EINVAL;
 4627		}
 4628		if (cur->active_spin_lock != reg->id) {
 4629			verbose(env, "bpf_spin_unlock of different lock\n");
 4630			return -EINVAL;
 4631		}
 4632		cur->active_spin_lock = 0;
 4633	}
 4634	return 0;
 4635}
 4636
 4637static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
 4638{
 4639	return type == ARG_PTR_TO_MEM ||
 4640	       type == ARG_PTR_TO_MEM_OR_NULL ||
 4641	       type == ARG_PTR_TO_UNINIT_MEM;
 4642}
 4643
 4644static bool arg_type_is_mem_size(enum bpf_arg_type type)
 4645{
 4646	return type == ARG_CONST_SIZE ||
 4647	       type == ARG_CONST_SIZE_OR_ZERO;
 4648}
 4649
 4650static bool arg_type_is_alloc_size(enum bpf_arg_type type)
 4651{
 4652	return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
 4653}
 4654
 4655static bool arg_type_is_int_ptr(enum bpf_arg_type type)
 4656{
 4657	return type == ARG_PTR_TO_INT ||
 4658	       type == ARG_PTR_TO_LONG;
 4659}
 4660
 4661static int int_ptr_type_to_size(enum bpf_arg_type type)
 4662{
 4663	if (type == ARG_PTR_TO_INT)
 4664		return sizeof(u32);
 4665	else if (type == ARG_PTR_TO_LONG)
 4666		return sizeof(u64);
 4667
 4668	return -EINVAL;
 4669}
 4670
 4671static int resolve_map_arg_type(struct bpf_verifier_env *env,
 4672				 const struct bpf_call_arg_meta *meta,
 4673				 enum bpf_arg_type *arg_type)
 4674{
 4675	if (!meta->map_ptr) {
 4676		/* kernel subsystem misconfigured verifier */
 4677		verbose(env, "invalid map_ptr to access map->type\n");
 4678		return -EACCES;
 4679	}
 4680
 4681	switch (meta->map_ptr->map_type) {
 4682	case BPF_MAP_TYPE_SOCKMAP:
 4683	case BPF_MAP_TYPE_SOCKHASH:
 4684		if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
 4685			*arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
 4686		} else {
 4687			verbose(env, "invalid arg_type for sockmap/sockhash\n");
 4688			return -EINVAL;
 4689		}
 4690		break;
 4691
 4692	default:
 4693		break;
 4694	}
 4695	return 0;
 4696}
 4697
 4698struct bpf_reg_types {
 4699	const enum bpf_reg_type types[10];
 4700	u32 *btf_id;
 4701};
 4702
 4703static const struct bpf_reg_types map_key_value_types = {
 4704	.types = {
 4705		PTR_TO_STACK,
 4706		PTR_TO_PACKET,
 4707		PTR_TO_PACKET_META,
 4708		PTR_TO_MAP_KEY,
 4709		PTR_TO_MAP_VALUE,
 4710	},
 4711};
 4712
 4713static const struct bpf_reg_types sock_types = {
 4714	.types = {
 4715		PTR_TO_SOCK_COMMON,
 4716		PTR_TO_SOCKET,
 4717		PTR_TO_TCP_SOCK,
 4718		PTR_TO_XDP_SOCK,
 4719	},
 4720};
 4721
 4722#ifdef CONFIG_NET
 4723static const struct bpf_reg_types btf_id_sock_common_types = {
 4724	.types = {
 4725		PTR_TO_SOCK_COMMON,
 4726		PTR_TO_SOCKET,
 4727		PTR_TO_TCP_SOCK,
 4728		PTR_TO_XDP_SOCK,
 4729		PTR_TO_BTF_ID,
 4730	},
 4731	.btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
 4732};
 4733#endif
 4734
 4735static const struct bpf_reg_types mem_types = {
 4736	.types = {
 4737		PTR_TO_STACK,
 4738		PTR_TO_PACKET,
 4739		PTR_TO_PACKET_META,
 4740		PTR_TO_MAP_KEY,
 4741		PTR_TO_MAP_VALUE,
 4742		PTR_TO_MEM,
 4743		PTR_TO_RDONLY_BUF,
 4744		PTR_TO_RDWR_BUF,
 4745	},
 4746};
 4747
 4748static const struct bpf_reg_types int_ptr_types = {
 4749	.types = {
 4750		PTR_TO_STACK,
 4751		PTR_TO_PACKET,
 4752		PTR_TO_PACKET_META,
 4753		PTR_TO_MAP_KEY,
 4754		PTR_TO_MAP_VALUE,
 4755	},
 4756};
 4757
 4758static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
 4759static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
 4760static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
 4761static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
 4762static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
 4763static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
 4764static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
 4765static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
 4766static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
 4767static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
 4768static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
 4769
 4770static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
 4771	[ARG_PTR_TO_MAP_KEY]		= &map_key_value_types,
 4772	[ARG_PTR_TO_MAP_VALUE]		= &map_key_value_types,
 4773	[ARG_PTR_TO_UNINIT_MAP_VALUE]	= &map_key_value_types,
 4774	[ARG_PTR_TO_MAP_VALUE_OR_NULL]	= &map_key_value_types,
 4775	[ARG_CONST_SIZE]		= &scalar_types,
 4776	[ARG_CONST_SIZE_OR_ZERO]	= &scalar_types,
 4777	[ARG_CONST_ALLOC_SIZE_OR_ZERO]	= &scalar_types,
 4778	[ARG_CONST_MAP_PTR]		= &const_map_ptr_types,
 4779	[ARG_PTR_TO_CTX]		= &context_types,
 4780	[ARG_PTR_TO_CTX_OR_NULL]	= &context_types,
 4781	[ARG_PTR_TO_SOCK_COMMON]	= &sock_types,
 4782#ifdef CONFIG_NET
 4783	[ARG_PTR_TO_BTF_ID_SOCK_COMMON]	= &btf_id_sock_common_types,
 4784#endif
 4785	[ARG_PTR_TO_SOCKET]		= &fullsock_types,
 4786	[ARG_PTR_TO_SOCKET_OR_NULL]	= &fullsock_types,
 4787	[ARG_PTR_TO_BTF_ID]		= &btf_ptr_types,
 4788	[ARG_PTR_TO_SPIN_LOCK]		= &spin_lock_types,
 4789	[ARG_PTR_TO_MEM]		= &mem_types,
 4790	[ARG_PTR_TO_MEM_OR_NULL]	= &mem_types,
 4791	[ARG_PTR_TO_UNINIT_MEM]		= &mem_types,
 4792	[ARG_PTR_TO_ALLOC_MEM]		= &alloc_mem_types,
 4793	[ARG_PTR_TO_ALLOC_MEM_OR_NULL]	= &alloc_mem_types,
 4794	[ARG_PTR_TO_INT]		= &int_ptr_types,
 4795	[ARG_PTR_TO_LONG]		= &int_ptr_types,
 4796	[ARG_PTR_TO_PERCPU_BTF_ID]	= &percpu_btf_ptr_types,
 4797	[ARG_PTR_TO_FUNC]		= &func_ptr_types,
 4798	[ARG_PTR_TO_STACK_OR_NULL]	= &stack_ptr_types,
 4799	[ARG_PTR_TO_CONST_STR]		= &const_str_ptr_types,
 4800};
 4801
 4802static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
 4803			  enum bpf_arg_type arg_type,
 4804			  const u32 *arg_btf_id)
 4805{
 4806	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
 4807	enum bpf_reg_type expected, type = reg->type;
 4808	const struct bpf_reg_types *compatible;
 4809	int i, j;
 4810
 4811	compatible = compatible_reg_types[arg_type];
 4812	if (!compatible) {
 4813		verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
 4814		return -EFAULT;
 4815	}
 4816
 4817	for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
 4818		expected = compatible->types[i];
 4819		if (expected == NOT_INIT)
 4820			break;
 4821
 4822		if (type == expected)
 4823			goto found;
 4824	}
 4825
 4826	verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
 4827	for (j = 0; j + 1 < i; j++)
 4828		verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
 4829	verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
 4830	return -EACCES;
 4831
 4832found:
 4833	if (type == PTR_TO_BTF_ID) {
 4834		if (!arg_btf_id) {
 4835			if (!compatible->btf_id) {
 4836				verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
 4837				return -EFAULT;
 4838			}
 4839			arg_btf_id = compatible->btf_id;
 4840		}
 4841
 4842		if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
 4843					  btf_vmlinux, *arg_btf_id)) {
 4844			verbose(env, "R%d is of type %s but %s is expected\n",
 4845				regno, kernel_type_name(reg->btf, reg->btf_id),
 4846				kernel_type_name(btf_vmlinux, *arg_btf_id));
 4847			return -EACCES;
 4848		}
 4849
 4850		if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
 4851			verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
 4852				regno);
 4853			return -EACCES;
 4854		}
 4855	}
 4856
 4857	return 0;
 4858}
 4859
 4860static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
 4861			  struct bpf_call_arg_meta *meta,
 4862			  const struct bpf_func_proto *fn)
 4863{
 4864	u32 regno = BPF_REG_1 + arg;
 4865	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
 4866	enum bpf_arg_type arg_type = fn->arg_type[arg];
 4867	enum bpf_reg_type type = reg->type;
 4868	int err = 0;
 4869
 4870	if (arg_type == ARG_DONTCARE)
 4871		return 0;
 4872
 4873	err = check_reg_arg(env, regno, SRC_OP);
 4874	if (err)
 4875		return err;
 4876
 4877	if (arg_type == ARG_ANYTHING) {
 4878		if (is_pointer_value(env, regno)) {
 4879			verbose(env, "R%d leaks addr into helper function\n",
 4880				regno);
 4881			return -EACCES;
 4882		}
 4883		return 0;
 4884	}
 4885
 4886	if (type_is_pkt_pointer(type) &&
 4887	    !may_access_direct_pkt_data(env, meta, BPF_READ)) {
 4888		verbose(env, "helper access to the packet is not allowed\n");
 4889		return -EACCES;
 4890	}
 4891
 4892	if (arg_type == ARG_PTR_TO_MAP_VALUE ||
 4893	    arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
 4894	    arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
 4895		err = resolve_map_arg_type(env, meta, &arg_type);
 4896		if (err)
 4897			return err;
 4898	}
 4899
 4900	if (register_is_null(reg) && arg_type_may_be_null(arg_type))
 4901		/* A NULL register has a SCALAR_VALUE type, so skip
 4902		 * type checking.
 4903		 */
 4904		goto skip_type_check;
 4905
 4906	err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
 4907	if (err)
 4908		return err;
 4909
 4910	if (type == PTR_TO_CTX) {
 4911		err = check_ctx_reg(env, reg, regno);
 4912		if (err < 0)
 4913			return err;
 4914	}
 4915
 4916skip_type_check:
 4917	if (reg->ref_obj_id) {
 4918		if (meta->ref_obj_id) {
 4919			verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
 4920				regno, reg->ref_obj_id,
 4921				meta->ref_obj_id);
 4922			return -EFAULT;
 4923		}
 4924		meta->ref_obj_id = reg->ref_obj_id;
 4925	}
 4926
 4927	if (arg_type == ARG_CONST_MAP_PTR) {
 4928		/* bpf_map_xxx(map_ptr) call: remember that map_ptr */
 4929		meta->map_ptr = reg->map_ptr;
 4930	} else if (arg_type == ARG_PTR_TO_MAP_KEY) {
 4931		/* bpf_map_xxx(..., map_ptr, ..., key) call:
 4932		 * check that [key, key + map->key_size) are within
 4933		 * stack limits and initialized
 4934		 */
 4935		if (!meta->map_ptr) {
 4936			/* in function declaration map_ptr must come before
 4937			 * map_key, so that it's verified and known before
 4938			 * we have to check map_key here. Otherwise it means
 4939			 * that kernel subsystem misconfigured verifier
 4940			 */
 4941			verbose(env, "invalid map_ptr to access map->key\n");
 4942			return -EACCES;
 4943		}
 4944		err = check_helper_mem_access(env, regno,
 4945					      meta->map_ptr->key_size, false,
 4946					      NULL);
 4947	} else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
 4948		   (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
 4949		    !register_is_null(reg)) ||
 4950		   arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
 4951		/* bpf_map_xxx(..., map_ptr, ..., value) call:
 4952		 * check [value, value + map->value_size) validity
 4953		 */
 4954		if (!meta->map_ptr) {
 4955			/* kernel subsystem misconfigured verifier */
 4956			verbose(env, "invalid map_ptr to access map->value\n");
 4957			return -EACCES;
 4958		}
 4959		meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
 4960		err = check_helper_mem_access(env, regno,
 4961					      meta->map_ptr->value_size, false,
 4962					      meta);
 4963	} else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
 4964		if (!reg->btf_id) {
 4965			verbose(env, "Helper has invalid btf_id in R%d\n", regno);
 4966			return -EACCES;
 4967		}
 4968		meta->ret_btf = reg->btf;
 4969		meta->ret_btf_id = reg->btf_id;
 4970	} else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
 4971		if (meta->func_id == BPF_FUNC_spin_lock) {
 4972			if (process_spin_lock(env, regno, true))
 4973				return -EACCES;
 4974		} else if (meta->func_id == BPF_FUNC_spin_unlock) {
 4975			if (process_spin_lock(env, regno, false))
 4976				return -EACCES;
 4977		} else {
 4978			verbose(env, "verifier internal error\n");
 4979			return -EFAULT;
 4980		}
 4981	} else if (arg_type == ARG_PTR_TO_FUNC) {
 4982		meta->subprogno = reg->subprogno;
 4983	} else if (arg_type_is_mem_ptr(arg_type)) {
 4984		/* The access to this pointer is only checked when we hit the
 4985		 * next is_mem_size argument below.
 4986		 */
 4987		meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
 4988	} else if (arg_type_is_mem_size(arg_type)) {
 4989		bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
 4990
 4991		/* This is used to refine r0 return value bounds for helpers
 4992		 * that enforce this value as an upper bound on return values.
 4993		 * See do_refine_retval_range() for helpers that can refine
 4994		 * the return value. C type of helper is u32 so we pull register
 4995		 * bound from umax_value however, if negative verifier errors
 4996		 * out. Only upper bounds can be learned because retval is an
 4997		 * int type and negative retvals are allowed.
 4998		 */
 4999		meta->msize_max_value = reg->umax_value;
 5000
 5001		/* The register is SCALAR_VALUE; the access check
 5002		 * happens using its boundaries.
 5003		 */
 5004		if (!tnum_is_const(reg->var_off))
 5005			/* For unprivileged variable accesses, disable raw
 5006			 * mode so that the program is required to
 5007			 * initialize all the memory that the helper could
 5008			 * just partially fill up.
 5009			 */
 5010			meta = NULL;
 5011
 5012		if (reg->smin_value < 0) {
 5013			verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
 5014				regno);
 5015			return -EACCES;
 5016		}
 5017
 5018		if (reg->umin_value == 0) {
 5019			err = check_helper_mem_access(env, regno - 1, 0,
 5020						      zero_size_allowed,
 5021						      meta);
 5022			if (err)
 5023				return err;
 5024		}
 5025
 5026		if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
 5027			verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
 5028				regno);
 5029			return -EACCES;
 5030		}
 5031		err = check_helper_mem_access(env, regno - 1,
 5032					      reg->umax_value,
 5033					      zero_size_allowed, meta);
 5034		if (!err)
 5035			err = mark_chain_precision(env, regno);
 5036	} else if (arg_type_is_alloc_size(arg_type)) {
 5037		if (!tnum_is_const(reg->var_off)) {
 5038			verbose(env, "R%d is not a known constant'\n",
 5039				regno);
 5040			return -EACCES;
 5041		}
 5042		meta->mem_size = reg->var_off.value;
 5043	} else if (arg_type_is_int_ptr(arg_type)) {
 5044		int size = int_ptr_type_to_size(arg_type);
 5045
 5046		err = check_helper_mem_access(env, regno, size, false, meta);
 5047		if (err)
 5048			return err;
 5049		err = check_ptr_alignment(env, reg, 0, size, true);
 5050	} else if (arg_type == ARG_PTR_TO_CONST_STR) {
 5051		struct bpf_map *map = reg->map_ptr;
 5052		int map_off;
 5053		u64 map_addr;
 5054		char *str_ptr;
 5055
 5056		if (!bpf_map_is_rdonly(map)) {
 5057			verbose(env, "R%d does not point to a readonly map'\n", regno);
 5058			return -EACCES;
 5059		}
 5060
 5061		if (!tnum_is_const(reg->var_off)) {
 5062			verbose(env, "R%d is not a constant address'\n", regno);
 5063			return -EACCES;
 5064		}
 5065
 5066		if (!map->ops->map_direct_value_addr) {
 5067			verbose(env, "no direct value access support for this map type\n");
 5068			return -EACCES;
 5069		}
 5070
 5071		err = check_map_access(env, regno, reg->off,
 5072				       map->value_size - reg->off, false);
 5073		if (err)
 5074			return err;
 5075
 5076		map_off = reg->off + reg->var_off.value;
 5077		err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
 5078		if (err) {
 5079			verbose(env, "direct value access on string failed\n");
 5080			return err;
 5081		}
 5082
 5083		str_ptr = (char *)(long)(map_addr);
 5084		if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
 5085			verbose(env, "string is not zero-terminated\n");
 5086			return -EINVAL;
 5087		}
 5088	}
 5089
 5090	return err;
 5091}
 5092
 5093static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
 5094{
 5095	enum bpf_attach_type eatype = env->prog->expected_attach_type;
 5096	enum bpf_prog_type type = resolve_prog_type(env->prog);
 5097
 5098	if (func_id != BPF_FUNC_map_update_elem)
 5099		return false;
 5100
 5101	/* It's not possible to get access to a locked struct sock in these
 5102	 * contexts, so updating is safe.
 5103	 */
 5104	switch (type) {
 5105	case BPF_PROG_TYPE_TRACING:
 5106		if (eatype == BPF_TRACE_ITER)
 5107			return true;
 5108		break;
 5109	case BPF_PROG_TYPE_SOCKET_FILTER:
 5110	case BPF_PROG_TYPE_SCHED_CLS:
 5111	case BPF_PROG_TYPE_SCHED_ACT:
 5112	case BPF_PROG_TYPE_XDP:
 5113	case BPF_PROG_TYPE_SK_REUSEPORT:
 5114	case BPF_PROG_TYPE_FLOW_DISSECTOR:
 5115	case BPF_PROG_TYPE_SK_LOOKUP:
 5116		return true;
 5117	default:
 5118		break;
 5119	}
 5120
 5121	verbose(env, "cannot update sockmap in this context\n");
 5122	return false;
 5123}
 5124
 5125static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
 5126{
 5127	return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
 5128}
 5129
 5130static int check_map_func_compatibility(struct bpf_verifier_env *env,
 5131					struct bpf_map *map, int func_id)
 5132{
 5133	if (!map)
 5134		return 0;
 5135
 5136	/* We need a two way check, first is from map perspective ... */
 5137	switch (map->map_type) {
 5138	case BPF_MAP_TYPE_PROG_ARRAY:
 5139		if (func_id != BPF_FUNC_tail_call)
 5140			goto error;
 5141		break;
 5142	case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
 5143		if (func_id != BPF_FUNC_perf_event_read &&
 5144		    func_id != BPF_FUNC_perf_event_output &&
 5145		    func_id != BPF_FUNC_skb_output &&
 5146		    func_id != BPF_FUNC_perf_event_read_value &&
 5147		    func_id != BPF_FUNC_xdp_output)
 5148			goto error;
 5149		break;
 5150	case BPF_MAP_TYPE_RINGBUF:
 5151		if (func_id != BPF_FUNC_ringbuf_output &&
 5152		    func_id != BPF_FUNC_ringbuf_reserve &&
 5153		    func_id != BPF_FUNC_ringbuf_query)
 5154			goto error;
 5155		break;
 5156	case BPF_MAP_TYPE_STACK_TRACE:
 5157		if (func_id != BPF_FUNC_get_stackid)
 5158			goto error;
 5159		break;
 5160	case BPF_MAP_TYPE_CGROUP_ARRAY:
 5161		if (func_id != BPF_FUNC_skb_under_cgroup &&
 5162		    func_id != BPF_FUNC_current_task_under_cgroup)
 5163			goto error;
 5164		break;
 5165	case BPF_MAP_TYPE_CGROUP_STORAGE:
 5166	case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
 5167		if (func_id != BPF_FUNC_get_local_storage)
 5168			goto error;
 5169		break;
 5170	case BPF_MAP_TYPE_DEVMAP:
 5171	case BPF_MAP_TYPE_DEVMAP_HASH:
 5172		if (func_id != BPF_FUNC_redirect_map &&
 5173		    func_id != BPF_FUNC_map_lookup_elem)
 5174			goto error;
 5175		break;
 5176	/* Restrict bpf side of cpumap and xskmap, open when use-cases
 5177	 * appear.
 5178	 */
 5179	case BPF_MAP_TYPE_CPUMAP:
 5180		if (func_id != BPF_FUNC_redirect_map)
 5181			goto error;
 5182		break;
 5183	case BPF_MAP_TYPE_XSKMAP:
 5184		if (func_id != BPF_FUNC_redirect_map &&
 5185		    func_id != BPF_FUNC_map_lookup_elem)
 5186			goto error;
 5187		break;
 5188	case BPF_MAP_TYPE_ARRAY_OF_MAPS:
 5189	case BPF_MAP_TYPE_HASH_OF_MAPS:
 5190		if (func_id != BPF_FUNC_map_lookup_elem)
 5191			goto error;
 5192		break;
 5193	case BPF_MAP_TYPE_SOCKMAP:
 5194		if (func_id != BPF_FUNC_sk_redirect_map &&
 5195		    func_id != BPF_FUNC_sock_map_update &&
 5196		    func_id != BPF_FUNC_map_delete_elem &&
 5197		    func_id != BPF_FUNC_msg_redirect_map &&
 5198		    func_id != BPF_FUNC_sk_select_reuseport &&
 5199		    func_id != BPF_FUNC_map_lookup_elem &&
 5200		    !may_update_sockmap(env, func_id))
 5201			goto error;
 5202		break;
 5203	case BPF_MAP_TYPE_SOCKHASH:
 5204		if (func_id != BPF_FUNC_sk_redirect_hash &&
 5205		    func_id != BPF_FUNC_sock_hash_update &&
 5206		    func_id != BPF_FUNC_map_delete_elem &&
 5207		    func_id != BPF_FUNC_msg_redirect_hash &&
 5208		    func_id != BPF_FUNC_sk_select_reuseport &&
 5209		    func_id != BPF_FUNC_map_lookup_elem &&
 5210		    !may_update_sockmap(env, func_id))
 5211			goto error;
 5212		break;
 5213	case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
 5214		if (func_id != BPF_FUNC_sk_select_reuseport)
 5215			goto error;
 5216		break;
 5217	case BPF_MAP_TYPE_QUEUE:
 5218	case BPF_MAP_TYPE_STACK:
 5219		if (func_id != BPF_FUNC_map_peek_elem &&
 5220		    func_id != BPF_FUNC_map_pop_elem &&
 5221		    func_id != BPF_FUNC_map_push_elem)
 5222			goto error;
 5223		break;
 5224	case BPF_MAP_TYPE_SK_STORAGE:
 5225		if (func_id != BPF_FUNC_sk_storage_get &&
 5226		    func_id != BPF_FUNC_sk_storage_delete)
 5227			goto error;
 5228		break;
 5229	case BPF_MAP_TYPE_INODE_STORAGE:
 5230		if (func_id != BPF_FUNC_inode_storage_get &&
 5231		    func_id != BPF_FUNC_inode_storage_delete)
 5232			goto error;
 5233		break;
 5234	case BPF_MAP_TYPE_TASK_STORAGE:
 5235		if (func_id != BPF_FUNC_task_storage_get &&
 5236		    func_id != BPF_FUNC_task_storage_delete)
 5237			goto error;
 5238		break;
 5239	default:
 5240		break;
 5241	}
 5242
 5243	/* ... and second from the function itself. */
 5244	switch (func_id) {
 5245	case BPF_FUNC_tail_call:
 5246		if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
 5247			goto error;
 5248		if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
 5249			verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
 5250			return -EINVAL;
 5251		}
 5252		break;
 5253	case BPF_FUNC_perf_event_read:
 5254	case BPF_FUNC_perf_event_output:
 5255	case BPF_FUNC_perf_event_read_value:
 5256	case BPF_FUNC_skb_output:
 5257	case BPF_FUNC_xdp_output:
 5258		if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
 5259			goto error;
 5260		break;
 5261	case BPF_FUNC_ringbuf_output:
 5262	case BPF_FUNC_ringbuf_reserve:
 5263	case BPF_FUNC_ringbuf_query:
 5264		if (map->map_type != BPF_MAP_TYPE_RINGBUF)
 5265			goto error;
 5266		break;
 5267	case BPF_FUNC_get_stackid:
 5268		if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
 5269			goto error;
 5270		break;
 5271	case BPF_FUNC_current_task_under_cgroup:
 5272	case BPF_FUNC_skb_under_cgroup:
 5273		if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
 5274			goto error;
 5275		break;
 5276	case BPF_FUNC_redirect_map:
 5277		if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
 5278		    map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
 5279		    map->map_type != BPF_MAP_TYPE_CPUMAP &&
 5280		    map->map_type != BPF_MAP_TYPE_XSKMAP)
 5281			goto error;
 5282		break;
 5283	case BPF_FUNC_sk_redirect_map:
 5284	case BPF_FUNC_msg_redirect_map:
 5285	case BPF_FUNC_sock_map_update:
 5286		if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
 5287			goto error;
 5288		break;
 5289	case BPF_FUNC_sk_redirect_hash:
 5290	case BPF_FUNC_msg_redirect_hash:
 5291	case BPF_FUNC_sock_hash_update:
 5292		if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
 5293			goto error;
 5294		break;
 5295	case BPF_FUNC_get_local_storage:
 5296		if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
 5297		    map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
 5298			goto error;
 5299		break;
 5300	case BPF_FUNC_sk_select_reuseport:
 5301		if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
 5302		    map->map_type != BPF_MAP_TYPE_SOCKMAP &&
 5303		    map->map_type != BPF_MAP_TYPE_SOCKHASH)
 5304			goto error;
 5305		break;
 5306	case BPF_FUNC_map_peek_elem:
 5307	case BPF_FUNC_map_pop_elem:
 5308	case BPF_FUNC_map_push_elem:
 5309		if (map->map_type != BPF_MAP_TYPE_QUEUE &&
 5310		    map->map_type != BPF_MAP_TYPE_STACK)
 5311			goto error;
 5312		break;
 5313	case BPF_FUNC_sk_storage_get:
 5314	case BPF_FUNC_sk_storage_delete:
 5315		if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
 5316			goto error;
 5317		break;
 5318	case BPF_FUNC_inode_storage_get:
 5319	case BPF_FUNC_inode_storage_delete:
 5320		if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
 5321			goto error;
 5322		break;
 5323	case BPF_FUNC_task_storage_get:
 5324	case BPF_FUNC_task_storage_delete:
 5325		if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
 5326			goto error;
 5327		break;
 5328	default:
 5329		break;
 5330	}
 5331
 5332	return 0;
 5333error:
 5334	verbose(env, "cannot pass map_type %d into func %s#%d\n",
 5335		map->map_type, func_id_name(func_id), func_id);
 5336	return -EINVAL;
 5337}
 5338
 5339static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
 5340{
 5341	int count = 0;
 5342
 5343	if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
 5344		count++;
 5345	if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
 5346		count++;
 5347	if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
 5348		count++;
 5349	if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
 5350		count++;
 5351	if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
 5352		count++;
 5353
 5354	/* We only support one arg being in raw mode at the moment,
 5355	 * which is sufficient for the helper functions we have
 5356	 * right now.
 5357	 */
 5358	return count <= 1;
 5359}
 5360
 5361static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
 5362				    enum bpf_arg_type arg_next)
 5363{
 5364	return (arg_type_is_mem_ptr(arg_curr) &&
 5365	        !arg_type_is_mem_size(arg_next)) ||
 5366	       (!arg_type_is_mem_ptr(arg_curr) &&
 5367		arg_type_is_mem_size(arg_next));
 5368}
 5369
 5370static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
 5371{
 5372	/* bpf_xxx(..., buf, len) call will access 'len'
 5373	 * bytes from memory 'buf'. Both arg types need
 5374	 * to be paired, so make sure there's no buggy
 5375	 * helper function specification.
 5376	 */
 5377	if (arg_type_is_mem_size(fn->arg1_type) ||
 5378	    arg_type_is_mem_ptr(fn->arg5_type)  ||
 5379	    check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
 5380	    check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
 5381	    check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
 5382	    check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
 5383		return false;
 5384
 5385	return true;
 5386}
 5387
 5388static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
 5389{
 5390	int count = 0;
 5391
 5392	if (arg_type_may_be_refcounted(fn->arg1_type))
 5393		count++;
 5394	if (arg_type_may_be_refcounted(fn->arg2_type))
 5395		count++;
 5396	if (arg_type_may_be_refcounted(fn->arg3_type))
 5397		count++;
 5398	if (arg_type_may_be_refcounted(fn->arg4_type))
 5399		count++;
 5400	if (arg_type_may_be_refcounted(fn->arg5_type))
 5401		count++;
 5402
 5403	/* A reference acquiring function cannot acquire
 5404	 * another refcounted ptr.
 5405	 */
 5406	if (may_be_acquire_function(func_id) && count)
 5407		return false;
 5408
 5409	/* We only support one arg being unreferenced at the moment,
 5410	 * which is sufficient for the helper functions we have right now.
 5411	 */
 5412	return count <= 1;
 5413}
 5414
 5415static bool check_btf_id_ok(const struct bpf_func_proto *fn)
 5416{
 5417	int i;
 5418
 5419	for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
 5420		if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
 5421			return false;
 5422
 5423		if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
 5424			return false;
 5425	}
 5426
 5427	return true;
 5428}
 5429
 5430static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
 5431{
 5432	return check_raw_mode_ok(fn) &&
 5433	       check_arg_pair_ok(fn) &&
 5434	       check_btf_id_ok(fn) &&
 5435	       check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
 5436}
 5437
 5438/* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
 5439 * are now invalid, so turn them into unknown SCALAR_VALUE.
 5440 */
 5441static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
 5442				     struct bpf_func_state *state)
 5443{
 5444	struct bpf_reg_state *regs = state->regs, *reg;
 5445	int i;
 5446
 5447	for (i = 0; i < MAX_BPF_REG; i++)
 5448		if (reg_is_pkt_pointer_any(&regs[i]))
 5449			mark_reg_unknown(env, regs, i);
 5450
 5451	bpf_for_each_spilled_reg(i, state, reg) {
 5452		if (!reg)
 5453			continue;
 5454		if (reg_is_pkt_pointer_any(reg))
 5455			__mark_reg_unknown(env, reg);
 5456	}
 5457}
 5458
 5459static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
 5460{
 5461	struct bpf_verifier_state *vstate = env->cur_state;
 5462	int i;
 5463
 5464	for (i = 0; i <= vstate->curframe; i++)
 5465		__clear_all_pkt_pointers(env, vstate->frame[i]);
 5466}
 5467
 5468enum {
 5469	AT_PKT_END = -1,
 5470	BEYOND_PKT_END = -2,
 5471};
 5472
 5473static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
 5474{
 5475	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 5476	struct bpf_reg_state *reg = &state->regs[regn];
 5477
 5478	if (reg->type != PTR_TO_PACKET)
 5479		/* PTR_TO_PACKET_META is not supported yet */
 5480		return;
 5481
 5482	/* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
 5483	 * How far beyond pkt_end it goes is unknown.
 5484	 * if (!range_open) it's the case of pkt >= pkt_end
 5485	 * if (range_open) it's the case of pkt > pkt_end
 5486	 * hence this pointer is at least 1 byte bigger than pkt_end
 5487	 */
 5488	if (range_open)
 5489		reg->range = BEYOND_PKT_END;
 5490	else
 5491		reg->range = AT_PKT_END;
 5492}
 5493
 5494static void release_reg_references(struct bpf_verifier_env *env,
 5495				   struct bpf_func_state *state,
 5496				   int ref_obj_id)
 5497{
 5498	struct bpf_reg_state *regs = state->regs, *reg;
 5499	int i;
 5500
 5501	for (i = 0; i < MAX_BPF_REG; i++)
 5502		if (regs[i].ref_obj_id == ref_obj_id)
 5503			mark_reg_unknown(env, regs, i);
 5504
 5505	bpf_for_each_spilled_reg(i, state, reg) {
 5506		if (!reg)
 5507			continue;
 5508		if (reg->ref_obj_id == ref_obj_id)
 5509			__mark_reg_unknown(env, reg);
 5510	}
 5511}
 5512
 5513/* The pointer with the specified id has released its reference to kernel
 5514 * resources. Identify all copies of the same pointer and clear the reference.
 5515 */
 5516static int release_reference(struct bpf_verifier_env *env,
 5517			     int ref_obj_id)
 5518{
 5519	struct bpf_verifier_state *vstate = env->cur_state;
 5520	int err;
 5521	int i;
 5522
 5523	err = release_reference_state(cur_func(env), ref_obj_id);
 5524	if (err)
 5525		return err;
 5526
 5527	for (i = 0; i <= vstate->curframe; i++)
 5528		release_reg_references(env, vstate->frame[i], ref_obj_id);
 5529
 5530	return 0;
 5531}
 5532
 5533static void clear_caller_saved_regs(struct bpf_verifier_env *env,
 5534				    struct bpf_reg_state *regs)
 5535{
 5536	int i;
 5537
 5538	/* after the call registers r0 - r5 were scratched */
 5539	for (i = 0; i < CALLER_SAVED_REGS; i++) {
 5540		mark_reg_not_init(env, regs, caller_saved[i]);
 5541		check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
 5542	}
 5543}
 5544
 5545typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
 5546				   struct bpf_func_state *caller,
 5547				   struct bpf_func_state *callee,
 5548				   int insn_idx);
 5549
 5550static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
 5551			     int *insn_idx, int subprog,
 5552			     set_callee_state_fn set_callee_state_cb)
 5553{
 5554	struct bpf_verifier_state *state = env->cur_state;
 5555	struct bpf_func_info_aux *func_info_aux;
 5556	struct bpf_func_state *caller, *callee;
 5557	int err;
 5558	bool is_global = false;
 5559
 5560	if (state->curframe + 1 >= MAX_CALL_FRAMES) {
 5561		verbose(env, "the call stack of %d frames is too deep\n",
 5562			state->curframe + 2);
 5563		return -E2BIG;
 5564	}
 5565
 5566	caller = state->frame[state->curframe];
 5567	if (state->frame[state->curframe + 1]) {
 5568		verbose(env, "verifier bug. Frame %d already allocated\n",
 5569			state->curframe + 1);
 5570		return -EFAULT;
 5571	}
 5572
 5573	func_info_aux = env->prog->aux->func_info_aux;
 5574	if (func_info_aux)
 5575		is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
 5576	err = btf_check_subprog_arg_match(env, subprog, caller->regs);
 5577	if (err == -EFAULT)
 5578		return err;
 5579	if (is_global) {
 5580		if (err) {
 5581			verbose(env, "Caller passes invalid args into func#%d\n",
 5582				subprog);
 5583			return err;
 5584		} else {
 5585			if (env->log.level & BPF_LOG_LEVEL)
 5586				verbose(env,
 5587					"Func#%d is global and valid. Skipping.\n",
 5588					subprog);
 5589			clear_caller_saved_regs(env, caller->regs);
 5590
 5591			/* All global functions return a 64-bit SCALAR_VALUE */
 5592			mark_reg_unknown(env, caller->regs, BPF_REG_0);
 5593			caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
 5594
 5595			/* continue with next insn after call */
 5596			return 0;
 5597		}
 5598	}
 5599
 5600	callee = kzalloc(sizeof(*callee), GFP_KERNEL);
 5601	if (!callee)
 5602		return -ENOMEM;
 5603	state->frame[state->curframe + 1] = callee;
 5604
 5605	/* callee cannot access r0, r6 - r9 for reading and has to write
 5606	 * into its own stack before reading from it.
 5607	 * callee can read/write into caller's stack
 5608	 */
 5609	init_func_state(env, callee,
 5610			/* remember the callsite, it will be used by bpf_exit */
 5611			*insn_idx /* callsite */,
 5612			state->curframe + 1 /* frameno within this callchain */,
 5613			subprog /* subprog number within this prog */);
 5614
 5615	/* Transfer references to the callee */
 5616	err = copy_reference_state(callee, caller);
 5617	if (err)
 5618		return err;
 5619
 5620	err = set_callee_state_cb(env, caller, callee, *insn_idx);
 5621	if (err)
 5622		return err;
 5623
 5624	clear_caller_saved_regs(env, caller->regs);
 5625
 5626	/* only increment it after check_reg_arg() finished */
 5627	state->curframe++;
 5628
 5629	/* and go analyze first insn of the callee */
 5630	*insn_idx = env->subprog_info[subprog].start - 1;
 5631
 5632	if (env->log.level & BPF_LOG_LEVEL) {
 5633		verbose(env, "caller:\n");
 5634		print_verifier_state(env, caller);
 5635		verbose(env, "callee:\n");
 5636		print_verifier_state(env, callee);
 5637	}
 5638	return 0;
 5639}
 5640
 5641int map_set_for_each_callback_args(struct bpf_verifier_env *env,
 5642				   struct bpf_func_state *caller,
 5643				   struct bpf_func_state *callee)
 5644{
 5645	/* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
 5646	 *      void *callback_ctx, u64 flags);
 5647	 * callback_fn(struct bpf_map *map, void *key, void *value,
 5648	 *      void *callback_ctx);
 5649	 */
 5650	callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
 5651
 5652	callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
 5653	__mark_reg_known_zero(&callee->regs[BPF_REG_2]);
 5654	callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
 5655
 5656	callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
 5657	__mark_reg_known_zero(&callee->regs[BPF_REG_3]);
 5658	callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
 5659
 5660	/* pointer to stack or null */
 5661	callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
 5662
 5663	/* unused */
 5664	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
 5665	return 0;
 5666}
 5667
 5668static int set_callee_state(struct bpf_verifier_env *env,
 5669			    struct bpf_func_state *caller,
 5670			    struct bpf_func_state *callee, int insn_idx)
 5671{
 5672	int i;
 5673
 5674	/* copy r1 - r5 args that callee can access.  The copy includes parent
 5675	 * pointers, which connects us up to the liveness chain
 5676	 */
 5677	for (i = BPF_REG_1; i <= BPF_REG_5; i++)
 5678		callee->regs[i] = caller->regs[i];
 5679	return 0;
 5680}
 5681
 5682static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
 5683			   int *insn_idx)
 5684{
 5685	int subprog, target_insn;
 5686
 5687	target_insn = *insn_idx + insn->imm + 1;
 5688	subprog = find_subprog(env, target_insn);
 5689	if (subprog < 0) {
 5690		verbose(env, "verifier bug. No program starts at insn %d\n",
 5691			target_insn);
 5692		return -EFAULT;
 5693	}
 5694
 5695	return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
 5696}
 5697
 5698static int set_map_elem_callback_state(struct bpf_verifier_env *env,
 5699				       struct bpf_func_state *caller,
 5700				       struct bpf_func_state *callee,
 5701				       int insn_idx)
 5702{
 5703	struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
 5704	struct bpf_map *map;
 5705	int err;
 5706
 5707	if (bpf_map_ptr_poisoned(insn_aux)) {
 5708		verbose(env, "tail_call abusing map_ptr\n");
 5709		return -EINVAL;
 5710	}
 5711
 5712	map = BPF_MAP_PTR(insn_aux->map_ptr_state);
 5713	if (!map->ops->map_set_for_each_callback_args ||
 5714	    !map->ops->map_for_each_callback) {
 5715		verbose(env, "callback function not allowed for map\n");
 5716		return -ENOTSUPP;
 5717	}
 5718
 5719	err = map->ops->map_set_for_each_callback_args(env, caller, callee);
 5720	if (err)
 5721		return err;
 5722
 5723	callee->in_callback_fn = true;
 5724	return 0;
 5725}
 5726
 5727static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
 5728{
 5729	struct bpf_verifier_state *state = env->cur_state;
 5730	struct bpf_func_state *caller, *callee;
 5731	struct bpf_reg_state *r0;
 5732	int err;
 5733
 5734	callee = state->frame[state->curframe];
 5735	r0 = &callee->regs[BPF_REG_0];
 5736	if (r0->type == PTR_TO_STACK) {
 5737		/* technically it's ok to return caller's stack pointer
 5738		 * (or caller's caller's pointer) back to the caller,
 5739		 * since these pointers are valid. Only current stack
 5740		 * pointer will be invalid as soon as function exits,
 5741		 * but let's be conservative
 5742		 */
 5743		verbose(env, "cannot return stack pointer to the caller\n");
 5744		return -EINVAL;
 5745	}
 5746
 5747	state->curframe--;
 5748	caller = state->frame[state->curframe];
 5749	if (callee->in_callback_fn) {
 5750		/* enforce R0 return value range [0, 1]. */
 5751		struct tnum range = tnum_range(0, 1);
 5752
 5753		if (r0->type != SCALAR_VALUE) {
 5754			verbose(env, "R0 not a scalar value\n");
 5755			return -EACCES;
 5756		}
 5757		if (!tnum_in(range, r0->var_off)) {
 5758			verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
 5759			return -EINVAL;
 5760		}
 5761	} else {
 5762		/* return to the caller whatever r0 had in the callee */
 5763		caller->regs[BPF_REG_0] = *r0;
 5764	}
 5765
 5766	/* Transfer references to the caller */
 5767	err = copy_reference_state(caller, callee);
 5768	if (err)
 5769		return err;
 5770
 5771	*insn_idx = callee->callsite + 1;
 5772	if (env->log.level & BPF_LOG_LEVEL) {
 5773		verbose(env, "returning from callee:\n");
 5774		print_verifier_state(env, callee);
 5775		verbose(env, "to caller at %d:\n", *insn_idx);
 5776		print_verifier_state(env, caller);
 5777	}
 5778	/* clear everything in the callee */
 5779	free_func_state(callee);
 5780	state->frame[state->curframe + 1] = NULL;
 5781	return 0;
 5782}
 5783
 5784static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
 5785				   int func_id,
 5786				   struct bpf_call_arg_meta *meta)
 5787{
 5788	struct bpf_reg_state *ret_reg = &regs[BPF_REG_0];
 5789
 5790	if (ret_type != RET_INTEGER ||
 5791	    (func_id != BPF_FUNC_get_stack &&
 5792	     func_id != BPF_FUNC_get_task_stack &&
 5793	     func_id != BPF_FUNC_probe_read_str &&
 5794	     func_id != BPF_FUNC_probe_read_kernel_str &&
 5795	     func_id != BPF_FUNC_probe_read_user_str))
 5796		return;
 5797
 5798	ret_reg->smax_value = meta->msize_max_value;
 5799	ret_reg->s32_max_value = meta->msize_max_value;
 5800	ret_reg->smin_value = -MAX_ERRNO;
 5801	ret_reg->s32_min_value = -MAX_ERRNO;
 5802	__reg_deduce_bounds(ret_reg);
 5803	__reg_bound_offset(ret_reg);
 5804	__update_reg_bounds(ret_reg);
 5805}
 5806
 5807static int
 5808record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
 5809		int func_id, int insn_idx)
 5810{
 5811	struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
 5812	struct bpf_map *map = meta->map_ptr;
 5813
 5814	if (func_id != BPF_FUNC_tail_call &&
 5815	    func_id != BPF_FUNC_map_lookup_elem &&
 5816	    func_id != BPF_FUNC_map_update_elem &&
 5817	    func_id != BPF_FUNC_map_delete_elem &&
 5818	    func_id != BPF_FUNC_map_push_elem &&
 5819	    func_id != BPF_FUNC_map_pop_elem &&
 5820	    func_id != BPF_FUNC_map_peek_elem &&
 5821	    func_id != BPF_FUNC_for_each_map_elem &&
 5822	    func_id != BPF_FUNC_redirect_map)
 5823		return 0;
 5824
 5825	if (map == NULL) {
 5826		verbose(env, "kernel subsystem misconfigured verifier\n");
 5827		return -EINVAL;
 5828	}
 5829
 5830	/* In case of read-only, some additional restrictions
 5831	 * need to be applied in order to prevent altering the
 5832	 * state of the map from program side.
 5833	 */
 5834	if ((map->map_flags & BPF_F_RDONLY_PROG) &&
 5835	    (func_id == BPF_FUNC_map_delete_elem ||
 5836	     func_id == BPF_FUNC_map_update_elem ||
 5837	     func_id == BPF_FUNC_map_push_elem ||
 5838	     func_id == BPF_FUNC_map_pop_elem)) {
 5839		verbose(env, "write into map forbidden\n");
 5840		return -EACCES;
 5841	}
 5842
 5843	if (!BPF_MAP_PTR(aux->map_ptr_state))
 5844		bpf_map_ptr_store(aux, meta->map_ptr,
 5845				  !meta->map_ptr->bypass_spec_v1);
 5846	else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
 5847		bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
 5848				  !meta->map_ptr->bypass_spec_v1);
 5849	return 0;
 5850}
 5851
 5852static int
 5853record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
 5854		int func_id, int insn_idx)
 5855{
 5856	struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
 5857	struct bpf_reg_state *regs = cur_regs(env), *reg;
 5858	struct bpf_map *map = meta->map_ptr;
 5859	struct tnum range;
 5860	u64 val;
 5861	int err;
 5862
 5863	if (func_id != BPF_FUNC_tail_call)
 5864		return 0;
 5865	if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
 5866		verbose(env, "kernel subsystem misconfigured verifier\n");
 5867		return -EINVAL;
 5868	}
 5869
 5870	range = tnum_range(0, map->max_entries - 1);
 5871	reg = &regs[BPF_REG_3];
 5872
 5873	if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
 5874		bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
 5875		return 0;
 5876	}
 5877
 5878	err = mark_chain_precision(env, BPF_REG_3);
 5879	if (err)
 5880		return err;
 5881
 5882	val = reg->var_off.value;
 5883	if (bpf_map_key_unseen(aux))
 5884		bpf_map_key_store(aux, val);
 5885	else if (!bpf_map_key_poisoned(aux) &&
 5886		  bpf_map_key_immediate(aux) != val)
 5887		bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
 5888	return 0;
 5889}
 5890
 5891static int check_reference_leak(struct bpf_verifier_env *env)
 5892{
 5893	struct bpf_func_state *state = cur_func(env);
 5894	int i;
 5895
 5896	for (i = 0; i < state->acquired_refs; i++) {
 5897		verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
 5898			state->refs[i].id, state->refs[i].insn_idx);
 5899	}
 5900	return state->acquired_refs ? -EINVAL : 0;
 5901}
 5902
 5903static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
 5904				   struct bpf_reg_state *regs)
 5905{
 5906	struct bpf_reg_state *fmt_reg = &regs[BPF_REG_3];
 5907	struct bpf_reg_state *data_len_reg = &regs[BPF_REG_5];
 5908	struct bpf_map *fmt_map = fmt_reg->map_ptr;
 5909	int err, fmt_map_off, num_args;
 5910	u64 fmt_addr;
 5911	char *fmt;
 5912
 5913	/* data must be an array of u64 */
 5914	if (data_len_reg->var_off.value % 8)
 5915		return -EINVAL;
 5916	num_args = data_len_reg->var_off.value / 8;
 5917
 5918	/* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
 5919	 * and map_direct_value_addr is set.
 5920	 */
 5921	fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
 5922	err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
 5923						  fmt_map_off);
 5924	if (err) {
 5925		verbose(env, "verifier bug\n");
 5926		return -EFAULT;
 5927	}
 5928	fmt = (char *)(long)fmt_addr + fmt_map_off;
 5929
 5930	/* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
 5931	 * can focus on validating the format specifiers.
 5932	 */
 5933	err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
 5934	if (err < 0)
 5935		verbose(env, "Invalid format string\n");
 5936
 5937	return err;
 5938}
 5939
 5940static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
 5941			     int *insn_idx_p)
 5942{
 5943	const struct bpf_func_proto *fn = NULL;
 5944	struct bpf_reg_state *regs;
 5945	struct bpf_call_arg_meta meta;
 5946	int insn_idx = *insn_idx_p;
 5947	bool changes_data;
 5948	int i, err, func_id;
 5949
 5950	/* find function prototype */
 5951	func_id = insn->imm;
 5952	if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
 5953		verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
 5954			func_id);
 5955		return -EINVAL;
 5956	}
 5957
 5958	if (env->ops->get_func_proto)
 5959		fn = env->ops->get_func_proto(func_id, env->prog);
 5960	if (!fn) {
 5961		verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
 5962			func_id);
 5963		return -EINVAL;
 5964	}
 5965
 5966	/* eBPF programs must be GPL compatible to use GPL-ed functions */
 5967	if (!env->prog->gpl_compatible && fn->gpl_only) {
 5968		verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
 5969		return -EINVAL;
 5970	}
 5971
 5972	if (fn->allowed && !fn->allowed(env->prog)) {
 5973		verbose(env, "helper call is not allowed in probe\n");
 5974		return -EINVAL;
 5975	}
 5976
 5977	/* With LD_ABS/IND some JITs save/restore skb from r1. */
 5978	changes_data = bpf_helper_changes_pkt_data(fn->func);
 5979	if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
 5980		verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
 5981			func_id_name(func_id), func_id);
 5982		return -EINVAL;
 5983	}
 5984
 5985	memset(&meta, 0, sizeof(meta));
 5986	meta.pkt_access = fn->pkt_access;
 5987
 5988	err = check_func_proto(fn, func_id);
 5989	if (err) {
 5990		verbose(env, "kernel subsystem misconfigured func %s#%d\n",
 5991			func_id_name(func_id), func_id);
 5992		return err;
 5993	}
 5994
 5995	meta.func_id = func_id;
 5996	/* check args */
 5997	for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
 5998		err = check_func_arg(env, i, &meta, fn);
 5999		if (err)
 6000			return err;
 6001	}
 6002
 6003	err = record_func_map(env, &meta, func_id, insn_idx);
 6004	if (err)
 6005		return err;
 6006
 6007	err = record_func_key(env, &meta, func_id, insn_idx);
 6008	if (err)
 6009		return err;
 6010
 6011	/* Mark slots with STACK_MISC in case of raw mode, stack offset
 6012	 * is inferred from register state.
 6013	 */
 6014	for (i = 0; i < meta.access_size; i++) {
 6015		err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
 6016				       BPF_WRITE, -1, false);
 6017		if (err)
 6018			return err;
 6019	}
 6020
 6021	if (func_id == BPF_FUNC_tail_call) {
 6022		err = check_reference_leak(env);
 6023		if (err) {
 6024			verbose(env, "tail_call would lead to reference leak\n");
 6025			return err;
 6026		}
 6027	} else if (is_release_function(func_id)) {
 6028		err = release_reference(env, meta.ref_obj_id);
 6029		if (err) {
 6030			verbose(env, "func %s#%d reference has not been acquired before\n",
 6031				func_id_name(func_id), func_id);
 6032			return err;
 6033		}
 6034	}
 6035
 6036	regs = cur_regs(env);
 6037
 6038	/* check that flags argument in get_local_storage(map, flags) is 0,
 6039	 * this is required because get_local_storage() can't return an error.
 6040	 */
 6041	if (func_id == BPF_FUNC_get_local_storage &&
 6042	    !register_is_null(&regs[BPF_REG_2])) {
 6043		verbose(env, "get_local_storage() doesn't support non-zero flags\n");
 6044		return -EINVAL;
 6045	}
 6046
 6047	if (func_id == BPF_FUNC_for_each_map_elem) {
 6048		err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
 6049					set_map_elem_callback_state);
 6050		if (err < 0)
 6051			return -EINVAL;
 6052	}
 6053
 6054	if (func_id == BPF_FUNC_snprintf) {
 6055		err = check_bpf_snprintf_call(env, regs);
 6056		if (err < 0)
 6057			return err;
 6058	}
 6059
 6060	/* reset caller saved regs */
 6061	for (i = 0; i < CALLER_SAVED_REGS; i++) {
 6062		mark_reg_not_init(env, regs, caller_saved[i]);
 6063		check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
 6064	}
 6065
 6066	/* helper call returns 64-bit value. */
 6067	regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
 6068
 6069	/* update return register (already marked as written above) */
 6070	if (fn->ret_type == RET_INTEGER) {
 6071		/* sets type to SCALAR_VALUE */
 6072		mark_reg_unknown(env, regs, BPF_REG_0);
 6073	} else if (fn->ret_type == RET_VOID) {
 6074		regs[BPF_REG_0].type = NOT_INIT;
 6075	} else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
 6076		   fn->ret_type == RET_PTR_TO_MAP_VALUE) {
 6077		/* There is no offset yet applied, variable or fixed */
 6078		mark_reg_known_zero(env, regs, BPF_REG_0);
 6079		/* remember map_ptr, so that check_map_access()
 6080		 * can check 'value_size' boundary of memory access
 6081		 * to map element returned from bpf_map_lookup_elem()
 6082		 */
 6083		if (meta.map_ptr == NULL) {
 6084			verbose(env,
 6085				"kernel subsystem misconfigured verifier\n");
 6086			return -EINVAL;
 6087		}
 6088		regs[BPF_REG_0].map_ptr = meta.map_ptr;
 6089		if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
 6090			regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
 6091			if (map_value_has_spin_lock(meta.map_ptr))
 6092				regs[BPF_REG_0].id = ++env->id_gen;
 6093		} else {
 6094			regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
 6095		}
 6096	} else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
 6097		mark_reg_known_zero(env, regs, BPF_REG_0);
 6098		regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
 6099	} else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
 6100		mark_reg_known_zero(env, regs, BPF_REG_0);
 6101		regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
 6102	} else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
 6103		mark_reg_known_zero(env, regs, BPF_REG_0);
 6104		regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
 6105	} else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
 6106		mark_reg_known_zero(env, regs, BPF_REG_0);
 6107		regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
 6108		regs[BPF_REG_0].mem_size = meta.mem_size;
 6109	} else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
 6110		   fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
 6111		const struct btf_type *t;
 6112
 6113		mark_reg_known_zero(env, regs, BPF_REG_0);
 6114		t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
 6115		if (!btf_type_is_struct(t)) {
 6116			u32 tsize;
 6117			const struct btf_type *ret;
 6118			const char *tname;
 6119
 6120			/* resolve the type size of ksym. */
 6121			ret = btf_resolve_size(meta.ret_btf, t, &tsize);
 6122			if (IS_ERR(ret)) {
 6123				tname = btf_name_by_offset(meta.ret_btf, t->name_off);
 6124				verbose(env, "unable to resolve the size of type '%s': %ld\n",
 6125					tname, PTR_ERR(ret));
 6126				return -EINVAL;
 6127			}
 6128			regs[BPF_REG_0].type =
 6129				fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
 6130				PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
 6131			regs[BPF_REG_0].mem_size = tsize;
 6132		} else {
 6133			regs[BPF_REG_0].type =
 6134				fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
 6135				PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
 6136			regs[BPF_REG_0].btf = meta.ret_btf;
 6137			regs[BPF_REG_0].btf_id = meta.ret_btf_id;
 6138		}
 6139	} else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
 6140		   fn->ret_type == RET_PTR_TO_BTF_ID) {
 6141		int ret_btf_id;
 6142
 6143		mark_reg_known_zero(env, regs, BPF_REG_0);
 6144		regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
 6145						     PTR_TO_BTF_ID :
 6146						     PTR_TO_BTF_ID_OR_NULL;
 6147		ret_btf_id = *fn->ret_btf_id;
 6148		if (ret_btf_id == 0) {
 6149			verbose(env, "invalid return type %d of func %s#%d\n",
 6150				fn->ret_type, func_id_name(func_id), func_id);
 6151			return -EINVAL;
 6152		}
 6153		/* current BPF helper definitions are only coming from
 6154		 * built-in code with type IDs from  vmlinux BTF
 6155		 */
 6156		regs[BPF_REG_0].btf = btf_vmlinux;
 6157		regs[BPF_REG_0].btf_id = ret_btf_id;
 6158	} else {
 6159		verbose(env, "unknown return type %d of func %s#%d\n",
 6160			fn->ret_type, func_id_name(func_id), func_id);
 6161		return -EINVAL;
 6162	}
 6163
 6164	if (reg_type_may_be_null(regs[BPF_REG_0].type))
 6165		regs[BPF_REG_0].id = ++env->id_gen;
 6166
 6167	if (is_ptr_cast_function(func_id)) {
 6168		/* For release_reference() */
 6169		regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
 6170	} else if (is_acquire_function(func_id, meta.map_ptr)) {
 6171		int id = acquire_reference_state(env, insn_idx);
 6172
 6173		if (id < 0)
 6174			return id;
 6175		/* For mark_ptr_or_null_reg() */
 6176		regs[BPF_REG_0].id = id;
 6177		/* For release_reference() */
 6178		regs[BPF_REG_0].ref_obj_id = id;
 6179	}
 6180
 6181	do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
 6182
 6183	err = check_map_func_compatibility(env, meta.map_ptr, func_id);
 6184	if (err)
 6185		return err;
 6186
 6187	if ((func_id == BPF_FUNC_get_stack ||
 6188	     func_id == BPF_FUNC_get_task_stack) &&
 6189	    !env->prog->has_callchain_buf) {
 6190		const char *err_str;
 6191
 6192#ifdef CONFIG_PERF_EVENTS
 6193		err = get_callchain_buffers(sysctl_perf_event_max_stack);
 6194		err_str = "cannot get callchain buffer for func %s#%d\n";
 6195#else
 6196		err = -ENOTSUPP;
 6197		err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
 6198#endif
 6199		if (err) {
 6200			verbose(env, err_str, func_id_name(func_id), func_id);
 6201			return err;
 6202		}
 6203
 6204		env->prog->has_callchain_buf = true;
 6205	}
 6206
 6207	if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
 6208		env->prog->call_get_stack = true;
 6209
 6210	if (changes_data)
 6211		clear_all_pkt_pointers(env);
 6212	return 0;
 6213}
 6214
 6215/* mark_btf_func_reg_size() is used when the reg size is determined by
 6216 * the BTF func_proto's return value size and argument.
 6217 */
 6218static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
 6219				   size_t reg_size)
 6220{
 6221	struct bpf_reg_state *reg = &cur_regs(env)[regno];
 6222
 6223	if (regno == BPF_REG_0) {
 6224		/* Function return value */
 6225		reg->live |= REG_LIVE_WRITTEN;
 6226		reg->subreg_def = reg_size == sizeof(u64) ?
 6227			DEF_NOT_SUBREG : env->insn_idx + 1;
 6228	} else {
 6229		/* Function argument */
 6230		if (reg_size == sizeof(u64)) {
 6231			mark_insn_zext(env, reg);
 6232			mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
 6233		} else {
 6234			mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
 6235		}
 6236	}
 6237}
 6238
 6239static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
 6240{
 6241	const struct btf_type *t, *func, *func_proto, *ptr_type;
 6242	struct bpf_reg_state *regs = cur_regs(env);
 6243	const char *func_name, *ptr_type_name;
 6244	u32 i, nargs, func_id, ptr_type_id;
 6245	const struct btf_param *args;
 6246	int err;
 6247
 6248	func_id = insn->imm;
 6249	func = btf_type_by_id(btf_vmlinux, func_id);
 6250	func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
 6251	func_proto = btf_type_by_id(btf_vmlinux, func->type);
 6252
 6253	if (!env->ops->check_kfunc_call ||
 6254	    !env->ops->check_kfunc_call(func_id)) {
 6255		verbose(env, "calling kernel function %s is not allowed\n",
 6256			func_name);
 6257		return -EACCES;
 6258	}
 6259
 6260	/* Check the arguments */
 6261	err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
 6262	if (err)
 6263		return err;
 6264
 6265	for (i = 0; i < CALLER_SAVED_REGS; i++)
 6266		mark_reg_not_init(env, regs, caller_saved[i]);
 6267
 6268	/* Check return type */
 6269	t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
 6270	if (btf_type_is_scalar(t)) {
 6271		mark_reg_unknown(env, regs, BPF_REG_0);
 6272		mark_btf_func_reg_size(env, BPF_REG_0, t->size);
 6273	} else if (btf_type_is_ptr(t)) {
 6274		ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
 6275						   &ptr_type_id);
 6276		if (!btf_type_is_struct(ptr_type)) {
 6277			ptr_type_name = btf_name_by_offset(btf_vmlinux,
 6278							   ptr_type->name_off);
 6279			verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
 6280				func_name, btf_type_str(ptr_type),
 6281				ptr_type_name);
 6282			return -EINVAL;
 6283		}
 6284		mark_reg_known_zero(env, regs, BPF_REG_0);
 6285		regs[BPF_REG_0].btf = btf_vmlinux;
 6286		regs[BPF_REG_0].type = PTR_TO_BTF_ID;
 6287		regs[BPF_REG_0].btf_id = ptr_type_id;
 6288		mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
 6289	} /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
 6290
 6291	nargs = btf_type_vlen(func_proto);
 6292	args = (const struct btf_param *)(func_proto + 1);
 6293	for (i = 0; i < nargs; i++) {
 6294		u32 regno = i + 1;
 6295
 6296		t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
 6297		if (btf_type_is_ptr(t))
 6298			mark_btf_func_reg_size(env, regno, sizeof(void *));
 6299		else
 6300			/* scalar. ensured by btf_check_kfunc_arg_match() */
 6301			mark_btf_func_reg_size(env, regno, t->size);
 6302	}
 6303
 6304	return 0;
 6305}
 6306
 6307static bool signed_add_overflows(s64 a, s64 b)
 6308{
 6309	/* Do the add in u64, where overflow is well-defined */
 6310	s64 res = (s64)((u64)a + (u64)b);
 6311
 6312	if (b < 0)
 6313		return res > a;
 6314	return res < a;
 6315}
 6316
 6317static bool signed_add32_overflows(s32 a, s32 b)
 6318{
 6319	/* Do the add in u32, where overflow is well-defined */
 6320	s32 res = (s32)((u32)a + (u32)b);
 6321
 6322	if (b < 0)
 6323		return res > a;
 6324	return res < a;
 6325}
 6326
 6327static bool signed_sub_overflows(s64 a, s64 b)
 6328{
 6329	/* Do the sub in u64, where overflow is well-defined */
 6330	s64 res = (s64)((u64)a - (u64)b);
 6331
 6332	if (b < 0)
 6333		return res < a;
 6334	return res > a;
 6335}
 6336
 6337static bool signed_sub32_overflows(s32 a, s32 b)
 6338{
 6339	/* Do the sub in u32, where overflow is well-defined */
 6340	s32 res = (s32)((u32)a - (u32)b);
 6341
 6342	if (b < 0)
 6343		return res < a;
 6344	return res > a;
 6345}
 6346
 6347static bool check_reg_sane_offset(struct bpf_verifier_env *env,
 6348				  const struct bpf_reg_state *reg,
 6349				  enum bpf_reg_type type)
 6350{
 6351	bool known = tnum_is_const(reg->var_off);
 6352	s64 val = reg->var_off.value;
 6353	s64 smin = reg->smin_value;
 6354
 6355	if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
 6356		verbose(env, "math between %s pointer and %lld is not allowed\n",
 6357			reg_type_str[type], val);
 6358		return false;
 6359	}
 6360
 6361	if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
 6362		verbose(env, "%s pointer offset %d is not allowed\n",
 6363			reg_type_str[type], reg->off);
 6364		return false;
 6365	}
 6366
 6367	if (smin == S64_MIN) {
 6368		verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
 6369			reg_type_str[type]);
 6370		return false;
 6371	}
 6372
 6373	if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
 6374		verbose(env, "value %lld makes %s pointer be out of bounds\n",
 6375			smin, reg_type_str[type]);
 6376		return false;
 6377	}
 6378
 6379	return true;
 6380}
 6381
 6382static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
 6383{
 6384	return &env->insn_aux_data[env->insn_idx];
 6385}
 6386
 6387enum {
 6388	REASON_BOUNDS	= -1,
 6389	REASON_TYPE	= -2,
 6390	REASON_PATHS	= -3,
 6391	REASON_LIMIT	= -4,
 6392	REASON_STACK	= -5,
 6393};
 6394
 6395static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
 6396			      u32 *alu_limit, bool mask_to_left)
 6397{
 6398	u32 max = 0, ptr_limit = 0;
 6399
 6400	switch (ptr_reg->type) {
 6401	case PTR_TO_STACK:
 6402		/* Offset 0 is out-of-bounds, but acceptable start for the
 6403		 * left direction, see BPF_REG_FP. Also, unknown scalar
 6404		 * offset where we would need to deal with min/max bounds is
 6405		 * currently prohibited for unprivileged.
 6406		 */
 6407		max = MAX_BPF_STACK + mask_to_left;
 6408		ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
 6409		break;
 6410	case PTR_TO_MAP_VALUE:
 6411		max = ptr_reg->map_ptr->value_size;
 6412		ptr_limit = (mask_to_left ?
 6413			     ptr_reg->smin_value :
 6414			     ptr_reg->umax_value) + ptr_reg->off;
 6415		break;
 6416	default:
 6417		return REASON_TYPE;
 6418	}
 6419
 6420	if (ptr_limit >= max)
 6421		return REASON_LIMIT;
 6422	*alu_limit = ptr_limit;
 6423	return 0;
 6424}
 6425
 6426static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
 6427				    const struct bpf_insn *insn)
 6428{
 6429	return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
 6430}
 6431
 6432static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
 6433				       u32 alu_state, u32 alu_limit)
 6434{
 6435	/* If we arrived here from different branches with different
 6436	 * state or limits to sanitize, then this won't work.
 6437	 */
 6438	if (aux->alu_state &&
 6439	    (aux->alu_state != alu_state ||
 6440	     aux->alu_limit != alu_limit))
 6441		return REASON_PATHS;
 6442
 6443	/* Corresponding fixup done in do_misc_fixups(). */
 6444	aux->alu_state = alu_state;
 6445	aux->alu_limit = alu_limit;
 6446	return 0;
 6447}
 6448
 6449static int sanitize_val_alu(struct bpf_verifier_env *env,
 6450			    struct bpf_insn *insn)
 6451{
 6452	struct bpf_insn_aux_data *aux = cur_aux(env);
 6453
 6454	if (can_skip_alu_sanitation(env, insn))
 6455		return 0;
 6456
 6457	return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
 6458}
 6459
 6460static bool sanitize_needed(u8 opcode)
 6461{
 6462	return opcode == BPF_ADD || opcode == BPF_SUB;
 6463}
 6464
 6465struct bpf_sanitize_info {
 6466	struct bpf_insn_aux_data aux;
 6467	bool mask_to_left;
 6468};
 6469
 6470static struct bpf_verifier_state *
 6471sanitize_speculative_path(struct bpf_verifier_env *env,
 6472			  const struct bpf_insn *insn,
 6473			  u32 next_idx, u32 curr_idx)
 6474{
 6475	struct bpf_verifier_state *branch;
 6476	struct bpf_reg_state *regs;
 6477
 6478	branch = push_stack(env, next_idx, curr_idx, true);
 6479	if (branch && insn) {
 6480		regs = branch->frame[branch->curframe]->regs;
 6481		if (BPF_SRC(insn->code) == BPF_K) {
 6482			mark_reg_unknown(env, regs, insn->dst_reg);
 6483		} else if (BPF_SRC(insn->code) == BPF_X) {
 6484			mark_reg_unknown(env, regs, insn->dst_reg);
 6485			mark_reg_unknown(env, regs, insn->src_reg);
 6486		}
 6487	}
 6488	return branch;
 6489}
 6490
 6491static int sanitize_ptr_alu(struct bpf_verifier_env *env,
 6492			    struct bpf_insn *insn,
 6493			    const struct bpf_reg_state *ptr_reg,
 6494			    const struct bpf_reg_state *off_reg,
 6495			    struct bpf_reg_state *dst_reg,
 6496			    struct bpf_sanitize_info *info,
 6497			    const bool commit_window)
 6498{
 6499	struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
 6500	struct bpf_verifier_state *vstate = env->cur_state;
 6501	bool off_is_imm = tnum_is_const(off_reg->var_off);
 6502	bool off_is_neg = off_reg->smin_value < 0;
 6503	bool ptr_is_dst_reg = ptr_reg == dst_reg;
 6504	u8 opcode = BPF_OP(insn->code);
 6505	u32 alu_state, alu_limit;
 6506	struct bpf_reg_state tmp;
 6507	bool ret;
 6508	int err;
 6509
 6510	if (can_skip_alu_sanitation(env, insn))
 6511		return 0;
 6512
 6513	/* We already marked aux for masking from non-speculative
 6514	 * paths, thus we got here in the first place. We only care
 6515	 * to explore bad access from here.
 6516	 */
 6517	if (vstate->speculative)
 6518		goto do_sim;
 6519
 6520	if (!commit_window) {
 6521		if (!tnum_is_const(off_reg->var_off) &&
 6522		    (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
 6523			return REASON_BOUNDS;
 6524
 6525		info->mask_to_left = (opcode == BPF_ADD &&  off_is_neg) ||
 6526				     (opcode == BPF_SUB && !off_is_neg);
 6527	}
 6528
 6529	err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
 6530	if (err < 0)
 6531		return err;
 6532
 6533	if (commit_window) {
 6534		/* In commit phase we narrow the masking window based on
 6535		 * the observed pointer move after the simulated operation.
 6536		 */
 6537		alu_state = info->aux.alu_state;
 6538		alu_limit = abs(info->aux.alu_limit - alu_limit);
 6539	} else {
 6540		alu_state  = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
 6541		alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
 6542		alu_state |= ptr_is_dst_reg ?
 6543			     BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
 6544
 6545		/* Limit pruning on unknown scalars to enable deep search for
 6546		 * potential masking differences from other program paths.
 6547		 */
 6548		if (!off_is_imm)
 6549			env->explore_alu_limits = true;
 6550	}
 6551
 6552	err = update_alu_sanitation_state(aux, alu_state, alu_limit);
 6553	if (err < 0)
 6554		return err;
 6555do_sim:
 6556	/* If we're in commit phase, we're done here given we already
 6557	 * pushed the truncated dst_reg into the speculative verification
 6558	 * stack.
 6559	 *
 6560	 * Also, when register is a known constant, we rewrite register-based
 6561	 * operation to immediate-based, and thus do not need masking (and as
 6562	 * a consequence, do not need to simulate the zero-truncation either).
 6563	 */
 6564	if (commit_window || off_is_imm)
 6565		return 0;
 6566
 6567	/* Simulate and find potential out-of-bounds access under
 6568	 * speculative execution from truncation as a result of
 6569	 * masking when off was not within expected range. If off
 6570	 * sits in dst, then we temporarily need to move ptr there
 6571	 * to simulate dst (== 0) +/-= ptr. Needed, for example,
 6572	 * for cases where we use K-based arithmetic in one direction
 6573	 * and truncated reg-based in the other in order to explore
 6574	 * bad access.
 6575	 */
 6576	if (!ptr_is_dst_reg) {
 6577		tmp = *dst_reg;
 6578		*dst_reg = *ptr_reg;
 6579	}
 6580	ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
 6581					env->insn_idx);
 6582	if (!ptr_is_dst_reg && ret)
 6583		*dst_reg = tmp;
 6584	return !ret ? REASON_STACK : 0;
 6585}
 6586
 6587static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
 6588{
 6589	struct bpf_verifier_state *vstate = env->cur_state;
 6590
 6591	/* If we simulate paths under speculation, we don't update the
 6592	 * insn as 'seen' such that when we verify unreachable paths in
 6593	 * the non-speculative domain, sanitize_dead_code() can still
 6594	 * rewrite/sanitize them.
 6595	 */
 6596	if (!vstate->speculative)
 6597		env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
 6598}
 6599
 6600static int sanitize_err(struct bpf_verifier_env *env,
 6601			const struct bpf_insn *insn, int reason,
 6602			const struct bpf_reg_state *off_reg,
 6603			const struct bpf_reg_state *dst_reg)
 6604{
 6605	static const char *err = "pointer arithmetic with it prohibited for !root";
 6606	const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
 6607	u32 dst = insn->dst_reg, src = insn->src_reg;
 6608
 6609	switch (reason) {
 6610	case REASON_BOUNDS:
 6611		verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
 6612			off_reg == dst_reg ? dst : src, err);
 6613		break;
 6614	case REASON_TYPE:
 6615		verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
 6616			off_reg == dst_reg ? src : dst, err);
 6617		break;
 6618	case REASON_PATHS:
 6619		verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
 6620			dst, op, err);
 6621		break;
 6622	case REASON_LIMIT:
 6623		verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
 6624			dst, op, err);
 6625		break;
 6626	case REASON_STACK:
 6627		verbose(env, "R%d could not be pushed for speculative verification, %s\n",
 6628			dst, err);
 6629		break;
 6630	default:
 6631		verbose(env, "verifier internal error: unknown reason (%d)\n",
 6632			reason);
 6633		break;
 6634	}
 6635
 6636	return -EACCES;
 6637}
 6638
 6639/* check that stack access falls within stack limits and that 'reg' doesn't
 6640 * have a variable offset.
 6641 *
 6642 * Variable offset is prohibited for unprivileged mode for simplicity since it
 6643 * requires corresponding support in Spectre masking for stack ALU.  See also
 6644 * retrieve_ptr_limit().
 6645 *
 6646 *
 6647 * 'off' includes 'reg->off'.
 6648 */
 6649static int check_stack_access_for_ptr_arithmetic(
 6650				struct bpf_verifier_env *env,
 6651				int regno,
 6652				const struct bpf_reg_state *reg,
 6653				int off)
 6654{
 6655	if (!tnum_is_const(reg->var_off)) {
 6656		char tn_buf[48];
 6657
 6658		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
 6659		verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
 6660			regno, tn_buf, off);
 6661		return -EACCES;
 6662	}
 6663
 6664	if (off >= 0 || off < -MAX_BPF_STACK) {
 6665		verbose(env, "R%d stack pointer arithmetic goes out of range, "
 6666			"prohibited for !root; off=%d\n", regno, off);
 6667		return -EACCES;
 6668	}
 6669
 6670	return 0;
 6671}
 6672
 6673static int sanitize_check_bounds(struct bpf_verifier_env *env,
 6674				 const struct bpf_insn *insn,
 6675				 const struct bpf_reg_state *dst_reg)
 6676{
 6677	u32 dst = insn->dst_reg;
 6678
 6679	/* For unprivileged we require that resulting offset must be in bounds
 6680	 * in order to be able to sanitize access later on.
 6681	 */
 6682	if (env->bypass_spec_v1)
 6683		return 0;
 6684
 6685	switch (dst_reg->type) {
 6686	case PTR_TO_STACK:
 6687		if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
 6688					dst_reg->off + dst_reg->var_off.value))
 6689			return -EACCES;
 6690		break;
 6691	case PTR_TO_MAP_VALUE:
 6692		if (check_map_access(env, dst, dst_reg->off, 1, false)) {
 6693			verbose(env, "R%d pointer arithmetic of map value goes out of range, "
 6694				"prohibited for !root\n", dst);
 6695			return -EACCES;
 6696		}
 6697		break;
 6698	default:
 6699		break;
 6700	}
 6701
 6702	return 0;
 6703}
 6704
 6705/* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
 6706 * Caller should also handle BPF_MOV case separately.
 6707 * If we return -EACCES, caller may want to try again treating pointer as a
 6708 * scalar.  So we only emit a diagnostic if !env->allow_ptr_leaks.
 6709 */
 6710static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
 6711				   struct bpf_insn *insn,
 6712				   const struct bpf_reg_state *ptr_reg,
 6713				   const struct bpf_reg_state *off_reg)
 6714{
 6715	struct bpf_verifier_state *vstate = env->cur_state;
 6716	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 6717	struct bpf_reg_state *regs = state->regs, *dst_reg;
 6718	bool known = tnum_is_const(off_reg->var_off);
 6719	s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
 6720	    smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
 6721	u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
 6722	    umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
 6723	struct bpf_sanitize_info info = {};
 6724	u8 opcode = BPF_OP(insn->code);
 6725	u32 dst = insn->dst_reg;
 6726	int ret;
 6727
 6728	dst_reg = &regs[dst];
 6729
 6730	if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
 6731	    smin_val > smax_val || umin_val > umax_val) {
 6732		/* Taint dst register if offset had invalid bounds derived from
 6733		 * e.g. dead branches.
 6734		 */
 6735		__mark_reg_unknown(env, dst_reg);
 6736		return 0;
 6737	}
 6738
 6739	if (BPF_CLASS(insn->code) != BPF_ALU64) {
 6740		/* 32-bit ALU ops on pointers produce (meaningless) scalars */
 6741		if (opcode == BPF_SUB && env->allow_ptr_leaks) {
 6742			__mark_reg_unknown(env, dst_reg);
 6743			return 0;
 6744		}
 6745
 6746		verbose(env,
 6747			"R%d 32-bit pointer arithmetic prohibited\n",
 6748			dst);
 6749		return -EACCES;
 6750	}
 6751
 6752	switch (ptr_reg->type) {
 6753	case PTR_TO_MAP_VALUE_OR_NULL:
 6754		verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
 6755			dst, reg_type_str[ptr_reg->type]);
 6756		return -EACCES;
 6757	case CONST_PTR_TO_MAP:
 6758		/* smin_val represents the known value */
 6759		if (known && smin_val == 0 && opcode == BPF_ADD)
 6760			break;
 6761		fallthrough;
 6762	case PTR_TO_PACKET_END:
 6763	case PTR_TO_SOCKET:
 6764	case PTR_TO_SOCKET_OR_NULL:
 6765	case PTR_TO_SOCK_COMMON:
 6766	case PTR_TO_SOCK_COMMON_OR_NULL:
 6767	case PTR_TO_TCP_SOCK:
 6768	case PTR_TO_TCP_SOCK_OR_NULL:
 6769	case PTR_TO_XDP_SOCK:
 6770		verbose(env, "R%d pointer arithmetic on %s prohibited\n",
 6771			dst, reg_type_str[ptr_reg->type]);
 6772		return -EACCES;
 6773	default:
 6774		break;
 6775	}
 6776
 6777	/* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
 6778	 * The id may be overwritten later if we create a new variable offset.
 6779	 */
 6780	dst_reg->type = ptr_reg->type;
 6781	dst_reg->id = ptr_reg->id;
 6782
 6783	if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
 6784	    !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
 6785		return -EINVAL;
 6786
 6787	/* pointer types do not carry 32-bit bounds at the moment. */
 6788	__mark_reg32_unbounded(dst_reg);
 6789
 6790	if (sanitize_needed(opcode)) {
 6791		ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
 6792				       &info, false);
 6793		if (ret < 0)
 6794			return sanitize_err(env, insn, ret, off_reg, dst_reg);
 6795	}
 6796
 6797	switch (opcode) {
 6798	case BPF_ADD:
 6799		/* We can take a fixed offset as long as it doesn't overflow
 6800		 * the s32 'off' field
 6801		 */
 6802		if (known && (ptr_reg->off + smin_val ==
 6803			      (s64)(s32)(ptr_reg->off + smin_val))) {
 6804			/* pointer += K.  Accumulate it into fixed offset */
 6805			dst_reg->smin_value = smin_ptr;
 6806			dst_reg->smax_value = smax_ptr;
 6807			dst_reg->umin_value = umin_ptr;
 6808			dst_reg->umax_value = umax_ptr;
 6809			dst_reg->var_off = ptr_reg->var_off;
 6810			dst_reg->off = ptr_reg->off + smin_val;
 6811			dst_reg->raw = ptr_reg->raw;
 6812			break;
 6813		}
 6814		/* A new variable offset is created.  Note that off_reg->off
 6815		 * == 0, since it's a scalar.
 6816		 * dst_reg gets the pointer type and since some positive
 6817		 * integer value was added to the pointer, give it a new 'id'
 6818		 * if it's a PTR_TO_PACKET.
 6819		 * this creates a new 'base' pointer, off_reg (variable) gets
 6820		 * added into the variable offset, and we copy the fixed offset
 6821		 * from ptr_reg.
 6822		 */
 6823		if (signed_add_overflows(smin_ptr, smin_val) ||
 6824		    signed_add_overflows(smax_ptr, smax_val)) {
 6825			dst_reg->smin_value = S64_MIN;
 6826			dst_reg->smax_value = S64_MAX;
 6827		} else {
 6828			dst_reg->smin_value = smin_ptr + smin_val;
 6829			dst_reg->smax_value = smax_ptr + smax_val;
 6830		}
 6831		if (umin_ptr + umin_val < umin_ptr ||
 6832		    umax_ptr + umax_val < umax_ptr) {
 6833			dst_reg->umin_value = 0;
 6834			dst_reg->umax_value = U64_MAX;
 6835		} else {
 6836			dst_reg->umin_value = umin_ptr + umin_val;
 6837			dst_reg->umax_value = umax_ptr + umax_val;
 6838		}
 6839		dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
 6840		dst_reg->off = ptr_reg->off;
 6841		dst_reg->raw = ptr_reg->raw;
 6842		if (reg_is_pkt_pointer(ptr_reg)) {
 6843			dst_reg->id = ++env->id_gen;
 6844			/* something was added to pkt_ptr, set range to zero */
 6845			memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
 6846		}
 6847		break;
 6848	case BPF_SUB:
 6849		if (dst_reg == off_reg) {
 6850			/* scalar -= pointer.  Creates an unknown scalar */
 6851			verbose(env, "R%d tried to subtract pointer from scalar\n",
 6852				dst);
 6853			return -EACCES;
 6854		}
 6855		/* We don't allow subtraction from FP, because (according to
 6856		 * test_verifier.c test "invalid fp arithmetic", JITs might not
 6857		 * be able to deal with it.
 6858		 */
 6859		if (ptr_reg->type == PTR_TO_STACK) {
 6860			verbose(env, "R%d subtraction from stack pointer prohibited\n",
 6861				dst);
 6862			return -EACCES;
 6863		}
 6864		if (known && (ptr_reg->off - smin_val ==
 6865			      (s64)(s32)(ptr_reg->off - smin_val))) {
 6866			/* pointer -= K.  Subtract it from fixed offset */
 6867			dst_reg->smin_value = smin_ptr;
 6868			dst_reg->smax_value = smax_ptr;
 6869			dst_reg->umin_value = umin_ptr;
 6870			dst_reg->umax_value = umax_ptr;
 6871			dst_reg->var_off = ptr_reg->var_off;
 6872			dst_reg->id = ptr_reg->id;
 6873			dst_reg->off = ptr_reg->off - smin_val;
 6874			dst_reg->raw = ptr_reg->raw;
 6875			break;
 6876		}
 6877		/* A new variable offset is created.  If the subtrahend is known
 6878		 * nonnegative, then any reg->range we had before is still good.
 6879		 */
 6880		if (signed_sub_overflows(smin_ptr, smax_val) ||
 6881		    signed_sub_overflows(smax_ptr, smin_val)) {
 6882			/* Overflow possible, we know nothing */
 6883			dst_reg->smin_value = S64_MIN;
 6884			dst_reg->smax_value = S64_MAX;
 6885		} else {
 6886			dst_reg->smin_value = smin_ptr - smax_val;
 6887			dst_reg->smax_value = smax_ptr - smin_val;
 6888		}
 6889		if (umin_ptr < umax_val) {
 6890			/* Overflow possible, we know nothing */
 6891			dst_reg->umin_value = 0;
 6892			dst_reg->umax_value = U64_MAX;
 6893		} else {
 6894			/* Cannot overflow (as long as bounds are consistent) */
 6895			dst_reg->umin_value = umin_ptr - umax_val;
 6896			dst_reg->umax_value = umax_ptr - umin_val;
 6897		}
 6898		dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
 6899		dst_reg->off = ptr_reg->off;
 6900		dst_reg->raw = ptr_reg->raw;
 6901		if (reg_is_pkt_pointer(ptr_reg)) {
 6902			dst_reg->id = ++env->id_gen;
 6903			/* something was added to pkt_ptr, set range to zero */
 6904			if (smin_val < 0)
 6905				memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
 6906		}
 6907		break;
 6908	case BPF_AND:
 6909	case BPF_OR:
 6910	case BPF_XOR:
 6911		/* bitwise ops on pointers are troublesome, prohibit. */
 6912		verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
 6913			dst, bpf_alu_string[opcode >> 4]);
 6914		return -EACCES;
 6915	default:
 6916		/* other operators (e.g. MUL,LSH) produce non-pointer results */
 6917		verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
 6918			dst, bpf_alu_string[opcode >> 4]);
 6919		return -EACCES;
 6920	}
 6921
 6922	if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
 6923		return -EINVAL;
 6924
 6925	__update_reg_bounds(dst_reg);
 6926	__reg_deduce_bounds(dst_reg);
 6927	__reg_bound_offset(dst_reg);
 6928
 6929	if (sanitize_check_bounds(env, insn, dst_reg) < 0)
 6930		return -EACCES;
 6931	if (sanitize_needed(opcode)) {
 6932		ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
 6933				       &info, true);
 6934		if (ret < 0)
 6935			return sanitize_err(env, insn, ret, off_reg, dst_reg);
 6936	}
 6937
 6938	return 0;
 6939}
 6940
 6941static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
 6942				 struct bpf_reg_state *src_reg)
 6943{
 6944	s32 smin_val = src_reg->s32_min_value;
 6945	s32 smax_val = src_reg->s32_max_value;
 6946	u32 umin_val = src_reg->u32_min_value;
 6947	u32 umax_val = src_reg->u32_max_value;
 6948
 6949	if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
 6950	    signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
 6951		dst_reg->s32_min_value = S32_MIN;
 6952		dst_reg->s32_max_value = S32_MAX;
 6953	} else {
 6954		dst_reg->s32_min_value += smin_val;
 6955		dst_reg->s32_max_value += smax_val;
 6956	}
 6957	if (dst_reg->u32_min_value + umin_val < umin_val ||
 6958	    dst_reg->u32_max_value + umax_val < umax_val) {
 6959		dst_reg->u32_min_value = 0;
 6960		dst_reg->u32_max_value = U32_MAX;
 6961	} else {
 6962		dst_reg->u32_min_value += umin_val;
 6963		dst_reg->u32_max_value += umax_val;
 6964	}
 6965}
 6966
 6967static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
 6968			       struct bpf_reg_state *src_reg)
 6969{
 6970	s64 smin_val = src_reg->smin_value;
 6971	s64 smax_val = src_reg->smax_value;
 6972	u64 umin_val = src_reg->umin_value;
 6973	u64 umax_val = src_reg->umax_value;
 6974
 6975	if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
 6976	    signed_add_overflows(dst_reg->smax_value, smax_val)) {
 6977		dst_reg->smin_value = S64_MIN;
 6978		dst_reg->smax_value = S64_MAX;
 6979	} else {
 6980		dst_reg->smin_value += smin_val;
 6981		dst_reg->smax_value += smax_val;
 6982	}
 6983	if (dst_reg->umin_value + umin_val < umin_val ||
 6984	    dst_reg->umax_value + umax_val < umax_val) {
 6985		dst_reg->umin_value = 0;
 6986		dst_reg->umax_value = U64_MAX;
 6987	} else {
 6988		dst_reg->umin_value += umin_val;
 6989		dst_reg->umax_value += umax_val;
 6990	}
 6991}
 6992
 6993static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
 6994				 struct bpf_reg_state *src_reg)
 6995{
 6996	s32 smin_val = src_reg->s32_min_value;
 6997	s32 smax_val = src_reg->s32_max_value;
 6998	u32 umin_val = src_reg->u32_min_value;
 6999	u32 umax_val = src_reg->u32_max_value;
 7000
 7001	if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
 7002	    signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
 7003		/* Overflow possible, we know nothing */
 7004		dst_reg->s32_min_value = S32_MIN;
 7005		dst_reg->s32_max_value = S32_MAX;
 7006	} else {
 7007		dst_reg->s32_min_value -= smax_val;
 7008		dst_reg->s32_max_value -= smin_val;
 7009	}
 7010	if (dst_reg->u32_min_value < umax_val) {
 7011		/* Overflow possible, we know nothing */
 7012		dst_reg->u32_min_value = 0;
 7013		dst_reg->u32_max_value = U32_MAX;
 7014	} else {
 7015		/* Cannot overflow (as long as bounds are consistent) */
 7016		dst_reg->u32_min_value -= umax_val;
 7017		dst_reg->u32_max_value -= umin_val;
 7018	}
 7019}
 7020
 7021static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
 7022			       struct bpf_reg_state *src_reg)
 7023{
 7024	s64 smin_val = src_reg->smin_value;
 7025	s64 smax_val = src_reg->smax_value;
 7026	u64 umin_val = src_reg->umin_value;
 7027	u64 umax_val = src_reg->umax_value;
 7028
 7029	if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
 7030	    signed_sub_overflows(dst_reg->smax_value, smin_val)) {
 7031		/* Overflow possible, we know nothing */
 7032		dst_reg->smin_value = S64_MIN;
 7033		dst_reg->smax_value = S64_MAX;
 7034	} else {
 7035		dst_reg->smin_value -= smax_val;
 7036		dst_reg->smax_value -= smin_val;
 7037	}
 7038	if (dst_reg->umin_value < umax_val) {
 7039		/* Overflow possible, we know nothing */
 7040		dst_reg->umin_value = 0;
 7041		dst_reg->umax_value = U64_MAX;
 7042	} else {
 7043		/* Cannot overflow (as long as bounds are consistent) */
 7044		dst_reg->umin_value -= umax_val;
 7045		dst_reg->umax_value -= umin_val;
 7046	}
 7047}
 7048
 7049static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
 7050				 struct bpf_reg_state *src_reg)
 7051{
 7052	s32 smin_val = src_reg->s32_min_value;
 7053	u32 umin_val = src_reg->u32_min_value;
 7054	u32 umax_val = src_reg->u32_max_value;
 7055
 7056	if (smin_val < 0 || dst_reg->s32_min_value < 0) {
 7057		/* Ain't nobody got time to multiply that sign */
 7058		__mark_reg32_unbounded(dst_reg);
 7059		return;
 7060	}
 7061	/* Both values are positive, so we can work with unsigned and
 7062	 * copy the result to signed (unless it exceeds S32_MAX).
 7063	 */
 7064	if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
 7065		/* Potential overflow, we know nothing */
 7066		__mark_reg32_unbounded(dst_reg);
 7067		return;
 7068	}
 7069	dst_reg->u32_min_value *= umin_val;
 7070	dst_reg->u32_max_value *= umax_val;
 7071	if (dst_reg->u32_max_value > S32_MAX) {
 7072		/* Overflow possible, we know nothing */
 7073		dst_reg->s32_min_value = S32_MIN;
 7074		dst_reg->s32_max_value = S32_MAX;
 7075	} else {
 7076		dst_reg->s32_min_value = dst_reg->u32_min_value;
 7077		dst_reg->s32_max_value = dst_reg->u32_max_value;
 7078	}
 7079}
 7080
 7081static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
 7082			       struct bpf_reg_state *src_reg)
 7083{
 7084	s64 smin_val = src_reg->smin_value;
 7085	u64 umin_val = src_reg->umin_value;
 7086	u64 umax_val = src_reg->umax_value;
 7087
 7088	if (smin_val < 0 || dst_reg->smin_value < 0) {
 7089		/* Ain't nobody got time to multiply that sign */
 7090		__mark_reg64_unbounded(dst_reg);
 7091		return;
 7092	}
 7093	/* Both values are positive, so we can work with unsigned and
 7094	 * copy the result to signed (unless it exceeds S64_MAX).
 7095	 */
 7096	if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
 7097		/* Potential overflow, we know nothing */
 7098		__mark_reg64_unbounded(dst_reg);
 7099		return;
 7100	}
 7101	dst_reg->umin_value *= umin_val;
 7102	dst_reg->umax_value *= umax_val;
 7103	if (dst_reg->umax_value > S64_MAX) {
 7104		/* Overflow possible, we know nothing */
 7105		dst_reg->smin_value = S64_MIN;
 7106		dst_reg->smax_value = S64_MAX;
 7107	} else {
 7108		dst_reg->smin_value = dst_reg->umin_value;
 7109		dst_reg->smax_value = dst_reg->umax_value;
 7110	}
 7111}
 7112
 7113static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
 7114				 struct bpf_reg_state *src_reg)
 7115{
 7116	bool src_known = tnum_subreg_is_const(src_reg->var_off);
 7117	bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
 7118	struct tnum var32_off = tnum_subreg(dst_reg->var_off);
 7119	s32 smin_val = src_reg->s32_min_value;
 7120	u32 umax_val = src_reg->u32_max_value;
 7121
 7122	if (src_known && dst_known) {
 7123		__mark_reg32_known(dst_reg, var32_off.value);
 7124		return;
 7125	}
 7126
 7127	/* We get our minimum from the var_off, since that's inherently
 7128	 * bitwise.  Our maximum is the minimum of the operands' maxima.
 7129	 */
 7130	dst_reg->u32_min_value = var32_off.value;
 7131	dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
 7132	if (dst_reg->s32_min_value < 0 || smin_val < 0) {
 7133		/* Lose signed bounds when ANDing negative numbers,
 7134		 * ain't nobody got time for that.
 7135		 */
 7136		dst_reg->s32_min_value = S32_MIN;
 7137		dst_reg->s32_max_value = S32_MAX;
 7138	} else {
 7139		/* ANDing two positives gives a positive, so safe to
 7140		 * cast result into s64.
 7141		 */
 7142		dst_reg->s32_min_value = dst_reg->u32_min_value;
 7143		dst_reg->s32_max_value = dst_reg->u32_max_value;
 7144	}
 7145}
 7146
 7147static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
 7148			       struct bpf_reg_state *src_reg)
 7149{
 7150	bool src_known = tnum_is_const(src_reg->var_off);
 7151	bool dst_known = tnum_is_const(dst_reg->var_off);
 7152	s64 smin_val = src_reg->smin_value;
 7153	u64 umax_val = src_reg->umax_value;
 7154
 7155	if (src_known && dst_known) {
 7156		__mark_reg_known(dst_reg, dst_reg->var_off.value);
 7157		return;
 7158	}
 7159
 7160	/* We get our minimum from the var_off, since that's inherently
 7161	 * bitwise.  Our maximum is the minimum of the operands' maxima.
 7162	 */
 7163	dst_reg->umin_value = dst_reg->var_off.value;
 7164	dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
 7165	if (dst_reg->smin_value < 0 || smin_val < 0) {
 7166		/* Lose signed bounds when ANDing negative numbers,
 7167		 * ain't nobody got time for that.
 7168		 */
 7169		dst_reg->smin_value = S64_MIN;
 7170		dst_reg->smax_value = S64_MAX;
 7171	} else {
 7172		/* ANDing two positives gives a positive, so safe to
 7173		 * cast result into s64.
 7174		 */
 7175		dst_reg->smin_value = dst_reg->umin_value;
 7176		dst_reg->smax_value = dst_reg->umax_value;
 7177	}
 7178	/* We may learn something more from the var_off */
 7179	__update_reg_bounds(dst_reg);
 7180}
 7181
 7182static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
 7183				struct bpf_reg_state *src_reg)
 7184{
 7185	bool src_known = tnum_subreg_is_const(src_reg->var_off);
 7186	bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
 7187	struct tnum var32_off = tnum_subreg(dst_reg->var_off);
 7188	s32 smin_val = src_reg->s32_min_value;
 7189	u32 umin_val = src_reg->u32_min_value;
 7190
 7191	if (src_known && dst_known) {
 7192		__mark_reg32_known(dst_reg, var32_off.value);
 7193		return;
 7194	}
 7195
 7196	/* We get our maximum from the var_off, and our minimum is the
 7197	 * maximum of the operands' minima
 7198	 */
 7199	dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
 7200	dst_reg->u32_max_value = var32_off.value | var32_off.mask;
 7201	if (dst_reg->s32_min_value < 0 || smin_val < 0) {
 7202		/* Lose signed bounds when ORing negative numbers,
 7203		 * ain't nobody got time for that.
 7204		 */
 7205		dst_reg->s32_min_value = S32_MIN;
 7206		dst_reg->s32_max_value = S32_MAX;
 7207	} else {
 7208		/* ORing two positives gives a positive, so safe to
 7209		 * cast result into s64.
 7210		 */
 7211		dst_reg->s32_min_value = dst_reg->u32_min_value;
 7212		dst_reg->s32_max_value = dst_reg->u32_max_value;
 7213	}
 7214}
 7215
 7216static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
 7217			      struct bpf_reg_state *src_reg)
 7218{
 7219	bool src_known = tnum_is_const(src_reg->var_off);
 7220	bool dst_known = tnum_is_const(dst_reg->var_off);
 7221	s64 smin_val = src_reg->smin_value;
 7222	u64 umin_val = src_reg->umin_value;
 7223
 7224	if (src_known && dst_known) {
 7225		__mark_reg_known(dst_reg, dst_reg->var_off.value);
 7226		return;
 7227	}
 7228
 7229	/* We get our maximum from the var_off, and our minimum is the
 7230	 * maximum of the operands' minima
 7231	 */
 7232	dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
 7233	dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
 7234	if (dst_reg->smin_value < 0 || smin_val < 0) {
 7235		/* Lose signed bounds when ORing negative numbers,
 7236		 * ain't nobody got time for that.
 7237		 */
 7238		dst_reg->smin_value = S64_MIN;
 7239		dst_reg->smax_value = S64_MAX;
 7240	} else {
 7241		/* ORing two positives gives a positive, so safe to
 7242		 * cast result into s64.
 7243		 */
 7244		dst_reg->smin_value = dst_reg->umin_value;
 7245		dst_reg->smax_value = dst_reg->umax_value;
 7246	}
 7247	/* We may learn something more from the var_off */
 7248	__update_reg_bounds(dst_reg);
 7249}
 7250
 7251static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
 7252				 struct bpf_reg_state *src_reg)
 7253{
 7254	bool src_known = tnum_subreg_is_const(src_reg->var_off);
 7255	bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
 7256	struct tnum var32_off = tnum_subreg(dst_reg->var_off);
 7257	s32 smin_val = src_reg->s32_min_value;
 7258
 7259	if (src_known && dst_known) {
 7260		__mark_reg32_known(dst_reg, var32_off.value);
 7261		return;
 7262	}
 7263
 7264	/* We get both minimum and maximum from the var32_off. */
 7265	dst_reg->u32_min_value = var32_off.value;
 7266	dst_reg->u32_max_value = var32_off.value | var32_off.mask;
 7267
 7268	if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
 7269		/* XORing two positive sign numbers gives a positive,
 7270		 * so safe to cast u32 result into s32.
 7271		 */
 7272		dst_reg->s32_min_value = dst_reg->u32_min_value;
 7273		dst_reg->s32_max_value = dst_reg->u32_max_value;
 7274	} else {
 7275		dst_reg->s32_min_value = S32_MIN;
 7276		dst_reg->s32_max_value = S32_MAX;
 7277	}
 7278}
 7279
 7280static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
 7281			       struct bpf_reg_state *src_reg)
 7282{
 7283	bool src_known = tnum_is_const(src_reg->var_off);
 7284	bool dst_known = tnum_is_const(dst_reg->var_off);
 7285	s64 smin_val = src_reg->smin_value;
 7286
 7287	if (src_known && dst_known) {
 7288		/* dst_reg->var_off.value has been updated earlier */
 7289		__mark_reg_known(dst_reg, dst_reg->var_off.value);
 7290		return;
 7291	}
 7292
 7293	/* We get both minimum and maximum from the var_off. */
 7294	dst_reg->umin_value = dst_reg->var_off.value;
 7295	dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
 7296
 7297	if (dst_reg->smin_value >= 0 && smin_val >= 0) {
 7298		/* XORing two positive sign numbers gives a positive,
 7299		 * so safe to cast u64 result into s64.
 7300		 */
 7301		dst_reg->smin_value = dst_reg->umin_value;
 7302		dst_reg->smax_value = dst_reg->umax_value;
 7303	} else {
 7304		dst_reg->smin_value = S64_MIN;
 7305		dst_reg->smax_value = S64_MAX;
 7306	}
 7307
 7308	__update_reg_bounds(dst_reg);
 7309}
 7310
 7311static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
 7312				   u64 umin_val, u64 umax_val)
 7313{
 7314	/* We lose all sign bit information (except what we can pick
 7315	 * up from var_off)
 7316	 */
 7317	dst_reg->s32_min_value = S32_MIN;
 7318	dst_reg->s32_max_value = S32_MAX;
 7319	/* If we might shift our top bit out, then we know nothing */
 7320	if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
 7321		dst_reg->u32_min_value = 0;
 7322		dst_reg->u32_max_value = U32_MAX;
 7323	} else {
 7324		dst_reg->u32_min_value <<= umin_val;
 7325		dst_reg->u32_max_value <<= umax_val;
 7326	}
 7327}
 7328
 7329static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
 7330				 struct bpf_reg_state *src_reg)
 7331{
 7332	u32 umax_val = src_reg->u32_max_value;
 7333	u32 umin_val = src_reg->u32_min_value;
 7334	/* u32 alu operation will zext upper bits */
 7335	struct tnum subreg = tnum_subreg(dst_reg->var_off);
 7336
 7337	__scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
 7338	dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
 7339	/* Not required but being careful mark reg64 bounds as unknown so
 7340	 * that we are forced to pick them up from tnum and zext later and
 7341	 * if some path skips this step we are still safe.
 7342	 */
 7343	__mark_reg64_unbounded(dst_reg);
 7344	__update_reg32_bounds(dst_reg);
 7345}
 7346
 7347static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
 7348				   u64 umin_val, u64 umax_val)
 7349{
 7350	/* Special case <<32 because it is a common compiler pattern to sign
 7351	 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
 7352	 * positive we know this shift will also be positive so we can track
 7353	 * bounds correctly. Otherwise we lose all sign bit information except
 7354	 * what we can pick up from var_off. Perhaps we can generalize this
 7355	 * later to shifts of any length.
 7356	 */
 7357	if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
 7358		dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
 7359	else
 7360		dst_reg->smax_value = S64_MAX;
 7361
 7362	if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
 7363		dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
 7364	else
 7365		dst_reg->smin_value = S64_MIN;
 7366
 7367	/* If we might shift our top bit out, then we know nothing */
 7368	if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
 7369		dst_reg->umin_value = 0;
 7370		dst_reg->umax_value = U64_MAX;
 7371	} else {
 7372		dst_reg->umin_value <<= umin_val;
 7373		dst_reg->umax_value <<= umax_val;
 7374	}
 7375}
 7376
 7377static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
 7378			       struct bpf_reg_state *src_reg)
 7379{
 7380	u64 umax_val = src_reg->umax_value;
 7381	u64 umin_val = src_reg->umin_value;
 7382
 7383	/* scalar64 calc uses 32bit unshifted bounds so must be called first */
 7384	__scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
 7385	__scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
 7386
 7387	dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
 7388	/* We may learn something more from the var_off */
 7389	__update_reg_bounds(dst_reg);
 7390}
 7391
 7392static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
 7393				 struct bpf_reg_state *src_reg)
 7394{
 7395	struct tnum subreg = tnum_subreg(dst_reg->var_off);
 7396	u32 umax_val = src_reg->u32_max_value;
 7397	u32 umin_val = src_reg->u32_min_value;
 7398
 7399	/* BPF_RSH is an unsigned shift.  If the value in dst_reg might
 7400	 * be negative, then either:
 7401	 * 1) src_reg might be zero, so the sign bit of the result is
 7402	 *    unknown, so we lose our signed bounds
 7403	 * 2) it's known negative, thus the unsigned bounds capture the
 7404	 *    signed bounds
 7405	 * 3) the signed bounds cross zero, so they tell us nothing
 7406	 *    about the result
 7407	 * If the value in dst_reg is known nonnegative, then again the
 7408	 * unsigned bounds capture the signed bounds.
 7409	 * Thus, in all cases it suffices to blow away our signed bounds
 7410	 * and rely on inferring new ones from the unsigned bounds and
 7411	 * var_off of the result.
 7412	 */
 7413	dst_reg->s32_min_value = S32_MIN;
 7414	dst_reg->s32_max_value = S32_MAX;
 7415
 7416	dst_reg->var_off = tnum_rshift(subreg, umin_val);
 7417	dst_reg->u32_min_value >>= umax_val;
 7418	dst_reg->u32_max_value >>= umin_val;
 7419
 7420	__mark_reg64_unbounded(dst_reg);
 7421	__update_reg32_bounds(dst_reg);
 7422}
 7423
 7424static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
 7425			       struct bpf_reg_state *src_reg)
 7426{
 7427	u64 umax_val = src_reg->umax_value;
 7428	u64 umin_val = src_reg->umin_value;
 7429
 7430	/* BPF_RSH is an unsigned shift.  If the value in dst_reg might
 7431	 * be negative, then either:
 7432	 * 1) src_reg might be zero, so the sign bit of the result is
 7433	 *    unknown, so we lose our signed bounds
 7434	 * 2) it's known negative, thus the unsigned bounds capture the
 7435	 *    signed bounds
 7436	 * 3) the signed bounds cross zero, so they tell us nothing
 7437	 *    about the result
 7438	 * If the value in dst_reg is known nonnegative, then again the
 7439	 * unsigned bounds capture the signed bounds.
 7440	 * Thus, in all cases it suffices to blow away our signed bounds
 7441	 * and rely on inferring new ones from the unsigned bounds and
 7442	 * var_off of the result.
 7443	 */
 7444	dst_reg->smin_value = S64_MIN;
 7445	dst_reg->smax_value = S64_MAX;
 7446	dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
 7447	dst_reg->umin_value >>= umax_val;
 7448	dst_reg->umax_value >>= umin_val;
 7449
 7450	/* Its not easy to operate on alu32 bounds here because it depends
 7451	 * on bits being shifted in. Take easy way out and mark unbounded
 7452	 * so we can recalculate later from tnum.
 7453	 */
 7454	__mark_reg32_unbounded(dst_reg);
 7455	__update_reg_bounds(dst_reg);
 7456}
 7457
 7458static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
 7459				  struct bpf_reg_state *src_reg)
 7460{
 7461	u64 umin_val = src_reg->u32_min_value;
 7462
 7463	/* Upon reaching here, src_known is true and
 7464	 * umax_val is equal to umin_val.
 7465	 */
 7466	dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
 7467	dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
 7468
 7469	dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
 7470
 7471	/* blow away the dst_reg umin_value/umax_value and rely on
 7472	 * dst_reg var_off to refine the result.
 7473	 */
 7474	dst_reg->u32_min_value = 0;
 7475	dst_reg->u32_max_value = U32_MAX;
 7476
 7477	__mark_reg64_unbounded(dst_reg);
 7478	__update_reg32_bounds(dst_reg);
 7479}
 7480
 7481static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
 7482				struct bpf_reg_state *src_reg)
 7483{
 7484	u64 umin_val = src_reg->umin_value;
 7485
 7486	/* Upon reaching here, src_known is true and umax_val is equal
 7487	 * to umin_val.
 7488	 */
 7489	dst_reg->smin_value >>= umin_val;
 7490	dst_reg->smax_value >>= umin_val;
 7491
 7492	dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
 7493
 7494	/* blow away the dst_reg umin_value/umax_value and rely on
 7495	 * dst_reg var_off to refine the result.
 7496	 */
 7497	dst_reg->umin_value = 0;
 7498	dst_reg->umax_value = U64_MAX;
 7499
 7500	/* Its not easy to operate on alu32 bounds here because it depends
 7501	 * on bits being shifted in from upper 32-bits. Take easy way out
 7502	 * and mark unbounded so we can recalculate later from tnum.
 7503	 */
 7504	__mark_reg32_unbounded(dst_reg);
 7505	__update_reg_bounds(dst_reg);
 7506}
 7507
 7508/* WARNING: This function does calculations on 64-bit values, but the actual
 7509 * execution may occur on 32-bit values. Therefore, things like bitshifts
 7510 * need extra checks in the 32-bit case.
 7511 */
 7512static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
 7513				      struct bpf_insn *insn,
 7514				      struct bpf_reg_state *dst_reg,
 7515				      struct bpf_reg_state src_reg)
 7516{
 7517	struct bpf_reg_state *regs = cur_regs(env);
 7518	u8 opcode = BPF_OP(insn->code);
 7519	bool src_known;
 7520	s64 smin_val, smax_val;
 7521	u64 umin_val, umax_val;
 7522	s32 s32_min_val, s32_max_val;
 7523	u32 u32_min_val, u32_max_val;
 7524	u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
 7525	bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
 7526	int ret;
 7527
 7528	smin_val = src_reg.smin_value;
 7529	smax_val = src_reg.smax_value;
 7530	umin_val = src_reg.umin_value;
 7531	umax_val = src_reg.umax_value;
 7532
 7533	s32_min_val = src_reg.s32_min_value;
 7534	s32_max_val = src_reg.s32_max_value;
 7535	u32_min_val = src_reg.u32_min_value;
 7536	u32_max_val = src_reg.u32_max_value;
 7537
 7538	if (alu32) {
 7539		src_known = tnum_subreg_is_const(src_reg.var_off);
 7540		if ((src_known &&
 7541		     (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
 7542		    s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
 7543			/* Taint dst register if offset had invalid bounds
 7544			 * derived from e.g. dead branches.
 7545			 */
 7546			__mark_reg_unknown(env, dst_reg);
 7547			return 0;
 7548		}
 7549	} else {
 7550		src_known = tnum_is_const(src_reg.var_off);
 7551		if ((src_known &&
 7552		     (smin_val != smax_val || umin_val != umax_val)) ||
 7553		    smin_val > smax_val || umin_val > umax_val) {
 7554			/* Taint dst register if offset had invalid bounds
 7555			 * derived from e.g. dead branches.
 7556			 */
 7557			__mark_reg_unknown(env, dst_reg);
 7558			return 0;
 7559		}
 7560	}
 7561
 7562	if (!src_known &&
 7563	    opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
 7564		__mark_reg_unknown(env, dst_reg);
 7565		return 0;
 7566	}
 7567
 7568	if (sanitize_needed(opcode)) {
 7569		ret = sanitize_val_alu(env, insn);
 7570		if (ret < 0)
 7571			return sanitize_err(env, insn, ret, NULL, NULL);
 7572	}
 7573
 7574	/* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
 7575	 * There are two classes of instructions: The first class we track both
 7576	 * alu32 and alu64 sign/unsigned bounds independently this provides the
 7577	 * greatest amount of precision when alu operations are mixed with jmp32
 7578	 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
 7579	 * and BPF_OR. This is possible because these ops have fairly easy to
 7580	 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
 7581	 * See alu32 verifier tests for examples. The second class of
 7582	 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
 7583	 * with regards to tracking sign/unsigned bounds because the bits may
 7584	 * cross subreg boundaries in the alu64 case. When this happens we mark
 7585	 * the reg unbounded in the subreg bound space and use the resulting
 7586	 * tnum to calculate an approximation of the sign/unsigned bounds.
 7587	 */
 7588	switch (opcode) {
 7589	case BPF_ADD:
 7590		scalar32_min_max_add(dst_reg, &src_reg);
 7591		scalar_min_max_add(dst_reg, &src_reg);
 7592		dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
 7593		break;
 7594	case BPF_SUB:
 7595		scalar32_min_max_sub(dst_reg, &src_reg);
 7596		scalar_min_max_sub(dst_reg, &src_reg);
 7597		dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
 7598		break;
 7599	case BPF_MUL:
 7600		dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
 7601		scalar32_min_max_mul(dst_reg, &src_reg);
 7602		scalar_min_max_mul(dst_reg, &src_reg);
 7603		break;
 7604	case BPF_AND:
 7605		dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
 7606		scalar32_min_max_and(dst_reg, &src_reg);
 7607		scalar_min_max_and(dst_reg, &src_reg);
 7608		break;
 7609	case BPF_OR:
 7610		dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
 7611		scalar32_min_max_or(dst_reg, &src_reg);
 7612		scalar_min_max_or(dst_reg, &src_reg);
 7613		break;
 7614	case BPF_XOR:
 7615		dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
 7616		scalar32_min_max_xor(dst_reg, &src_reg);
 7617		scalar_min_max_xor(dst_reg, &src_reg);
 7618		break;
 7619	case BPF_LSH:
 7620		if (umax_val >= insn_bitness) {
 7621			/* Shifts greater than 31 or 63 are undefined.
 7622			 * This includes shifts by a negative number.
 7623			 */
 7624			mark_reg_unknown(env, regs, insn->dst_reg);
 7625			break;
 7626		}
 7627		if (alu32)
 7628			scalar32_min_max_lsh(dst_reg, &src_reg);
 7629		else
 7630			scalar_min_max_lsh(dst_reg, &src_reg);
 7631		break;
 7632	case BPF_RSH:
 7633		if (umax_val >= insn_bitness) {
 7634			/* Shifts greater than 31 or 63 are undefined.
 7635			 * This includes shifts by a negative number.
 7636			 */
 7637			mark_reg_unknown(env, regs, insn->dst_reg);
 7638			break;
 7639		}
 7640		if (alu32)
 7641			scalar32_min_max_rsh(dst_reg, &src_reg);
 7642		else
 7643			scalar_min_max_rsh(dst_reg, &src_reg);
 7644		break;
 7645	case BPF_ARSH:
 7646		if (umax_val >= insn_bitness) {
 7647			/* Shifts greater than 31 or 63 are undefined.
 7648			 * This includes shifts by a negative number.
 7649			 */
 7650			mark_reg_unknown(env, regs, insn->dst_reg);
 7651			break;
 7652		}
 7653		if (alu32)
 7654			scalar32_min_max_arsh(dst_reg, &src_reg);
 7655		else
 7656			scalar_min_max_arsh(dst_reg, &src_reg);
 7657		break;
 7658	default:
 7659		mark_reg_unknown(env, regs, insn->dst_reg);
 7660		break;
 7661	}
 7662
 7663	/* ALU32 ops are zero extended into 64bit register */
 7664	if (alu32)
 7665		zext_32_to_64(dst_reg);
 7666
 7667	__update_reg_bounds(dst_reg);
 7668	__reg_deduce_bounds(dst_reg);
 7669	__reg_bound_offset(dst_reg);
 7670	return 0;
 7671}
 7672
 7673/* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
 7674 * and var_off.
 7675 */
 7676static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
 7677				   struct bpf_insn *insn)
 7678{
 7679	struct bpf_verifier_state *vstate = env->cur_state;
 7680	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 7681	struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
 7682	struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
 7683	u8 opcode = BPF_OP(insn->code);
 7684	int err;
 7685
 7686	dst_reg = &regs[insn->dst_reg];
 7687	src_reg = NULL;
 7688	if (dst_reg->type != SCALAR_VALUE)
 7689		ptr_reg = dst_reg;
 7690	else
 7691		/* Make sure ID is cleared otherwise dst_reg min/max could be
 7692		 * incorrectly propagated into other registers by find_equal_scalars()
 7693		 */
 7694		dst_reg->id = 0;
 7695	if (BPF_SRC(insn->code) == BPF_X) {
 7696		src_reg = &regs[insn->src_reg];
 7697		if (src_reg->type != SCALAR_VALUE) {
 7698			if (dst_reg->type != SCALAR_VALUE) {
 7699				/* Combining two pointers by any ALU op yields
 7700				 * an arbitrary scalar. Disallow all math except
 7701				 * pointer subtraction
 7702				 */
 7703				if (opcode == BPF_SUB && env->allow_ptr_leaks) {
 7704					mark_reg_unknown(env, regs, insn->dst_reg);
 7705					return 0;
 7706				}
 7707				verbose(env, "R%d pointer %s pointer prohibited\n",
 7708					insn->dst_reg,
 7709					bpf_alu_string[opcode >> 4]);
 7710				return -EACCES;
 7711			} else {
 7712				/* scalar += pointer
 7713				 * This is legal, but we have to reverse our
 7714				 * src/dest handling in computing the range
 7715				 */
 7716				err = mark_chain_precision(env, insn->dst_reg);
 7717				if (err)
 7718					return err;
 7719				return adjust_ptr_min_max_vals(env, insn,
 7720							       src_reg, dst_reg);
 7721			}
 7722		} else if (ptr_reg) {
 7723			/* pointer += scalar */
 7724			err = mark_chain_precision(env, insn->src_reg);
 7725			if (err)
 7726				return err;
 7727			return adjust_ptr_min_max_vals(env, insn,
 7728						       dst_reg, src_reg);
 7729		}
 7730	} else {
 7731		/* Pretend the src is a reg with a known value, since we only
 7732		 * need to be able to read from this state.
 7733		 */
 7734		off_reg.type = SCALAR_VALUE;
 7735		__mark_reg_known(&off_reg, insn->imm);
 7736		src_reg = &off_reg;
 7737		if (ptr_reg) /* pointer += K */
 7738			return adjust_ptr_min_max_vals(env, insn,
 7739						       ptr_reg, src_reg);
 7740	}
 7741
 7742	/* Got here implies adding two SCALAR_VALUEs */
 7743	if (WARN_ON_ONCE(ptr_reg)) {
 7744		print_verifier_state(env, state);
 7745		verbose(env, "verifier internal error: unexpected ptr_reg\n");
 7746		return -EINVAL;
 7747	}
 7748	if (WARN_ON(!src_reg)) {
 7749		print_verifier_state(env, state);
 7750		verbose(env, "verifier internal error: no src_reg\n");
 7751		return -EINVAL;
 7752	}
 7753	return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
 7754}
 7755
 7756/* check validity of 32-bit and 64-bit arithmetic operations */
 7757static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
 7758{
 7759	struct bpf_reg_state *regs = cur_regs(env);
 7760	u8 opcode = BPF_OP(insn->code);
 7761	int err;
 7762
 7763	if (opcode == BPF_END || opcode == BPF_NEG) {
 7764		if (opcode == BPF_NEG) {
 7765			if (BPF_SRC(insn->code) != 0 ||
 7766			    insn->src_reg != BPF_REG_0 ||
 7767			    insn->off != 0 || insn->imm != 0) {
 7768				verbose(env, "BPF_NEG uses reserved fields\n");
 7769				return -EINVAL;
 7770			}
 7771		} else {
 7772			if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
 7773			    (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
 7774			    BPF_CLASS(insn->code) == BPF_ALU64) {
 7775				verbose(env, "BPF_END uses reserved fields\n");
 7776				return -EINVAL;
 7777			}
 7778		}
 7779
 7780		/* check src operand */
 7781		err = check_reg_arg(env, insn->dst_reg, SRC_OP);
 7782		if (err)
 7783			return err;
 7784
 7785		if (is_pointer_value(env, insn->dst_reg)) {
 7786			verbose(env, "R%d pointer arithmetic prohibited\n",
 7787				insn->dst_reg);
 7788			return -EACCES;
 7789		}
 7790
 7791		/* check dest operand */
 7792		err = check_reg_arg(env, insn->dst_reg, DST_OP);
 7793		if (err)
 7794			return err;
 7795
 7796	} else if (opcode == BPF_MOV) {
 7797
 7798		if (BPF_SRC(insn->code) == BPF_X) {
 7799			if (insn->imm != 0 || insn->off != 0) {
 7800				verbose(env, "BPF_MOV uses reserved fields\n");
 7801				return -EINVAL;
 7802			}
 7803
 7804			/* check src operand */
 7805			err = check_reg_arg(env, insn->src_reg, SRC_OP);
 7806			if (err)
 7807				return err;
 7808		} else {
 7809			if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
 7810				verbose(env, "BPF_MOV uses reserved fields\n");
 7811				return -EINVAL;
 7812			}
 7813		}
 7814
 7815		/* check dest operand, mark as required later */
 7816		err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
 7817		if (err)
 7818			return err;
 7819
 7820		if (BPF_SRC(insn->code) == BPF_X) {
 7821			struct bpf_reg_state *src_reg = regs + insn->src_reg;
 7822			struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
 7823
 7824			if (BPF_CLASS(insn->code) == BPF_ALU64) {
 7825				/* case: R1 = R2
 7826				 * copy register state to dest reg
 7827				 */
 7828				if (src_reg->type == SCALAR_VALUE && !src_reg->id)
 7829					/* Assign src and dst registers the same ID
 7830					 * that will be used by find_equal_scalars()
 7831					 * to propagate min/max range.
 7832					 */
 7833					src_reg->id = ++env->id_gen;
 7834				*dst_reg = *src_reg;
 7835				dst_reg->live |= REG_LIVE_WRITTEN;
 7836				dst_reg->subreg_def = DEF_NOT_SUBREG;
 7837			} else {
 7838				/* R1 = (u32) R2 */
 7839				if (is_pointer_value(env, insn->src_reg)) {
 7840					verbose(env,
 7841						"R%d partial copy of pointer\n",
 7842						insn->src_reg);
 7843					return -EACCES;
 7844				} else if (src_reg->type == SCALAR_VALUE) {
 7845					*dst_reg = *src_reg;
 7846					/* Make sure ID is cleared otherwise
 7847					 * dst_reg min/max could be incorrectly
 7848					 * propagated into src_reg by find_equal_scalars()
 7849					 */
 7850					dst_reg->id = 0;
 7851					dst_reg->live |= REG_LIVE_WRITTEN;
 7852					dst_reg->subreg_def = env->insn_idx + 1;
 7853				} else {
 7854					mark_reg_unknown(env, regs,
 7855							 insn->dst_reg);
 7856				}
 7857				zext_32_to_64(dst_reg);
 7858			}
 7859		} else {
 7860			/* case: R = imm
 7861			 * remember the value we stored into this reg
 7862			 */
 7863			/* clear any state __mark_reg_known doesn't set */
 7864			mark_reg_unknown(env, regs, insn->dst_reg);
 7865			regs[insn->dst_reg].type = SCALAR_VALUE;
 7866			if (BPF_CLASS(insn->code) == BPF_ALU64) {
 7867				__mark_reg_known(regs + insn->dst_reg,
 7868						 insn->imm);
 7869			} else {
 7870				__mark_reg_known(regs + insn->dst_reg,
 7871						 (u32)insn->imm);
 7872			}
 7873		}
 7874
 7875	} else if (opcode > BPF_END) {
 7876		verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
 7877		return -EINVAL;
 7878
 7879	} else {	/* all other ALU ops: and, sub, xor, add, ... */
 7880
 7881		if (BPF_SRC(insn->code) == BPF_X) {
 7882			if (insn->imm != 0 || insn->off != 0) {
 7883				verbose(env, "BPF_ALU uses reserved fields\n");
 7884				return -EINVAL;
 7885			}
 7886			/* check src1 operand */
 7887			err = check_reg_arg(env, insn->src_reg, SRC_OP);
 7888			if (err)
 7889				return err;
 7890		} else {
 7891			if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
 7892				verbose(env, "BPF_ALU uses reserved fields\n");
 7893				return -EINVAL;
 7894			}
 7895		}
 7896
 7897		/* check src2 operand */
 7898		err = check_reg_arg(env, insn->dst_reg, SRC_OP);
 7899		if (err)
 7900			return err;
 7901
 7902		if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
 7903		    BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
 7904			verbose(env, "div by zero\n");
 7905			return -EINVAL;
 7906		}
 7907
 7908		if ((opcode == BPF_LSH || opcode == BPF_RSH ||
 7909		     opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
 7910			int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
 7911
 7912			if (insn->imm < 0 || insn->imm >= size) {
 7913				verbose(env, "invalid shift %d\n", insn->imm);
 7914				return -EINVAL;
 7915			}
 7916		}
 7917
 7918		/* check dest operand */
 7919		err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
 7920		if (err)
 7921			return err;
 7922
 7923		return adjust_reg_min_max_vals(env, insn);
 7924	}
 7925
 7926	return 0;
 7927}
 7928
 7929static void __find_good_pkt_pointers(struct bpf_func_state *state,
 7930				     struct bpf_reg_state *dst_reg,
 7931				     enum bpf_reg_type type, int new_range)
 7932{
 7933	struct bpf_reg_state *reg;
 7934	int i;
 7935
 7936	for (i = 0; i < MAX_BPF_REG; i++) {
 7937		reg = &state->regs[i];
 7938		if (reg->type == type && reg->id == dst_reg->id)
 7939			/* keep the maximum range already checked */
 7940			reg->range = max(reg->range, new_range);
 7941	}
 7942
 7943	bpf_for_each_spilled_reg(i, state, reg) {
 7944		if (!reg)
 7945			continue;
 7946		if (reg->type == type && reg->id == dst_reg->id)
 7947			reg->range = max(reg->range, new_range);
 7948	}
 7949}
 7950
 7951static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
 7952				   struct bpf_reg_state *dst_reg,
 7953				   enum bpf_reg_type type,
 7954				   bool range_right_open)
 7955{
 7956	int new_range, i;
 7957
 7958	if (dst_reg->off < 0 ||
 7959	    (dst_reg->off == 0 && range_right_open))
 7960		/* This doesn't give us any range */
 7961		return;
 7962
 7963	if (dst_reg->umax_value > MAX_PACKET_OFF ||
 7964	    dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
 7965		/* Risk of overflow.  For instance, ptr + (1<<63) may be less
 7966		 * than pkt_end, but that's because it's also less than pkt.
 7967		 */
 7968		return;
 7969
 7970	new_range = dst_reg->off;
 7971	if (range_right_open)
 7972		new_range--;
 7973
 7974	/* Examples for register markings:
 7975	 *
 7976	 * pkt_data in dst register:
 7977	 *
 7978	 *   r2 = r3;
 7979	 *   r2 += 8;
 7980	 *   if (r2 > pkt_end) goto <handle exception>
 7981	 *   <access okay>
 7982	 *
 7983	 *   r2 = r3;
 7984	 *   r2 += 8;
 7985	 *   if (r2 < pkt_end) goto <access okay>
 7986	 *   <handle exception>
 7987	 *
 7988	 *   Where:
 7989	 *     r2 == dst_reg, pkt_end == src_reg
 7990	 *     r2=pkt(id=n,off=8,r=0)
 7991	 *     r3=pkt(id=n,off=0,r=0)
 7992	 *
 7993	 * pkt_data in src register:
 7994	 *
 7995	 *   r2 = r3;
 7996	 *   r2 += 8;
 7997	 *   if (pkt_end >= r2) goto <access okay>
 7998	 *   <handle exception>
 7999	 *
 8000	 *   r2 = r3;
 8001	 *   r2 += 8;
 8002	 *   if (pkt_end <= r2) goto <handle exception>
 8003	 *   <access okay>
 8004	 *
 8005	 *   Where:
 8006	 *     pkt_end == dst_reg, r2 == src_reg
 8007	 *     r2=pkt(id=n,off=8,r=0)
 8008	 *     r3=pkt(id=n,off=0,r=0)
 8009	 *
 8010	 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
 8011	 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
 8012	 * and [r3, r3 + 8-1) respectively is safe to access depending on
 8013	 * the check.
 8014	 */
 8015
 8016	/* If our ids match, then we must have the same max_value.  And we
 8017	 * don't care about the other reg's fixed offset, since if it's too big
 8018	 * the range won't allow anything.
 8019	 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
 8020	 */
 8021	for (i = 0; i <= vstate->curframe; i++)
 8022		__find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
 8023					 new_range);
 8024}
 8025
 8026static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
 8027{
 8028	struct tnum subreg = tnum_subreg(reg->var_off);
 8029	s32 sval = (s32)val;
 8030
 8031	switch (opcode) {
 8032	case BPF_JEQ:
 8033		if (tnum_is_const(subreg))
 8034			return !!tnum_equals_const(subreg, val);
 8035		break;
 8036	case BPF_JNE:
 8037		if (tnum_is_const(subreg))
 8038			return !tnum_equals_const(subreg, val);
 8039		break;
 8040	case BPF_JSET:
 8041		if ((~subreg.mask & subreg.value) & val)
 8042			return 1;
 8043		if (!((subreg.mask | subreg.value) & val))
 8044			return 0;
 8045		break;
 8046	case BPF_JGT:
 8047		if (reg->u32_min_value > val)
 8048			return 1;
 8049		else if (reg->u32_max_value <= val)
 8050			return 0;
 8051		break;
 8052	case BPF_JSGT:
 8053		if (reg->s32_min_value > sval)
 8054			return 1;
 8055		else if (reg->s32_max_value <= sval)
 8056			return 0;
 8057		break;
 8058	case BPF_JLT:
 8059		if (reg->u32_max_value < val)
 8060			return 1;
 8061		else if (reg->u32_min_value >= val)
 8062			return 0;
 8063		break;
 8064	case BPF_JSLT:
 8065		if (reg->s32_max_value < sval)
 8066			return 1;
 8067		else if (reg->s32_min_value >= sval)
 8068			return 0;
 8069		break;
 8070	case BPF_JGE:
 8071		if (reg->u32_min_value >= val)
 8072			return 1;
 8073		else if (reg->u32_max_value < val)
 8074			return 0;
 8075		break;
 8076	case BPF_JSGE:
 8077		if (reg->s32_min_value >= sval)
 8078			return 1;
 8079		else if (reg->s32_max_value < sval)
 8080			return 0;
 8081		break;
 8082	case BPF_JLE:
 8083		if (reg->u32_max_value <= val)
 8084			return 1;
 8085		else if (reg->u32_min_value > val)
 8086			return 0;
 8087		break;
 8088	case BPF_JSLE:
 8089		if (reg->s32_max_value <= sval)
 8090			return 1;
 8091		else if (reg->s32_min_value > sval)
 8092			return 0;
 8093		break;
 8094	}
 8095
 8096	return -1;
 8097}
 8098
 8099
 8100static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
 8101{
 8102	s64 sval = (s64)val;
 8103
 8104	switch (opcode) {
 8105	case BPF_JEQ:
 8106		if (tnum_is_const(reg->var_off))
 8107			return !!tnum_equals_const(reg->var_off, val);
 8108		break;
 8109	case BPF_JNE:
 8110		if (tnum_is_const(reg->var_off))
 8111			return !tnum_equals_const(reg->var_off, val);
 8112		break;
 8113	case BPF_JSET:
 8114		if ((~reg->var_off.mask & reg->var_off.value) & val)
 8115			return 1;
 8116		if (!((reg->var_off.mask | reg->var_off.value) & val))
 8117			return 0;
 8118		break;
 8119	case BPF_JGT:
 8120		if (reg->umin_value > val)
 8121			return 1;
 8122		else if (reg->umax_value <= val)
 8123			return 0;
 8124		break;
 8125	case BPF_JSGT:
 8126		if (reg->smin_value > sval)
 8127			return 1;
 8128		else if (reg->smax_value <= sval)
 8129			return 0;
 8130		break;
 8131	case BPF_JLT:
 8132		if (reg->umax_value < val)
 8133			return 1;
 8134		else if (reg->umin_value >= val)
 8135			return 0;
 8136		break;
 8137	case BPF_JSLT:
 8138		if (reg->smax_value < sval)
 8139			return 1;
 8140		else if (reg->smin_value >= sval)
 8141			return 0;
 8142		break;
 8143	case BPF_JGE:
 8144		if (reg->umin_value >= val)
 8145			return 1;
 8146		else if (reg->umax_value < val)
 8147			return 0;
 8148		break;
 8149	case BPF_JSGE:
 8150		if (reg->smin_value >= sval)
 8151			return 1;
 8152		else if (reg->smax_value < sval)
 8153			return 0;
 8154		break;
 8155	case BPF_JLE:
 8156		if (reg->umax_value <= val)
 8157			return 1;
 8158		else if (reg->umin_value > val)
 8159			return 0;
 8160		break;
 8161	case BPF_JSLE:
 8162		if (reg->smax_value <= sval)
 8163			return 1;
 8164		else if (reg->smin_value > sval)
 8165			return 0;
 8166		break;
 8167	}
 8168
 8169	return -1;
 8170}
 8171
 8172/* compute branch direction of the expression "if (reg opcode val) goto target;"
 8173 * and return:
 8174 *  1 - branch will be taken and "goto target" will be executed
 8175 *  0 - branch will not be taken and fall-through to next insn
 8176 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
 8177 *      range [0,10]
 8178 */
 8179static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
 8180			   bool is_jmp32)
 8181{
 8182	if (__is_pointer_value(false, reg)) {
 8183		if (!reg_type_not_null(reg->type))
 8184			return -1;
 8185
 8186		/* If pointer is valid tests against zero will fail so we can
 8187		 * use this to direct branch taken.
 8188		 */
 8189		if (val != 0)
 8190			return -1;
 8191
 8192		switch (opcode) {
 8193		case BPF_JEQ:
 8194			return 0;
 8195		case BPF_JNE:
 8196			return 1;
 8197		default:
 8198			return -1;
 8199		}
 8200	}
 8201
 8202	if (is_jmp32)
 8203		return is_branch32_taken(reg, val, opcode);
 8204	return is_branch64_taken(reg, val, opcode);
 8205}
 8206
 8207static int flip_opcode(u32 opcode)
 8208{
 8209	/* How can we transform "a <op> b" into "b <op> a"? */
 8210	static const u8 opcode_flip[16] = {
 8211		/* these stay the same */
 8212		[BPF_JEQ  >> 4] = BPF_JEQ,
 8213		[BPF_JNE  >> 4] = BPF_JNE,
 8214		[BPF_JSET >> 4] = BPF_JSET,
 8215		/* these swap "lesser" and "greater" (L and G in the opcodes) */
 8216		[BPF_JGE  >> 4] = BPF_JLE,
 8217		[BPF_JGT  >> 4] = BPF_JLT,
 8218		[BPF_JLE  >> 4] = BPF_JGE,
 8219		[BPF_JLT  >> 4] = BPF_JGT,
 8220		[BPF_JSGE >> 4] = BPF_JSLE,
 8221		[BPF_JSGT >> 4] = BPF_JSLT,
 8222		[BPF_JSLE >> 4] = BPF_JSGE,
 8223		[BPF_JSLT >> 4] = BPF_JSGT
 8224	};
 8225	return opcode_flip[opcode >> 4];
 8226}
 8227
 8228static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
 8229				   struct bpf_reg_state *src_reg,
 8230				   u8 opcode)
 8231{
 8232	struct bpf_reg_state *pkt;
 8233
 8234	if (src_reg->type == PTR_TO_PACKET_END) {
 8235		pkt = dst_reg;
 8236	} else if (dst_reg->type == PTR_TO_PACKET_END) {
 8237		pkt = src_reg;
 8238		opcode = flip_opcode(opcode);
 8239	} else {
 8240		return -1;
 8241	}
 8242
 8243	if (pkt->range >= 0)
 8244		return -1;
 8245
 8246	switch (opcode) {
 8247	case BPF_JLE:
 8248		/* pkt <= pkt_end */
 8249		fallthrough;
 8250	case BPF_JGT:
 8251		/* pkt > pkt_end */
 8252		if (pkt->range == BEYOND_PKT_END)
 8253			/* pkt has at last one extra byte beyond pkt_end */
 8254			return opcode == BPF_JGT;
 8255		break;
 8256	case BPF_JLT:
 8257		/* pkt < pkt_end */
 8258		fallthrough;
 8259	case BPF_JGE:
 8260		/* pkt >= pkt_end */
 8261		if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
 8262			return opcode == BPF_JGE;
 8263		break;
 8264	}
 8265	return -1;
 8266}
 8267
 8268/* Adjusts the register min/max values in the case that the dst_reg is the
 8269 * variable register that we are working on, and src_reg is a constant or we're
 8270 * simply doing a BPF_K check.
 8271 * In JEQ/JNE cases we also adjust the var_off values.
 8272 */
 8273static void reg_set_min_max(struct bpf_reg_state *true_reg,
 8274			    struct bpf_reg_state *false_reg,
 8275			    u64 val, u32 val32,
 8276			    u8 opcode, bool is_jmp32)
 8277{
 8278	struct tnum false_32off = tnum_subreg(false_reg->var_off);
 8279	struct tnum false_64off = false_reg->var_off;
 8280	struct tnum true_32off = tnum_subreg(true_reg->var_off);
 8281	struct tnum true_64off = true_reg->var_off;
 8282	s64 sval = (s64)val;
 8283	s32 sval32 = (s32)val32;
 8284
 8285	/* If the dst_reg is a pointer, we can't learn anything about its
 8286	 * variable offset from the compare (unless src_reg were a pointer into
 8287	 * the same object, but we don't bother with that.
 8288	 * Since false_reg and true_reg have the same type by construction, we
 8289	 * only need to check one of them for pointerness.
 8290	 */
 8291	if (__is_pointer_value(false, false_reg))
 8292		return;
 8293
 8294	switch (opcode) {
 8295	case BPF_JEQ:
 8296	case BPF_JNE:
 8297	{
 8298		struct bpf_reg_state *reg =
 8299			opcode == BPF_JEQ ? true_reg : false_reg;
 8300
 8301		/* JEQ/JNE comparison doesn't change the register equivalence.
 8302		 * r1 = r2;
 8303		 * if (r1 == 42) goto label;
 8304		 * ...
 8305		 * label: // here both r1 and r2 are known to be 42.
 8306		 *
 8307		 * Hence when marking register as known preserve it's ID.
 8308		 */
 8309		if (is_jmp32)
 8310			__mark_reg32_known(reg, val32);
 8311		else
 8312			___mark_reg_known(reg, val);
 8313		break;
 8314	}
 8315	case BPF_JSET:
 8316		if (is_jmp32) {
 8317			false_32off = tnum_and(false_32off, tnum_const(~val32));
 8318			if (is_power_of_2(val32))
 8319				true_32off = tnum_or(true_32off,
 8320						     tnum_const(val32));
 8321		} else {
 8322			false_64off = tnum_and(false_64off, tnum_const(~val));
 8323			if (is_power_of_2(val))
 8324				true_64off = tnum_or(true_64off,
 8325						     tnum_const(val));
 8326		}
 8327		break;
 8328	case BPF_JGE:
 8329	case BPF_JGT:
 8330	{
 8331		if (is_jmp32) {
 8332			u32 false_umax = opcode == BPF_JGT ? val32  : val32 - 1;
 8333			u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
 8334
 8335			false_reg->u32_max_value = min(false_reg->u32_max_value,
 8336						       false_umax);
 8337			true_reg->u32_min_value = max(true_reg->u32_min_value,
 8338						      true_umin);
 8339		} else {
 8340			u64 false_umax = opcode == BPF_JGT ? val    : val - 1;
 8341			u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
 8342
 8343			false_reg->umax_value = min(false_reg->umax_value, false_umax);
 8344			true_reg->umin_value = max(true_reg->umin_value, true_umin);
 8345		}
 8346		break;
 8347	}
 8348	case BPF_JSGE:
 8349	case BPF_JSGT:
 8350	{
 8351		if (is_jmp32) {
 8352			s32 false_smax = opcode == BPF_JSGT ? sval32    : sval32 - 1;
 8353			s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
 8354
 8355			false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
 8356			true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
 8357		} else {
 8358			s64 false_smax = opcode == BPF_JSGT ? sval    : sval - 1;
 8359			s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
 8360
 8361			false_reg->smax_value = min(false_reg->smax_value, false_smax);
 8362			true_reg->smin_value = max(true_reg->smin_value, true_smin);
 8363		}
 8364		break;
 8365	}
 8366	case BPF_JLE:
 8367	case BPF_JLT:
 8368	{
 8369		if (is_jmp32) {
 8370			u32 false_umin = opcode == BPF_JLT ? val32  : val32 + 1;
 8371			u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
 8372
 8373			false_reg->u32_min_value = max(false_reg->u32_min_value,
 8374						       false_umin);
 8375			true_reg->u32_max_value = min(true_reg->u32_max_value,
 8376						      true_umax);
 8377		} else {
 8378			u64 false_umin = opcode == BPF_JLT ? val    : val + 1;
 8379			u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
 8380
 8381			false_reg->umin_value = max(false_reg->umin_value, false_umin);
 8382			true_reg->umax_value = min(true_reg->umax_value, true_umax);
 8383		}
 8384		break;
 8385	}
 8386	case BPF_JSLE:
 8387	case BPF_JSLT:
 8388	{
 8389		if (is_jmp32) {
 8390			s32 false_smin = opcode == BPF_JSLT ? sval32    : sval32 + 1;
 8391			s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
 8392
 8393			false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
 8394			true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
 8395		} else {
 8396			s64 false_smin = opcode == BPF_JSLT ? sval    : sval + 1;
 8397			s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
 8398
 8399			false_reg->smin_value = max(false_reg->smin_value, false_smin);
 8400			true_reg->smax_value = min(true_reg->smax_value, true_smax);
 8401		}
 8402		break;
 8403	}
 8404	default:
 8405		return;
 8406	}
 8407
 8408	if (is_jmp32) {
 8409		false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
 8410					     tnum_subreg(false_32off));
 8411		true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
 8412					    tnum_subreg(true_32off));
 8413		__reg_combine_32_into_64(false_reg);
 8414		__reg_combine_32_into_64(true_reg);
 8415	} else {
 8416		false_reg->var_off = false_64off;
 8417		true_reg->var_off = true_64off;
 8418		__reg_combine_64_into_32(false_reg);
 8419		__reg_combine_64_into_32(true_reg);
 8420	}
 8421}
 8422
 8423/* Same as above, but for the case that dst_reg holds a constant and src_reg is
 8424 * the variable reg.
 8425 */
 8426static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
 8427				struct bpf_reg_state *false_reg,
 8428				u64 val, u32 val32,
 8429				u8 opcode, bool is_jmp32)
 8430{
 8431	opcode = flip_opcode(opcode);
 8432	/* This uses zero as "not present in table"; luckily the zero opcode,
 8433	 * BPF_JA, can't get here.
 8434	 */
 8435	if (opcode)
 8436		reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
 8437}
 8438
 8439/* Regs are known to be equal, so intersect their min/max/var_off */
 8440static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
 8441				  struct bpf_reg_state *dst_reg)
 8442{
 8443	src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
 8444							dst_reg->umin_value);
 8445	src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
 8446							dst_reg->umax_value);
 8447	src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
 8448							dst_reg->smin_value);
 8449	src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
 8450							dst_reg->smax_value);
 8451	src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
 8452							     dst_reg->var_off);
 8453	/* We might have learned new bounds from the var_off. */
 8454	__update_reg_bounds(src_reg);
 8455	__update_reg_bounds(dst_reg);
 8456	/* We might have learned something about the sign bit. */
 8457	__reg_deduce_bounds(src_reg);
 8458	__reg_deduce_bounds(dst_reg);
 8459	/* We might have learned some bits from the bounds. */
 8460	__reg_bound_offset(src_reg);
 8461	__reg_bound_offset(dst_reg);
 8462	/* Intersecting with the old var_off might have improved our bounds
 8463	 * slightly.  e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
 8464	 * then new var_off is (0; 0x7f...fc) which improves our umax.
 8465	 */
 8466	__update_reg_bounds(src_reg);
 8467	__update_reg_bounds(dst_reg);
 8468}
 8469
 8470static void reg_combine_min_max(struct bpf_reg_state *true_src,
 8471				struct bpf_reg_state *true_dst,
 8472				struct bpf_reg_state *false_src,
 8473				struct bpf_reg_state *false_dst,
 8474				u8 opcode)
 8475{
 8476	switch (opcode) {
 8477	case BPF_JEQ:
 8478		__reg_combine_min_max(true_src, true_dst);
 8479		break;
 8480	case BPF_JNE:
 8481		__reg_combine_min_max(false_src, false_dst);
 8482		break;
 8483	}
 8484}
 8485
 8486static void mark_ptr_or_null_reg(struct bpf_func_state *state,
 8487				 struct bpf_reg_state *reg, u32 id,
 8488				 bool is_null)
 8489{
 8490	if (reg_type_may_be_null(reg->type) && reg->id == id &&
 8491	    !WARN_ON_ONCE(!reg->id)) {
 8492		/* Old offset (both fixed and variable parts) should
 8493		 * have been known-zero, because we don't allow pointer
 8494		 * arithmetic on pointers that might be NULL.
 8495		 */
 8496		if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
 8497				 !tnum_equals_const(reg->var_off, 0) ||
 8498				 reg->off)) {
 8499			__mark_reg_known_zero(reg);
 8500			reg->off = 0;
 8501		}
 8502		if (is_null) {
 8503			reg->type = SCALAR_VALUE;
 8504			/* We don't need id and ref_obj_id from this point
 8505			 * onwards anymore, thus we should better reset it,
 8506			 * so that state pruning has chances to take effect.
 8507			 */
 8508			reg->id = 0;
 8509			reg->ref_obj_id = 0;
 8510
 8511			return;
 8512		}
 8513
 8514		mark_ptr_not_null_reg(reg);
 8515
 8516		if (!reg_may_point_to_spin_lock(reg)) {
 8517			/* For not-NULL ptr, reg->ref_obj_id will be reset
 8518			 * in release_reg_references().
 8519			 *
 8520			 * reg->id is still used by spin_lock ptr. Other
 8521			 * than spin_lock ptr type, reg->id can be reset.
 8522			 */
 8523			reg->id = 0;
 8524		}
 8525	}
 8526}
 8527
 8528static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
 8529				    bool is_null)
 8530{
 8531	struct bpf_reg_state *reg;
 8532	int i;
 8533
 8534	for (i = 0; i < MAX_BPF_REG; i++)
 8535		mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
 8536
 8537	bpf_for_each_spilled_reg(i, state, reg) {
 8538		if (!reg)
 8539			continue;
 8540		mark_ptr_or_null_reg(state, reg, id, is_null);
 8541	}
 8542}
 8543
 8544/* The logic is similar to find_good_pkt_pointers(), both could eventually
 8545 * be folded together at some point.
 8546 */
 8547static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
 8548				  bool is_null)
 8549{
 8550	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 8551	struct bpf_reg_state *regs = state->regs;
 8552	u32 ref_obj_id = regs[regno].ref_obj_id;
 8553	u32 id = regs[regno].id;
 8554	int i;
 8555
 8556	if (ref_obj_id && ref_obj_id == id && is_null)
 8557		/* regs[regno] is in the " == NULL" branch.
 8558		 * No one could have freed the reference state before
 8559		 * doing the NULL check.
 8560		 */
 8561		WARN_ON_ONCE(release_reference_state(state, id));
 8562
 8563	for (i = 0; i <= vstate->curframe; i++)
 8564		__mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
 8565}
 8566
 8567static bool try_match_pkt_pointers(const struct bpf_insn *insn,
 8568				   struct bpf_reg_state *dst_reg,
 8569				   struct bpf_reg_state *src_reg,
 8570				   struct bpf_verifier_state *this_branch,
 8571				   struct bpf_verifier_state *other_branch)
 8572{
 8573	if (BPF_SRC(insn->code) != BPF_X)
 8574		return false;
 8575
 8576	/* Pointers are always 64-bit. */
 8577	if (BPF_CLASS(insn->code) == BPF_JMP32)
 8578		return false;
 8579
 8580	switch (BPF_OP(insn->code)) {
 8581	case BPF_JGT:
 8582		if ((dst_reg->type == PTR_TO_PACKET &&
 8583		     src_reg->type == PTR_TO_PACKET_END) ||
 8584		    (dst_reg->type == PTR_TO_PACKET_META &&
 8585		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
 8586			/* pkt_data' > pkt_end, pkt_meta' > pkt_data */
 8587			find_good_pkt_pointers(this_branch, dst_reg,
 8588					       dst_reg->type, false);
 8589			mark_pkt_end(other_branch, insn->dst_reg, true);
 8590		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
 8591			    src_reg->type == PTR_TO_PACKET) ||
 8592			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
 8593			    src_reg->type == PTR_TO_PACKET_META)) {
 8594			/* pkt_end > pkt_data', pkt_data > pkt_meta' */
 8595			find_good_pkt_pointers(other_branch, src_reg,
 8596					       src_reg->type, true);
 8597			mark_pkt_end(this_branch, insn->src_reg, false);
 8598		} else {
 8599			return false;
 8600		}
 8601		break;
 8602	case BPF_JLT:
 8603		if ((dst_reg->type == PTR_TO_PACKET &&
 8604		     src_reg->type == PTR_TO_PACKET_END) ||
 8605		    (dst_reg->type == PTR_TO_PACKET_META &&
 8606		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
 8607			/* pkt_data' < pkt_end, pkt_meta' < pkt_data */
 8608			find_good_pkt_pointers(other_branch, dst_reg,
 8609					       dst_reg->type, true);
 8610			mark_pkt_end(this_branch, insn->dst_reg, false);
 8611		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
 8612			    src_reg->type == PTR_TO_PACKET) ||
 8613			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
 8614			    src_reg->type == PTR_TO_PACKET_META)) {
 8615			/* pkt_end < pkt_data', pkt_data > pkt_meta' */
 8616			find_good_pkt_pointers(this_branch, src_reg,
 8617					       src_reg->type, false);
 8618			mark_pkt_end(other_branch, insn->src_reg, true);
 8619		} else {
 8620			return false;
 8621		}
 8622		break;
 8623	case BPF_JGE:
 8624		if ((dst_reg->type == PTR_TO_PACKET &&
 8625		     src_reg->type == PTR_TO_PACKET_END) ||
 8626		    (dst_reg->type == PTR_TO_PACKET_META &&
 8627		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
 8628			/* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
 8629			find_good_pkt_pointers(this_branch, dst_reg,
 8630					       dst_reg->type, true);
 8631			mark_pkt_end(other_branch, insn->dst_reg, false);
 8632		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
 8633			    src_reg->type == PTR_TO_PACKET) ||
 8634			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
 8635			    src_reg->type == PTR_TO_PACKET_META)) {
 8636			/* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
 8637			find_good_pkt_pointers(other_branch, src_reg,
 8638					       src_reg->type, false);
 8639			mark_pkt_end(this_branch, insn->src_reg, true);
 8640		} else {
 8641			return false;
 8642		}
 8643		break;
 8644	case BPF_JLE:
 8645		if ((dst_reg->type == PTR_TO_PACKET &&
 8646		     src_reg->type == PTR_TO_PACKET_END) ||
 8647		    (dst_reg->type == PTR_TO_PACKET_META &&
 8648		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
 8649			/* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
 8650			find_good_pkt_pointers(other_branch, dst_reg,
 8651					       dst_reg->type, false);
 8652			mark_pkt_end(this_branch, insn->dst_reg, true);
 8653		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
 8654			    src_reg->type == PTR_TO_PACKET) ||
 8655			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
 8656			    src_reg->type == PTR_TO_PACKET_META)) {
 8657			/* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
 8658			find_good_pkt_pointers(this_branch, src_reg,
 8659					       src_reg->type, true);
 8660			mark_pkt_end(other_branch, insn->src_reg, false);
 8661		} else {
 8662			return false;
 8663		}
 8664		break;
 8665	default:
 8666		return false;
 8667	}
 8668
 8669	return true;
 8670}
 8671
 8672static void find_equal_scalars(struct bpf_verifier_state *vstate,
 8673			       struct bpf_reg_state *known_reg)
 8674{
 8675	struct bpf_func_state *state;
 8676	struct bpf_reg_state *reg;
 8677	int i, j;
 8678
 8679	for (i = 0; i <= vstate->curframe; i++) {
 8680		state = vstate->frame[i];
 8681		for (j = 0; j < MAX_BPF_REG; j++) {
 8682			reg = &state->regs[j];
 8683			if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
 8684				*reg = *known_reg;
 8685		}
 8686
 8687		bpf_for_each_spilled_reg(j, state, reg) {
 8688			if (!reg)
 8689				continue;
 8690			if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
 8691				*reg = *known_reg;
 8692		}
 8693	}
 8694}
 8695
 8696static int check_cond_jmp_op(struct bpf_verifier_env *env,
 8697			     struct bpf_insn *insn, int *insn_idx)
 8698{
 8699	struct bpf_verifier_state *this_branch = env->cur_state;
 8700	struct bpf_verifier_state *other_branch;
 8701	struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
 8702	struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
 8703	u8 opcode = BPF_OP(insn->code);
 8704	bool is_jmp32;
 8705	int pred = -1;
 8706	int err;
 8707
 8708	/* Only conditional jumps are expected to reach here. */
 8709	if (opcode == BPF_JA || opcode > BPF_JSLE) {
 8710		verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
 8711		return -EINVAL;
 8712	}
 8713
 8714	if (BPF_SRC(insn->code) == BPF_X) {
 8715		if (insn->imm != 0) {
 8716			verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
 8717			return -EINVAL;
 8718		}
 8719
 8720		/* check src1 operand */
 8721		err = check_reg_arg(env, insn->src_reg, SRC_OP);
 8722		if (err)
 8723			return err;
 8724
 8725		if (is_pointer_value(env, insn->src_reg)) {
 8726			verbose(env, "R%d pointer comparison prohibited\n",
 8727				insn->src_reg);
 8728			return -EACCES;
 8729		}
 8730		src_reg = &regs[insn->src_reg];
 8731	} else {
 8732		if (insn->src_reg != BPF_REG_0) {
 8733			verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
 8734			return -EINVAL;
 8735		}
 8736	}
 8737
 8738	/* check src2 operand */
 8739	err = check_reg_arg(env, insn->dst_reg, SRC_OP);
 8740	if (err)
 8741		return err;
 8742
 8743	dst_reg = &regs[insn->dst_reg];
 8744	is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
 8745
 8746	if (BPF_SRC(insn->code) == BPF_K) {
 8747		pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
 8748	} else if (src_reg->type == SCALAR_VALUE &&
 8749		   is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
 8750		pred = is_branch_taken(dst_reg,
 8751				       tnum_subreg(src_reg->var_off).value,
 8752				       opcode,
 8753				       is_jmp32);
 8754	} else if (src_reg->type == SCALAR_VALUE &&
 8755		   !is_jmp32 && tnum_is_const(src_reg->var_off)) {
 8756		pred = is_branch_taken(dst_reg,
 8757				       src_reg->var_off.value,
 8758				       opcode,
 8759				       is_jmp32);
 8760	} else if (reg_is_pkt_pointer_any(dst_reg) &&
 8761		   reg_is_pkt_pointer_any(src_reg) &&
 8762		   !is_jmp32) {
 8763		pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
 8764	}
 8765
 8766	if (pred >= 0) {
 8767		/* If we get here with a dst_reg pointer type it is because
 8768		 * above is_branch_taken() special cased the 0 comparison.
 8769		 */
 8770		if (!__is_pointer_value(false, dst_reg))
 8771			err = mark_chain_precision(env, insn->dst_reg);
 8772		if (BPF_SRC(insn->code) == BPF_X && !err &&
 8773		    !__is_pointer_value(false, src_reg))
 8774			err = mark_chain_precision(env, insn->src_reg);
 8775		if (err)
 8776			return err;
 8777	}
 8778
 8779	if (pred == 1) {
 8780		/* Only follow the goto, ignore fall-through. If needed, push
 8781		 * the fall-through branch for simulation under speculative
 8782		 * execution.
 8783		 */
 8784		if (!env->bypass_spec_v1 &&
 8785		    !sanitize_speculative_path(env, insn, *insn_idx + 1,
 8786					       *insn_idx))
 8787			return -EFAULT;
 8788		*insn_idx += insn->off;
 8789		return 0;
 8790	} else if (pred == 0) {
 8791		/* Only follow the fall-through branch, since that's where the
 8792		 * program will go. If needed, push the goto branch for
 8793		 * simulation under speculative execution.
 8794		 */
 8795		if (!env->bypass_spec_v1 &&
 8796		    !sanitize_speculative_path(env, insn,
 8797					       *insn_idx + insn->off + 1,
 8798					       *insn_idx))
 8799			return -EFAULT;
 8800		return 0;
 8801	}
 8802
 8803	other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
 8804				  false);
 8805	if (!other_branch)
 8806		return -EFAULT;
 8807	other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
 8808
 8809	/* detect if we are comparing against a constant value so we can adjust
 8810	 * our min/max values for our dst register.
 8811	 * this is only legit if both are scalars (or pointers to the same
 8812	 * object, I suppose, but we don't support that right now), because
 8813	 * otherwise the different base pointers mean the offsets aren't
 8814	 * comparable.
 8815	 */
 8816	if (BPF_SRC(insn->code) == BPF_X) {
 8817		struct bpf_reg_state *src_reg = &regs[insn->src_reg];
 8818
 8819		if (dst_reg->type == SCALAR_VALUE &&
 8820		    src_reg->type == SCALAR_VALUE) {
 8821			if (tnum_is_const(src_reg->var_off) ||
 8822			    (is_jmp32 &&
 8823			     tnum_is_const(tnum_subreg(src_reg->var_off))))
 8824				reg_set_min_max(&other_branch_regs[insn->dst_reg],
 8825						dst_reg,
 8826						src_reg->var_off.value,
 8827						tnum_subreg(src_reg->var_off).value,
 8828						opcode, is_jmp32);
 8829			else if (tnum_is_const(dst_reg->var_off) ||
 8830				 (is_jmp32 &&
 8831				  tnum_is_const(tnum_subreg(dst_reg->var_off))))
 8832				reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
 8833						    src_reg,
 8834						    dst_reg->var_off.value,
 8835						    tnum_subreg(dst_reg->var_off).value,
 8836						    opcode, is_jmp32);
 8837			else if (!is_jmp32 &&
 8838				 (opcode == BPF_JEQ || opcode == BPF_JNE))
 8839				/* Comparing for equality, we can combine knowledge */
 8840				reg_combine_min_max(&other_branch_regs[insn->src_reg],
 8841						    &other_branch_regs[insn->dst_reg],
 8842						    src_reg, dst_reg, opcode);
 8843			if (src_reg->id &&
 8844			    !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
 8845				find_equal_scalars(this_branch, src_reg);
 8846				find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
 8847			}
 8848
 8849		}
 8850	} else if (dst_reg->type == SCALAR_VALUE) {
 8851		reg_set_min_max(&other_branch_regs[insn->dst_reg],
 8852					dst_reg, insn->imm, (u32)insn->imm,
 8853					opcode, is_jmp32);
 8854	}
 8855
 8856	if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
 8857	    !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
 8858		find_equal_scalars(this_branch, dst_reg);
 8859		find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
 8860	}
 8861
 8862	/* detect if R == 0 where R is returned from bpf_map_lookup_elem().
 8863	 * NOTE: these optimizations below are related with pointer comparison
 8864	 *       which will never be JMP32.
 8865	 */
 8866	if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
 8867	    insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
 8868	    reg_type_may_be_null(dst_reg->type)) {
 8869		/* Mark all identical registers in each branch as either
 8870		 * safe or unknown depending R == 0 or R != 0 conditional.
 8871		 */
 8872		mark_ptr_or_null_regs(this_branch, insn->dst_reg,
 8873				      opcode == BPF_JNE);
 8874		mark_ptr_or_null_regs(other_branch, insn->dst_reg,
 8875				      opcode == BPF_JEQ);
 8876	} else if (!try_match_pkt_pointers(insn, dst_reg, &regs[insn->src_reg],
 8877					   this_branch, other_branch) &&
 8878		   is_pointer_value(env, insn->dst_reg)) {
 8879		verbose(env, "R%d pointer comparison prohibited\n",
 8880			insn->dst_reg);
 8881		return -EACCES;
 8882	}
 8883	if (env->log.level & BPF_LOG_LEVEL)
 8884		print_verifier_state(env, this_branch->frame[this_branch->curframe]);
 8885	return 0;
 8886}
 8887
 8888/* verify BPF_LD_IMM64 instruction */
 8889static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
 8890{
 8891	struct bpf_insn_aux_data *aux = cur_aux(env);
 8892	struct bpf_reg_state *regs = cur_regs(env);
 8893	struct bpf_reg_state *dst_reg;
 8894	struct bpf_map *map;
 8895	int err;
 8896
 8897	if (BPF_SIZE(insn->code) != BPF_DW) {
 8898		verbose(env, "invalid BPF_LD_IMM insn\n");
 8899		return -EINVAL;
 8900	}
 8901	if (insn->off != 0) {
 8902		verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
 8903		return -EINVAL;
 8904	}
 8905
 8906	err = check_reg_arg(env, insn->dst_reg, DST_OP);
 8907	if (err)
 8908		return err;
 8909
 8910	dst_reg = &regs[insn->dst_reg];
 8911	if (insn->src_reg == 0) {
 8912		u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
 8913
 8914		dst_reg->type = SCALAR_VALUE;
 8915		__mark_reg_known(&regs[insn->dst_reg], imm);
 8916		return 0;
 8917	}
 8918
 8919	if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
 8920		mark_reg_known_zero(env, regs, insn->dst_reg);
 8921
 8922		dst_reg->type = aux->btf_var.reg_type;
 8923		switch (dst_reg->type) {
 8924		case PTR_TO_MEM:
 8925			dst_reg->mem_size = aux->btf_var.mem_size;
 8926			break;
 8927		case PTR_TO_BTF_ID:
 8928		case PTR_TO_PERCPU_BTF_ID:
 8929			dst_reg->btf = aux->btf_var.btf;
 8930			dst_reg->btf_id = aux->btf_var.btf_id;
 8931			break;
 8932		default:
 8933			verbose(env, "bpf verifier is misconfigured\n");
 8934			return -EFAULT;
 8935		}
 8936		return 0;
 8937	}
 8938
 8939	if (insn->src_reg == BPF_PSEUDO_FUNC) {
 8940		struct bpf_prog_aux *aux = env->prog->aux;
 8941		u32 subprogno = insn[1].imm;
 8942
 8943		if (!aux->func_info) {
 8944			verbose(env, "missing btf func_info\n");
 8945			return -EINVAL;
 8946		}
 8947		if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
 8948			verbose(env, "callback function not static\n");
 8949			return -EINVAL;
 8950		}
 8951
 8952		dst_reg->type = PTR_TO_FUNC;
 8953		dst_reg->subprogno = subprogno;
 8954		return 0;
 8955	}
 8956
 8957	map = env->used_maps[aux->map_index];
 8958	mark_reg_known_zero(env, regs, insn->dst_reg);
 8959	dst_reg->map_ptr = map;
 8960
 8961	if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
 8962	    insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
 8963		dst_reg->type = PTR_TO_MAP_VALUE;
 8964		dst_reg->off = aux->map_off;
 8965		if (map_value_has_spin_lock(map))
 8966			dst_reg->id = ++env->id_gen;
 8967	} else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
 8968		   insn->src_reg == BPF_PSEUDO_MAP_IDX) {
 8969		dst_reg->type = CONST_PTR_TO_MAP;
 8970	} else {
 8971		verbose(env, "bpf verifier is misconfigured\n");
 8972		return -EINVAL;
 8973	}
 8974
 8975	return 0;
 8976}
 8977
 8978static bool may_access_skb(enum bpf_prog_type type)
 8979{
 8980	switch (type) {
 8981	case BPF_PROG_TYPE_SOCKET_FILTER:
 8982	case BPF_PROG_TYPE_SCHED_CLS:
 8983	case BPF_PROG_TYPE_SCHED_ACT:
 8984		return true;
 8985	default:
 8986		return false;
 8987	}
 8988}
 8989
 8990/* verify safety of LD_ABS|LD_IND instructions:
 8991 * - they can only appear in the programs where ctx == skb
 8992 * - since they are wrappers of function calls, they scratch R1-R5 registers,
 8993 *   preserve R6-R9, and store return value into R0
 8994 *
 8995 * Implicit input:
 8996 *   ctx == skb == R6 == CTX
 8997 *
 8998 * Explicit input:
 8999 *   SRC == any register
 9000 *   IMM == 32-bit immediate
 9001 *
 9002 * Output:
 9003 *   R0 - 8/16/32-bit skb data converted to cpu endianness
 9004 */
 9005static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
 9006{
 9007	struct bpf_reg_state *regs = cur_regs(env);
 9008	static const int ctx_reg = BPF_REG_6;
 9009	u8 mode = BPF_MODE(insn->code);
 9010	int i, err;
 9011
 9012	if (!may_access_skb(resolve_prog_type(env->prog))) {
 9013		verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
 9014		return -EINVAL;
 9015	}
 9016
 9017	if (!env->ops->gen_ld_abs) {
 9018		verbose(env, "bpf verifier is misconfigured\n");
 9019		return -EINVAL;
 9020	}
 9021
 9022	if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
 9023	    BPF_SIZE(insn->code) == BPF_DW ||
 9024	    (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
 9025		verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
 9026		return -EINVAL;
 9027	}
 9028
 9029	/* check whether implicit source operand (register R6) is readable */
 9030	err = check_reg_arg(env, ctx_reg, SRC_OP);
 9031	if (err)
 9032		return err;
 9033
 9034	/* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
 9035	 * gen_ld_abs() may terminate the program at runtime, leading to
 9036	 * reference leak.
 9037	 */
 9038	err = check_reference_leak(env);
 9039	if (err) {
 9040		verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
 9041		return err;
 9042	}
 9043
 9044	if (env->cur_state->active_spin_lock) {
 9045		verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
 9046		return -EINVAL;
 9047	}
 9048
 9049	if (regs[ctx_reg].type != PTR_TO_CTX) {
 9050		verbose(env,
 9051			"at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
 9052		return -EINVAL;
 9053	}
 9054
 9055	if (mode == BPF_IND) {
 9056		/* check explicit source operand */
 9057		err = check_reg_arg(env, insn->src_reg, SRC_OP);
 9058		if (err)
 9059			return err;
 9060	}
 9061
 9062	err = check_ctx_reg(env, &regs[ctx_reg], ctx_reg);
 9063	if (err < 0)
 9064		return err;
 9065
 9066	/* reset caller saved regs to unreadable */
 9067	for (i = 0; i < CALLER_SAVED_REGS; i++) {
 9068		mark_reg_not_init(env, regs, caller_saved[i]);
 9069		check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
 9070	}
 9071
 9072	/* mark destination R0 register as readable, since it contains
 9073	 * the value fetched from the packet.
 9074	 * Already marked as written above.
 9075	 */
 9076	mark_reg_unknown(env, regs, BPF_REG_0);
 9077	/* ld_abs load up to 32-bit skb data. */
 9078	regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
 9079	return 0;
 9080}
 9081
 9082static int check_return_code(struct bpf_verifier_env *env)
 9083{
 9084	struct tnum enforce_attach_type_range = tnum_unknown;
 9085	const struct bpf_prog *prog = env->prog;
 9086	struct bpf_reg_state *reg;
 9087	struct tnum range = tnum_range(0, 1);
 9088	enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
 9089	int err;
 9090	const bool is_subprog = env->cur_state->frame[0]->subprogno;
 9091
 9092	/* LSM and struct_ops func-ptr's return type could be "void" */
 9093	if (!is_subprog &&
 9094	    (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
 9095	     prog_type == BPF_PROG_TYPE_LSM) &&
 9096	    !prog->aux->attach_func_proto->type)
 9097		return 0;
 9098
 9099	/* eBPF calling convention is such that R0 is used
 9100	 * to return the value from eBPF program.
 9101	 * Make sure that it's readable at this time
 9102	 * of bpf_exit, which means that program wrote
 9103	 * something into it earlier
 9104	 */
 9105	err = check_reg_arg(env, BPF_REG_0, SRC_OP);
 9106	if (err)
 9107		return err;
 9108
 9109	if (is_pointer_value(env, BPF_REG_0)) {
 9110		verbose(env, "R0 leaks addr as return value\n");
 9111		return -EACCES;
 9112	}
 9113
 9114	reg = cur_regs(env) + BPF_REG_0;
 9115	if (is_subprog) {
 9116		if (reg->type != SCALAR_VALUE) {
 9117			verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
 9118				reg_type_str[reg->type]);
 9119			return -EINVAL;
 9120		}
 9121		return 0;
 9122	}
 9123
 9124	switch (prog_type) {
 9125	case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
 9126		if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
 9127		    env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
 9128		    env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
 9129		    env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
 9130		    env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
 9131		    env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
 9132			range = tnum_range(1, 1);
 9133		if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
 9134		    env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
 9135			range = tnum_range(0, 3);
 9136		break;
 9137	case BPF_PROG_TYPE_CGROUP_SKB:
 9138		if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
 9139			range = tnum_range(0, 3);
 9140			enforce_attach_type_range = tnum_range(2, 3);
 9141		}
 9142		break;
 9143	case BPF_PROG_TYPE_CGROUP_SOCK:
 9144	case BPF_PROG_TYPE_SOCK_OPS:
 9145	case BPF_PROG_TYPE_CGROUP_DEVICE:
 9146	case BPF_PROG_TYPE_CGROUP_SYSCTL:
 9147	case BPF_PROG_TYPE_CGROUP_SOCKOPT:
 9148		break;
 9149	case BPF_PROG_TYPE_RAW_TRACEPOINT:
 9150		if (!env->prog->aux->attach_btf_id)
 9151			return 0;
 9152		range = tnum_const(0);
 9153		break;
 9154	case BPF_PROG_TYPE_TRACING:
 9155		switch (env->prog->expected_attach_type) {
 9156		case BPF_TRACE_FENTRY:
 9157		case BPF_TRACE_FEXIT:
 9158			range = tnum_const(0);
 9159			break;
 9160		case BPF_TRACE_RAW_TP:
 9161		case BPF_MODIFY_RETURN:
 9162			return 0;
 9163		case BPF_TRACE_ITER:
 9164			break;
 9165		default:
 9166			return -ENOTSUPP;
 9167		}
 9168		break;
 9169	case BPF_PROG_TYPE_SK_LOOKUP:
 9170		range = tnum_range(SK_DROP, SK_PASS);
 9171		break;
 9172	case BPF_PROG_TYPE_EXT:
 9173		/* freplace program can return anything as its return value
 9174		 * depends on the to-be-replaced kernel func or bpf program.
 9175		 */
 9176	default:
 9177		return 0;
 9178	}
 9179
 9180	if (reg->type != SCALAR_VALUE) {
 9181		verbose(env, "At program exit the register R0 is not a known value (%s)\n",
 9182			reg_type_str[reg->type]);
 9183		return -EINVAL;
 9184	}
 9185
 9186	if (!tnum_in(range, reg->var_off)) {
 9187		verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
 9188		return -EINVAL;
 9189	}
 9190
 9191	if (!tnum_is_unknown(enforce_attach_type_range) &&
 9192	    tnum_in(enforce_attach_type_range, reg->var_off))
 9193		env->prog->enforce_expected_attach_type = 1;
 9194	return 0;
 9195}
 9196
 9197/* non-recursive DFS pseudo code
 9198 * 1  procedure DFS-iterative(G,v):
 9199 * 2      label v as discovered
 9200 * 3      let S be a stack
 9201 * 4      S.push(v)
 9202 * 5      while S is not empty
 9203 * 6            t <- S.pop()
 9204 * 7            if t is what we're looking for:
 9205 * 8                return t
 9206 * 9            for all edges e in G.adjacentEdges(t) do
 9207 * 10               if edge e is already labelled
 9208 * 11                   continue with the next edge
 9209 * 12               w <- G.adjacentVertex(t,e)
 9210 * 13               if vertex w is not discovered and not explored
 9211 * 14                   label e as tree-edge
 9212 * 15                   label w as discovered
 9213 * 16                   S.push(w)
 9214 * 17                   continue at 5
 9215 * 18               else if vertex w is discovered
 9216 * 19                   label e as back-edge
 9217 * 20               else
 9218 * 21                   // vertex w is explored
 9219 * 22                   label e as forward- or cross-edge
 9220 * 23           label t as explored
 9221 * 24           S.pop()
 9222 *
 9223 * convention:
 9224 * 0x10 - discovered
 9225 * 0x11 - discovered and fall-through edge labelled
 9226 * 0x12 - discovered and fall-through and branch edges labelled
 9227 * 0x20 - explored
 9228 */
 9229
 9230enum {
 9231	DISCOVERED = 0x10,
 9232	EXPLORED = 0x20,
 9233	FALLTHROUGH = 1,
 9234	BRANCH = 2,
 9235};
 9236
 9237static u32 state_htab_size(struct bpf_verifier_env *env)
 9238{
 9239	return env->prog->len;
 9240}
 9241
 9242static struct bpf_verifier_state_list **explored_state(
 9243					struct bpf_verifier_env *env,
 9244					int idx)
 9245{
 9246	struct bpf_verifier_state *cur = env->cur_state;
 9247	struct bpf_func_state *state = cur->frame[cur->curframe];
 9248
 9249	return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
 9250}
 9251
 9252static void init_explored_state(struct bpf_verifier_env *env, int idx)
 9253{
 9254	env->insn_aux_data[idx].prune_point = true;
 9255}
 9256
 9257enum {
 9258	DONE_EXPLORING = 0,
 9259	KEEP_EXPLORING = 1,
 9260};
 9261
 9262/* t, w, e - match pseudo-code above:
 9263 * t - index of current instruction
 9264 * w - next instruction
 9265 * e - edge
 9266 */
 9267static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
 9268		     bool loop_ok)
 9269{
 9270	int *insn_stack = env->cfg.insn_stack;
 9271	int *insn_state = env->cfg.insn_state;
 9272
 9273	if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
 9274		return DONE_EXPLORING;
 9275
 9276	if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
 9277		return DONE_EXPLORING;
 9278
 9279	if (w < 0 || w >= env->prog->len) {
 9280		verbose_linfo(env, t, "%d: ", t);
 9281		verbose(env, "jump out of range from insn %d to %d\n", t, w);
 9282		return -EINVAL;
 9283	}
 9284
 9285	if (e == BRANCH)
 9286		/* mark branch target for state pruning */
 9287		init_explored_state(env, w);
 9288
 9289	if (insn_state[w] == 0) {
 9290		/* tree-edge */
 9291		insn_state[t] = DISCOVERED | e;
 9292		insn_state[w] = DISCOVERED;
 9293		if (env->cfg.cur_stack >= env->prog->len)
 9294			return -E2BIG;
 9295		insn_stack[env->cfg.cur_stack++] = w;
 9296		return KEEP_EXPLORING;
 9297	} else if ((insn_state[w] & 0xF0) == DISCOVERED) {
 9298		if (loop_ok && env->bpf_capable)
 9299			return DONE_EXPLORING;
 9300		verbose_linfo(env, t, "%d: ", t);
 9301		verbose_linfo(env, w, "%d: ", w);
 9302		verbose(env, "back-edge from insn %d to %d\n", t, w);
 9303		return -EINVAL;
 9304	} else if (insn_state[w] == EXPLORED) {
 9305		/* forward- or cross-edge */
 9306		insn_state[t] = DISCOVERED | e;
 9307	} else {
 9308		verbose(env, "insn state internal bug\n");
 9309		return -EFAULT;
 9310	}
 9311	return DONE_EXPLORING;
 9312}
 9313
 9314static int visit_func_call_insn(int t, int insn_cnt,
 9315				struct bpf_insn *insns,
 9316				struct bpf_verifier_env *env,
 9317				bool visit_callee)
 9318{
 9319	int ret;
 9320
 9321	ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
 9322	if (ret)
 9323		return ret;
 9324
 9325	if (t + 1 < insn_cnt)
 9326		init_explored_state(env, t + 1);
 9327	if (visit_callee) {
 9328		init_explored_state(env, t);
 9329		ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
 9330				env, false);
 9331	}
 9332	return ret;
 9333}
 9334
 9335/* Visits the instruction at index t and returns one of the following:
 9336 *  < 0 - an error occurred
 9337 *  DONE_EXPLORING - the instruction was fully explored
 9338 *  KEEP_EXPLORING - there is still work to be done before it is fully explored
 9339 */
 9340static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
 9341{
 9342	struct bpf_insn *insns = env->prog->insnsi;
 9343	int ret;
 9344
 9345	if (bpf_pseudo_func(insns + t))
 9346		return visit_func_call_insn(t, insn_cnt, insns, env, true);
 9347
 9348	/* All non-branch instructions have a single fall-through edge. */
 9349	if (BPF_CLASS(insns[t].code) != BPF_JMP &&
 9350	    BPF_CLASS(insns[t].code) != BPF_JMP32)
 9351		return push_insn(t, t + 1, FALLTHROUGH, env, false);
 9352
 9353	switch (BPF_OP(insns[t].code)) {
 9354	case BPF_EXIT:
 9355		return DONE_EXPLORING;
 9356
 9357	case BPF_CALL:
 9358		return visit_func_call_insn(t, insn_cnt, insns, env,
 9359					    insns[t].src_reg == BPF_PSEUDO_CALL);
 9360
 9361	case BPF_JA:
 9362		if (BPF_SRC(insns[t].code) != BPF_K)
 9363			return -EINVAL;
 9364
 9365		/* unconditional jump with single edge */
 9366		ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
 9367				true);
 9368		if (ret)
 9369			return ret;
 9370
 9371		/* unconditional jmp is not a good pruning point,
 9372		 * but it's marked, since backtracking needs
 9373		 * to record jmp history in is_state_visited().
 9374		 */
 9375		init_explored_state(env, t + insns[t].off + 1);
 9376		/* tell verifier to check for equivalent states
 9377		 * after every call and jump
 9378		 */
 9379		if (t + 1 < insn_cnt)
 9380			init_explored_state(env, t + 1);
 9381
 9382		return ret;
 9383
 9384	default:
 9385		/* conditional jump with two edges */
 9386		init_explored_state(env, t);
 9387		ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
 9388		if (ret)
 9389			return ret;
 9390
 9391		return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
 9392	}
 9393}
 9394
 9395/* non-recursive depth-first-search to detect loops in BPF program
 9396 * loop == back-edge in directed graph
 9397 */
 9398static int check_cfg(struct bpf_verifier_env *env)
 9399{
 9400	int insn_cnt = env->prog->len;
 9401	int *insn_stack, *insn_state;
 9402	int ret = 0;
 9403	int i;
 9404
 9405	insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
 9406	if (!insn_state)
 9407		return -ENOMEM;
 9408
 9409	insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
 9410	if (!insn_stack) {
 9411		kvfree(insn_state);
 9412		return -ENOMEM;
 9413	}
 9414
 9415	insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
 9416	insn_stack[0] = 0; /* 0 is the first instruction */
 9417	env->cfg.cur_stack = 1;
 9418
 9419	while (env->cfg.cur_stack > 0) {
 9420		int t = insn_stack[env->cfg.cur_stack - 1];
 9421
 9422		ret = visit_insn(t, insn_cnt, env);
 9423		switch (ret) {
 9424		case DONE_EXPLORING:
 9425			insn_state[t] = EXPLORED;
 9426			env->cfg.cur_stack--;
 9427			break;
 9428		case KEEP_EXPLORING:
 9429			break;
 9430		default:
 9431			if (ret > 0) {
 9432				verbose(env, "visit_insn internal bug\n");
 9433				ret = -EFAULT;
 9434			}
 9435			goto err_free;
 9436		}
 9437	}
 9438
 9439	if (env->cfg.cur_stack < 0) {
 9440		verbose(env, "pop stack internal bug\n");
 9441		ret = -EFAULT;
 9442		goto err_free;
 9443	}
 9444
 9445	for (i = 0; i < insn_cnt; i++) {
 9446		if (insn_state[i] != EXPLORED) {
 9447			verbose(env, "unreachable insn %d\n", i);
 9448			ret = -EINVAL;
 9449			goto err_free;
 9450		}
 9451	}
 9452	ret = 0; /* cfg looks good */
 9453
 9454err_free:
 9455	kvfree(insn_state);
 9456	kvfree(insn_stack);
 9457	env->cfg.insn_state = env->cfg.insn_stack = NULL;
 9458	return ret;
 9459}
 9460
 9461static int check_abnormal_return(struct bpf_verifier_env *env)
 9462{
 9463	int i;
 9464
 9465	for (i = 1; i < env->subprog_cnt; i++) {
 9466		if (env->subprog_info[i].has_ld_abs) {
 9467			verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
 9468			return -EINVAL;
 9469		}
 9470		if (env->subprog_info[i].has_tail_call) {
 9471			verbose(env, "tail_call is not allowed in subprogs without BTF\n");
 9472			return -EINVAL;
 9473		}
 9474	}
 9475	return 0;
 9476}
 9477
 9478/* The minimum supported BTF func info size */
 9479#define MIN_BPF_FUNCINFO_SIZE	8
 9480#define MAX_FUNCINFO_REC_SIZE	252
 9481
 9482static int check_btf_func(struct bpf_verifier_env *env,
 9483			  const union bpf_attr *attr,
 9484			  bpfptr_t uattr)
 9485{
 9486	const struct btf_type *type, *func_proto, *ret_type;
 9487	u32 i, nfuncs, urec_size, min_size;
 9488	u32 krec_size = sizeof(struct bpf_func_info);
 9489	struct bpf_func_info *krecord;
 9490	struct bpf_func_info_aux *info_aux = NULL;
 9491	struct bpf_prog *prog;
 9492	const struct btf *btf;
 9493	bpfptr_t urecord;
 9494	u32 prev_offset = 0;
 9495	bool scalar_return;
 9496	int ret = -ENOMEM;
 9497
 9498	nfuncs = attr->func_info_cnt;
 9499	if (!nfuncs) {
 9500		if (check_abnormal_return(env))
 9501			return -EINVAL;
 9502		return 0;
 9503	}
 9504
 9505	if (nfuncs != env->subprog_cnt) {
 9506		verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
 9507		return -EINVAL;
 9508	}
 9509
 9510	urec_size = attr->func_info_rec_size;
 9511	if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
 9512	    urec_size > MAX_FUNCINFO_REC_SIZE ||
 9513	    urec_size % sizeof(u32)) {
 9514		verbose(env, "invalid func info rec size %u\n", urec_size);
 9515		return -EINVAL;
 9516	}
 9517
 9518	prog = env->prog;
 9519	btf = prog->aux->btf;
 9520
 9521	urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
 9522	min_size = min_t(u32, krec_size, urec_size);
 9523
 9524	krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
 9525	if (!krecord)
 9526		return -ENOMEM;
 9527	info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
 9528	if (!info_aux)
 9529		goto err_free;
 9530
 9531	for (i = 0; i < nfuncs; i++) {
 9532		ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
 9533		if (ret) {
 9534			if (ret == -E2BIG) {
 9535				verbose(env, "nonzero tailing record in func info");
 9536				/* set the size kernel expects so loader can zero
 9537				 * out the rest of the record.
 9538				 */
 9539				if (copy_to_bpfptr_offset(uattr,
 9540							  offsetof(union bpf_attr, func_info_rec_size),
 9541							  &min_size, sizeof(min_size)))
 9542					ret = -EFAULT;
 9543			}
 9544			goto err_free;
 9545		}
 9546
 9547		if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
 9548			ret = -EFAULT;
 9549			goto err_free;
 9550		}
 9551
 9552		/* check insn_off */
 9553		ret = -EINVAL;
 9554		if (i == 0) {
 9555			if (krecord[i].insn_off) {
 9556				verbose(env,
 9557					"nonzero insn_off %u for the first func info record",
 9558					krecord[i].insn_off);
 9559				goto err_free;
 9560			}
 9561		} else if (krecord[i].insn_off <= prev_offset) {
 9562			verbose(env,
 9563				"same or smaller insn offset (%u) than previous func info record (%u)",
 9564				krecord[i].insn_off, prev_offset);
 9565			goto err_free;
 9566		}
 9567
 9568		if (env->subprog_info[i].start != krecord[i].insn_off) {
 9569			verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
 9570			goto err_free;
 9571		}
 9572
 9573		/* check type_id */
 9574		type = btf_type_by_id(btf, krecord[i].type_id);
 9575		if (!type || !btf_type_is_func(type)) {
 9576			verbose(env, "invalid type id %d in func info",
 9577				krecord[i].type_id);
 9578			goto err_free;
 9579		}
 9580		info_aux[i].linkage = BTF_INFO_VLEN(type->info);
 9581
 9582		func_proto = btf_type_by_id(btf, type->type);
 9583		if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
 9584			/* btf_func_check() already verified it during BTF load */
 9585			goto err_free;
 9586		ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
 9587		scalar_return =
 9588			btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
 9589		if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
 9590			verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
 9591			goto err_free;
 9592		}
 9593		if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
 9594			verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
 9595			goto err_free;
 9596		}
 9597
 9598		prev_offset = krecord[i].insn_off;
 9599		bpfptr_add(&urecord, urec_size);
 9600	}
 9601
 9602	prog->aux->func_info = krecord;
 9603	prog->aux->func_info_cnt = nfuncs;
 9604	prog->aux->func_info_aux = info_aux;
 9605	return 0;
 9606
 9607err_free:
 9608	kvfree(krecord);
 9609	kfree(info_aux);
 9610	return ret;
 9611}
 9612
 9613static void adjust_btf_func(struct bpf_verifier_env *env)
 9614{
 9615	struct bpf_prog_aux *aux = env->prog->aux;
 9616	int i;
 9617
 9618	if (!aux->func_info)
 9619		return;
 9620
 9621	for (i = 0; i < env->subprog_cnt; i++)
 9622		aux->func_info[i].insn_off = env->subprog_info[i].start;
 9623}
 9624
 9625#define MIN_BPF_LINEINFO_SIZE	(offsetof(struct bpf_line_info, line_col) + \
 9626		sizeof(((struct bpf_line_info *)(0))->line_col))
 9627#define MAX_LINEINFO_REC_SIZE	MAX_FUNCINFO_REC_SIZE
 9628
 9629static int check_btf_line(struct bpf_verifier_env *env,
 9630			  const union bpf_attr *attr,
 9631			  bpfptr_t uattr)
 9632{
 9633	u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
 9634	struct bpf_subprog_info *sub;
 9635	struct bpf_line_info *linfo;
 9636	struct bpf_prog *prog;
 9637	const struct btf *btf;
 9638	bpfptr_t ulinfo;
 9639	int err;
 9640
 9641	nr_linfo = attr->line_info_cnt;
 9642	if (!nr_linfo)
 9643		return 0;
 9644	if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
 9645		return -EINVAL;
 9646
 9647	rec_size = attr->line_info_rec_size;
 9648	if (rec_size < MIN_BPF_LINEINFO_SIZE ||
 9649	    rec_size > MAX_LINEINFO_REC_SIZE ||
 9650	    rec_size & (sizeof(u32) - 1))
 9651		return -EINVAL;
 9652
 9653	/* Need to zero it in case the userspace may
 9654	 * pass in a smaller bpf_line_info object.
 9655	 */
 9656	linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
 9657			 GFP_KERNEL | __GFP_NOWARN);
 9658	if (!linfo)
 9659		return -ENOMEM;
 9660
 9661	prog = env->prog;
 9662	btf = prog->aux->btf;
 9663
 9664	s = 0;
 9665	sub = env->subprog_info;
 9666	ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
 9667	expected_size = sizeof(struct bpf_line_info);
 9668	ncopy = min_t(u32, expected_size, rec_size);
 9669	for (i = 0; i < nr_linfo; i++) {
 9670		err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
 9671		if (err) {
 9672			if (err == -E2BIG) {
 9673				verbose(env, "nonzero tailing record in line_info");
 9674				if (copy_to_bpfptr_offset(uattr,
 9675							  offsetof(union bpf_attr, line_info_rec_size),
 9676							  &expected_size, sizeof(expected_size)))
 9677					err = -EFAULT;
 9678			}
 9679			goto err_free;
 9680		}
 9681
 9682		if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
 9683			err = -EFAULT;
 9684			goto err_free;
 9685		}
 9686
 9687		/*
 9688		 * Check insn_off to ensure
 9689		 * 1) strictly increasing AND
 9690		 * 2) bounded by prog->len
 9691		 *
 9692		 * The linfo[0].insn_off == 0 check logically falls into
 9693		 * the later "missing bpf_line_info for func..." case
 9694		 * because the first linfo[0].insn_off must be the
 9695		 * first sub also and the first sub must have
 9696		 * subprog_info[0].start == 0.
 9697		 */
 9698		if ((i && linfo[i].insn_off <= prev_offset) ||
 9699		    linfo[i].insn_off >= prog->len) {
 9700			verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
 9701				i, linfo[i].insn_off, prev_offset,
 9702				prog->len);
 9703			err = -EINVAL;
 9704			goto err_free;
 9705		}
 9706
 9707		if (!prog->insnsi[linfo[i].insn_off].code) {
 9708			verbose(env,
 9709				"Invalid insn code at line_info[%u].insn_off\n",
 9710				i);
 9711			err = -EINVAL;
 9712			goto err_free;
 9713		}
 9714
 9715		if (!btf_name_by_offset(btf, linfo[i].line_off) ||
 9716		    !btf_name_by_offset(btf, linfo[i].file_name_off)) {
 9717			verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
 9718			err = -EINVAL;
 9719			goto err_free;
 9720		}
 9721
 9722		if (s != env->subprog_cnt) {
 9723			if (linfo[i].insn_off == sub[s].start) {
 9724				sub[s].linfo_idx = i;
 9725				s++;
 9726			} else if (sub[s].start < linfo[i].insn_off) {
 9727				verbose(env, "missing bpf_line_info for func#%u\n", s);
 9728				err = -EINVAL;
 9729				goto err_free;
 9730			}
 9731		}
 9732
 9733		prev_offset = linfo[i].insn_off;
 9734		bpfptr_add(&ulinfo, rec_size);
 9735	}
 9736
 9737	if (s != env->subprog_cnt) {
 9738		verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
 9739			env->subprog_cnt - s, s);
 9740		err = -EINVAL;
 9741		goto err_free;
 9742	}
 9743
 9744	prog->aux->linfo = linfo;
 9745	prog->aux->nr_linfo = nr_linfo;
 9746
 9747	return 0;
 9748
 9749err_free:
 9750	kvfree(linfo);
 9751	return err;
 9752}
 9753
 9754static int check_btf_info(struct bpf_verifier_env *env,
 9755			  const union bpf_attr *attr,
 9756			  bpfptr_t uattr)
 9757{
 9758	struct btf *btf;
 9759	int err;
 9760
 9761	if (!attr->func_info_cnt && !attr->line_info_cnt) {
 9762		if (check_abnormal_return(env))
 9763			return -EINVAL;
 9764		return 0;
 9765	}
 9766
 9767	btf = btf_get_by_fd(attr->prog_btf_fd);
 9768	if (IS_ERR(btf))
 9769		return PTR_ERR(btf);
 9770	if (btf_is_kernel(btf)) {
 9771		btf_put(btf);
 9772		return -EACCES;
 9773	}
 9774	env->prog->aux->btf = btf;
 9775
 9776	err = check_btf_func(env, attr, uattr);
 9777	if (err)
 9778		return err;
 9779
 9780	err = check_btf_line(env, attr, uattr);
 9781	if (err)
 9782		return err;
 9783
 9784	return 0;
 9785}
 9786
 9787/* check %cur's range satisfies %old's */
 9788static bool range_within(struct bpf_reg_state *old,
 9789			 struct bpf_reg_state *cur)
 9790{
 9791	return old->umin_value <= cur->umin_value &&
 9792	       old->umax_value >= cur->umax_value &&
 9793	       old->smin_value <= cur->smin_value &&
 9794	       old->smax_value >= cur->smax_value &&
 9795	       old->u32_min_value <= cur->u32_min_value &&
 9796	       old->u32_max_value >= cur->u32_max_value &&
 9797	       old->s32_min_value <= cur->s32_min_value &&
 9798	       old->s32_max_value >= cur->s32_max_value;
 9799}
 9800
 9801/* If in the old state two registers had the same id, then they need to have
 9802 * the same id in the new state as well.  But that id could be different from
 9803 * the old state, so we need to track the mapping from old to new ids.
 9804 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
 9805 * regs with old id 5 must also have new id 9 for the new state to be safe.  But
 9806 * regs with a different old id could still have new id 9, we don't care about
 9807 * that.
 9808 * So we look through our idmap to see if this old id has been seen before.  If
 9809 * so, we require the new id to match; otherwise, we add the id pair to the map.
 9810 */
 9811static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
 9812{
 9813	unsigned int i;
 9814
 9815	for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
 9816		if (!idmap[i].old) {
 9817			/* Reached an empty slot; haven't seen this id before */
 9818			idmap[i].old = old_id;
 9819			idmap[i].cur = cur_id;
 9820			return true;
 9821		}
 9822		if (idmap[i].old == old_id)
 9823			return idmap[i].cur == cur_id;
 9824	}
 9825	/* We ran out of idmap slots, which should be impossible */
 9826	WARN_ON_ONCE(1);
 9827	return false;
 9828}
 9829
 9830static void clean_func_state(struct bpf_verifier_env *env,
 9831			     struct bpf_func_state *st)
 9832{
 9833	enum bpf_reg_liveness live;
 9834	int i, j;
 9835
 9836	for (i = 0; i < BPF_REG_FP; i++) {
 9837		live = st->regs[i].live;
 9838		/* liveness must not touch this register anymore */
 9839		st->regs[i].live |= REG_LIVE_DONE;
 9840		if (!(live & REG_LIVE_READ))
 9841			/* since the register is unused, clear its state
 9842			 * to make further comparison simpler
 9843			 */
 9844			__mark_reg_not_init(env, &st->regs[i]);
 9845	}
 9846
 9847	for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
 9848		live = st->stack[i].spilled_ptr.live;
 9849		/* liveness must not touch this stack slot anymore */
 9850		st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
 9851		if (!(live & REG_LIVE_READ)) {
 9852			__mark_reg_not_init(env, &st->stack[i].spilled_ptr);
 9853			for (j = 0; j < BPF_REG_SIZE; j++)
 9854				st->stack[i].slot_type[j] = STACK_INVALID;
 9855		}
 9856	}
 9857}
 9858
 9859static void clean_verifier_state(struct bpf_verifier_env *env,
 9860				 struct bpf_verifier_state *st)
 9861{
 9862	int i;
 9863
 9864	if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
 9865		/* all regs in this state in all frames were already marked */
 9866		return;
 9867
 9868	for (i = 0; i <= st->curframe; i++)
 9869		clean_func_state(env, st->frame[i]);
 9870}
 9871
 9872/* the parentage chains form a tree.
 9873 * the verifier states are added to state lists at given insn and
 9874 * pushed into state stack for future exploration.
 9875 * when the verifier reaches bpf_exit insn some of the verifer states
 9876 * stored in the state lists have their final liveness state already,
 9877 * but a lot of states will get revised from liveness point of view when
 9878 * the verifier explores other branches.
 9879 * Example:
 9880 * 1: r0 = 1
 9881 * 2: if r1 == 100 goto pc+1
 9882 * 3: r0 = 2
 9883 * 4: exit
 9884 * when the verifier reaches exit insn the register r0 in the state list of
 9885 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
 9886 * of insn 2 and goes exploring further. At the insn 4 it will walk the
 9887 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
 9888 *
 9889 * Since the verifier pushes the branch states as it sees them while exploring
 9890 * the program the condition of walking the branch instruction for the second
 9891 * time means that all states below this branch were already explored and
 9892 * their final liveness marks are already propagated.
 9893 * Hence when the verifier completes the search of state list in is_state_visited()
 9894 * we can call this clean_live_states() function to mark all liveness states
 9895 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
 9896 * will not be used.
 9897 * This function also clears the registers and stack for states that !READ
 9898 * to simplify state merging.
 9899 *
 9900 * Important note here that walking the same branch instruction in the callee
 9901 * doesn't meant that the states are DONE. The verifier has to compare
 9902 * the callsites
 9903 */
 9904static void clean_live_states(struct bpf_verifier_env *env, int insn,
 9905			      struct bpf_verifier_state *cur)
 9906{
 9907	struct bpf_verifier_state_list *sl;
 9908	int i;
 9909
 9910	sl = *explored_state(env, insn);
 9911	while (sl) {
 9912		if (sl->state.branches)
 9913			goto next;
 9914		if (sl->state.insn_idx != insn ||
 9915		    sl->state.curframe != cur->curframe)
 9916			goto next;
 9917		for (i = 0; i <= cur->curframe; i++)
 9918			if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
 9919				goto next;
 9920		clean_verifier_state(env, &sl->state);
 9921next:
 9922		sl = sl->next;
 9923	}
 9924}
 9925
 9926/* Returns true if (rold safe implies rcur safe) */
 9927static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
 9928		    struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
 9929{
 9930	bool equal;
 9931
 9932	if (!(rold->live & REG_LIVE_READ))
 9933		/* explored state didn't use this */
 9934		return true;
 9935
 9936	equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
 9937
 9938	if (rold->type == PTR_TO_STACK)
 9939		/* two stack pointers are equal only if they're pointing to
 9940		 * the same stack frame, since fp-8 in foo != fp-8 in bar
 9941		 */
 9942		return equal && rold->frameno == rcur->frameno;
 9943
 9944	if (equal)
 9945		return true;
 9946
 9947	if (rold->type == NOT_INIT)
 9948		/* explored state can't have used this */
 9949		return true;
 9950	if (rcur->type == NOT_INIT)
 9951		return false;
 9952	switch (rold->type) {
 9953	case SCALAR_VALUE:
 9954		if (env->explore_alu_limits)
 9955			return false;
 9956		if (rcur->type == SCALAR_VALUE) {
 9957			if (!rold->precise && !rcur->precise)
 9958				return true;
 9959			/* new val must satisfy old val knowledge */
 9960			return range_within(rold, rcur) &&
 9961			       tnum_in(rold->var_off, rcur->var_off);
 9962		} else {
 9963			/* We're trying to use a pointer in place of a scalar.
 9964			 * Even if the scalar was unbounded, this could lead to
 9965			 * pointer leaks because scalars are allowed to leak
 9966			 * while pointers are not. We could make this safe in
 9967			 * special cases if root is calling us, but it's
 9968			 * probably not worth the hassle.
 9969			 */
 9970			return false;
 9971		}
 9972	case PTR_TO_MAP_KEY:
 9973	case PTR_TO_MAP_VALUE:
 9974		/* If the new min/max/var_off satisfy the old ones and
 9975		 * everything else matches, we are OK.
 9976		 * 'id' is not compared, since it's only used for maps with
 9977		 * bpf_spin_lock inside map element and in such cases if
 9978		 * the rest of the prog is valid for one map element then
 9979		 * it's valid for all map elements regardless of the key
 9980		 * used in bpf_map_lookup()
 9981		 */
 9982		return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
 9983		       range_within(rold, rcur) &&
 9984		       tnum_in(rold->var_off, rcur->var_off);
 9985	case PTR_TO_MAP_VALUE_OR_NULL:
 9986		/* a PTR_TO_MAP_VALUE could be safe to use as a
 9987		 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
 9988		 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
 9989		 * checked, doing so could have affected others with the same
 9990		 * id, and we can't check for that because we lost the id when
 9991		 * we converted to a PTR_TO_MAP_VALUE.
 9992		 */
 9993		if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
 9994			return false;
 9995		if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
 9996			return false;
 9997		/* Check our ids match any regs they're supposed to */
 9998		return check_ids(rold->id, rcur->id, idmap);
 9999	case PTR_TO_PACKET_META:
10000	case PTR_TO_PACKET:
10001		if (rcur->type != rold->type)
10002			return false;
10003		/* We must have at least as much range as the old ptr
10004		 * did, so that any accesses which were safe before are
10005		 * still safe.  This is true even if old range < old off,
10006		 * since someone could have accessed through (ptr - k), or
10007		 * even done ptr -= k in a register, to get a safe access.
10008		 */
10009		if (rold->range > rcur->range)
10010			return false;
10011		/* If the offsets don't match, we can't trust our alignment;
10012		 * nor can we be sure that we won't fall out of range.
10013		 */
10014		if (rold->off != rcur->off)
10015			return false;
10016		/* id relations must be preserved */
10017		if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10018			return false;
10019		/* new val must satisfy old val knowledge */
10020		return range_within(rold, rcur) &&
10021		       tnum_in(rold->var_off, rcur->var_off);
10022	case PTR_TO_CTX:
10023	case CONST_PTR_TO_MAP:
10024	case PTR_TO_PACKET_END:
10025	case PTR_TO_FLOW_KEYS:
10026	case PTR_TO_SOCKET:
10027	case PTR_TO_SOCKET_OR_NULL:
10028	case PTR_TO_SOCK_COMMON:
10029	case PTR_TO_SOCK_COMMON_OR_NULL:
10030	case PTR_TO_TCP_SOCK:
10031	case PTR_TO_TCP_SOCK_OR_NULL:
10032	case PTR_TO_XDP_SOCK:
10033		/* Only valid matches are exact, which memcmp() above
10034		 * would have accepted
10035		 */
10036	default:
10037		/* Don't know what's going on, just say it's not safe */
10038		return false;
10039	}
10040
10041	/* Shouldn't get here; if we do, say it's not safe */
10042	WARN_ON_ONCE(1);
10043	return false;
10044}
10045
10046static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10047		      struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10048{
10049	int i, spi;
10050
10051	/* walk slots of the explored stack and ignore any additional
10052	 * slots in the current stack, since explored(safe) state
10053	 * didn't use them
10054	 */
10055	for (i = 0; i < old->allocated_stack; i++) {
10056		spi = i / BPF_REG_SIZE;
10057
10058		if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10059			i += BPF_REG_SIZE - 1;
10060			/* explored state didn't use this */
10061			continue;
10062		}
10063
10064		if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10065			continue;
10066
10067		/* explored stack has more populated slots than current stack
10068		 * and these slots were used
10069		 */
10070		if (i >= cur->allocated_stack)
10071			return false;
10072
10073		/* if old state was safe with misc data in the stack
10074		 * it will be safe with zero-initialized stack.
10075		 * The opposite is not true
10076		 */
10077		if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10078		    cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10079			continue;
10080		if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10081		    cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10082			/* Ex: old explored (safe) state has STACK_SPILL in
10083			 * this stack slot, but current has STACK_MISC ->
10084			 * this verifier states are not equivalent,
10085			 * return false to continue verification of this path
10086			 */
10087			return false;
10088		if (i % BPF_REG_SIZE)
10089			continue;
10090		if (old->stack[spi].slot_type[0] != STACK_SPILL)
10091			continue;
10092		if (!regsafe(env, &old->stack[spi].spilled_ptr,
10093			     &cur->stack[spi].spilled_ptr, idmap))
10094			/* when explored and current stack slot are both storing
10095			 * spilled registers, check that stored pointers types
10096			 * are the same as well.
10097			 * Ex: explored safe path could have stored
10098			 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10099			 * but current path has stored:
10100			 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10101			 * such verifier states are not equivalent.
10102			 * return false to continue verification of this path
10103			 */
10104			return false;
10105	}
10106	return true;
10107}
10108
10109static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10110{
10111	if (old->acquired_refs != cur->acquired_refs)
10112		return false;
10113	return !memcmp(old->refs, cur->refs,
10114		       sizeof(*old->refs) * old->acquired_refs);
10115}
10116
10117/* compare two verifier states
10118 *
10119 * all states stored in state_list are known to be valid, since
10120 * verifier reached 'bpf_exit' instruction through them
10121 *
10122 * this function is called when verifier exploring different branches of
10123 * execution popped from the state stack. If it sees an old state that has
10124 * more strict register state and more strict stack state then this execution
10125 * branch doesn't need to be explored further, since verifier already
10126 * concluded that more strict state leads to valid finish.
10127 *
10128 * Therefore two states are equivalent if register state is more conservative
10129 * and explored stack state is more conservative than the current one.
10130 * Example:
10131 *       explored                   current
10132 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10133 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10134 *
10135 * In other words if current stack state (one being explored) has more
10136 * valid slots than old one that already passed validation, it means
10137 * the verifier can stop exploring and conclude that current state is valid too
10138 *
10139 * Similarly with registers. If explored state has register type as invalid
10140 * whereas register type in current state is meaningful, it means that
10141 * the current state will reach 'bpf_exit' instruction safely
10142 */
10143static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10144			      struct bpf_func_state *cur)
10145{
10146	int i;
10147
10148	memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10149	for (i = 0; i < MAX_BPF_REG; i++)
10150		if (!regsafe(env, &old->regs[i], &cur->regs[i],
10151			     env->idmap_scratch))
10152			return false;
10153
10154	if (!stacksafe(env, old, cur, env->idmap_scratch))
10155		return false;
10156
10157	if (!refsafe(old, cur))
10158		return false;
10159
10160	return true;
10161}
10162
10163static bool states_equal(struct bpf_verifier_env *env,
10164			 struct bpf_verifier_state *old,
10165			 struct bpf_verifier_state *cur)
10166{
10167	int i;
10168
10169	if (old->curframe != cur->curframe)
10170		return false;
10171
10172	/* Verification state from speculative execution simulation
10173	 * must never prune a non-speculative execution one.
10174	 */
10175	if (old->speculative && !cur->speculative)
10176		return false;
10177
10178	if (old->active_spin_lock != cur->active_spin_lock)
10179		return false;
10180
10181	/* for states to be equal callsites have to be the same
10182	 * and all frame states need to be equivalent
10183	 */
10184	for (i = 0; i <= old->curframe; i++) {
10185		if (old->frame[i]->callsite != cur->frame[i]->callsite)
10186			return false;
10187		if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10188			return false;
10189	}
10190	return true;
10191}
10192
10193/* Return 0 if no propagation happened. Return negative error code if error
10194 * happened. Otherwise, return the propagated bit.
10195 */
10196static int propagate_liveness_reg(struct bpf_verifier_env *env,
10197				  struct bpf_reg_state *reg,
10198				  struct bpf_reg_state *parent_reg)
10199{
10200	u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10201	u8 flag = reg->live & REG_LIVE_READ;
10202	int err;
10203
10204	/* When comes here, read flags of PARENT_REG or REG could be any of
10205	 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10206	 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10207	 */
10208	if (parent_flag == REG_LIVE_READ64 ||
10209	    /* Or if there is no read flag from REG. */
10210	    !flag ||
10211	    /* Or if the read flag from REG is the same as PARENT_REG. */
10212	    parent_flag == flag)
10213		return 0;
10214
10215	err = mark_reg_read(env, reg, parent_reg, flag);
10216	if (err)
10217		return err;
10218
10219	return flag;
10220}
10221
10222/* A write screens off any subsequent reads; but write marks come from the
10223 * straight-line code between a state and its parent.  When we arrive at an
10224 * equivalent state (jump target or such) we didn't arrive by the straight-line
10225 * code, so read marks in the state must propagate to the parent regardless
10226 * of the state's write marks. That's what 'parent == state->parent' comparison
10227 * in mark_reg_read() is for.
10228 */
10229static int propagate_liveness(struct bpf_verifier_env *env,
10230			      const struct bpf_verifier_state *vstate,
10231			      struct bpf_verifier_state *vparent)
10232{
10233	struct bpf_reg_state *state_reg, *parent_reg;
10234	struct bpf_func_state *state, *parent;
10235	int i, frame, err = 0;
10236
10237	if (vparent->curframe != vstate->curframe) {
10238		WARN(1, "propagate_live: parent frame %d current frame %d\n",
10239		     vparent->curframe, vstate->curframe);
10240		return -EFAULT;
10241	}
10242	/* Propagate read liveness of registers... */
10243	BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10244	for (frame = 0; frame <= vstate->curframe; frame++) {
10245		parent = vparent->frame[frame];
10246		state = vstate->frame[frame];
10247		parent_reg = parent->regs;
10248		state_reg = state->regs;
10249		/* We don't need to worry about FP liveness, it's read-only */
10250		for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10251			err = propagate_liveness_reg(env, &state_reg[i],
10252						     &parent_reg[i]);
10253			if (err < 0)
10254				return err;
10255			if (err == REG_LIVE_READ64)
10256				mark_insn_zext(env, &parent_reg[i]);
10257		}
10258
10259		/* Propagate stack slots. */
10260		for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10261			    i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10262			parent_reg = &parent->stack[i].spilled_ptr;
10263			state_reg = &state->stack[i].spilled_ptr;
10264			err = propagate_liveness_reg(env, state_reg,
10265						     parent_reg);
10266			if (err < 0)
10267				return err;
10268		}
10269	}
10270	return 0;
10271}
10272
10273/* find precise scalars in the previous equivalent state and
10274 * propagate them into the current state
10275 */
10276static int propagate_precision(struct bpf_verifier_env *env,
10277			       const struct bpf_verifier_state *old)
10278{
10279	struct bpf_reg_state *state_reg;
10280	struct bpf_func_state *state;
10281	int i, err = 0;
10282
10283	state = old->frame[old->curframe];
10284	state_reg = state->regs;
10285	for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10286		if (state_reg->type != SCALAR_VALUE ||
10287		    !state_reg->precise)
10288			continue;
10289		if (env->log.level & BPF_LOG_LEVEL2)
10290			verbose(env, "propagating r%d\n", i);
10291		err = mark_chain_precision(env, i);
10292		if (err < 0)
10293			return err;
10294	}
10295
10296	for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10297		if (state->stack[i].slot_type[0] != STACK_SPILL)
10298			continue;
10299		state_reg = &state->stack[i].spilled_ptr;
10300		if (state_reg->type != SCALAR_VALUE ||
10301		    !state_reg->precise)
10302			continue;
10303		if (env->log.level & BPF_LOG_LEVEL2)
10304			verbose(env, "propagating fp%d\n",
10305				(-i - 1) * BPF_REG_SIZE);
10306		err = mark_chain_precision_stack(env, i);
10307		if (err < 0)
10308			return err;
10309	}
10310	return 0;
10311}
10312
10313static bool states_maybe_looping(struct bpf_verifier_state *old,
10314				 struct bpf_verifier_state *cur)
10315{
10316	struct bpf_func_state *fold, *fcur;
10317	int i, fr = cur->curframe;
10318
10319	if (old->curframe != fr)
10320		return false;
10321
10322	fold = old->frame[fr];
10323	fcur = cur->frame[fr];
10324	for (i = 0; i < MAX_BPF_REG; i++)
10325		if (memcmp(&fold->regs[i], &fcur->regs[i],
10326			   offsetof(struct bpf_reg_state, parent)))
10327			return false;
10328	return true;
10329}
10330
10331
10332static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10333{
10334	struct bpf_verifier_state_list *new_sl;
10335	struct bpf_verifier_state_list *sl, **pprev;
10336	struct bpf_verifier_state *cur = env->cur_state, *new;
10337	int i, j, err, states_cnt = 0;
10338	bool add_new_state = env->test_state_freq ? true : false;
10339
10340	cur->last_insn_idx = env->prev_insn_idx;
10341	if (!env->insn_aux_data[insn_idx].prune_point)
10342		/* this 'insn_idx' instruction wasn't marked, so we will not
10343		 * be doing state search here
10344		 */
10345		return 0;
10346
10347	/* bpf progs typically have pruning point every 4 instructions
10348	 * http://vger.kernel.org/bpfconf2019.html#session-1
10349	 * Do not add new state for future pruning if the verifier hasn't seen
10350	 * at least 2 jumps and at least 8 instructions.
10351	 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10352	 * In tests that amounts to up to 50% reduction into total verifier
10353	 * memory consumption and 20% verifier time speedup.
10354	 */
10355	if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10356	    env->insn_processed - env->prev_insn_processed >= 8)
10357		add_new_state = true;
10358
10359	pprev = explored_state(env, insn_idx);
10360	sl = *pprev;
10361
10362	clean_live_states(env, insn_idx, cur);
10363
10364	while (sl) {
10365		states_cnt++;
10366		if (sl->state.insn_idx != insn_idx)
10367			goto next;
10368		if (sl->state.branches) {
10369			if (states_maybe_looping(&sl->state, cur) &&
10370			    states_equal(env, &sl->state, cur)) {
10371				verbose_linfo(env, insn_idx, "; ");
10372				verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10373				return -EINVAL;
10374			}
10375			/* if the verifier is processing a loop, avoid adding new state
10376			 * too often, since different loop iterations have distinct
10377			 * states and may not help future pruning.
10378			 * This threshold shouldn't be too low to make sure that
10379			 * a loop with large bound will be rejected quickly.
10380			 * The most abusive loop will be:
10381			 * r1 += 1
10382			 * if r1 < 1000000 goto pc-2
10383			 * 1M insn_procssed limit / 100 == 10k peak states.
10384			 * This threshold shouldn't be too high either, since states
10385			 * at the end of the loop are likely to be useful in pruning.
10386			 */
10387			if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10388			    env->insn_processed - env->prev_insn_processed < 100)
10389				add_new_state = false;
10390			goto miss;
10391		}
10392		if (states_equal(env, &sl->state, cur)) {
10393			sl->hit_cnt++;
10394			/* reached equivalent register/stack state,
10395			 * prune the search.
10396			 * Registers read by the continuation are read by us.
10397			 * If we have any write marks in env->cur_state, they
10398			 * will prevent corresponding reads in the continuation
10399			 * from reaching our parent (an explored_state).  Our
10400			 * own state will get the read marks recorded, but
10401			 * they'll be immediately forgotten as we're pruning
10402			 * this state and will pop a new one.
10403			 */
10404			err = propagate_liveness(env, &sl->state, cur);
10405
10406			/* if previous state reached the exit with precision and
10407			 * current state is equivalent to it (except precsion marks)
10408			 * the precision needs to be propagated back in
10409			 * the current state.
10410			 */
10411			err = err ? : push_jmp_history(env, cur);
10412			err = err ? : propagate_precision(env, &sl->state);
10413			if (err)
10414				return err;
10415			return 1;
10416		}
10417miss:
10418		/* when new state is not going to be added do not increase miss count.
10419		 * Otherwise several loop iterations will remove the state
10420		 * recorded earlier. The goal of these heuristics is to have
10421		 * states from some iterations of the loop (some in the beginning
10422		 * and some at the end) to help pruning.
10423		 */
10424		if (add_new_state)
10425			sl->miss_cnt++;
10426		/* heuristic to determine whether this state is beneficial
10427		 * to keep checking from state equivalence point of view.
10428		 * Higher numbers increase max_states_per_insn and verification time,
10429		 * but do not meaningfully decrease insn_processed.
10430		 */
10431		if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10432			/* the state is unlikely to be useful. Remove it to
10433			 * speed up verification
10434			 */
10435			*pprev = sl->next;
10436			if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10437				u32 br = sl->state.branches;
10438
10439				WARN_ONCE(br,
10440					  "BUG live_done but branches_to_explore %d\n",
10441					  br);
10442				free_verifier_state(&sl->state, false);
10443				kfree(sl);
10444				env->peak_states--;
10445			} else {
10446				/* cannot free this state, since parentage chain may
10447				 * walk it later. Add it for free_list instead to
10448				 * be freed at the end of verification
10449				 */
10450				sl->next = env->free_list;
10451				env->free_list = sl;
10452			}
10453			sl = *pprev;
10454			continue;
10455		}
10456next:
10457		pprev = &sl->next;
10458		sl = *pprev;
10459	}
10460
10461	if (env->max_states_per_insn < states_cnt)
10462		env->max_states_per_insn = states_cnt;
10463
10464	if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10465		return push_jmp_history(env, cur);
10466
10467	if (!add_new_state)
10468		return push_jmp_history(env, cur);
10469
10470	/* There were no equivalent states, remember the current one.
10471	 * Technically the current state is not proven to be safe yet,
10472	 * but it will either reach outer most bpf_exit (which means it's safe)
10473	 * or it will be rejected. When there are no loops the verifier won't be
10474	 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10475	 * again on the way to bpf_exit.
10476	 * When looping the sl->state.branches will be > 0 and this state
10477	 * will not be considered for equivalence until branches == 0.
10478	 */
10479	new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10480	if (!new_sl)
10481		return -ENOMEM;
10482	env->total_states++;
10483	env->peak_states++;
10484	env->prev_jmps_processed = env->jmps_processed;
10485	env->prev_insn_processed = env->insn_processed;
10486
10487	/* add new state to the head of linked list */
10488	new = &new_sl->state;
10489	err = copy_verifier_state(new, cur);
10490	if (err) {
10491		free_verifier_state(new, false);
10492		kfree(new_sl);
10493		return err;
10494	}
10495	new->insn_idx = insn_idx;
10496	WARN_ONCE(new->branches != 1,
10497		  "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10498
10499	cur->parent = new;
10500	cur->first_insn_idx = insn_idx;
10501	clear_jmp_history(cur);
10502	new_sl->next = *explored_state(env, insn_idx);
10503	*explored_state(env, insn_idx) = new_sl;
10504	/* connect new state to parentage chain. Current frame needs all
10505	 * registers connected. Only r6 - r9 of the callers are alive (pushed
10506	 * to the stack implicitly by JITs) so in callers' frames connect just
10507	 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10508	 * the state of the call instruction (with WRITTEN set), and r0 comes
10509	 * from callee with its full parentage chain, anyway.
10510	 */
10511	/* clear write marks in current state: the writes we did are not writes
10512	 * our child did, so they don't screen off its reads from us.
10513	 * (There are no read marks in current state, because reads always mark
10514	 * their parent and current state never has children yet.  Only
10515	 * explored_states can get read marks.)
10516	 */
10517	for (j = 0; j <= cur->curframe; j++) {
10518		for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10519			cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10520		for (i = 0; i < BPF_REG_FP; i++)
10521			cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10522	}
10523
10524	/* all stack frames are accessible from callee, clear them all */
10525	for (j = 0; j <= cur->curframe; j++) {
10526		struct bpf_func_state *frame = cur->frame[j];
10527		struct bpf_func_state *newframe = new->frame[j];
10528
10529		for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10530			frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10531			frame->stack[i].spilled_ptr.parent =
10532						&newframe->stack[i].spilled_ptr;
10533		}
10534	}
10535	return 0;
10536}
10537
10538/* Return true if it's OK to have the same insn return a different type. */
10539static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10540{
10541	switch (type) {
10542	case PTR_TO_CTX:
10543	case PTR_TO_SOCKET:
10544	case PTR_TO_SOCKET_OR_NULL:
10545	case PTR_TO_SOCK_COMMON:
10546	case PTR_TO_SOCK_COMMON_OR_NULL:
10547	case PTR_TO_TCP_SOCK:
10548	case PTR_TO_TCP_SOCK_OR_NULL:
10549	case PTR_TO_XDP_SOCK:
10550	case PTR_TO_BTF_ID:
10551	case PTR_TO_BTF_ID_OR_NULL:
10552		return false;
10553	default:
10554		return true;
10555	}
10556}
10557
10558/* If an instruction was previously used with particular pointer types, then we
10559 * need to be careful to avoid cases such as the below, where it may be ok
10560 * for one branch accessing the pointer, but not ok for the other branch:
10561 *
10562 * R1 = sock_ptr
10563 * goto X;
10564 * ...
10565 * R1 = some_other_valid_ptr;
10566 * goto X;
10567 * ...
10568 * R2 = *(u32 *)(R1 + 0);
10569 */
10570static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10571{
10572	return src != prev && (!reg_type_mismatch_ok(src) ||
10573			       !reg_type_mismatch_ok(prev));
10574}
10575
10576static int do_check(struct bpf_verifier_env *env)
10577{
10578	bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10579	struct bpf_verifier_state *state = env->cur_state;
10580	struct bpf_insn *insns = env->prog->insnsi;
10581	struct bpf_reg_state *regs;
10582	int insn_cnt = env->prog->len;
10583	bool do_print_state = false;
10584	int prev_insn_idx = -1;
10585
10586	for (;;) {
10587		struct bpf_insn *insn;
10588		u8 class;
10589		int err;
10590
10591		env->prev_insn_idx = prev_insn_idx;
10592		if (env->insn_idx >= insn_cnt) {
10593			verbose(env, "invalid insn idx %d insn_cnt %d\n",
10594				env->insn_idx, insn_cnt);
10595			return -EFAULT;
10596		}
10597
10598		insn = &insns[env->insn_idx];
10599		class = BPF_CLASS(insn->code);
10600
10601		if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10602			verbose(env,
10603				"BPF program is too large. Processed %d insn\n",
10604				env->insn_processed);
10605			return -E2BIG;
10606		}
10607
10608		err = is_state_visited(env, env->insn_idx);
10609		if (err < 0)
10610			return err;
10611		if (err == 1) {
10612			/* found equivalent state, can prune the search */
10613			if (env->log.level & BPF_LOG_LEVEL) {
10614				if (do_print_state)
10615					verbose(env, "\nfrom %d to %d%s: safe\n",
10616						env->prev_insn_idx, env->insn_idx,
10617						env->cur_state->speculative ?
10618						" (speculative execution)" : "");
10619				else
10620					verbose(env, "%d: safe\n", env->insn_idx);
10621			}
10622			goto process_bpf_exit;
10623		}
10624
10625		if (signal_pending(current))
10626			return -EAGAIN;
10627
10628		if (need_resched())
10629			cond_resched();
10630
10631		if (env->log.level & BPF_LOG_LEVEL2 ||
10632		    (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10633			if (env->log.level & BPF_LOG_LEVEL2)
10634				verbose(env, "%d:", env->insn_idx);
10635			else
10636				verbose(env, "\nfrom %d to %d%s:",
10637					env->prev_insn_idx, env->insn_idx,
10638					env->cur_state->speculative ?
10639					" (speculative execution)" : "");
10640			print_verifier_state(env, state->frame[state->curframe]);
10641			do_print_state = false;
10642		}
10643
10644		if (env->log.level & BPF_LOG_LEVEL) {
10645			const struct bpf_insn_cbs cbs = {
10646				.cb_call	= disasm_kfunc_name,
10647				.cb_print	= verbose,
10648				.private_data	= env,
10649			};
10650
10651			verbose_linfo(env, env->insn_idx, "; ");
10652			verbose(env, "%d: ", env->insn_idx);
10653			print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10654		}
10655
10656		if (bpf_prog_is_dev_bound(env->prog->aux)) {
10657			err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10658							   env->prev_insn_idx);
10659			if (err)
10660				return err;
10661		}
10662
10663		regs = cur_regs(env);
10664		sanitize_mark_insn_seen(env);
10665		prev_insn_idx = env->insn_idx;
10666
10667		if (class == BPF_ALU || class == BPF_ALU64) {
10668			err = check_alu_op(env, insn);
10669			if (err)
10670				return err;
10671
10672		} else if (class == BPF_LDX) {
10673			enum bpf_reg_type *prev_src_type, src_reg_type;
10674
10675			/* check for reserved fields is already done */
10676
10677			/* check src operand */
10678			err = check_reg_arg(env, insn->src_reg, SRC_OP);
10679			if (err)
10680				return err;
10681
10682			err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10683			if (err)
10684				return err;
10685
10686			src_reg_type = regs[insn->src_reg].type;
10687
10688			/* check that memory (src_reg + off) is readable,
10689			 * the state of dst_reg will be updated by this func
10690			 */
10691			err = check_mem_access(env, env->insn_idx, insn->src_reg,
10692					       insn->off, BPF_SIZE(insn->code),
10693					       BPF_READ, insn->dst_reg, false);
10694			if (err)
10695				return err;
10696
10697			prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10698
10699			if (*prev_src_type == NOT_INIT) {
10700				/* saw a valid insn
10701				 * dst_reg = *(u32 *)(src_reg + off)
10702				 * save type to validate intersecting paths
10703				 */
10704				*prev_src_type = src_reg_type;
10705
10706			} else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10707				/* ABuser program is trying to use the same insn
10708				 * dst_reg = *(u32*) (src_reg + off)
10709				 * with different pointer types:
10710				 * src_reg == ctx in one branch and
10711				 * src_reg == stack|map in some other branch.
10712				 * Reject it.
10713				 */
10714				verbose(env, "same insn cannot be used with different pointers\n");
10715				return -EINVAL;
10716			}
10717
10718		} else if (class == BPF_STX) {
10719			enum bpf_reg_type *prev_dst_type, dst_reg_type;
10720
10721			if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10722				err = check_atomic(env, env->insn_idx, insn);
10723				if (err)
10724					return err;
10725				env->insn_idx++;
10726				continue;
10727			}
10728
10729			if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10730				verbose(env, "BPF_STX uses reserved fields\n");
10731				return -EINVAL;
10732			}
10733
10734			/* check src1 operand */
10735			err = check_reg_arg(env, insn->src_reg, SRC_OP);
10736			if (err)
10737				return err;
10738			/* check src2 operand */
10739			err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10740			if (err)
10741				return err;
10742
10743			dst_reg_type = regs[insn->dst_reg].type;
10744
10745			/* check that memory (dst_reg + off) is writeable */
10746			err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10747					       insn->off, BPF_SIZE(insn->code),
10748					       BPF_WRITE, insn->src_reg, false);
10749			if (err)
10750				return err;
10751
10752			prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10753
10754			if (*prev_dst_type == NOT_INIT) {
10755				*prev_dst_type = dst_reg_type;
10756			} else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
10757				verbose(env, "same insn cannot be used with different pointers\n");
10758				return -EINVAL;
10759			}
10760
10761		} else if (class == BPF_ST) {
10762			if (BPF_MODE(insn->code) != BPF_MEM ||
10763			    insn->src_reg != BPF_REG_0) {
10764				verbose(env, "BPF_ST uses reserved fields\n");
10765				return -EINVAL;
10766			}
10767			/* check src operand */
10768			err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10769			if (err)
10770				return err;
10771
10772			if (is_ctx_reg(env, insn->dst_reg)) {
10773				verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
10774					insn->dst_reg,
10775					reg_type_str[reg_state(env, insn->dst_reg)->type]);
10776				return -EACCES;
10777			}
10778
10779			/* check that memory (dst_reg + off) is writeable */
10780			err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10781					       insn->off, BPF_SIZE(insn->code),
10782					       BPF_WRITE, -1, false);
10783			if (err)
10784				return err;
10785
10786		} else if (class == BPF_JMP || class == BPF_JMP32) {
10787			u8 opcode = BPF_OP(insn->code);
10788
10789			env->jmps_processed++;
10790			if (opcode == BPF_CALL) {
10791				if (BPF_SRC(insn->code) != BPF_K ||
10792				    insn->off != 0 ||
10793				    (insn->src_reg != BPF_REG_0 &&
10794				     insn->src_reg != BPF_PSEUDO_CALL &&
10795				     insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
10796				    insn->dst_reg != BPF_REG_0 ||
10797				    class == BPF_JMP32) {
10798					verbose(env, "BPF_CALL uses reserved fields\n");
10799					return -EINVAL;
10800				}
10801
10802				if (env->cur_state->active_spin_lock &&
10803				    (insn->src_reg == BPF_PSEUDO_CALL ||
10804				     insn->imm != BPF_FUNC_spin_unlock)) {
10805					verbose(env, "function calls are not allowed while holding a lock\n");
10806					return -EINVAL;
10807				}
10808				if (insn->src_reg == BPF_PSEUDO_CALL)
10809					err = check_func_call(env, insn, &env->insn_idx);
10810				else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
10811					err = check_kfunc_call(env, insn);
10812				else
10813					err = check_helper_call(env, insn, &env->insn_idx);
10814				if (err)
10815					return err;
10816			} else if (opcode == BPF_JA) {
10817				if (BPF_SRC(insn->code) != BPF_K ||
10818				    insn->imm != 0 ||
10819				    insn->src_reg != BPF_REG_0 ||
10820				    insn->dst_reg != BPF_REG_0 ||
10821				    class == BPF_JMP32) {
10822					verbose(env, "BPF_JA uses reserved fields\n");
10823					return -EINVAL;
10824				}
10825
10826				env->insn_idx += insn->off + 1;
10827				continue;
10828
10829			} else if (opcode == BPF_EXIT) {
10830				if (BPF_SRC(insn->code) != BPF_K ||
10831				    insn->imm != 0 ||
10832				    insn->src_reg != BPF_REG_0 ||
10833				    insn->dst_reg != BPF_REG_0 ||
10834				    class == BPF_JMP32) {
10835					verbose(env, "BPF_EXIT uses reserved fields\n");
10836					return -EINVAL;
10837				}
10838
10839				if (env->cur_state->active_spin_lock) {
10840					verbose(env, "bpf_spin_unlock is missing\n");
10841					return -EINVAL;
10842				}
10843
10844				if (state->curframe) {
10845					/* exit from nested function */
10846					err = prepare_func_exit(env, &env->insn_idx);
10847					if (err)
10848						return err;
10849					do_print_state = true;
10850					continue;
10851				}
10852
10853				err = check_reference_leak(env);
10854				if (err)
10855					return err;
10856
10857				err = check_return_code(env);
10858				if (err)
10859					return err;
10860process_bpf_exit:
10861				update_branch_counts(env, env->cur_state);
10862				err = pop_stack(env, &prev_insn_idx,
10863						&env->insn_idx, pop_log);
10864				if (err < 0) {
10865					if (err != -ENOENT)
10866						return err;
10867					break;
10868				} else {
10869					do_print_state = true;
10870					continue;
10871				}
10872			} else {
10873				err = check_cond_jmp_op(env, insn, &env->insn_idx);
10874				if (err)
10875					return err;
10876			}
10877		} else if (class == BPF_LD) {
10878			u8 mode = BPF_MODE(insn->code);
10879
10880			if (mode == BPF_ABS || mode == BPF_IND) {
10881				err = check_ld_abs(env, insn);
10882				if (err)
10883					return err;
10884
10885			} else if (mode == BPF_IMM) {
10886				err = check_ld_imm(env, insn);
10887				if (err)
10888					return err;
10889
10890				env->insn_idx++;
10891				sanitize_mark_insn_seen(env);
10892			} else {
10893				verbose(env, "invalid BPF_LD mode\n");
10894				return -EINVAL;
10895			}
10896		} else {
10897			verbose(env, "unknown insn class %d\n", class);
10898			return -EINVAL;
10899		}
10900
10901		env->insn_idx++;
10902	}
10903
10904	return 0;
10905}
10906
10907static int find_btf_percpu_datasec(struct btf *btf)
10908{
10909	const struct btf_type *t;
10910	const char *tname;
10911	int i, n;
10912
10913	/*
10914	 * Both vmlinux and module each have their own ".data..percpu"
10915	 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
10916	 * types to look at only module's own BTF types.
10917	 */
10918	n = btf_nr_types(btf);
10919	if (btf_is_module(btf))
10920		i = btf_nr_types(btf_vmlinux);
10921	else
10922		i = 1;
10923
10924	for(; i < n; i++) {
10925		t = btf_type_by_id(btf, i);
10926		if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
10927			continue;
10928
10929		tname = btf_name_by_offset(btf, t->name_off);
10930		if (!strcmp(tname, ".data..percpu"))
10931			return i;
10932	}
10933
10934	return -ENOENT;
10935}
10936
10937/* replace pseudo btf_id with kernel symbol address */
10938static int check_pseudo_btf_id(struct bpf_verifier_env *env,
10939			       struct bpf_insn *insn,
10940			       struct bpf_insn_aux_data *aux)
10941{
10942	const struct btf_var_secinfo *vsi;
10943	const struct btf_type *datasec;
10944	struct btf_mod_pair *btf_mod;
10945	const struct btf_type *t;
10946	const char *sym_name;
10947	bool percpu = false;
10948	u32 type, id = insn->imm;
10949	struct btf *btf;
10950	s32 datasec_id;
10951	u64 addr;
10952	int i, btf_fd, err;
10953
10954	btf_fd = insn[1].imm;
10955	if (btf_fd) {
10956		btf = btf_get_by_fd(btf_fd);
10957		if (IS_ERR(btf)) {
10958			verbose(env, "invalid module BTF object FD specified.\n");
10959			return -EINVAL;
10960		}
10961	} else {
10962		if (!btf_vmlinux) {
10963			verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
10964			return -EINVAL;
10965		}
10966		btf = btf_vmlinux;
10967		btf_get(btf);
10968	}
10969
10970	t = btf_type_by_id(btf, id);
10971	if (!t) {
10972		verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
10973		err = -ENOENT;
10974		goto err_put;
10975	}
10976
10977	if (!btf_type_is_var(t)) {
10978		verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
10979		err = -EINVAL;
10980		goto err_put;
10981	}
10982
10983	sym_name = btf_name_by_offset(btf, t->name_off);
10984	addr = kallsyms_lookup_name(sym_name);
10985	if (!addr) {
10986		verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
10987			sym_name);
10988		err = -ENOENT;
10989		goto err_put;
10990	}
10991
10992	datasec_id = find_btf_percpu_datasec(btf);
10993	if (datasec_id > 0) {
10994		datasec = btf_type_by_id(btf, datasec_id);
10995		for_each_vsi(i, datasec, vsi) {
10996			if (vsi->type == id) {
10997				percpu = true;
10998				break;
10999			}
11000		}
11001	}
11002
11003	insn[0].imm = (u32)addr;
11004	insn[1].imm = addr >> 32;
11005
11006	type = t->type;
11007	t = btf_type_skip_modifiers(btf, type, NULL);
11008	if (percpu) {
11009		aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11010		aux->btf_var.btf = btf;
11011		aux->btf_var.btf_id = type;
11012	} else if (!btf_type_is_struct(t)) {
11013		const struct btf_type *ret;
11014		const char *tname;
11015		u32 tsize;
11016
11017		/* resolve the type size of ksym. */
11018		ret = btf_resolve_size(btf, t, &tsize);
11019		if (IS_ERR(ret)) {
11020			tname = btf_name_by_offset(btf, t->name_off);
11021			verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11022				tname, PTR_ERR(ret));
11023			err = -EINVAL;
11024			goto err_put;
11025		}
11026		aux->btf_var.reg_type = PTR_TO_MEM;
11027		aux->btf_var.mem_size = tsize;
11028	} else {
11029		aux->btf_var.reg_type = PTR_TO_BTF_ID;
11030		aux->btf_var.btf = btf;
11031		aux->btf_var.btf_id = type;
11032	}
11033
11034	/* check whether we recorded this BTF (and maybe module) already */
11035	for (i = 0; i < env->used_btf_cnt; i++) {
11036		if (env->used_btfs[i].btf == btf) {
11037			btf_put(btf);
11038			return 0;
11039		}
11040	}
11041
11042	if (env->used_btf_cnt >= MAX_USED_BTFS) {
11043		err = -E2BIG;
11044		goto err_put;
11045	}
11046
11047	btf_mod = &env->used_btfs[env->used_btf_cnt];
11048	btf_mod->btf = btf;
11049	btf_mod->module = NULL;
11050
11051	/* if we reference variables from kernel module, bump its refcount */
11052	if (btf_is_module(btf)) {
11053		btf_mod->module = btf_try_get_module(btf);
11054		if (!btf_mod->module) {
11055			err = -ENXIO;
11056			goto err_put;
11057		}
11058	}
11059
11060	env->used_btf_cnt++;
11061
11062	return 0;
11063err_put:
11064	btf_put(btf);
11065	return err;
11066}
11067
11068static int check_map_prealloc(struct bpf_map *map)
11069{
11070	return (map->map_type != BPF_MAP_TYPE_HASH &&
11071		map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11072		map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11073		!(map->map_flags & BPF_F_NO_PREALLOC);
11074}
11075
11076static bool is_tracing_prog_type(enum bpf_prog_type type)
11077{
11078	switch (type) {
11079	case BPF_PROG_TYPE_KPROBE:
11080	case BPF_PROG_TYPE_TRACEPOINT:
11081	case BPF_PROG_TYPE_PERF_EVENT:
11082	case BPF_PROG_TYPE_RAW_TRACEPOINT:
11083		return true;
11084	default:
11085		return false;
11086	}
11087}
11088
11089static bool is_preallocated_map(struct bpf_map *map)
11090{
11091	if (!check_map_prealloc(map))
11092		return false;
11093	if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11094		return false;
11095	return true;
11096}
11097
11098static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11099					struct bpf_map *map,
11100					struct bpf_prog *prog)
11101
11102{
11103	enum bpf_prog_type prog_type = resolve_prog_type(prog);
11104	/*
11105	 * Validate that trace type programs use preallocated hash maps.
11106	 *
11107	 * For programs attached to PERF events this is mandatory as the
11108	 * perf NMI can hit any arbitrary code sequence.
11109	 *
11110	 * All other trace types using preallocated hash maps are unsafe as
11111	 * well because tracepoint or kprobes can be inside locked regions
11112	 * of the memory allocator or at a place where a recursion into the
11113	 * memory allocator would see inconsistent state.
11114	 *
11115	 * On RT enabled kernels run-time allocation of all trace type
11116	 * programs is strictly prohibited due to lock type constraints. On
11117	 * !RT kernels it is allowed for backwards compatibility reasons for
11118	 * now, but warnings are emitted so developers are made aware of
11119	 * the unsafety and can fix their programs before this is enforced.
11120	 */
11121	if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11122		if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11123			verbose(env, "perf_event programs can only use preallocated hash map\n");
11124			return -EINVAL;
11125		}
11126		if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11127			verbose(env, "trace type programs can only use preallocated hash map\n");
11128			return -EINVAL;
11129		}
11130		WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11131		verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11132	}
11133
11134	if (map_value_has_spin_lock(map)) {
11135		if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11136			verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11137			return -EINVAL;
11138		}
11139
11140		if (is_tracing_prog_type(prog_type)) {
11141			verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11142			return -EINVAL;
11143		}
11144
11145		if (prog->aux->sleepable) {
11146			verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11147			return -EINVAL;
11148		}
11149	}
11150
11151	if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11152	    !bpf_offload_prog_map_match(prog, map)) {
11153		verbose(env, "offload device mismatch between prog and map\n");
11154		return -EINVAL;
11155	}
11156
11157	if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11158		verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11159		return -EINVAL;
11160	}
11161
11162	if (prog->aux->sleepable)
11163		switch (map->map_type) {
11164		case BPF_MAP_TYPE_HASH:
11165		case BPF_MAP_TYPE_LRU_HASH:
11166		case BPF_MAP_TYPE_ARRAY:
11167		case BPF_MAP_TYPE_PERCPU_HASH:
11168		case BPF_MAP_TYPE_PERCPU_ARRAY:
11169		case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11170		case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11171		case BPF_MAP_TYPE_HASH_OF_MAPS:
11172			if (!is_preallocated_map(map)) {
11173				verbose(env,
11174					"Sleepable programs can only use preallocated maps\n");
11175				return -EINVAL;
11176			}
11177			break;
11178		case BPF_MAP_TYPE_RINGBUF:
11179			break;
11180		default:
11181			verbose(env,
11182				"Sleepable programs can only use array, hash, and ringbuf maps\n");
11183			return -EINVAL;
11184		}
11185
11186	return 0;
11187}
11188
11189static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11190{
11191	return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11192		map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11193}
11194
11195/* find and rewrite pseudo imm in ld_imm64 instructions:
11196 *
11197 * 1. if it accesses map FD, replace it with actual map pointer.
11198 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11199 *
11200 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11201 */
11202static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11203{
11204	struct bpf_insn *insn = env->prog->insnsi;
11205	int insn_cnt = env->prog->len;
11206	int i, j, err;
11207
11208	err = bpf_prog_calc_tag(env->prog);
11209	if (err)
11210		return err;
11211
11212	for (i = 0; i < insn_cnt; i++, insn++) {
11213		if (BPF_CLASS(insn->code) == BPF_LDX &&
11214		    (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11215			verbose(env, "BPF_LDX uses reserved fields\n");
11216			return -EINVAL;
11217		}
11218
11219		if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11220			struct bpf_insn_aux_data *aux;
11221			struct bpf_map *map;
11222			struct fd f;
11223			u64 addr;
11224			u32 fd;
11225
11226			if (i == insn_cnt - 1 || insn[1].code != 0 ||
11227			    insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11228			    insn[1].off != 0) {
11229				verbose(env, "invalid bpf_ld_imm64 insn\n");
11230				return -EINVAL;
11231			}
11232
11233			if (insn[0].src_reg == 0)
11234				/* valid generic load 64-bit imm */
11235				goto next_insn;
11236
11237			if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11238				aux = &env->insn_aux_data[i];
11239				err = check_pseudo_btf_id(env, insn, aux);
11240				if (err)
11241					return err;
11242				goto next_insn;
11243			}
11244
11245			if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11246				aux = &env->insn_aux_data[i];
11247				aux->ptr_type = PTR_TO_FUNC;
11248				goto next_insn;
11249			}
11250
11251			/* In final convert_pseudo_ld_imm64() step, this is
11252			 * converted into regular 64-bit imm load insn.
11253			 */
11254			switch (insn[0].src_reg) {
11255			case BPF_PSEUDO_MAP_VALUE:
11256			case BPF_PSEUDO_MAP_IDX_VALUE:
11257				break;
11258			case BPF_PSEUDO_MAP_FD:
11259			case BPF_PSEUDO_MAP_IDX:
11260				if (insn[1].imm == 0)
11261					break;
11262				fallthrough;
11263			default:
11264				verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11265				return -EINVAL;
11266			}
11267
11268			switch (insn[0].src_reg) {
11269			case BPF_PSEUDO_MAP_IDX_VALUE:
11270			case BPF_PSEUDO_MAP_IDX:
11271				if (bpfptr_is_null(env->fd_array)) {
11272					verbose(env, "fd_idx without fd_array is invalid\n");
11273					return -EPROTO;
11274				}
11275				if (copy_from_bpfptr_offset(&fd, env->fd_array,
11276							    insn[0].imm * sizeof(fd),
11277							    sizeof(fd)))
11278					return -EFAULT;
11279				break;
11280			default:
11281				fd = insn[0].imm;
11282				break;
11283			}
11284
11285			f = fdget(fd);
11286			map = __bpf_map_get(f);
11287			if (IS_ERR(map)) {
11288				verbose(env, "fd %d is not pointing to valid bpf_map\n",
11289					insn[0].imm);
11290				return PTR_ERR(map);
11291			}
11292
11293			err = check_map_prog_compatibility(env, map, env->prog);
11294			if (err) {
11295				fdput(f);
11296				return err;
11297			}
11298
11299			aux = &env->insn_aux_data[i];
11300			if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11301			    insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11302				addr = (unsigned long)map;
11303			} else {
11304				u32 off = insn[1].imm;
11305
11306				if (off >= BPF_MAX_VAR_OFF) {
11307					verbose(env, "direct value offset of %u is not allowed\n", off);
11308					fdput(f);
11309					return -EINVAL;
11310				}
11311
11312				if (!map->ops->map_direct_value_addr) {
11313					verbose(env, "no direct value access support for this map type\n");
11314					fdput(f);
11315					return -EINVAL;
11316				}
11317
11318				err = map->ops->map_direct_value_addr(map, &addr, off);
11319				if (err) {
11320					verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11321						map->value_size, off);
11322					fdput(f);
11323					return err;
11324				}
11325
11326				aux->map_off = off;
11327				addr += off;
11328			}
11329
11330			insn[0].imm = (u32)addr;
11331			insn[1].imm = addr >> 32;
11332
11333			/* check whether we recorded this map already */
11334			for (j = 0; j < env->used_map_cnt; j++) {
11335				if (env->used_maps[j] == map) {
11336					aux->map_index = j;
11337					fdput(f);
11338					goto next_insn;
11339				}
11340			}
11341
11342			if (env->used_map_cnt >= MAX_USED_MAPS) {
11343				fdput(f);
11344				return -E2BIG;
11345			}
11346
11347			/* hold the map. If the program is rejected by verifier,
11348			 * the map will be released by release_maps() or it
11349			 * will be used by the valid program until it's unloaded
11350			 * and all maps are released in free_used_maps()
11351			 */
11352			bpf_map_inc(map);
11353
11354			aux->map_index = env->used_map_cnt;
11355			env->used_maps[env->used_map_cnt++] = map;
11356
11357			if (bpf_map_is_cgroup_storage(map) &&
11358			    bpf_cgroup_storage_assign(env->prog->aux, map)) {
11359				verbose(env, "only one cgroup storage of each type is allowed\n");
11360				fdput(f);
11361				return -EBUSY;
11362			}
11363
11364			fdput(f);
11365next_insn:
11366			insn++;
11367			i++;
11368			continue;
11369		}
11370
11371		/* Basic sanity check before we invest more work here. */
11372		if (!bpf_opcode_in_insntable(insn->code)) {
11373			verbose(env, "unknown opcode %02x\n", insn->code);
11374			return -EINVAL;
11375		}
11376	}
11377
11378	/* now all pseudo BPF_LD_IMM64 instructions load valid
11379	 * 'struct bpf_map *' into a register instead of user map_fd.
11380	 * These pointers will be used later by verifier to validate map access.
11381	 */
11382	return 0;
11383}
11384
11385/* drop refcnt of maps used by the rejected program */
11386static void release_maps(struct bpf_verifier_env *env)
11387{
11388	__bpf_free_used_maps(env->prog->aux, env->used_maps,
11389			     env->used_map_cnt);
11390}
11391
11392/* drop refcnt of maps used by the rejected program */
11393static void release_btfs(struct bpf_verifier_env *env)
11394{
11395	__bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11396			     env->used_btf_cnt);
11397}
11398
11399/* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11400static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11401{
11402	struct bpf_insn *insn = env->prog->insnsi;
11403	int insn_cnt = env->prog->len;
11404	int i;
11405
11406	for (i = 0; i < insn_cnt; i++, insn++) {
11407		if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11408			continue;
11409		if (insn->src_reg == BPF_PSEUDO_FUNC)
11410			continue;
11411		insn->src_reg = 0;
11412	}
11413}
11414
11415/* single env->prog->insni[off] instruction was replaced with the range
11416 * insni[off, off + cnt).  Adjust corresponding insn_aux_data by copying
11417 * [0, off) and [off, end) to new locations, so the patched range stays zero
11418 */
11419static void adjust_insn_aux_data(struct bpf_verifier_env *env,
11420				 struct bpf_insn_aux_data *new_data,
11421				 struct bpf_prog *new_prog, u32 off, u32 cnt)
11422{
11423	struct bpf_insn_aux_data *old_data = env->insn_aux_data;
11424	struct bpf_insn *insn = new_prog->insnsi;
11425	u32 old_seen = old_data[off].seen;
11426	u32 prog_len;
11427	int i;
11428
11429	/* aux info at OFF always needs adjustment, no matter fast path
11430	 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11431	 * original insn at old prog.
11432	 */
11433	old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11434
11435	if (cnt == 1)
11436		return;
11437	prog_len = new_prog->len;
11438
11439	memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11440	memcpy(new_data + off + cnt - 1, old_data + off,
11441	       sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11442	for (i = off; i < off + cnt - 1; i++) {
11443		/* Expand insni[off]'s seen count to the patched range. */
11444		new_data[i].seen = old_seen;
11445		new_data[i].zext_dst = insn_has_def32(env, insn + i);
11446	}
11447	env->insn_aux_data = new_data;
11448	vfree(old_data);
11449}
11450
11451static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11452{
11453	int i;
11454
11455	if (len == 1)
11456		return;
11457	/* NOTE: fake 'exit' subprog should be updated as well. */
11458	for (i = 0; i <= env->subprog_cnt; i++) {
11459		if (env->subprog_info[i].start <= off)
11460			continue;
11461		env->subprog_info[i].start += len - 1;
11462	}
11463}
11464
11465static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11466{
11467	struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11468	int i, sz = prog->aux->size_poke_tab;
11469	struct bpf_jit_poke_descriptor *desc;
11470
11471	for (i = 0; i < sz; i++) {
11472		desc = &tab[i];
11473		if (desc->insn_idx <= off)
11474			continue;
11475		desc->insn_idx += len - 1;
11476	}
11477}
11478
11479static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11480					    const struct bpf_insn *patch, u32 len)
11481{
11482	struct bpf_prog *new_prog;
11483	struct bpf_insn_aux_data *new_data = NULL;
11484
11485	if (len > 1) {
11486		new_data = vzalloc(array_size(env->prog->len + len - 1,
11487					      sizeof(struct bpf_insn_aux_data)));
11488		if (!new_data)
11489			return NULL;
11490	}
11491
11492	new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11493	if (IS_ERR(new_prog)) {
11494		if (PTR_ERR(new_prog) == -ERANGE)
11495			verbose(env,
11496				"insn %d cannot be patched due to 16-bit range\n",
11497				env->insn_aux_data[off].orig_idx);
11498		vfree(new_data);
11499		return NULL;
11500	}
11501	adjust_insn_aux_data(env, new_data, new_prog, off, len);
11502	adjust_subprog_starts(env, off, len);
11503	adjust_poke_descs(new_prog, off, len);
11504	return new_prog;
11505}
11506
11507static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11508					      u32 off, u32 cnt)
11509{
11510	int i, j;
11511
11512	/* find first prog starting at or after off (first to remove) */
11513	for (i = 0; i < env->subprog_cnt; i++)
11514		if (env->subprog_info[i].start >= off)
11515			break;
11516	/* find first prog starting at or after off + cnt (first to stay) */
11517	for (j = i; j < env->subprog_cnt; j++)
11518		if (env->subprog_info[j].start >= off + cnt)
11519			break;
11520	/* if j doesn't start exactly at off + cnt, we are just removing
11521	 * the front of previous prog
11522	 */
11523	if (env->subprog_info[j].start != off + cnt)
11524		j--;
11525
11526	if (j > i) {
11527		struct bpf_prog_aux *aux = env->prog->aux;
11528		int move;
11529
11530		/* move fake 'exit' subprog as well */
11531		move = env->subprog_cnt + 1 - j;
11532
11533		memmove(env->subprog_info + i,
11534			env->subprog_info + j,
11535			sizeof(*env->subprog_info) * move);
11536		env->subprog_cnt -= j - i;
11537
11538		/* remove func_info */
11539		if (aux->func_info) {
11540			move = aux->func_info_cnt - j;
11541
11542			memmove(aux->func_info + i,
11543				aux->func_info + j,
11544				sizeof(*aux->func_info) * move);
11545			aux->func_info_cnt -= j - i;
11546			/* func_info->insn_off is set after all code rewrites,
11547			 * in adjust_btf_func() - no need to adjust
11548			 */
11549		}
11550	} else {
11551		/* convert i from "first prog to remove" to "first to adjust" */
11552		if (env->subprog_info[i].start == off)
11553			i++;
11554	}
11555
11556	/* update fake 'exit' subprog as well */
11557	for (; i <= env->subprog_cnt; i++)
11558		env->subprog_info[i].start -= cnt;
11559
11560	return 0;
11561}
11562
11563static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11564				      u32 cnt)
11565{
11566	struct bpf_prog *prog = env->prog;
11567	u32 i, l_off, l_cnt, nr_linfo;
11568	struct bpf_line_info *linfo;
11569
11570	nr_linfo = prog->aux->nr_linfo;
11571	if (!nr_linfo)
11572		return 0;
11573
11574	linfo = prog->aux->linfo;
11575
11576	/* find first line info to remove, count lines to be removed */
11577	for (i = 0; i < nr_linfo; i++)
11578		if (linfo[i].insn_off >= off)
11579			break;
11580
11581	l_off = i;
11582	l_cnt = 0;
11583	for (; i < nr_linfo; i++)
11584		if (linfo[i].insn_off < off + cnt)
11585			l_cnt++;
11586		else
11587			break;
11588
11589	/* First live insn doesn't match first live linfo, it needs to "inherit"
11590	 * last removed linfo.  prog is already modified, so prog->len == off
11591	 * means no live instructions after (tail of the program was removed).
11592	 */
11593	if (prog->len != off && l_cnt &&
11594	    (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11595		l_cnt--;
11596		linfo[--i].insn_off = off + cnt;
11597	}
11598
11599	/* remove the line info which refer to the removed instructions */
11600	if (l_cnt) {
11601		memmove(linfo + l_off, linfo + i,
11602			sizeof(*linfo) * (nr_linfo - i));
11603
11604		prog->aux->nr_linfo -= l_cnt;
11605		nr_linfo = prog->aux->nr_linfo;
11606	}
11607
11608	/* pull all linfo[i].insn_off >= off + cnt in by cnt */
11609	for (i = l_off; i < nr_linfo; i++)
11610		linfo[i].insn_off -= cnt;
11611
11612	/* fix up all subprogs (incl. 'exit') which start >= off */
11613	for (i = 0; i <= env->subprog_cnt; i++)
11614		if (env->subprog_info[i].linfo_idx > l_off) {
11615			/* program may have started in the removed region but
11616			 * may not be fully removed
11617			 */
11618			if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11619				env->subprog_info[i].linfo_idx -= l_cnt;
11620			else
11621				env->subprog_info[i].linfo_idx = l_off;
11622		}
11623
11624	return 0;
11625}
11626
11627static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11628{
11629	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11630	unsigned int orig_prog_len = env->prog->len;
11631	int err;
11632
11633	if (bpf_prog_is_dev_bound(env->prog->aux))
11634		bpf_prog_offload_remove_insns(env, off, cnt);
11635
11636	err = bpf_remove_insns(env->prog, off, cnt);
11637	if (err)
11638		return err;
11639
11640	err = adjust_subprog_starts_after_remove(env, off, cnt);
11641	if (err)
11642		return err;
11643
11644	err = bpf_adj_linfo_after_remove(env, off, cnt);
11645	if (err)
11646		return err;
11647
11648	memmove(aux_data + off,	aux_data + off + cnt,
11649		sizeof(*aux_data) * (orig_prog_len - off - cnt));
11650
11651	return 0;
11652}
11653
11654/* The verifier does more data flow analysis than llvm and will not
11655 * explore branches that are dead at run time. Malicious programs can
11656 * have dead code too. Therefore replace all dead at-run-time code
11657 * with 'ja -1'.
11658 *
11659 * Just nops are not optimal, e.g. if they would sit at the end of the
11660 * program and through another bug we would manage to jump there, then
11661 * we'd execute beyond program memory otherwise. Returning exception
11662 * code also wouldn't work since we can have subprogs where the dead
11663 * code could be located.
11664 */
11665static void sanitize_dead_code(struct bpf_verifier_env *env)
11666{
11667	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11668	struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11669	struct bpf_insn *insn = env->prog->insnsi;
11670	const int insn_cnt = env->prog->len;
11671	int i;
11672
11673	for (i = 0; i < insn_cnt; i++) {
11674		if (aux_data[i].seen)
11675			continue;
11676		memcpy(insn + i, &trap, sizeof(trap));
11677		aux_data[i].zext_dst = false;
11678	}
11679}
11680
11681static bool insn_is_cond_jump(u8 code)
11682{
11683	u8 op;
11684
11685	if (BPF_CLASS(code) == BPF_JMP32)
11686		return true;
11687
11688	if (BPF_CLASS(code) != BPF_JMP)
11689		return false;
11690
11691	op = BPF_OP(code);
11692	return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11693}
11694
11695static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11696{
11697	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11698	struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11699	struct bpf_insn *insn = env->prog->insnsi;
11700	const int insn_cnt = env->prog->len;
11701	int i;
11702
11703	for (i = 0; i < insn_cnt; i++, insn++) {
11704		if (!insn_is_cond_jump(insn->code))
11705			continue;
11706
11707		if (!aux_data[i + 1].seen)
11708			ja.off = insn->off;
11709		else if (!aux_data[i + 1 + insn->off].seen)
11710			ja.off = 0;
11711		else
11712			continue;
11713
11714		if (bpf_prog_is_dev_bound(env->prog->aux))
11715			bpf_prog_offload_replace_insn(env, i, &ja);
11716
11717		memcpy(insn, &ja, sizeof(ja));
11718	}
11719}
11720
11721static int opt_remove_dead_code(struct bpf_verifier_env *env)
11722{
11723	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11724	int insn_cnt = env->prog->len;
11725	int i, err;
11726
11727	for (i = 0; i < insn_cnt; i++) {
11728		int j;
11729
11730		j = 0;
11731		while (i + j < insn_cnt && !aux_data[i + j].seen)
11732			j++;
11733		if (!j)
11734			continue;
11735
11736		err = verifier_remove_insns(env, i, j);
11737		if (err)
11738			return err;
11739		insn_cnt = env->prog->len;
11740	}
11741
11742	return 0;
11743}
11744
11745static int opt_remove_nops(struct bpf_verifier_env *env)
11746{
11747	const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11748	struct bpf_insn *insn = env->prog->insnsi;
11749	int insn_cnt = env->prog->len;
11750	int i, err;
11751
11752	for (i = 0; i < insn_cnt; i++) {
11753		if (memcmp(&insn[i], &ja, sizeof(ja)))
11754			continue;
11755
11756		err = verifier_remove_insns(env, i, 1);
11757		if (err)
11758			return err;
11759		insn_cnt--;
11760		i--;
11761	}
11762
11763	return 0;
11764}
11765
11766static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
11767					 const union bpf_attr *attr)
11768{
11769	struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
11770	struct bpf_insn_aux_data *aux = env->insn_aux_data;
11771	int i, patch_len, delta = 0, len = env->prog->len;
11772	struct bpf_insn *insns = env->prog->insnsi;
11773	struct bpf_prog *new_prog;
11774	bool rnd_hi32;
11775
11776	rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
11777	zext_patch[1] = BPF_ZEXT_REG(0);
11778	rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
11779	rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
11780	rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
11781	for (i = 0; i < len; i++) {
11782		int adj_idx = i + delta;
11783		struct bpf_insn insn;
11784		int load_reg;
11785
11786		insn = insns[adj_idx];
11787		load_reg = insn_def_regno(&insn);
11788		if (!aux[adj_idx].zext_dst) {
11789			u8 code, class;
11790			u32 imm_rnd;
11791
11792			if (!rnd_hi32)
11793				continue;
11794
11795			code = insn.code;
11796			class = BPF_CLASS(code);
11797			if (load_reg == -1)
11798				continue;
11799
11800			/* NOTE: arg "reg" (the fourth one) is only used for
11801			 *       BPF_STX + SRC_OP, so it is safe to pass NULL
11802			 *       here.
11803			 */
11804			if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
11805				if (class == BPF_LD &&
11806				    BPF_MODE(code) == BPF_IMM)
11807					i++;
11808				continue;
11809			}
11810
11811			/* ctx load could be transformed into wider load. */
11812			if (class == BPF_LDX &&
11813			    aux[adj_idx].ptr_type == PTR_TO_CTX)
11814				continue;
11815
11816			imm_rnd = get_random_int();
11817			rnd_hi32_patch[0] = insn;
11818			rnd_hi32_patch[1].imm = imm_rnd;
11819			rnd_hi32_patch[3].dst_reg = load_reg;
11820			patch = rnd_hi32_patch;
11821			patch_len = 4;
11822			goto apply_patch_buffer;
11823		}
11824
11825		/* Add in an zero-extend instruction if a) the JIT has requested
11826		 * it or b) it's a CMPXCHG.
11827		 *
11828		 * The latter is because: BPF_CMPXCHG always loads a value into
11829		 * R0, therefore always zero-extends. However some archs'
11830		 * equivalent instruction only does this load when the
11831		 * comparison is successful. This detail of CMPXCHG is
11832		 * orthogonal to the general zero-extension behaviour of the
11833		 * CPU, so it's treated independently of bpf_jit_needs_zext.
11834		 */
11835		if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
11836			continue;
11837
11838		if (WARN_ON(load_reg == -1)) {
11839			verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
11840			return -EFAULT;
11841		}
11842
11843		zext_patch[0] = insn;
11844		zext_patch[1].dst_reg = load_reg;
11845		zext_patch[1].src_reg = load_reg;
11846		patch = zext_patch;
11847		patch_len = 2;
11848apply_patch_buffer:
11849		new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
11850		if (!new_prog)
11851			return -ENOMEM;
11852		env->prog = new_prog;
11853		insns = new_prog->insnsi;
11854		aux = env->insn_aux_data;
11855		delta += patch_len - 1;
11856	}
11857
11858	return 0;
11859}
11860
11861/* convert load instructions that access fields of a context type into a
11862 * sequence of instructions that access fields of the underlying structure:
11863 *     struct __sk_buff    -> struct sk_buff
11864 *     struct bpf_sock_ops -> struct sock
11865 */
11866static int convert_ctx_accesses(struct bpf_verifier_env *env)
11867{
11868	const struct bpf_verifier_ops *ops = env->ops;
11869	int i, cnt, size, ctx_field_size, delta = 0;
11870	const int insn_cnt = env->prog->len;
11871	struct bpf_insn insn_buf[16], *insn;
11872	u32 target_size, size_default, off;
11873	struct bpf_prog *new_prog;
11874	enum bpf_access_type type;
11875	bool is_narrower_load;
11876
11877	if (ops->gen_prologue || env->seen_direct_write) {
11878		if (!ops->gen_prologue) {
11879			verbose(env, "bpf verifier is misconfigured\n");
11880			return -EINVAL;
11881		}
11882		cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
11883					env->prog);
11884		if (cnt >= ARRAY_SIZE(insn_buf)) {
11885			verbose(env, "bpf verifier is misconfigured\n");
11886			return -EINVAL;
11887		} else if (cnt) {
11888			new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
11889			if (!new_prog)
11890				return -ENOMEM;
11891
11892			env->prog = new_prog;
11893			delta += cnt - 1;
11894		}
11895	}
11896
11897	if (bpf_prog_is_dev_bound(env->prog->aux))
11898		return 0;
11899
11900	insn = env->prog->insnsi + delta;
11901
11902	for (i = 0; i < insn_cnt; i++, insn++) {
11903		bpf_convert_ctx_access_t convert_ctx_access;
11904		bool ctx_access;
11905
11906		if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
11907		    insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
11908		    insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
11909		    insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
11910			type = BPF_READ;
11911			ctx_access = true;
11912		} else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
11913			   insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
11914			   insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
11915			   insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
11916			   insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
11917			   insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
11918			   insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
11919			   insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
11920			type = BPF_WRITE;
11921			ctx_access = BPF_CLASS(insn->code) == BPF_STX;
11922		} else {
11923			continue;
11924		}
11925
11926		if (type == BPF_WRITE &&
11927		    env->insn_aux_data[i + delta].sanitize_stack_spill) {
11928			struct bpf_insn patch[] = {
11929				*insn,
11930				BPF_ST_NOSPEC(),
11931			};
11932
11933			cnt = ARRAY_SIZE(patch);
11934			new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
11935			if (!new_prog)
11936				return -ENOMEM;
11937
11938			delta    += cnt - 1;
11939			env->prog = new_prog;
11940			insn      = new_prog->insnsi + i + delta;
11941			continue;
11942		}
11943
11944		if (!ctx_access)
11945			continue;
11946
11947		switch (env->insn_aux_data[i + delta].ptr_type) {
11948		case PTR_TO_CTX:
11949			if (!ops->convert_ctx_access)
11950				continue;
11951			convert_ctx_access = ops->convert_ctx_access;
11952			break;
11953		case PTR_TO_SOCKET:
11954		case PTR_TO_SOCK_COMMON:
11955			convert_ctx_access = bpf_sock_convert_ctx_access;
11956			break;
11957		case PTR_TO_TCP_SOCK:
11958			convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
11959			break;
11960		case PTR_TO_XDP_SOCK:
11961			convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
11962			break;
11963		case PTR_TO_BTF_ID:
11964			if (type == BPF_READ) {
11965				insn->code = BPF_LDX | BPF_PROBE_MEM |
11966					BPF_SIZE((insn)->code);
11967				env->prog->aux->num_exentries++;
11968			} else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
11969				verbose(env, "Writes through BTF pointers are not allowed\n");
11970				return -EINVAL;
11971			}
11972			continue;
11973		default:
11974			continue;
11975		}
11976
11977		ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
11978		size = BPF_LDST_BYTES(insn);
11979
11980		/* If the read access is a narrower load of the field,
11981		 * convert to a 4/8-byte load, to minimum program type specific
11982		 * convert_ctx_access changes. If conversion is successful,
11983		 * we will apply proper mask to the result.
11984		 */
11985		is_narrower_load = size < ctx_field_size;
11986		size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
11987		off = insn->off;
11988		if (is_narrower_load) {
11989			u8 size_code;
11990
11991			if (type == BPF_WRITE) {
11992				verbose(env, "bpf verifier narrow ctx access misconfigured\n");
11993				return -EINVAL;
11994			}
11995
11996			size_code = BPF_H;
11997			if (ctx_field_size == 4)
11998				size_code = BPF_W;
11999			else if (ctx_field_size == 8)
12000				size_code = BPF_DW;
12001
12002			insn->off = off & ~(size_default - 1);
12003			insn->code = BPF_LDX | BPF_MEM | size_code;
12004		}
12005
12006		target_size = 0;
12007		cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12008					 &target_size);
12009		if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12010		    (ctx_field_size && !target_size)) {
12011			verbose(env, "bpf verifier is misconfigured\n");
12012			return -EINVAL;
12013		}
12014
12015		if (is_narrower_load && size < target_size) {
12016			u8 shift = bpf_ctx_narrow_access_offset(
12017				off, size, size_default) * 8;
12018			if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
12019				verbose(env, "bpf verifier narrow ctx load misconfigured\n");
12020				return -EINVAL;
12021			}
12022			if (ctx_field_size <= 4) {
12023				if (shift)
12024					insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12025									insn->dst_reg,
12026									shift);
12027				insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12028								(1 << size * 8) - 1);
12029			} else {
12030				if (shift)
12031					insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12032									insn->dst_reg,
12033									shift);
12034				insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12035								(1ULL << size * 8) - 1);
12036			}
12037		}
12038
12039		new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12040		if (!new_prog)
12041			return -ENOMEM;
12042
12043		delta += cnt - 1;
12044
12045		/* keep walking new program and skip insns we just inserted */
12046		env->prog = new_prog;
12047		insn      = new_prog->insnsi + i + delta;
12048	}
12049
12050	return 0;
12051}
12052
12053static int jit_subprogs(struct bpf_verifier_env *env)
12054{
12055	struct bpf_prog *prog = env->prog, **func, *tmp;
12056	int i, j, subprog_start, subprog_end = 0, len, subprog;
12057	struct bpf_map *map_ptr;
12058	struct bpf_insn *insn;
12059	void *old_bpf_func;
12060	int err, num_exentries;
12061
12062	if (env->subprog_cnt <= 1)
12063		return 0;
12064
12065	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12066		if (bpf_pseudo_func(insn)) {
12067			env->insn_aux_data[i].call_imm = insn->imm;
12068			/* subprog is encoded in insn[1].imm */
12069			continue;
12070		}
12071
12072		if (!bpf_pseudo_call(insn))
12073			continue;
12074		/* Upon error here we cannot fall back to interpreter but
12075		 * need a hard reject of the program. Thus -EFAULT is
12076		 * propagated in any case.
12077		 */
12078		subprog = find_subprog(env, i + insn->imm + 1);
12079		if (subprog < 0) {
12080			WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12081				  i + insn->imm + 1);
12082			return -EFAULT;
12083		}
12084		/* temporarily remember subprog id inside insn instead of
12085		 * aux_data, since next loop will split up all insns into funcs
12086		 */
12087		insn->off = subprog;
12088		/* remember original imm in case JIT fails and fallback
12089		 * to interpreter will be needed
12090		 */
12091		env->insn_aux_data[i].call_imm = insn->imm;
12092		/* point imm to __bpf_call_base+1 from JITs point of view */
12093		insn->imm = 1;
12094	}
12095
12096	err = bpf_prog_alloc_jited_linfo(prog);
12097	if (err)
12098		goto out_undo_insn;
12099
12100	err = -ENOMEM;
12101	func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12102	if (!func)
12103		goto out_undo_insn;
12104
12105	for (i = 0; i < env->subprog_cnt; i++) {
12106		subprog_start = subprog_end;
12107		subprog_end = env->subprog_info[i + 1].start;
12108
12109		len = subprog_end - subprog_start;
12110		/* BPF_PROG_RUN doesn't call subprogs directly,
12111		 * hence main prog stats include the runtime of subprogs.
12112		 * subprogs don't have IDs and not reachable via prog_get_next_id
12113		 * func[i]->stats will never be accessed and stays NULL
12114		 */
12115		func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12116		if (!func[i])
12117			goto out_free;
12118		memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12119		       len * sizeof(struct bpf_insn));
12120		func[i]->type = prog->type;
12121		func[i]->len = len;
12122		if (bpf_prog_calc_tag(func[i]))
12123			goto out_free;
12124		func[i]->is_func = 1;
12125		func[i]->aux->func_idx = i;
12126		/* Below members will be freed only at prog->aux */
12127		func[i]->aux->btf = prog->aux->btf;
12128		func[i]->aux->func_info = prog->aux->func_info;
12129		func[i]->aux->poke_tab = prog->aux->poke_tab;
12130		func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12131
12132		for (j = 0; j < prog->aux->size_poke_tab; j++) {
12133			struct bpf_jit_poke_descriptor *poke;
12134
12135			poke = &prog->aux->poke_tab[j];
12136			if (poke->insn_idx < subprog_end &&
12137			    poke->insn_idx >= subprog_start)
12138				poke->aux = func[i]->aux;
12139		}
12140
12141		/* Use bpf_prog_F_tag to indicate functions in stack traces.
12142		 * Long term would need debug info to populate names
12143		 */
12144		func[i]->aux->name[0] = 'F';
12145		func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12146		func[i]->jit_requested = 1;
12147		func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12148		func[i]->aux->linfo = prog->aux->linfo;
12149		func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12150		func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12151		func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12152		num_exentries = 0;
12153		insn = func[i]->insnsi;
12154		for (j = 0; j < func[i]->len; j++, insn++) {
12155			if (BPF_CLASS(insn->code) == BPF_LDX &&
12156			    BPF_MODE(insn->code) == BPF_PROBE_MEM)
12157				num_exentries++;
12158		}
12159		func[i]->aux->num_exentries = num_exentries;
12160		func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12161		func[i] = bpf_int_jit_compile(func[i]);
12162		if (!func[i]->jited) {
12163			err = -ENOTSUPP;
12164			goto out_free;
12165		}
12166		cond_resched();
12167	}
12168
12169	/* at this point all bpf functions were successfully JITed
12170	 * now populate all bpf_calls with correct addresses and
12171	 * run last pass of JIT
12172	 */
12173	for (i = 0; i < env->subprog_cnt; i++) {
12174		insn = func[i]->insnsi;
12175		for (j = 0; j < func[i]->len; j++, insn++) {
12176			if (bpf_pseudo_func(insn)) {
12177				subprog = insn[1].imm;
12178				insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12179				insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12180				continue;
12181			}
12182			if (!bpf_pseudo_call(insn))
12183				continue;
12184			subprog = insn->off;
12185			insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12186				    __bpf_call_base;
12187		}
12188
12189		/* we use the aux data to keep a list of the start addresses
12190		 * of the JITed images for each function in the program
12191		 *
12192		 * for some architectures, such as powerpc64, the imm field
12193		 * might not be large enough to hold the offset of the start
12194		 * address of the callee's JITed image from __bpf_call_base
12195		 *
12196		 * in such cases, we can lookup the start address of a callee
12197		 * by using its subprog id, available from the off field of
12198		 * the call instruction, as an index for this list
12199		 */
12200		func[i]->aux->func = func;
12201		func[i]->aux->func_cnt = env->subprog_cnt;
12202	}
12203	for (i = 0; i < env->subprog_cnt; i++) {
12204		old_bpf_func = func[i]->bpf_func;
12205		tmp = bpf_int_jit_compile(func[i]);
12206		if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12207			verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12208			err = -ENOTSUPP;
12209			goto out_free;
12210		}
12211		cond_resched();
12212	}
12213
12214	/* finally lock prog and jit images for all functions and
12215	 * populate kallsysm
12216	 */
12217	for (i = 0; i < env->subprog_cnt; i++) {
12218		bpf_prog_lock_ro(func[i]);
12219		bpf_prog_kallsyms_add(func[i]);
12220	}
12221
12222	/* Last step: make now unused interpreter insns from main
12223	 * prog consistent for later dump requests, so they can
12224	 * later look the same as if they were interpreted only.
12225	 */
12226	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12227		if (bpf_pseudo_func(insn)) {
12228			insn[0].imm = env->insn_aux_data[i].call_imm;
12229			insn[1].imm = find_subprog(env, i + insn[0].imm + 1);
12230			continue;
12231		}
12232		if (!bpf_pseudo_call(insn))
12233			continue;
12234		insn->off = env->insn_aux_data[i].call_imm;
12235		subprog = find_subprog(env, i + insn->off + 1);
12236		insn->imm = subprog;
12237	}
12238
12239	prog->jited = 1;
12240	prog->bpf_func = func[0]->bpf_func;
12241	prog->aux->func = func;
12242	prog->aux->func_cnt = env->subprog_cnt;
12243	bpf_prog_jit_attempt_done(prog);
12244	return 0;
12245out_free:
12246	/* We failed JIT'ing, so at this point we need to unregister poke
12247	 * descriptors from subprogs, so that kernel is not attempting to
12248	 * patch it anymore as we're freeing the subprog JIT memory.
12249	 */
12250	for (i = 0; i < prog->aux->size_poke_tab; i++) {
12251		map_ptr = prog->aux->poke_tab[i].tail_call.map;
12252		map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12253	}
12254	/* At this point we're guaranteed that poke descriptors are not
12255	 * live anymore. We can just unlink its descriptor table as it's
12256	 * released with the main prog.
12257	 */
12258	for (i = 0; i < env->subprog_cnt; i++) {
12259		if (!func[i])
12260			continue;
12261		func[i]->aux->poke_tab = NULL;
12262		bpf_jit_free(func[i]);
12263	}
12264	kfree(func);
12265out_undo_insn:
12266	/* cleanup main prog to be interpreted */
12267	prog->jit_requested = 0;
12268	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12269		if (!bpf_pseudo_call(insn))
12270			continue;
12271		insn->off = 0;
12272		insn->imm = env->insn_aux_data[i].call_imm;
12273	}
12274	bpf_prog_jit_attempt_done(prog);
12275	return err;
12276}
12277
12278static int fixup_call_args(struct bpf_verifier_env *env)
12279{
12280#ifndef CONFIG_BPF_JIT_ALWAYS_ON
12281	struct bpf_prog *prog = env->prog;
12282	struct bpf_insn *insn = prog->insnsi;
12283	bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12284	int i, depth;
12285#endif
12286	int err = 0;
12287
12288	if (env->prog->jit_requested &&
12289	    !bpf_prog_is_dev_bound(env->prog->aux)) {
12290		err = jit_subprogs(env);
12291		if (err == 0)
12292			return 0;
12293		if (err == -EFAULT)
12294			return err;
12295	}
12296#ifndef CONFIG_BPF_JIT_ALWAYS_ON
12297	if (has_kfunc_call) {
12298		verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12299		return -EINVAL;
12300	}
12301	if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12302		/* When JIT fails the progs with bpf2bpf calls and tail_calls
12303		 * have to be rejected, since interpreter doesn't support them yet.
12304		 */
12305		verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12306		return -EINVAL;
12307	}
12308	for (i = 0; i < prog->len; i++, insn++) {
12309		if (bpf_pseudo_func(insn)) {
12310			/* When JIT fails the progs with callback calls
12311			 * have to be rejected, since interpreter doesn't support them yet.
12312			 */
12313			verbose(env, "callbacks are not allowed in non-JITed programs\n");
12314			return -EINVAL;
12315		}
12316
12317		if (!bpf_pseudo_call(insn))
12318			continue;
12319		depth = get_callee_stack_depth(env, insn, i);
12320		if (depth < 0)
12321			return depth;
12322		bpf_patch_call_args(insn, depth);
12323	}
12324	err = 0;
12325#endif
12326	return err;
12327}
12328
12329static int fixup_kfunc_call(struct bpf_verifier_env *env,
12330			    struct bpf_insn *insn)
12331{
12332	const struct bpf_kfunc_desc *desc;
12333
12334	/* insn->imm has the btf func_id. Replace it with
12335	 * an address (relative to __bpf_base_call).
12336	 */
12337	desc = find_kfunc_desc(env->prog, insn->imm);
12338	if (!desc) {
12339		verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12340			insn->imm);
12341		return -EFAULT;
12342	}
12343
12344	insn->imm = desc->imm;
12345
12346	return 0;
12347}
12348
12349/* Do various post-verification rewrites in a single program pass.
12350 * These rewrites simplify JIT and interpreter implementations.
12351 */
12352static int do_misc_fixups(struct bpf_verifier_env *env)
12353{
12354	struct bpf_prog *prog = env->prog;
12355	bool expect_blinding = bpf_jit_blinding_enabled(prog);
12356	struct bpf_insn *insn = prog->insnsi;
12357	const struct bpf_func_proto *fn;
12358	const int insn_cnt = prog->len;
12359	const struct bpf_map_ops *ops;
12360	struct bpf_insn_aux_data *aux;
12361	struct bpf_insn insn_buf[16];
12362	struct bpf_prog *new_prog;
12363	struct bpf_map *map_ptr;
12364	int i, ret, cnt, delta = 0;
12365
12366	for (i = 0; i < insn_cnt; i++, insn++) {
12367		/* Make divide-by-zero exceptions impossible. */
12368		if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12369		    insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12370		    insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12371		    insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12372			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12373			bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12374			struct bpf_insn *patchlet;
12375			struct bpf_insn chk_and_div[] = {
12376				/* [R,W]x div 0 -> 0 */
12377				BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12378					     BPF_JNE | BPF_K, insn->src_reg,
12379					     0, 2, 0),
12380				BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12381				BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12382				*insn,
12383			};
12384			struct bpf_insn chk_and_mod[] = {
12385				/* [R,W]x mod 0 -> [R,W]x */
12386				BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12387					     BPF_JEQ | BPF_K, insn->src_reg,
12388					     0, 1 + (is64 ? 0 : 1), 0),
12389				*insn,
12390				BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12391				BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12392			};
12393
12394			patchlet = isdiv ? chk_and_div : chk_and_mod;
12395			cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12396				      ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12397
12398			new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12399			if (!new_prog)
12400				return -ENOMEM;
12401
12402			delta    += cnt - 1;
12403			env->prog = prog = new_prog;
12404			insn      = new_prog->insnsi + i + delta;
12405			continue;
12406		}
12407
12408		/* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12409		if (BPF_CLASS(insn->code) == BPF_LD &&
12410		    (BPF_MODE(insn->code) == BPF_ABS ||
12411		     BPF_MODE(insn->code) == BPF_IND)) {
12412			cnt = env->ops->gen_ld_abs(insn, insn_buf);
12413			if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12414				verbose(env, "bpf verifier is misconfigured\n");
12415				return -EINVAL;
12416			}
12417
12418			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12419			if (!new_prog)
12420				return -ENOMEM;
12421
12422			delta    += cnt - 1;
12423			env->prog = prog = new_prog;
12424			insn      = new_prog->insnsi + i + delta;
12425			continue;
12426		}
12427
12428		/* Rewrite pointer arithmetic to mitigate speculation attacks. */
12429		if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12430		    insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12431			const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12432			const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12433			struct bpf_insn *patch = &insn_buf[0];
12434			bool issrc, isneg, isimm;
12435			u32 off_reg;
12436
12437			aux = &env->insn_aux_data[i + delta];
12438			if (!aux->alu_state ||
12439			    aux->alu_state == BPF_ALU_NON_POINTER)
12440				continue;
12441
12442			isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12443			issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12444				BPF_ALU_SANITIZE_SRC;
12445			isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12446
12447			off_reg = issrc ? insn->src_reg : insn->dst_reg;
12448			if (isimm) {
12449				*patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12450			} else {
12451				if (isneg)
12452					*patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12453				*patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12454				*patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12455				*patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12456				*patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12457				*patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12458				*patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12459			}
12460			if (!issrc)
12461				*patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12462			insn->src_reg = BPF_REG_AX;
12463			if (isneg)
12464				insn->code = insn->code == code_add ?
12465					     code_sub : code_add;
12466			*patch++ = *insn;
12467			if (issrc && isneg && !isimm)
12468				*patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12469			cnt = patch - insn_buf;
12470
12471			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12472			if (!new_prog)
12473				return -ENOMEM;
12474
12475			delta    += cnt - 1;
12476			env->prog = prog = new_prog;
12477			insn      = new_prog->insnsi + i + delta;
12478			continue;
12479		}
12480
12481		if (insn->code != (BPF_JMP | BPF_CALL))
12482			continue;
12483		if (insn->src_reg == BPF_PSEUDO_CALL)
12484			continue;
12485		if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12486			ret = fixup_kfunc_call(env, insn);
12487			if (ret)
12488				return ret;
12489			continue;
12490		}
12491
12492		if (insn->imm == BPF_FUNC_get_route_realm)
12493			prog->dst_needed = 1;
12494		if (insn->imm == BPF_FUNC_get_prandom_u32)
12495			bpf_user_rnd_init_once();
12496		if (insn->imm == BPF_FUNC_override_return)
12497			prog->kprobe_override = 1;
12498		if (insn->imm == BPF_FUNC_tail_call) {
12499			/* If we tail call into other programs, we
12500			 * cannot make any assumptions since they can
12501			 * be replaced dynamically during runtime in
12502			 * the program array.
12503			 */
12504			prog->cb_access = 1;
12505			if (!allow_tail_call_in_subprogs(env))
12506				prog->aux->stack_depth = MAX_BPF_STACK;
12507			prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12508
12509			/* mark bpf_tail_call as different opcode to avoid
12510			 * conditional branch in the interpreter for every normal
12511			 * call and to prevent accidental JITing by JIT compiler
12512			 * that doesn't support bpf_tail_call yet
12513			 */
12514			insn->imm = 0;
12515			insn->code = BPF_JMP | BPF_TAIL_CALL;
12516
12517			aux = &env->insn_aux_data[i + delta];
12518			if (env->bpf_capable && !expect_blinding &&
12519			    prog->jit_requested &&
12520			    !bpf_map_key_poisoned(aux) &&
12521			    !bpf_map_ptr_poisoned(aux) &&
12522			    !bpf_map_ptr_unpriv(aux)) {
12523				struct bpf_jit_poke_descriptor desc = {
12524					.reason = BPF_POKE_REASON_TAIL_CALL,
12525					.tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12526					.tail_call.key = bpf_map_key_immediate(aux),
12527					.insn_idx = i + delta,
12528				};
12529
12530				ret = bpf_jit_add_poke_descriptor(prog, &desc);
12531				if (ret < 0) {
12532					verbose(env, "adding tail call poke descriptor failed\n");
12533					return ret;
12534				}
12535
12536				insn->imm = ret + 1;
12537				continue;
12538			}
12539
12540			if (!bpf_map_ptr_unpriv(aux))
12541				continue;
12542
12543			/* instead of changing every JIT dealing with tail_call
12544			 * emit two extra insns:
12545			 * if (index >= max_entries) goto out;
12546			 * index &= array->index_mask;
12547			 * to avoid out-of-bounds cpu speculation
12548			 */
12549			if (bpf_map_ptr_poisoned(aux)) {
12550				verbose(env, "tail_call abusing map_ptr\n");
12551				return -EINVAL;
12552			}
12553
12554			map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12555			insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12556						  map_ptr->max_entries, 2);
12557			insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12558						    container_of(map_ptr,
12559								 struct bpf_array,
12560								 map)->index_mask);
12561			insn_buf[2] = *insn;
12562			cnt = 3;
12563			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12564			if (!new_prog)
12565				return -ENOMEM;
12566
12567			delta    += cnt - 1;
12568			env->prog = prog = new_prog;
12569			insn      = new_prog->insnsi + i + delta;
12570			continue;
12571		}
12572
12573		/* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12574		 * and other inlining handlers are currently limited to 64 bit
12575		 * only.
12576		 */
12577		if (prog->jit_requested && BITS_PER_LONG == 64 &&
12578		    (insn->imm == BPF_FUNC_map_lookup_elem ||
12579		     insn->imm == BPF_FUNC_map_update_elem ||
12580		     insn->imm == BPF_FUNC_map_delete_elem ||
12581		     insn->imm == BPF_FUNC_map_push_elem   ||
12582		     insn->imm == BPF_FUNC_map_pop_elem    ||
12583		     insn->imm == BPF_FUNC_map_peek_elem   ||
12584		     insn->imm == BPF_FUNC_redirect_map)) {
12585			aux = &env->insn_aux_data[i + delta];
12586			if (bpf_map_ptr_poisoned(aux))
12587				goto patch_call_imm;
12588
12589			map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12590			ops = map_ptr->ops;
12591			if (insn->imm == BPF_FUNC_map_lookup_elem &&
12592			    ops->map_gen_lookup) {
12593				cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12594				if (cnt == -EOPNOTSUPP)
12595					goto patch_map_ops_generic;
12596				if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12597					verbose(env, "bpf verifier is misconfigured\n");
12598					return -EINVAL;
12599				}
12600
12601				new_prog = bpf_patch_insn_data(env, i + delta,
12602							       insn_buf, cnt);
12603				if (!new_prog)
12604					return -ENOMEM;
12605
12606				delta    += cnt - 1;
12607				env->prog = prog = new_prog;
12608				insn      = new_prog->insnsi + i + delta;
12609				continue;
12610			}
12611
12612			BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12613				     (void *(*)(struct bpf_map *map, void *key))NULL));
12614			BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12615				     (int (*)(struct bpf_map *map, void *key))NULL));
12616			BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12617				     (int (*)(struct bpf_map *map, void *key, void *value,
12618					      u64 flags))NULL));
12619			BUILD_BUG_ON(!__same_type(ops->map_push_elem,
12620				     (int (*)(struct bpf_map *map, void *value,
12621					      u64 flags))NULL));
12622			BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
12623				     (int (*)(struct bpf_map *map, void *value))NULL));
12624			BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
12625				     (int (*)(struct bpf_map *map, void *value))NULL));
12626			BUILD_BUG_ON(!__same_type(ops->map_redirect,
12627				     (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
12628
12629patch_map_ops_generic:
12630			switch (insn->imm) {
12631			case BPF_FUNC_map_lookup_elem:
12632				insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
12633					    __bpf_call_base;
12634				continue;
12635			case BPF_FUNC_map_update_elem:
12636				insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
12637					    __bpf_call_base;
12638				continue;
12639			case BPF_FUNC_map_delete_elem:
12640				insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
12641					    __bpf_call_base;
12642				continue;
12643			case BPF_FUNC_map_push_elem:
12644				insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
12645					    __bpf_call_base;
12646				continue;
12647			case BPF_FUNC_map_pop_elem:
12648				insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
12649					    __bpf_call_base;
12650				continue;
12651			case BPF_FUNC_map_peek_elem:
12652				insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
12653					    __bpf_call_base;
12654				continue;
12655			case BPF_FUNC_redirect_map:
12656				insn->imm = BPF_CAST_CALL(ops->map_redirect) -
12657					    __bpf_call_base;
12658				continue;
12659			}
12660
12661			goto patch_call_imm;
12662		}
12663
12664		/* Implement bpf_jiffies64 inline. */
12665		if (prog->jit_requested && BITS_PER_LONG == 64 &&
12666		    insn->imm == BPF_FUNC_jiffies64) {
12667			struct bpf_insn ld_jiffies_addr[2] = {
12668				BPF_LD_IMM64(BPF_REG_0,
12669					     (unsigned long)&jiffies),
12670			};
12671
12672			insn_buf[0] = ld_jiffies_addr[0];
12673			insn_buf[1] = ld_jiffies_addr[1];
12674			insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
12675						  BPF_REG_0, 0);
12676			cnt = 3;
12677
12678			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
12679						       cnt);
12680			if (!new_prog)
12681				return -ENOMEM;
12682
12683			delta    += cnt - 1;
12684			env->prog = prog = new_prog;
12685			insn      = new_prog->insnsi + i + delta;
12686			continue;
12687		}
12688
12689patch_call_imm:
12690		fn = env->ops->get_func_proto(insn->imm, env->prog);
12691		/* all functions that have prototype and verifier allowed
12692		 * programs to call them, must be real in-kernel functions
12693		 */
12694		if (!fn->func) {
12695			verbose(env,
12696				"kernel subsystem misconfigured func %s#%d\n",
12697				func_id_name(insn->imm), insn->imm);
12698			return -EFAULT;
12699		}
12700		insn->imm = fn->func - __bpf_call_base;
12701	}
12702
12703	/* Since poke tab is now finalized, publish aux to tracker. */
12704	for (i = 0; i < prog->aux->size_poke_tab; i++) {
12705		map_ptr = prog->aux->poke_tab[i].tail_call.map;
12706		if (!map_ptr->ops->map_poke_track ||
12707		    !map_ptr->ops->map_poke_untrack ||
12708		    !map_ptr->ops->map_poke_run) {
12709			verbose(env, "bpf verifier is misconfigured\n");
12710			return -EINVAL;
12711		}
12712
12713		ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
12714		if (ret < 0) {
12715			verbose(env, "tracking tail call prog failed\n");
12716			return ret;
12717		}
12718	}
12719
12720	sort_kfunc_descs_by_imm(env->prog);
12721
12722	return 0;
12723}
12724
12725static void free_states(struct bpf_verifier_env *env)
12726{
12727	struct bpf_verifier_state_list *sl, *sln;
12728	int i;
12729
12730	sl = env->free_list;
12731	while (sl) {
12732		sln = sl->next;
12733		free_verifier_state(&sl->state, false);
12734		kfree(sl);
12735		sl = sln;
12736	}
12737	env->free_list = NULL;
12738
12739	if (!env->explored_states)
12740		return;
12741
12742	for (i = 0; i < state_htab_size(env); i++) {
12743		sl = env->explored_states[i];
12744
12745		while (sl) {
12746			sln = sl->next;
12747			free_verifier_state(&sl->state, false);
12748			kfree(sl);
12749			sl = sln;
12750		}
12751		env->explored_states[i] = NULL;
12752	}
12753}
12754
12755static int do_check_common(struct bpf_verifier_env *env, int subprog)
12756{
12757	bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12758	struct bpf_verifier_state *state;
12759	struct bpf_reg_state *regs;
12760	int ret, i;
12761
12762	env->prev_linfo = NULL;
12763	env->pass_cnt++;
12764
12765	state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
12766	if (!state)
12767		return -ENOMEM;
12768	state->curframe = 0;
12769	state->speculative = false;
12770	state->branches = 1;
12771	state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
12772	if (!state->frame[0]) {
12773		kfree(state);
12774		return -ENOMEM;
12775	}
12776	env->cur_state = state;
12777	init_func_state(env, state->frame[0],
12778			BPF_MAIN_FUNC /* callsite */,
12779			0 /* frameno */,
12780			subprog);
12781
12782	regs = state->frame[state->curframe]->regs;
12783	if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
12784		ret = btf_prepare_func_args(env, subprog, regs);
12785		if (ret)
12786			goto out;
12787		for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
12788			if (regs[i].type == PTR_TO_CTX)
12789				mark_reg_known_zero(env, regs, i);
12790			else if (regs[i].type == SCALAR_VALUE)
12791				mark_reg_unknown(env, regs, i);
12792			else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
12793				const u32 mem_size = regs[i].mem_size;
12794
12795				mark_reg_known_zero(env, regs, i);
12796				regs[i].mem_size = mem_size;
12797				regs[i].id = ++env->id_gen;
12798			}
12799		}
12800	} else {
12801		/* 1st arg to a function */
12802		regs[BPF_REG_1].type = PTR_TO_CTX;
12803		mark_reg_known_zero(env, regs, BPF_REG_1);
12804		ret = btf_check_subprog_arg_match(env, subprog, regs);
12805		if (ret == -EFAULT)
12806			/* unlikely verifier bug. abort.
12807			 * ret == 0 and ret < 0 are sadly acceptable for
12808			 * main() function due to backward compatibility.
12809			 * Like socket filter program may be written as:
12810			 * int bpf_prog(struct pt_regs *ctx)
12811			 * and never dereference that ctx in the program.
12812			 * 'struct pt_regs' is a type mismatch for socket
12813			 * filter that should be using 'struct __sk_buff'.
12814			 */
12815			goto out;
12816	}
12817
12818	ret = do_check(env);
12819out:
12820	/* check for NULL is necessary, since cur_state can be freed inside
12821	 * do_check() under memory pressure.
12822	 */
12823	if (env->cur_state) {
12824		free_verifier_state(env->cur_state, true);
12825		env->cur_state = NULL;
12826	}
12827	while (!pop_stack(env, NULL, NULL, false));
12828	if (!ret && pop_log)
12829		bpf_vlog_reset(&env->log, 0);
12830	free_states(env);
12831	return ret;
12832}
12833
12834/* Verify all global functions in a BPF program one by one based on their BTF.
12835 * All global functions must pass verification. Otherwise the whole program is rejected.
12836 * Consider:
12837 * int bar(int);
12838 * int foo(int f)
12839 * {
12840 *    return bar(f);
12841 * }
12842 * int bar(int b)
12843 * {
12844 *    ...
12845 * }
12846 * foo() will be verified first for R1=any_scalar_value. During verification it
12847 * will be assumed that bar() already verified successfully and call to bar()
12848 * from foo() will be checked for type match only. Later bar() will be verified
12849 * independently to check that it's safe for R1=any_scalar_value.
12850 */
12851static int do_check_subprogs(struct bpf_verifier_env *env)
12852{
12853	struct bpf_prog_aux *aux = env->prog->aux;
12854	int i, ret;
12855
12856	if (!aux->func_info)
12857		return 0;
12858
12859	for (i = 1; i < env->subprog_cnt; i++) {
12860		if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
12861			continue;
12862		env->insn_idx = env->subprog_info[i].start;
12863		WARN_ON_ONCE(env->insn_idx == 0);
12864		ret = do_check_common(env, i);
12865		if (ret) {
12866			return ret;
12867		} else if (env->log.level & BPF_LOG_LEVEL) {
12868			verbose(env,
12869				"Func#%d is safe for any args that match its prototype\n",
12870				i);
12871		}
12872	}
12873	return 0;
12874}
12875
12876static int do_check_main(struct bpf_verifier_env *env)
12877{
12878	int ret;
12879
12880	env->insn_idx = 0;
12881	ret = do_check_common(env, 0);
12882	if (!ret)
12883		env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
12884	return ret;
12885}
12886
12887
12888static void print_verification_stats(struct bpf_verifier_env *env)
12889{
12890	int i;
12891
12892	if (env->log.level & BPF_LOG_STATS) {
12893		verbose(env, "verification time %lld usec\n",
12894			div_u64(env->verification_time, 1000));
12895		verbose(env, "stack depth ");
12896		for (i = 0; i < env->subprog_cnt; i++) {
12897			u32 depth = env->subprog_info[i].stack_depth;
12898
12899			verbose(env, "%d", depth);
12900			if (i + 1 < env->subprog_cnt)
12901				verbose(env, "+");
12902		}
12903		verbose(env, "\n");
12904	}
12905	verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
12906		"total_states %d peak_states %d mark_read %d\n",
12907		env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
12908		env->max_states_per_insn, env->total_states,
12909		env->peak_states, env->longest_mark_read_walk);
12910}
12911
12912static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
12913{
12914	const struct btf_type *t, *func_proto;
12915	const struct bpf_struct_ops *st_ops;
12916	const struct btf_member *member;
12917	struct bpf_prog *prog = env->prog;
12918	u32 btf_id, member_idx;
12919	const char *mname;
12920
12921	if (!prog->gpl_compatible) {
12922		verbose(env, "struct ops programs must have a GPL compatible license\n");
12923		return -EINVAL;
12924	}
12925
12926	btf_id = prog->aux->attach_btf_id;
12927	st_ops = bpf_struct_ops_find(btf_id);
12928	if (!st_ops) {
12929		verbose(env, "attach_btf_id %u is not a supported struct\n",
12930			btf_id);
12931		return -ENOTSUPP;
12932	}
12933
12934	t = st_ops->type;
12935	member_idx = prog->expected_attach_type;
12936	if (member_idx >= btf_type_vlen(t)) {
12937		verbose(env, "attach to invalid member idx %u of struct %s\n",
12938			member_idx, st_ops->name);
12939		return -EINVAL;
12940	}
12941
12942	member = &btf_type_member(t)[member_idx];
12943	mname = btf_name_by_offset(btf_vmlinux, member->name_off);
12944	func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
12945					       NULL);
12946	if (!func_proto) {
12947		verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
12948			mname, member_idx, st_ops->name);
12949		return -EINVAL;
12950	}
12951
12952	if (st_ops->check_member) {
12953		int err = st_ops->check_member(t, member);
12954
12955		if (err) {
12956			verbose(env, "attach to unsupported member %s of struct %s\n",
12957				mname, st_ops->name);
12958			return err;
12959		}
12960	}
12961
12962	prog->aux->attach_func_proto = func_proto;
12963	prog->aux->attach_func_name = mname;
12964	env->ops = st_ops->verifier_ops;
12965
12966	return 0;
12967}
12968#define SECURITY_PREFIX "security_"
12969
12970static int check_attach_modify_return(unsigned long addr, const char *func_name)
12971{
12972	if (within_error_injection_list(addr) ||
12973	    !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
12974		return 0;
12975
12976	return -EINVAL;
12977}
12978
12979/* list of non-sleepable functions that are otherwise on
12980 * ALLOW_ERROR_INJECTION list
12981 */
12982BTF_SET_START(btf_non_sleepable_error_inject)
12983/* Three functions below can be called from sleepable and non-sleepable context.
12984 * Assume non-sleepable from bpf safety point of view.
12985 */
12986BTF_ID(func, __add_to_page_cache_locked)
12987BTF_ID(func, should_fail_alloc_page)
12988BTF_ID(func, should_failslab)
12989BTF_SET_END(btf_non_sleepable_error_inject)
12990
12991static int check_non_sleepable_error_inject(u32 btf_id)
12992{
12993	return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
12994}
12995
12996int bpf_check_attach_target(struct bpf_verifier_log *log,
12997			    const struct bpf_prog *prog,
12998			    const struct bpf_prog *tgt_prog,
12999			    u32 btf_id,
13000			    struct bpf_attach_target_info *tgt_info)
13001{
13002	bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13003	const char prefix[] = "btf_trace_";
13004	int ret = 0, subprog = -1, i;
13005	const struct btf_type *t;
13006	bool conservative = true;
13007	const char *tname;
13008	struct btf *btf;
13009	long addr = 0;
13010
13011	if (!btf_id) {
13012		bpf_log(log, "Tracing programs must provide btf_id\n");
13013		return -EINVAL;
13014	}
13015	btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13016	if (!btf) {
13017		bpf_log(log,
13018			"FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13019		return -EINVAL;
13020	}
13021	t = btf_type_by_id(btf, btf_id);
13022	if (!t) {
13023		bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13024		return -EINVAL;
13025	}
13026	tname = btf_name_by_offset(btf, t->name_off);
13027	if (!tname) {
13028		bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13029		return -EINVAL;
13030	}
13031	if (tgt_prog) {
13032		struct bpf_prog_aux *aux = tgt_prog->aux;
13033
13034		for (i = 0; i < aux->func_info_cnt; i++)
13035			if (aux->func_info[i].type_id == btf_id) {
13036				subprog = i;
13037				break;
13038			}
13039		if (subprog == -1) {
13040			bpf_log(log, "Subprog %s doesn't exist\n", tname);
13041			return -EINVAL;
13042		}
13043		conservative = aux->func_info_aux[subprog].unreliable;
13044		if (prog_extension) {
13045			if (conservative) {
13046				bpf_log(log,
13047					"Cannot replace static functions\n");
13048				return -EINVAL;
13049			}
13050			if (!prog->jit_requested) {
13051				bpf_log(log,
13052					"Extension programs should be JITed\n");
13053				return -EINVAL;
13054			}
13055		}
13056		if (!tgt_prog->jited) {
13057			bpf_log(log, "Can attach to only JITed progs\n");
13058			return -EINVAL;
13059		}
13060		if (tgt_prog->type == prog->type) {
13061			/* Cannot fentry/fexit another fentry/fexit program.
13062			 * Cannot attach program extension to another extension.
13063			 * It's ok to attach fentry/fexit to extension program.
13064			 */
13065			bpf_log(log, "Cannot recursively attach\n");
13066			return -EINVAL;
13067		}
13068		if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13069		    prog_extension &&
13070		    (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13071		     tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13072			/* Program extensions can extend all program types
13073			 * except fentry/fexit. The reason is the following.
13074			 * The fentry/fexit programs are used for performance
13075			 * analysis, stats and can be attached to any program
13076			 * type except themselves. When extension program is
13077			 * replacing XDP function it is necessary to allow
13078			 * performance analysis of all functions. Both original
13079			 * XDP program and its program extension. Hence
13080			 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13081			 * allowed. If extending of fentry/fexit was allowed it
13082			 * would be possible to create long call chain
13083			 * fentry->extension->fentry->extension beyond
13084			 * reasonable stack size. Hence extending fentry is not
13085			 * allowed.
13086			 */
13087			bpf_log(log, "Cannot extend fentry/fexit\n");
13088			return -EINVAL;
13089		}
13090	} else {
13091		if (prog_extension) {
13092			bpf_log(log, "Cannot replace kernel functions\n");
13093			return -EINVAL;
13094		}
13095	}
13096
13097	switch (prog->expected_attach_type) {
13098	case BPF_TRACE_RAW_TP:
13099		if (tgt_prog) {
13100			bpf_log(log,
13101				"Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13102			return -EINVAL;
13103		}
13104		if (!btf_type_is_typedef(t)) {
13105			bpf_log(log, "attach_btf_id %u is not a typedef\n",
13106				btf_id);
13107			return -EINVAL;
13108		}
13109		if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13110			bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13111				btf_id, tname);
13112			return -EINVAL;
13113		}
13114		tname += sizeof(prefix) - 1;
13115		t = btf_type_by_id(btf, t->type);
13116		if (!btf_type_is_ptr(t))
13117			/* should never happen in valid vmlinux build */
13118			return -EINVAL;
13119		t = btf_type_by_id(btf, t->type);
13120		if (!btf_type_is_func_proto(t))
13121			/* should never happen in valid vmlinux build */
13122			return -EINVAL;
13123
13124		break;
13125	case BPF_TRACE_ITER:
13126		if (!btf_type_is_func(t)) {
13127			bpf_log(log, "attach_btf_id %u is not a function\n",
13128				btf_id);
13129			return -EINVAL;
13130		}
13131		t = btf_type_by_id(btf, t->type);
13132		if (!btf_type_is_func_proto(t))
13133			return -EINVAL;
13134		ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13135		if (ret)
13136			return ret;
13137		break;
13138	default:
13139		if (!prog_extension)
13140			return -EINVAL;
13141		fallthrough;
13142	case BPF_MODIFY_RETURN:
13143	case BPF_LSM_MAC:
13144	case BPF_TRACE_FENTRY:
13145	case BPF_TRACE_FEXIT:
13146		if (!btf_type_is_func(t)) {
13147			bpf_log(log, "attach_btf_id %u is not a function\n",
13148				btf_id);
13149			return -EINVAL;
13150		}
13151		if (prog_extension &&
13152		    btf_check_type_match(log, prog, btf, t))
13153			return -EINVAL;
13154		t = btf_type_by_id(btf, t->type);
13155		if (!btf_type_is_func_proto(t))
13156			return -EINVAL;
13157
13158		if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13159		    (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13160		     prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13161			return -EINVAL;
13162
13163		if (tgt_prog && conservative)
13164			t = NULL;
13165
13166		ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13167		if (ret < 0)
13168			return ret;
13169
13170		if (tgt_prog) {
13171			if (subprog == 0)
13172				addr = (long) tgt_prog->bpf_func;
13173			else
13174				addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13175		} else {
13176			addr = kallsyms_lookup_name(tname);
13177			if (!addr) {
13178				bpf_log(log,
13179					"The address of function %s cannot be found\n",
13180					tname);
13181				return -ENOENT;
13182			}
13183		}
13184
13185		if (prog->aux->sleepable) {
13186			ret = -EINVAL;
13187			switch (prog->type) {
13188			case BPF_PROG_TYPE_TRACING:
13189				/* fentry/fexit/fmod_ret progs can be sleepable only if they are
13190				 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13191				 */
13192				if (!check_non_sleepable_error_inject(btf_id) &&
13193				    within_error_injection_list(addr))
13194					ret = 0;
13195				break;
13196			case BPF_PROG_TYPE_LSM:
13197				/* LSM progs check that they are attached to bpf_lsm_*() funcs.
13198				 * Only some of them are sleepable.
13199				 */
13200				if (bpf_lsm_is_sleepable_hook(btf_id))
13201					ret = 0;
13202				break;
13203			default:
13204				break;
13205			}
13206			if (ret) {
13207				bpf_log(log, "%s is not sleepable\n", tname);
13208				return ret;
13209			}
13210		} else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13211			if (tgt_prog) {
13212				bpf_log(log, "can't modify return codes of BPF programs\n");
13213				return -EINVAL;
13214			}
13215			ret = check_attach_modify_return(addr, tname);
13216			if (ret) {
13217				bpf_log(log, "%s() is not modifiable\n", tname);
13218				return ret;
13219			}
13220		}
13221
13222		break;
13223	}
13224	tgt_info->tgt_addr = addr;
13225	tgt_info->tgt_name = tname;
13226	tgt_info->tgt_type = t;
13227	return 0;
13228}
13229
13230BTF_SET_START(btf_id_deny)
13231BTF_ID_UNUSED
13232#ifdef CONFIG_SMP
13233BTF_ID(func, migrate_disable)
13234BTF_ID(func, migrate_enable)
13235#endif
13236#if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13237BTF_ID(func, rcu_read_unlock_strict)
13238#endif
13239BTF_SET_END(btf_id_deny)
13240
13241static int check_attach_btf_id(struct bpf_verifier_env *env)
13242{
13243	struct bpf_prog *prog = env->prog;
13244	struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13245	struct bpf_attach_target_info tgt_info = {};
13246	u32 btf_id = prog->aux->attach_btf_id;
13247	struct bpf_trampoline *tr;
13248	int ret;
13249	u64 key;
13250
13251	if (prog->type == BPF_PROG_TYPE_SYSCALL) {
13252		if (prog->aux->sleepable)
13253			/* attach_btf_id checked to be zero already */
13254			return 0;
13255		verbose(env, "Syscall programs can only be sleepable\n");
13256		return -EINVAL;
13257	}
13258
13259	if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13260	    prog->type != BPF_PROG_TYPE_LSM) {
13261		verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13262		return -EINVAL;
13263	}
13264
13265	if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13266		return check_struct_ops_btf_id(env);
13267
13268	if (prog->type != BPF_PROG_TYPE_TRACING &&
13269	    prog->type != BPF_PROG_TYPE_LSM &&
13270	    prog->type != BPF_PROG_TYPE_EXT)
13271		return 0;
13272
13273	ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13274	if (ret)
13275		return ret;
13276
13277	if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13278		/* to make freplace equivalent to their targets, they need to
13279		 * inherit env->ops and expected_attach_type for the rest of the
13280		 * verification
13281		 */
13282		env->ops = bpf_verifier_ops[tgt_prog->type];
13283		prog->expected_attach_type = tgt_prog->expected_attach_type;
13284	}
13285
13286	/* store info about the attachment target that will be used later */
13287	prog->aux->attach_func_proto = tgt_info.tgt_type;
13288	prog->aux->attach_func_name = tgt_info.tgt_name;
13289
13290	if (tgt_prog) {
13291		prog->aux->saved_dst_prog_type = tgt_prog->type;
13292		prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13293	}
13294
13295	if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13296		prog->aux->attach_btf_trace = true;
13297		return 0;
13298	} else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13299		if (!bpf_iter_prog_supported(prog))
13300			return -EINVAL;
13301		return 0;
13302	}
13303
13304	if (prog->type == BPF_PROG_TYPE_LSM) {
13305		ret = bpf_lsm_verify_prog(&env->log, prog);
13306		if (ret < 0)
13307			return ret;
13308	} else if (prog->type == BPF_PROG_TYPE_TRACING &&
13309		   btf_id_set_contains(&btf_id_deny, btf_id)) {
13310		return -EINVAL;
13311	}
13312
13313	key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13314	tr = bpf_trampoline_get(key, &tgt_info);
13315	if (!tr)
13316		return -ENOMEM;
13317
13318	prog->aux->dst_trampoline = tr;
13319	return 0;
13320}
13321
13322struct btf *bpf_get_btf_vmlinux(void)
13323{
13324	if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13325		mutex_lock(&bpf_verifier_lock);
13326		if (!btf_vmlinux)
13327			btf_vmlinux = btf_parse_vmlinux();
13328		mutex_unlock(&bpf_verifier_lock);
13329	}
13330	return btf_vmlinux;
13331}
13332
13333int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
13334{
13335	u64 start_time = ktime_get_ns();
13336	struct bpf_verifier_env *env;
13337	struct bpf_verifier_log *log;
13338	int i, len, ret = -EINVAL;
13339	bool is_priv;
13340
13341	/* no program is valid */
13342	if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13343		return -EINVAL;
13344
13345	/* 'struct bpf_verifier_env' can be global, but since it's not small,
13346	 * allocate/free it every time bpf_check() is called
13347	 */
13348	env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13349	if (!env)
13350		return -ENOMEM;
13351	log = &env->log;
13352
13353	len = (*prog)->len;
13354	env->insn_aux_data =
13355		vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13356	ret = -ENOMEM;
13357	if (!env->insn_aux_data)
13358		goto err_free_env;
13359	for (i = 0; i < len; i++)
13360		env->insn_aux_data[i].orig_idx = i;
13361	env->prog = *prog;
13362	env->ops = bpf_verifier_ops[env->prog->type];
13363	env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
13364	is_priv = bpf_capable();
13365
13366	bpf_get_btf_vmlinux();
13367
13368	/* grab the mutex to protect few globals used by verifier */
13369	if (!is_priv)
13370		mutex_lock(&bpf_verifier_lock);
13371
13372	if (attr->log_level || attr->log_buf || attr->log_size) {
13373		/* user requested verbose verifier output
13374		 * and supplied buffer to store the verification trace
13375		 */
13376		log->level = attr->log_level;
13377		log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13378		log->len_total = attr->log_size;
13379
13380		ret = -EINVAL;
13381		/* log attributes have to be sane */
13382		if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
13383		    !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
13384			goto err_unlock;
13385	}
13386
13387	if (IS_ERR(btf_vmlinux)) {
13388		/* Either gcc or pahole or kernel are broken. */
13389		verbose(env, "in-kernel BTF is malformed\n");
13390		ret = PTR_ERR(btf_vmlinux);
13391		goto skip_full_check;
13392	}
13393
13394	env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13395	if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13396		env->strict_alignment = true;
13397	if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13398		env->strict_alignment = false;
13399
13400	env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13401	env->allow_uninit_stack = bpf_allow_uninit_stack();
13402	env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13403	env->bypass_spec_v1 = bpf_bypass_spec_v1();
13404	env->bypass_spec_v4 = bpf_bypass_spec_v4();
13405	env->bpf_capable = bpf_capable();
13406
13407	if (is_priv)
13408		env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13409
13410	env->explored_states = kvcalloc(state_htab_size(env),
13411				       sizeof(struct bpf_verifier_state_list *),
13412				       GFP_USER);
13413	ret = -ENOMEM;
13414	if (!env->explored_states)
13415		goto skip_full_check;
13416
13417	ret = add_subprog_and_kfunc(env);
13418	if (ret < 0)
13419		goto skip_full_check;
13420
13421	ret = check_subprogs(env);
13422	if (ret < 0)
13423		goto skip_full_check;
13424
13425	ret = check_btf_info(env, attr, uattr);
13426	if (ret < 0)
13427		goto skip_full_check;
13428
13429	ret = check_attach_btf_id(env);
13430	if (ret)
13431		goto skip_full_check;
13432
13433	ret = resolve_pseudo_ldimm64(env);
13434	if (ret < 0)
13435		goto skip_full_check;
13436
13437	if (bpf_prog_is_dev_bound(env->prog->aux)) {
13438		ret = bpf_prog_offload_verifier_prep(env->prog);
13439		if (ret)
13440			goto skip_full_check;
13441	}
13442
13443	ret = check_cfg(env);
13444	if (ret < 0)
13445		goto skip_full_check;
13446
13447	ret = do_check_subprogs(env);
13448	ret = ret ?: do_check_main(env);
13449
13450	if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13451		ret = bpf_prog_offload_finalize(env);
13452
13453skip_full_check:
13454	kvfree(env->explored_states);
13455
13456	if (ret == 0)
13457		ret = check_max_stack_depth(env);
13458
13459	/* instruction rewrites happen after this point */
13460	if (is_priv) {
13461		if (ret == 0)
13462			opt_hard_wire_dead_code_branches(env);
13463		if (ret == 0)
13464			ret = opt_remove_dead_code(env);
13465		if (ret == 0)
13466			ret = opt_remove_nops(env);
13467	} else {
13468		if (ret == 0)
13469			sanitize_dead_code(env);
13470	}
13471
13472	if (ret == 0)
13473		/* program is valid, convert *(u32*)(ctx + off) accesses */
13474		ret = convert_ctx_accesses(env);
13475
13476	if (ret == 0)
13477		ret = do_misc_fixups(env);
13478
13479	/* do 32-bit optimization after insn patching has done so those patched
13480	 * insns could be handled correctly.
13481	 */
13482	if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13483		ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13484		env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13485								     : false;
13486	}
13487
13488	if (ret == 0)
13489		ret = fixup_call_args(env);
13490
13491	env->verification_time = ktime_get_ns() - start_time;
13492	print_verification_stats(env);
13493
13494	if (log->level && bpf_verifier_log_full(log))
13495		ret = -ENOSPC;
13496	if (log->level && !log->ubuf) {
13497		ret = -EFAULT;
13498		goto err_release_maps;
13499	}
13500
13501	if (ret)
13502		goto err_release_maps;
13503
13504	if (env->used_map_cnt) {
13505		/* if program passed verifier, update used_maps in bpf_prog_info */
13506		env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13507							  sizeof(env->used_maps[0]),
13508							  GFP_KERNEL);
13509
13510		if (!env->prog->aux->used_maps) {
13511			ret = -ENOMEM;
13512			goto err_release_maps;
13513		}
13514
13515		memcpy(env->prog->aux->used_maps, env->used_maps,
13516		       sizeof(env->used_maps[0]) * env->used_map_cnt);
13517		env->prog->aux->used_map_cnt = env->used_map_cnt;
13518	}
13519	if (env->used_btf_cnt) {
13520		/* if program passed verifier, update used_btfs in bpf_prog_aux */
13521		env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13522							  sizeof(env->used_btfs[0]),
13523							  GFP_KERNEL);
13524		if (!env->prog->aux->used_btfs) {
13525			ret = -ENOMEM;
13526			goto err_release_maps;
13527		}
13528
13529		memcpy(env->prog->aux->used_btfs, env->used_btfs,
13530		       sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13531		env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13532	}
13533	if (env->used_map_cnt || env->used_btf_cnt) {
13534		/* program is valid. Convert pseudo bpf_ld_imm64 into generic
13535		 * bpf_ld_imm64 instructions
13536		 */
13537		convert_pseudo_ld_imm64(env);
13538	}
13539
13540	adjust_btf_func(env);
13541
13542err_release_maps:
13543	if (!env->prog->aux->used_maps)
13544		/* if we didn't copy map pointers into bpf_prog_info, release
13545		 * them now. Otherwise free_used_maps() will release them.
13546		 */
13547		release_maps(env);
13548	if (!env->prog->aux->used_btfs)
13549		release_btfs(env);
13550
13551	/* extension progs temporarily inherit the attach_type of their targets
13552	   for verification purposes, so set it back to zero before returning
13553	 */
13554	if (env->prog->type == BPF_PROG_TYPE_EXT)
13555		env->prog->expected_attach_type = 0;
13556
13557	*prog = env->prog;
13558err_unlock:
13559	if (!is_priv)
13560		mutex_unlock(&bpf_verifier_lock);
13561	vfree(env->insn_aux_data);
13562err_free_env:
13563	kfree(env);
13564	return ret;
13565}