Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 *  Kernel Probes (KProbes)
   4 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   5 * Copyright (C) IBM Corporation, 2002, 2004
   6 *
   7 * 2002-Oct	Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
   8 *		Probes initial implementation ( includes contributions from
   9 *		Rusty Russell).
  10 * 2004-July	Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
  11 *		interface to access function arguments.
  12 * 2004-Oct	Jim Keniston <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  13 *		<prasanna@in.ibm.com> adapted for x86_64 from i386.
  14 * 2005-Mar	Roland McGrath <roland@redhat.com>
  15 *		Fixed to handle %rip-relative addressing mode correctly.
  16 * 2005-May	Hien Nguyen <hien@us.ibm.com>, Jim Keniston
  17 *		<jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  18 *		<prasanna@in.ibm.com> added function-return probes.
  19 * 2005-May	Rusty Lynch <rusty.lynch@intel.com>
  20 *		Added function return probes functionality
  21 * 2006-Feb	Masami Hiramatsu <hiramatu@sdl.hitachi.co.jp> added
  22 *		kprobe-booster and kretprobe-booster for i386.
  23 * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com> added kprobe-booster
  24 *		and kretprobe-booster for x86-64
  25 * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com>, Arjan van de Ven
  26 *		<arjan@infradead.org> and Jim Keniston <jkenisto@us.ibm.com>
  27 *		unified x86 kprobes code.
  28 */
  29#include <linux/kprobes.h>
  30#include <linux/ptrace.h>
  31#include <linux/string.h>
  32#include <linux/slab.h>
  33#include <linux/hardirq.h>
  34#include <linux/preempt.h>
  35#include <linux/sched/debug.h>
  36#include <linux/extable.h>
  37#include <linux/kdebug.h>
  38#include <linux/kallsyms.h>
  39#include <linux/ftrace.h>
  40#include <linux/frame.h>
  41#include <linux/kasan.h>
  42#include <linux/moduleloader.h>
  43
  44#include <asm/text-patching.h>
  45#include <asm/cacheflush.h>
  46#include <asm/desc.h>
  47#include <asm/pgtable.h>
  48#include <linux/uaccess.h>
  49#include <asm/alternative.h>
  50#include <asm/insn.h>
  51#include <asm/debugreg.h>
  52#include <asm/set_memory.h>
  53
  54#include "common.h"
  55
 
 
  56DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  57DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  58
  59#define stack_addr(regs) ((unsigned long *)regs->sp)
  60
  61#define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
  62	(((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  63	  (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  64	  (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  65	  (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  66	 << (row % 32))
  67	/*
  68	 * Undefined/reserved opcodes, conditional jump, Opcode Extension
  69	 * Groups, and some special opcodes can not boost.
  70	 * This is non-const and volatile to keep gcc from statically
  71	 * optimizing it out, as variable_test_bit makes gcc think only
  72	 * *(unsigned long*) is used.
  73	 */
  74static volatile u32 twobyte_is_boostable[256 / 32] = {
  75	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  76	/*      ----------------------------------------------          */
  77	W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
  78	W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
  79	W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
  80	W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
  81	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
  82	W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
  83	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
  84	W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
  85	W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
  86	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
  87	W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
  88	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
  89	W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
  90	W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
  91	W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
  92	W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
  93	/*      -----------------------------------------------         */
  94	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  95};
  96#undef W
  97
  98struct kretprobe_blackpoint kretprobe_blacklist[] = {
  99	{"__switch_to", }, /* This function switches only current task, but
 100			      doesn't switch kernel stack.*/
 101	{NULL, NULL}	/* Terminator */
 102};
 103
 104const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
 105
 106static nokprobe_inline void
 107__synthesize_relative_insn(void *dest, void *from, void *to, u8 op)
 108{
 109	struct __arch_relative_insn {
 110		u8 op;
 111		s32 raddr;
 112	} __packed *insn;
 113
 114	insn = (struct __arch_relative_insn *)dest;
 115	insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
 116	insn->op = op;
 117}
 118
 119/* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
 120void synthesize_reljump(void *dest, void *from, void *to)
 121{
 122	__synthesize_relative_insn(dest, from, to, RELATIVEJUMP_OPCODE);
 123}
 124NOKPROBE_SYMBOL(synthesize_reljump);
 125
 126/* Insert a call instruction at address 'from', which calls address 'to'.*/
 127void synthesize_relcall(void *dest, void *from, void *to)
 128{
 129	__synthesize_relative_insn(dest, from, to, RELATIVECALL_OPCODE);
 130}
 131NOKPROBE_SYMBOL(synthesize_relcall);
 132
 133/*
 134 * Skip the prefixes of the instruction.
 135 */
 136static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
 137{
 138	insn_attr_t attr;
 139
 140	attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 141	while (inat_is_legacy_prefix(attr)) {
 142		insn++;
 143		attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 144	}
 145#ifdef CONFIG_X86_64
 146	if (inat_is_rex_prefix(attr))
 147		insn++;
 148#endif
 149	return insn;
 150}
 151NOKPROBE_SYMBOL(skip_prefixes);
 152
 153/*
 154 * Returns non-zero if INSN is boostable.
 155 * RIP relative instructions are adjusted at copying time in 64 bits mode
 156 */
 157int can_boost(struct insn *insn, void *addr)
 158{
 159	kprobe_opcode_t opcode;
 160
 161	if (search_exception_tables((unsigned long)addr))
 162		return 0;	/* Page fault may occur on this address. */
 163
 164	/* 2nd-byte opcode */
 165	if (insn->opcode.nbytes == 2)
 166		return test_bit(insn->opcode.bytes[1],
 167				(unsigned long *)twobyte_is_boostable);
 168
 169	if (insn->opcode.nbytes != 1)
 170		return 0;
 171
 172	/* Can't boost Address-size override prefix */
 173	if (unlikely(inat_is_address_size_prefix(insn->attr)))
 174		return 0;
 175
 176	opcode = insn->opcode.bytes[0];
 177
 178	switch (opcode & 0xf0) {
 179	case 0x60:
 180		/* can't boost "bound" */
 181		return (opcode != 0x62);
 182	case 0x70:
 183		return 0; /* can't boost conditional jump */
 184	case 0x90:
 185		return opcode != 0x9a;	/* can't boost call far */
 186	case 0xc0:
 187		/* can't boost software-interruptions */
 188		return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 189	case 0xd0:
 190		/* can boost AA* and XLAT */
 191		return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 192	case 0xe0:
 193		/* can boost in/out and absolute jmps */
 194		return ((opcode & 0x04) || opcode == 0xea);
 195	case 0xf0:
 196		/* clear and set flags are boostable */
 197		return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 198	default:
 199		/* CS override prefix and call are not boostable */
 200		return (opcode != 0x2e && opcode != 0x9a);
 201	}
 202}
 203
 204static unsigned long
 205__recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
 206{
 207	struct kprobe *kp;
 208	unsigned long faddr;
 209
 210	kp = get_kprobe((void *)addr);
 211	faddr = ftrace_location(addr);
 212	/*
 213	 * Addresses inside the ftrace location are refused by
 214	 * arch_check_ftrace_location(). Something went terribly wrong
 215	 * if such an address is checked here.
 216	 */
 217	if (WARN_ON(faddr && faddr != addr))
 218		return 0UL;
 219	/*
 220	 * Use the current code if it is not modified by Kprobe
 221	 * and it cannot be modified by ftrace.
 222	 */
 223	if (!kp && !faddr)
 224		return addr;
 225
 226	/*
 227	 * Basically, kp->ainsn.insn has an original instruction.
 228	 * However, RIP-relative instruction can not do single-stepping
 229	 * at different place, __copy_instruction() tweaks the displacement of
 230	 * that instruction. In that case, we can't recover the instruction
 231	 * from the kp->ainsn.insn.
 232	 *
 233	 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
 234	 * of the first byte of the probed instruction, which is overwritten
 235	 * by int3. And the instruction at kp->addr is not modified by kprobes
 236	 * except for the first byte, we can recover the original instruction
 237	 * from it and kp->opcode.
 238	 *
 239	 * In case of Kprobes using ftrace, we do not have a copy of
 240	 * the original instruction. In fact, the ftrace location might
 241	 * be modified at anytime and even could be in an inconsistent state.
 242	 * Fortunately, we know that the original code is the ideal 5-byte
 243	 * long NOP.
 244	 */
 245	if (probe_kernel_read(buf, (void *)addr,
 246		MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
 247		return 0UL;
 248
 249	if (faddr)
 250		memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
 251	else
 252		buf[0] = kp->opcode;
 253	return (unsigned long)buf;
 254}
 255
 256/*
 257 * Recover the probed instruction at addr for further analysis.
 258 * Caller must lock kprobes by kprobe_mutex, or disable preemption
 259 * for preventing to release referencing kprobes.
 260 * Returns zero if the instruction can not get recovered (or access failed).
 261 */
 262unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
 263{
 264	unsigned long __addr;
 265
 266	__addr = __recover_optprobed_insn(buf, addr);
 267	if (__addr != addr)
 268		return __addr;
 269
 270	return __recover_probed_insn(buf, addr);
 271}
 272
 273/* Check if paddr is at an instruction boundary */
 274static int can_probe(unsigned long paddr)
 275{
 276	unsigned long addr, __addr, offset = 0;
 277	struct insn insn;
 278	kprobe_opcode_t buf[MAX_INSN_SIZE];
 279
 280	if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
 281		return 0;
 282
 283	/* Decode instructions */
 284	addr = paddr - offset;
 285	while (addr < paddr) {
 286		/*
 287		 * Check if the instruction has been modified by another
 288		 * kprobe, in which case we replace the breakpoint by the
 289		 * original instruction in our buffer.
 290		 * Also, jump optimization will change the breakpoint to
 291		 * relative-jump. Since the relative-jump itself is
 292		 * normally used, we just go through if there is no kprobe.
 293		 */
 294		__addr = recover_probed_instruction(buf, addr);
 295		if (!__addr)
 296			return 0;
 297		kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
 298		insn_get_length(&insn);
 299
 300		/*
 301		 * Another debugging subsystem might insert this breakpoint.
 302		 * In that case, we can't recover it.
 303		 */
 304		if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 305			return 0;
 306		addr += insn.length;
 307	}
 308
 309	return (addr == paddr);
 310}
 311
 312/*
 313 * Returns non-zero if opcode modifies the interrupt flag.
 314 */
 315static int is_IF_modifier(kprobe_opcode_t *insn)
 316{
 317	/* Skip prefixes */
 318	insn = skip_prefixes(insn);
 319
 320	switch (*insn) {
 321	case 0xfa:		/* cli */
 322	case 0xfb:		/* sti */
 323	case 0xcf:		/* iret/iretd */
 324	case 0x9d:		/* popf/popfd */
 325		return 1;
 326	}
 327
 328	return 0;
 329}
 330
 331/*
 332 * Copy an instruction with recovering modified instruction by kprobes
 333 * and adjust the displacement if the instruction uses the %rip-relative
 334 * addressing mode. Note that since @real will be the final place of copied
 335 * instruction, displacement must be adjust by @real, not @dest.
 336 * This returns the length of copied instruction, or 0 if it has an error.
 337 */
 338int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn)
 339{
 340	kprobe_opcode_t buf[MAX_INSN_SIZE];
 341	unsigned long recovered_insn =
 342		recover_probed_instruction(buf, (unsigned long)src);
 343
 344	if (!recovered_insn || !insn)
 345		return 0;
 346
 347	/* This can access kernel text if given address is not recovered */
 348	if (probe_kernel_read(dest, (void *)recovered_insn, MAX_INSN_SIZE))
 349		return 0;
 350
 351	kernel_insn_init(insn, dest, MAX_INSN_SIZE);
 352	insn_get_length(insn);
 353
 354	/* Another subsystem puts a breakpoint, failed to recover */
 355	if (insn->opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 356		return 0;
 357
 358	/* We should not singlestep on the exception masking instructions */
 359	if (insn_masking_exception(insn))
 360		return 0;
 361
 362#ifdef CONFIG_X86_64
 363	/* Only x86_64 has RIP relative instructions */
 364	if (insn_rip_relative(insn)) {
 365		s64 newdisp;
 366		u8 *disp;
 367		/*
 368		 * The copied instruction uses the %rip-relative addressing
 369		 * mode.  Adjust the displacement for the difference between
 370		 * the original location of this instruction and the location
 371		 * of the copy that will actually be run.  The tricky bit here
 372		 * is making sure that the sign extension happens correctly in
 373		 * this calculation, since we need a signed 32-bit result to
 374		 * be sign-extended to 64 bits when it's added to the %rip
 375		 * value and yield the same 64-bit result that the sign-
 376		 * extension of the original signed 32-bit displacement would
 377		 * have given.
 378		 */
 379		newdisp = (u8 *) src + (s64) insn->displacement.value
 380			  - (u8 *) real;
 381		if ((s64) (s32) newdisp != newdisp) {
 382			pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
 
 
 383			return 0;
 384		}
 385		disp = (u8 *) dest + insn_offset_displacement(insn);
 386		*(s32 *) disp = (s32) newdisp;
 387	}
 388#endif
 389	return insn->length;
 390}
 391
 392/* Prepare reljump right after instruction to boost */
 393static int prepare_boost(kprobe_opcode_t *buf, struct kprobe *p,
 394			  struct insn *insn)
 395{
 396	int len = insn->length;
 397
 398	if (can_boost(insn, p->addr) &&
 399	    MAX_INSN_SIZE - len >= RELATIVEJUMP_SIZE) {
 400		/*
 401		 * These instructions can be executed directly if it
 402		 * jumps back to correct address.
 403		 */
 404		synthesize_reljump(buf + len, p->ainsn.insn + len,
 405				   p->addr + insn->length);
 406		len += RELATIVEJUMP_SIZE;
 407		p->ainsn.boostable = true;
 408	} else {
 409		p->ainsn.boostable = false;
 410	}
 411
 412	return len;
 413}
 414
 415/* Make page to RO mode when allocate it */
 416void *alloc_insn_page(void)
 417{
 418	void *page;
 419
 420	page = module_alloc(PAGE_SIZE);
 421	if (!page)
 422		return NULL;
 423
 424	set_vm_flush_reset_perms(page);
 425	/*
 426	 * First make the page read-only, and only then make it executable to
 427	 * prevent it from being W+X in between.
 428	 */
 429	set_memory_ro((unsigned long)page, 1);
 430
 431	/*
 432	 * TODO: Once additional kernel code protection mechanisms are set, ensure
 433	 * that the page was not maliciously altered and it is still zeroed.
 434	 */
 435	set_memory_x((unsigned long)page, 1);
 436
 437	return page;
 438}
 439
 440/* Recover page to RW mode before releasing it */
 441void free_insn_page(void *page)
 442{
 
 
 443	module_memfree(page);
 444}
 445
 446static int arch_copy_kprobe(struct kprobe *p)
 447{
 448	struct insn insn;
 449	kprobe_opcode_t buf[MAX_INSN_SIZE];
 450	int len;
 451
 452	/* Copy an instruction with recovering if other optprobe modifies it.*/
 453	len = __copy_instruction(buf, p->addr, p->ainsn.insn, &insn);
 454	if (!len)
 455		return -EINVAL;
 456
 457	/*
 458	 * __copy_instruction can modify the displacement of the instruction,
 459	 * but it doesn't affect boostable check.
 460	 */
 461	len = prepare_boost(buf, p, &insn);
 462
 463	/* Check whether the instruction modifies Interrupt Flag or not */
 464	p->ainsn.if_modifier = is_IF_modifier(buf);
 465
 466	/* Also, displacement change doesn't affect the first byte */
 467	p->opcode = buf[0];
 468
 469	/* OK, write back the instruction(s) into ROX insn buffer */
 470	text_poke(p->ainsn.insn, buf, len);
 471
 472	return 0;
 473}
 474
 475int arch_prepare_kprobe(struct kprobe *p)
 476{
 477	int ret;
 478
 479	if (alternatives_text_reserved(p->addr, p->addr))
 480		return -EINVAL;
 481
 482	if (!can_probe((unsigned long)p->addr))
 483		return -EILSEQ;
 484	/* insn: must be on special executable page on x86. */
 485	p->ainsn.insn = get_insn_slot();
 486	if (!p->ainsn.insn)
 487		return -ENOMEM;
 488
 489	ret = arch_copy_kprobe(p);
 490	if (ret) {
 491		free_insn_slot(p->ainsn.insn, 0);
 492		p->ainsn.insn = NULL;
 493	}
 494
 495	return ret;
 496}
 497
 498void arch_arm_kprobe(struct kprobe *p)
 499{
 500	text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 501}
 502
 503void arch_disarm_kprobe(struct kprobe *p)
 504{
 505	text_poke(p->addr, &p->opcode, 1);
 506}
 507
 508void arch_remove_kprobe(struct kprobe *p)
 509{
 510	if (p->ainsn.insn) {
 511		free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
 512		p->ainsn.insn = NULL;
 513	}
 514}
 515
 516static nokprobe_inline void
 517save_previous_kprobe(struct kprobe_ctlblk *kcb)
 518{
 519	kcb->prev_kprobe.kp = kprobe_running();
 520	kcb->prev_kprobe.status = kcb->kprobe_status;
 521	kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
 522	kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
 523}
 524
 525static nokprobe_inline void
 526restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 527{
 528	__this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
 529	kcb->kprobe_status = kcb->prev_kprobe.status;
 530	kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
 531	kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
 532}
 533
 534static nokprobe_inline void
 535set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 536		   struct kprobe_ctlblk *kcb)
 537{
 538	__this_cpu_write(current_kprobe, p);
 539	kcb->kprobe_saved_flags = kcb->kprobe_old_flags
 540		= (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 541	if (p->ainsn.if_modifier)
 542		kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
 543}
 544
 545static nokprobe_inline void clear_btf(void)
 546{
 547	if (test_thread_flag(TIF_BLOCKSTEP)) {
 548		unsigned long debugctl = get_debugctlmsr();
 549
 550		debugctl &= ~DEBUGCTLMSR_BTF;
 551		update_debugctlmsr(debugctl);
 552	}
 553}
 554
 555static nokprobe_inline void restore_btf(void)
 556{
 557	if (test_thread_flag(TIF_BLOCKSTEP)) {
 558		unsigned long debugctl = get_debugctlmsr();
 559
 560		debugctl |= DEBUGCTLMSR_BTF;
 561		update_debugctlmsr(debugctl);
 562	}
 563}
 564
 565void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
 566{
 567	unsigned long *sara = stack_addr(regs);
 568
 569	ri->ret_addr = (kprobe_opcode_t *) *sara;
 570	ri->fp = sara;
 571
 572	/* Replace the return addr with trampoline addr */
 573	*sara = (unsigned long) &kretprobe_trampoline;
 574}
 575NOKPROBE_SYMBOL(arch_prepare_kretprobe);
 576
 577static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
 578			     struct kprobe_ctlblk *kcb, int reenter)
 579{
 580	if (setup_detour_execution(p, regs, reenter))
 581		return;
 582
 583#if !defined(CONFIG_PREEMPTION)
 584	if (p->ainsn.boostable && !p->post_handler) {
 585		/* Boost up -- we can execute copied instructions directly */
 586		if (!reenter)
 587			reset_current_kprobe();
 588		/*
 589		 * Reentering boosted probe doesn't reset current_kprobe,
 590		 * nor set current_kprobe, because it doesn't use single
 591		 * stepping.
 592		 */
 593		regs->ip = (unsigned long)p->ainsn.insn;
 
 594		return;
 595	}
 596#endif
 597	if (reenter) {
 598		save_previous_kprobe(kcb);
 599		set_current_kprobe(p, regs, kcb);
 600		kcb->kprobe_status = KPROBE_REENTER;
 601	} else
 602		kcb->kprobe_status = KPROBE_HIT_SS;
 603	/* Prepare real single stepping */
 604	clear_btf();
 605	regs->flags |= X86_EFLAGS_TF;
 606	regs->flags &= ~X86_EFLAGS_IF;
 607	/* single step inline if the instruction is an int3 */
 608	if (p->opcode == BREAKPOINT_INSTRUCTION)
 609		regs->ip = (unsigned long)p->addr;
 610	else
 611		regs->ip = (unsigned long)p->ainsn.insn;
 612}
 613NOKPROBE_SYMBOL(setup_singlestep);
 614
 615/*
 616 * We have reentered the kprobe_handler(), since another probe was hit while
 617 * within the handler. We save the original kprobes variables and just single
 618 * step on the instruction of the new probe without calling any user handlers.
 619 */
 620static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
 621			  struct kprobe_ctlblk *kcb)
 622{
 623	switch (kcb->kprobe_status) {
 624	case KPROBE_HIT_SSDONE:
 625	case KPROBE_HIT_ACTIVE:
 626	case KPROBE_HIT_SS:
 627		kprobes_inc_nmissed_count(p);
 628		setup_singlestep(p, regs, kcb, 1);
 629		break;
 630	case KPROBE_REENTER:
 631		/* A probe has been hit in the codepath leading up to, or just
 632		 * after, single-stepping of a probed instruction. This entire
 633		 * codepath should strictly reside in .kprobes.text section.
 634		 * Raise a BUG or we'll continue in an endless reentering loop
 635		 * and eventually a stack overflow.
 636		 */
 637		pr_err("Unrecoverable kprobe detected.\n");
 
 638		dump_kprobe(p);
 639		BUG();
 640	default:
 641		/* impossible cases */
 642		WARN_ON(1);
 643		return 0;
 644	}
 645
 646	return 1;
 647}
 648NOKPROBE_SYMBOL(reenter_kprobe);
 649
 650/*
 651 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 652 * remain disabled throughout this function.
 653 */
 654int kprobe_int3_handler(struct pt_regs *regs)
 655{
 656	kprobe_opcode_t *addr;
 657	struct kprobe *p;
 658	struct kprobe_ctlblk *kcb;
 659
 660	if (user_mode(regs))
 661		return 0;
 662
 663	addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 664	/*
 665	 * We don't want to be preempted for the entire duration of kprobe
 666	 * processing. Since int3 and debug trap disables irqs and we clear
 667	 * IF while singlestepping, it must be no preemptible.
 
 668	 */
 
 669
 670	kcb = get_kprobe_ctlblk();
 671	p = get_kprobe(addr);
 672
 673	if (p) {
 674		if (kprobe_running()) {
 675			if (reenter_kprobe(p, regs, kcb))
 676				return 1;
 677		} else {
 678			set_current_kprobe(p, regs, kcb);
 679			kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 680
 681			/*
 682			 * If we have no pre-handler or it returned 0, we
 683			 * continue with normal processing.  If we have a
 684			 * pre-handler and it returned non-zero, that means
 685			 * user handler setup registers to exit to another
 686			 * instruction, we must skip the single stepping.
 
 687			 */
 688			if (!p->pre_handler || !p->pre_handler(p, regs))
 689				setup_singlestep(p, regs, kcb, 0);
 690			else
 691				reset_current_kprobe();
 692			return 1;
 693		}
 694	} else if (*addr != BREAKPOINT_INSTRUCTION) {
 695		/*
 696		 * The breakpoint instruction was removed right
 697		 * after we hit it.  Another cpu has removed
 698		 * either a probepoint or a debugger breakpoint
 699		 * at this address.  In either case, no further
 700		 * handling of this interrupt is appropriate.
 701		 * Back up over the (now missing) int3 and run
 702		 * the original instruction.
 703		 */
 704		regs->ip = (unsigned long)addr;
 
 705		return 1;
 
 
 
 
 
 
 
 706	} /* else: not a kprobe fault; let the kernel handle it */
 707
 
 708	return 0;
 709}
 710NOKPROBE_SYMBOL(kprobe_int3_handler);
 711
 712/*
 713 * When a retprobed function returns, this code saves registers and
 714 * calls trampoline_handler() runs, which calls the kretprobe's handler.
 715 */
 716asm(
 717	".text\n"
 718	".global kretprobe_trampoline\n"
 719	".type kretprobe_trampoline, @function\n"
 720	"kretprobe_trampoline:\n"
 721	/* We don't bother saving the ss register */
 722#ifdef CONFIG_X86_64
 
 723	"	pushq %rsp\n"
 724	"	pushfq\n"
 725	SAVE_REGS_STRING
 726	"	movq %rsp, %rdi\n"
 727	"	call trampoline_handler\n"
 728	/* Replace saved sp with true return address. */
 729	"	movq %rax, 19*8(%rsp)\n"
 730	RESTORE_REGS_STRING
 731	"	popfq\n"
 732#else
 733	"	pushl %esp\n"
 734	"	pushfl\n"
 735	SAVE_REGS_STRING
 736	"	movl %esp, %eax\n"
 737	"	call trampoline_handler\n"
 738	/* Replace saved sp with true return address. */
 739	"	movl %eax, 15*4(%esp)\n"
 
 
 
 740	RESTORE_REGS_STRING
 741	"	popfl\n"
 742#endif
 743	"	ret\n"
 744	".size kretprobe_trampoline, .-kretprobe_trampoline\n"
 745);
 746NOKPROBE_SYMBOL(kretprobe_trampoline);
 747STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
 748
 749static struct kprobe kretprobe_kprobe = {
 750	.addr = (void *)kretprobe_trampoline,
 751};
 752
 753/*
 754 * Called from kretprobe_trampoline
 755 */
 756__used __visible void *trampoline_handler(struct pt_regs *regs)
 757{
 758	struct kprobe_ctlblk *kcb;
 759	struct kretprobe_instance *ri = NULL;
 760	struct hlist_head *head, empty_rp;
 761	struct hlist_node *tmp;
 762	unsigned long flags, orig_ret_address = 0;
 763	unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 764	kprobe_opcode_t *correct_ret_addr = NULL;
 765	void *frame_pointer;
 766	bool skipped = false;
 767
 768	preempt_disable();
 769
 770	/*
 771	 * Set a dummy kprobe for avoiding kretprobe recursion.
 772	 * Since kretprobe never run in kprobe handler, kprobe must not
 773	 * be running at this point.
 774	 */
 775	kcb = get_kprobe_ctlblk();
 776	__this_cpu_write(current_kprobe, &kretprobe_kprobe);
 777	kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 778
 779	INIT_HLIST_HEAD(&empty_rp);
 780	kretprobe_hash_lock(current, &head, &flags);
 781	/* fixup registers */
 
 782	regs->cs = __KERNEL_CS;
 783#ifdef CONFIG_X86_32
 784	regs->cs |= get_kernel_rpl();
 785	regs->gs = 0;
 786#endif
 787	/* We use pt_regs->sp for return address holder. */
 788	frame_pointer = &regs->sp;
 789	regs->ip = trampoline_address;
 790	regs->orig_ax = ~0UL;
 791
 792	/*
 793	 * It is possible to have multiple instances associated with a given
 794	 * task either because multiple functions in the call path have
 795	 * return probes installed on them, and/or more than one
 796	 * return probe was registered for a target function.
 797	 *
 798	 * We can handle this because:
 799	 *     - instances are always pushed into the head of the list
 800	 *     - when multiple return probes are registered for the same
 801	 *	 function, the (chronologically) first instance's ret_addr
 802	 *	 will be the real return address, and all the rest will
 803	 *	 point to kretprobe_trampoline.
 804	 */
 805	hlist_for_each_entry(ri, head, hlist) {
 806		if (ri->task != current)
 807			/* another task is sharing our hash bucket */
 808			continue;
 809		/*
 810		 * Return probes must be pushed on this hash list correct
 811		 * order (same as return order) so that it can be popped
 812		 * correctly. However, if we find it is pushed it incorrect
 813		 * order, this means we find a function which should not be
 814		 * probed, because the wrong order entry is pushed on the
 815		 * path of processing other kretprobe itself.
 816		 */
 817		if (ri->fp != frame_pointer) {
 818			if (!skipped)
 819				pr_warn("kretprobe is stacked incorrectly. Trying to fixup.\n");
 820			skipped = true;
 821			continue;
 822		}
 823
 824		orig_ret_address = (unsigned long)ri->ret_addr;
 825		if (skipped)
 826			pr_warn("%ps must be blacklisted because of incorrect kretprobe order\n",
 827				ri->rp->kp.addr);
 828
 829		if (orig_ret_address != trampoline_address)
 830			/*
 831			 * This is the real return address. Any other
 832			 * instances associated with this task are for
 833			 * other calls deeper on the call stack
 834			 */
 835			break;
 836	}
 837
 838	kretprobe_assert(ri, orig_ret_address, trampoline_address);
 839
 840	correct_ret_addr = ri->ret_addr;
 841	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
 842		if (ri->task != current)
 843			/* another task is sharing our hash bucket */
 844			continue;
 845		if (ri->fp != frame_pointer)
 846			continue;
 847
 848		orig_ret_address = (unsigned long)ri->ret_addr;
 849		if (ri->rp && ri->rp->handler) {
 850			__this_cpu_write(current_kprobe, &ri->rp->kp);
 
 851			ri->ret_addr = correct_ret_addr;
 852			ri->rp->handler(ri, regs);
 853			__this_cpu_write(current_kprobe, &kretprobe_kprobe);
 854		}
 855
 856		recycle_rp_inst(ri, &empty_rp);
 857
 858		if (orig_ret_address != trampoline_address)
 859			/*
 860			 * This is the real return address. Any other
 861			 * instances associated with this task are for
 862			 * other calls deeper on the call stack
 863			 */
 864			break;
 865	}
 866
 867	kretprobe_hash_unlock(current, &flags);
 868
 869	__this_cpu_write(current_kprobe, NULL);
 870	preempt_enable();
 871
 872	hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
 873		hlist_del(&ri->hlist);
 874		kfree(ri);
 875	}
 876	return (void *)orig_ret_address;
 877}
 878NOKPROBE_SYMBOL(trampoline_handler);
 879
 880/*
 881 * Called after single-stepping.  p->addr is the address of the
 882 * instruction whose first byte has been replaced by the "int 3"
 883 * instruction.  To avoid the SMP problems that can occur when we
 884 * temporarily put back the original opcode to single-step, we
 885 * single-stepped a copy of the instruction.  The address of this
 886 * copy is p->ainsn.insn.
 887 *
 888 * This function prepares to return from the post-single-step
 889 * interrupt.  We have to fix up the stack as follows:
 890 *
 891 * 0) Except in the case of absolute or indirect jump or call instructions,
 892 * the new ip is relative to the copied instruction.  We need to make
 893 * it relative to the original instruction.
 894 *
 895 * 1) If the single-stepped instruction was pushfl, then the TF and IF
 896 * flags are set in the just-pushed flags, and may need to be cleared.
 897 *
 898 * 2) If the single-stepped instruction was a call, the return address
 899 * that is atop the stack is the address following the copied instruction.
 900 * We need to make it the address following the original instruction.
 901 *
 902 * If this is the first time we've single-stepped the instruction at
 903 * this probepoint, and the instruction is boostable, boost it: add a
 904 * jump instruction after the copied instruction, that jumps to the next
 905 * instruction after the probepoint.
 906 */
 907static void resume_execution(struct kprobe *p, struct pt_regs *regs,
 908			     struct kprobe_ctlblk *kcb)
 909{
 910	unsigned long *tos = stack_addr(regs);
 911	unsigned long copy_ip = (unsigned long)p->ainsn.insn;
 912	unsigned long orig_ip = (unsigned long)p->addr;
 913	kprobe_opcode_t *insn = p->ainsn.insn;
 914
 915	/* Skip prefixes */
 916	insn = skip_prefixes(insn);
 917
 918	regs->flags &= ~X86_EFLAGS_TF;
 919	switch (*insn) {
 920	case 0x9c:	/* pushfl */
 921		*tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
 922		*tos |= kcb->kprobe_old_flags;
 923		break;
 924	case 0xc2:	/* iret/ret/lret */
 925	case 0xc3:
 926	case 0xca:
 927	case 0xcb:
 928	case 0xcf:
 929	case 0xea:	/* jmp absolute -- ip is correct */
 930		/* ip is already adjusted, no more changes required */
 931		p->ainsn.boostable = true;
 932		goto no_change;
 933	case 0xe8:	/* call relative - Fix return addr */
 934		*tos = orig_ip + (*tos - copy_ip);
 935		break;
 936#ifdef CONFIG_X86_32
 937	case 0x9a:	/* call absolute -- same as call absolute, indirect */
 938		*tos = orig_ip + (*tos - copy_ip);
 939		goto no_change;
 940#endif
 941	case 0xff:
 942		if ((insn[1] & 0x30) == 0x10) {
 943			/*
 944			 * call absolute, indirect
 945			 * Fix return addr; ip is correct.
 946			 * But this is not boostable
 947			 */
 948			*tos = orig_ip + (*tos - copy_ip);
 949			goto no_change;
 950		} else if (((insn[1] & 0x31) == 0x20) ||
 951			   ((insn[1] & 0x31) == 0x21)) {
 952			/*
 953			 * jmp near and far, absolute indirect
 954			 * ip is correct. And this is boostable
 955			 */
 956			p->ainsn.boostable = true;
 957			goto no_change;
 958		}
 959	default:
 960		break;
 961	}
 962
 963	regs->ip += orig_ip - copy_ip;
 964
 965no_change:
 966	restore_btf();
 967}
 968NOKPROBE_SYMBOL(resume_execution);
 969
 970/*
 971 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 972 * remain disabled throughout this function.
 973 */
 974int kprobe_debug_handler(struct pt_regs *regs)
 975{
 976	struct kprobe *cur = kprobe_running();
 977	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 978
 979	if (!cur)
 980		return 0;
 981
 982	resume_execution(cur, regs, kcb);
 983	regs->flags |= kcb->kprobe_saved_flags;
 984
 985	if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 986		kcb->kprobe_status = KPROBE_HIT_SSDONE;
 987		cur->post_handler(cur, regs, 0);
 988	}
 989
 990	/* Restore back the original saved kprobes variables and continue. */
 991	if (kcb->kprobe_status == KPROBE_REENTER) {
 992		restore_previous_kprobe(kcb);
 993		goto out;
 994	}
 995	reset_current_kprobe();
 996out:
 
 
 997	/*
 998	 * if somebody else is singlestepping across a probe point, flags
 999	 * will have TF set, in which case, continue the remaining processing
1000	 * of do_debug, as if this is not a probe hit.
1001	 */
1002	if (regs->flags & X86_EFLAGS_TF)
1003		return 0;
1004
1005	return 1;
1006}
1007NOKPROBE_SYMBOL(kprobe_debug_handler);
1008
1009int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
1010{
1011	struct kprobe *cur = kprobe_running();
1012	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1013
1014	if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1015		/* This must happen on single-stepping */
1016		WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1017			kcb->kprobe_status != KPROBE_REENTER);
1018		/*
1019		 * We are here because the instruction being single
1020		 * stepped caused a page fault. We reset the current
1021		 * kprobe and the ip points back to the probe address
1022		 * and allow the page fault handler to continue as a
1023		 * normal page fault.
1024		 */
1025		regs->ip = (unsigned long)cur->addr;
1026		/*
1027		 * Trap flag (TF) has been set here because this fault
1028		 * happened where the single stepping will be done.
1029		 * So clear it by resetting the current kprobe:
1030		 */
1031		regs->flags &= ~X86_EFLAGS_TF;
1032
1033		/*
1034		 * If the TF flag was set before the kprobe hit,
1035		 * don't touch it:
1036		 */
1037		regs->flags |= kcb->kprobe_old_flags;
1038
1039		if (kcb->kprobe_status == KPROBE_REENTER)
1040			restore_previous_kprobe(kcb);
1041		else
1042			reset_current_kprobe();
 
1043	} else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
1044		   kcb->kprobe_status == KPROBE_HIT_SSDONE) {
1045		/*
1046		 * We increment the nmissed count for accounting,
1047		 * we can also use npre/npostfault count for accounting
1048		 * these specific fault cases.
1049		 */
1050		kprobes_inc_nmissed_count(cur);
1051
1052		/*
1053		 * We come here because instructions in the pre/post
1054		 * handler caused the page_fault, this could happen
1055		 * if handler tries to access user space by
1056		 * copy_from_user(), get_user() etc. Let the
1057		 * user-specified handler try to fix it first.
1058		 */
1059		if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1060			return 1;
 
 
 
 
 
 
 
 
 
 
 
 
1061	}
1062
1063	return 0;
1064}
1065NOKPROBE_SYMBOL(kprobe_fault_handler);
1066
1067int __init arch_populate_kprobe_blacklist(void)
 
 
 
 
1068{
1069	int ret;
 
1070
1071	ret = kprobe_add_area_blacklist((unsigned long)__irqentry_text_start,
1072					 (unsigned long)__irqentry_text_end);
1073	if (ret)
1074		return ret;
1075
1076	return kprobe_add_area_blacklist((unsigned long)__entry_text_start,
1077					 (unsigned long)__entry_text_end);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1078}
1079
1080int __init arch_init_kprobes(void)
1081{
1082	return 0;
1083}
1084
1085int arch_trampoline_kprobe(struct kprobe *p)
1086{
1087	return 0;
1088}
v4.17
 
   1/*
   2 *  Kernel Probes (KProbes)
   3 *
   4 * This program is free software; you can redistribute it and/or modify
   5 * it under the terms of the GNU General Public License as published by
   6 * the Free Software Foundation; either version 2 of the License, or
   7 * (at your option) any later version.
   8 *
   9 * This program is distributed in the hope that it will be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write to the Free Software
  16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17 *
  18 * Copyright (C) IBM Corporation, 2002, 2004
  19 *
  20 * 2002-Oct	Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
  21 *		Probes initial implementation ( includes contributions from
  22 *		Rusty Russell).
  23 * 2004-July	Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
  24 *		interface to access function arguments.
  25 * 2004-Oct	Jim Keniston <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  26 *		<prasanna@in.ibm.com> adapted for x86_64 from i386.
  27 * 2005-Mar	Roland McGrath <roland@redhat.com>
  28 *		Fixed to handle %rip-relative addressing mode correctly.
  29 * 2005-May	Hien Nguyen <hien@us.ibm.com>, Jim Keniston
  30 *		<jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  31 *		<prasanna@in.ibm.com> added function-return probes.
  32 * 2005-May	Rusty Lynch <rusty.lynch@intel.com>
  33 *		Added function return probes functionality
  34 * 2006-Feb	Masami Hiramatsu <hiramatu@sdl.hitachi.co.jp> added
  35 *		kprobe-booster and kretprobe-booster for i386.
  36 * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com> added kprobe-booster
  37 *		and kretprobe-booster for x86-64
  38 * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com>, Arjan van de Ven
  39 *		<arjan@infradead.org> and Jim Keniston <jkenisto@us.ibm.com>
  40 *		unified x86 kprobes code.
  41 */
  42#include <linux/kprobes.h>
  43#include <linux/ptrace.h>
  44#include <linux/string.h>
  45#include <linux/slab.h>
  46#include <linux/hardirq.h>
  47#include <linux/preempt.h>
  48#include <linux/sched/debug.h>
  49#include <linux/extable.h>
  50#include <linux/kdebug.h>
  51#include <linux/kallsyms.h>
  52#include <linux/ftrace.h>
  53#include <linux/frame.h>
  54#include <linux/kasan.h>
  55#include <linux/moduleloader.h>
  56
  57#include <asm/text-patching.h>
  58#include <asm/cacheflush.h>
  59#include <asm/desc.h>
  60#include <asm/pgtable.h>
  61#include <linux/uaccess.h>
  62#include <asm/alternative.h>
  63#include <asm/insn.h>
  64#include <asm/debugreg.h>
  65#include <asm/set_memory.h>
  66
  67#include "common.h"
  68
  69void jprobe_return_end(void);
  70
  71DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  72DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  73
  74#define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
  75
  76#define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
  77	(((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  78	  (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  79	  (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  80	  (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  81	 << (row % 32))
  82	/*
  83	 * Undefined/reserved opcodes, conditional jump, Opcode Extension
  84	 * Groups, and some special opcodes can not boost.
  85	 * This is non-const and volatile to keep gcc from statically
  86	 * optimizing it out, as variable_test_bit makes gcc think only
  87	 * *(unsigned long*) is used.
  88	 */
  89static volatile u32 twobyte_is_boostable[256 / 32] = {
  90	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  91	/*      ----------------------------------------------          */
  92	W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
  93	W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
  94	W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
  95	W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
  96	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
  97	W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
  98	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
  99	W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
 100	W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
 101	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
 102	W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
 103	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
 104	W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
 105	W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
 106	W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
 107	W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
 108	/*      -----------------------------------------------         */
 109	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
 110};
 111#undef W
 112
 113struct kretprobe_blackpoint kretprobe_blacklist[] = {
 114	{"__switch_to", }, /* This function switches only current task, but
 115			      doesn't switch kernel stack.*/
 116	{NULL, NULL}	/* Terminator */
 117};
 118
 119const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
 120
 121static nokprobe_inline void
 122__synthesize_relative_insn(void *dest, void *from, void *to, u8 op)
 123{
 124	struct __arch_relative_insn {
 125		u8 op;
 126		s32 raddr;
 127	} __packed *insn;
 128
 129	insn = (struct __arch_relative_insn *)dest;
 130	insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
 131	insn->op = op;
 132}
 133
 134/* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
 135void synthesize_reljump(void *dest, void *from, void *to)
 136{
 137	__synthesize_relative_insn(dest, from, to, RELATIVEJUMP_OPCODE);
 138}
 139NOKPROBE_SYMBOL(synthesize_reljump);
 140
 141/* Insert a call instruction at address 'from', which calls address 'to'.*/
 142void synthesize_relcall(void *dest, void *from, void *to)
 143{
 144	__synthesize_relative_insn(dest, from, to, RELATIVECALL_OPCODE);
 145}
 146NOKPROBE_SYMBOL(synthesize_relcall);
 147
 148/*
 149 * Skip the prefixes of the instruction.
 150 */
 151static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
 152{
 153	insn_attr_t attr;
 154
 155	attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 156	while (inat_is_legacy_prefix(attr)) {
 157		insn++;
 158		attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 159	}
 160#ifdef CONFIG_X86_64
 161	if (inat_is_rex_prefix(attr))
 162		insn++;
 163#endif
 164	return insn;
 165}
 166NOKPROBE_SYMBOL(skip_prefixes);
 167
 168/*
 169 * Returns non-zero if INSN is boostable.
 170 * RIP relative instructions are adjusted at copying time in 64 bits mode
 171 */
 172int can_boost(struct insn *insn, void *addr)
 173{
 174	kprobe_opcode_t opcode;
 175
 176	if (search_exception_tables((unsigned long)addr))
 177		return 0;	/* Page fault may occur on this address. */
 178
 179	/* 2nd-byte opcode */
 180	if (insn->opcode.nbytes == 2)
 181		return test_bit(insn->opcode.bytes[1],
 182				(unsigned long *)twobyte_is_boostable);
 183
 184	if (insn->opcode.nbytes != 1)
 185		return 0;
 186
 187	/* Can't boost Address-size override prefix */
 188	if (unlikely(inat_is_address_size_prefix(insn->attr)))
 189		return 0;
 190
 191	opcode = insn->opcode.bytes[0];
 192
 193	switch (opcode & 0xf0) {
 194	case 0x60:
 195		/* can't boost "bound" */
 196		return (opcode != 0x62);
 197	case 0x70:
 198		return 0; /* can't boost conditional jump */
 199	case 0x90:
 200		return opcode != 0x9a;	/* can't boost call far */
 201	case 0xc0:
 202		/* can't boost software-interruptions */
 203		return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 204	case 0xd0:
 205		/* can boost AA* and XLAT */
 206		return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 207	case 0xe0:
 208		/* can boost in/out and absolute jmps */
 209		return ((opcode & 0x04) || opcode == 0xea);
 210	case 0xf0:
 211		/* clear and set flags are boostable */
 212		return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 213	default:
 214		/* CS override prefix and call are not boostable */
 215		return (opcode != 0x2e && opcode != 0x9a);
 216	}
 217}
 218
 219static unsigned long
 220__recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
 221{
 222	struct kprobe *kp;
 223	unsigned long faddr;
 224
 225	kp = get_kprobe((void *)addr);
 226	faddr = ftrace_location(addr);
 227	/*
 228	 * Addresses inside the ftrace location are refused by
 229	 * arch_check_ftrace_location(). Something went terribly wrong
 230	 * if such an address is checked here.
 231	 */
 232	if (WARN_ON(faddr && faddr != addr))
 233		return 0UL;
 234	/*
 235	 * Use the current code if it is not modified by Kprobe
 236	 * and it cannot be modified by ftrace.
 237	 */
 238	if (!kp && !faddr)
 239		return addr;
 240
 241	/*
 242	 * Basically, kp->ainsn.insn has an original instruction.
 243	 * However, RIP-relative instruction can not do single-stepping
 244	 * at different place, __copy_instruction() tweaks the displacement of
 245	 * that instruction. In that case, we can't recover the instruction
 246	 * from the kp->ainsn.insn.
 247	 *
 248	 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
 249	 * of the first byte of the probed instruction, which is overwritten
 250	 * by int3. And the instruction at kp->addr is not modified by kprobes
 251	 * except for the first byte, we can recover the original instruction
 252	 * from it and kp->opcode.
 253	 *
 254	 * In case of Kprobes using ftrace, we do not have a copy of
 255	 * the original instruction. In fact, the ftrace location might
 256	 * be modified at anytime and even could be in an inconsistent state.
 257	 * Fortunately, we know that the original code is the ideal 5-byte
 258	 * long NOP.
 259	 */
 260	if (probe_kernel_read(buf, (void *)addr,
 261		MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
 262		return 0UL;
 263
 264	if (faddr)
 265		memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
 266	else
 267		buf[0] = kp->opcode;
 268	return (unsigned long)buf;
 269}
 270
 271/*
 272 * Recover the probed instruction at addr for further analysis.
 273 * Caller must lock kprobes by kprobe_mutex, or disable preemption
 274 * for preventing to release referencing kprobes.
 275 * Returns zero if the instruction can not get recovered (or access failed).
 276 */
 277unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
 278{
 279	unsigned long __addr;
 280
 281	__addr = __recover_optprobed_insn(buf, addr);
 282	if (__addr != addr)
 283		return __addr;
 284
 285	return __recover_probed_insn(buf, addr);
 286}
 287
 288/* Check if paddr is at an instruction boundary */
 289static int can_probe(unsigned long paddr)
 290{
 291	unsigned long addr, __addr, offset = 0;
 292	struct insn insn;
 293	kprobe_opcode_t buf[MAX_INSN_SIZE];
 294
 295	if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
 296		return 0;
 297
 298	/* Decode instructions */
 299	addr = paddr - offset;
 300	while (addr < paddr) {
 301		/*
 302		 * Check if the instruction has been modified by another
 303		 * kprobe, in which case we replace the breakpoint by the
 304		 * original instruction in our buffer.
 305		 * Also, jump optimization will change the breakpoint to
 306		 * relative-jump. Since the relative-jump itself is
 307		 * normally used, we just go through if there is no kprobe.
 308		 */
 309		__addr = recover_probed_instruction(buf, addr);
 310		if (!__addr)
 311			return 0;
 312		kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
 313		insn_get_length(&insn);
 314
 315		/*
 316		 * Another debugging subsystem might insert this breakpoint.
 317		 * In that case, we can't recover it.
 318		 */
 319		if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 320			return 0;
 321		addr += insn.length;
 322	}
 323
 324	return (addr == paddr);
 325}
 326
 327/*
 328 * Returns non-zero if opcode modifies the interrupt flag.
 329 */
 330static int is_IF_modifier(kprobe_opcode_t *insn)
 331{
 332	/* Skip prefixes */
 333	insn = skip_prefixes(insn);
 334
 335	switch (*insn) {
 336	case 0xfa:		/* cli */
 337	case 0xfb:		/* sti */
 338	case 0xcf:		/* iret/iretd */
 339	case 0x9d:		/* popf/popfd */
 340		return 1;
 341	}
 342
 343	return 0;
 344}
 345
 346/*
 347 * Copy an instruction with recovering modified instruction by kprobes
 348 * and adjust the displacement if the instruction uses the %rip-relative
 349 * addressing mode. Note that since @real will be the final place of copied
 350 * instruction, displacement must be adjust by @real, not @dest.
 351 * This returns the length of copied instruction, or 0 if it has an error.
 352 */
 353int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn)
 354{
 355	kprobe_opcode_t buf[MAX_INSN_SIZE];
 356	unsigned long recovered_insn =
 357		recover_probed_instruction(buf, (unsigned long)src);
 358
 359	if (!recovered_insn || !insn)
 360		return 0;
 361
 362	/* This can access kernel text if given address is not recovered */
 363	if (probe_kernel_read(dest, (void *)recovered_insn, MAX_INSN_SIZE))
 364		return 0;
 365
 366	kernel_insn_init(insn, dest, MAX_INSN_SIZE);
 367	insn_get_length(insn);
 368
 369	/* Another subsystem puts a breakpoint, failed to recover */
 370	if (insn->opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 371		return 0;
 372
 373	/* We should not singlestep on the exception masking instructions */
 374	if (insn_masking_exception(insn))
 375		return 0;
 376
 377#ifdef CONFIG_X86_64
 378	/* Only x86_64 has RIP relative instructions */
 379	if (insn_rip_relative(insn)) {
 380		s64 newdisp;
 381		u8 *disp;
 382		/*
 383		 * The copied instruction uses the %rip-relative addressing
 384		 * mode.  Adjust the displacement for the difference between
 385		 * the original location of this instruction and the location
 386		 * of the copy that will actually be run.  The tricky bit here
 387		 * is making sure that the sign extension happens correctly in
 388		 * this calculation, since we need a signed 32-bit result to
 389		 * be sign-extended to 64 bits when it's added to the %rip
 390		 * value and yield the same 64-bit result that the sign-
 391		 * extension of the original signed 32-bit displacement would
 392		 * have given.
 393		 */
 394		newdisp = (u8 *) src + (s64) insn->displacement.value
 395			  - (u8 *) real;
 396		if ((s64) (s32) newdisp != newdisp) {
 397			pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
 398			pr_err("\tSrc: %p, Dest: %p, old disp: %x\n",
 399				src, real, insn->displacement.value);
 400			return 0;
 401		}
 402		disp = (u8 *) dest + insn_offset_displacement(insn);
 403		*(s32 *) disp = (s32) newdisp;
 404	}
 405#endif
 406	return insn->length;
 407}
 408
 409/* Prepare reljump right after instruction to boost */
 410static int prepare_boost(kprobe_opcode_t *buf, struct kprobe *p,
 411			  struct insn *insn)
 412{
 413	int len = insn->length;
 414
 415	if (can_boost(insn, p->addr) &&
 416	    MAX_INSN_SIZE - len >= RELATIVEJUMP_SIZE) {
 417		/*
 418		 * These instructions can be executed directly if it
 419		 * jumps back to correct address.
 420		 */
 421		synthesize_reljump(buf + len, p->ainsn.insn + len,
 422				   p->addr + insn->length);
 423		len += RELATIVEJUMP_SIZE;
 424		p->ainsn.boostable = true;
 425	} else {
 426		p->ainsn.boostable = false;
 427	}
 428
 429	return len;
 430}
 431
 432/* Make page to RO mode when allocate it */
 433void *alloc_insn_page(void)
 434{
 435	void *page;
 436
 437	page = module_alloc(PAGE_SIZE);
 438	if (page)
 439		set_memory_ro((unsigned long)page & PAGE_MASK, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 440
 441	return page;
 442}
 443
 444/* Recover page to RW mode before releasing it */
 445void free_insn_page(void *page)
 446{
 447	set_memory_nx((unsigned long)page & PAGE_MASK, 1);
 448	set_memory_rw((unsigned long)page & PAGE_MASK, 1);
 449	module_memfree(page);
 450}
 451
 452static int arch_copy_kprobe(struct kprobe *p)
 453{
 454	struct insn insn;
 455	kprobe_opcode_t buf[MAX_INSN_SIZE];
 456	int len;
 457
 458	/* Copy an instruction with recovering if other optprobe modifies it.*/
 459	len = __copy_instruction(buf, p->addr, p->ainsn.insn, &insn);
 460	if (!len)
 461		return -EINVAL;
 462
 463	/*
 464	 * __copy_instruction can modify the displacement of the instruction,
 465	 * but it doesn't affect boostable check.
 466	 */
 467	len = prepare_boost(buf, p, &insn);
 468
 469	/* Check whether the instruction modifies Interrupt Flag or not */
 470	p->ainsn.if_modifier = is_IF_modifier(buf);
 471
 472	/* Also, displacement change doesn't affect the first byte */
 473	p->opcode = buf[0];
 474
 475	/* OK, write back the instruction(s) into ROX insn buffer */
 476	text_poke(p->ainsn.insn, buf, len);
 477
 478	return 0;
 479}
 480
 481int arch_prepare_kprobe(struct kprobe *p)
 482{
 483	int ret;
 484
 485	if (alternatives_text_reserved(p->addr, p->addr))
 486		return -EINVAL;
 487
 488	if (!can_probe((unsigned long)p->addr))
 489		return -EILSEQ;
 490	/* insn: must be on special executable page on x86. */
 491	p->ainsn.insn = get_insn_slot();
 492	if (!p->ainsn.insn)
 493		return -ENOMEM;
 494
 495	ret = arch_copy_kprobe(p);
 496	if (ret) {
 497		free_insn_slot(p->ainsn.insn, 0);
 498		p->ainsn.insn = NULL;
 499	}
 500
 501	return ret;
 502}
 503
 504void arch_arm_kprobe(struct kprobe *p)
 505{
 506	text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 507}
 508
 509void arch_disarm_kprobe(struct kprobe *p)
 510{
 511	text_poke(p->addr, &p->opcode, 1);
 512}
 513
 514void arch_remove_kprobe(struct kprobe *p)
 515{
 516	if (p->ainsn.insn) {
 517		free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
 518		p->ainsn.insn = NULL;
 519	}
 520}
 521
 522static nokprobe_inline void
 523save_previous_kprobe(struct kprobe_ctlblk *kcb)
 524{
 525	kcb->prev_kprobe.kp = kprobe_running();
 526	kcb->prev_kprobe.status = kcb->kprobe_status;
 527	kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
 528	kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
 529}
 530
 531static nokprobe_inline void
 532restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 533{
 534	__this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
 535	kcb->kprobe_status = kcb->prev_kprobe.status;
 536	kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
 537	kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
 538}
 539
 540static nokprobe_inline void
 541set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 542		   struct kprobe_ctlblk *kcb)
 543{
 544	__this_cpu_write(current_kprobe, p);
 545	kcb->kprobe_saved_flags = kcb->kprobe_old_flags
 546		= (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 547	if (p->ainsn.if_modifier)
 548		kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
 549}
 550
 551static nokprobe_inline void clear_btf(void)
 552{
 553	if (test_thread_flag(TIF_BLOCKSTEP)) {
 554		unsigned long debugctl = get_debugctlmsr();
 555
 556		debugctl &= ~DEBUGCTLMSR_BTF;
 557		update_debugctlmsr(debugctl);
 558	}
 559}
 560
 561static nokprobe_inline void restore_btf(void)
 562{
 563	if (test_thread_flag(TIF_BLOCKSTEP)) {
 564		unsigned long debugctl = get_debugctlmsr();
 565
 566		debugctl |= DEBUGCTLMSR_BTF;
 567		update_debugctlmsr(debugctl);
 568	}
 569}
 570
 571void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
 572{
 573	unsigned long *sara = stack_addr(regs);
 574
 575	ri->ret_addr = (kprobe_opcode_t *) *sara;
 
 576
 577	/* Replace the return addr with trampoline addr */
 578	*sara = (unsigned long) &kretprobe_trampoline;
 579}
 580NOKPROBE_SYMBOL(arch_prepare_kretprobe);
 581
 582static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
 583			     struct kprobe_ctlblk *kcb, int reenter)
 584{
 585	if (setup_detour_execution(p, regs, reenter))
 586		return;
 587
 588#if !defined(CONFIG_PREEMPT)
 589	if (p->ainsn.boostable && !p->post_handler) {
 590		/* Boost up -- we can execute copied instructions directly */
 591		if (!reenter)
 592			reset_current_kprobe();
 593		/*
 594		 * Reentering boosted probe doesn't reset current_kprobe,
 595		 * nor set current_kprobe, because it doesn't use single
 596		 * stepping.
 597		 */
 598		regs->ip = (unsigned long)p->ainsn.insn;
 599		preempt_enable_no_resched();
 600		return;
 601	}
 602#endif
 603	if (reenter) {
 604		save_previous_kprobe(kcb);
 605		set_current_kprobe(p, regs, kcb);
 606		kcb->kprobe_status = KPROBE_REENTER;
 607	} else
 608		kcb->kprobe_status = KPROBE_HIT_SS;
 609	/* Prepare real single stepping */
 610	clear_btf();
 611	regs->flags |= X86_EFLAGS_TF;
 612	regs->flags &= ~X86_EFLAGS_IF;
 613	/* single step inline if the instruction is an int3 */
 614	if (p->opcode == BREAKPOINT_INSTRUCTION)
 615		regs->ip = (unsigned long)p->addr;
 616	else
 617		regs->ip = (unsigned long)p->ainsn.insn;
 618}
 619NOKPROBE_SYMBOL(setup_singlestep);
 620
 621/*
 622 * We have reentered the kprobe_handler(), since another probe was hit while
 623 * within the handler. We save the original kprobes variables and just single
 624 * step on the instruction of the new probe without calling any user handlers.
 625 */
 626static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
 627			  struct kprobe_ctlblk *kcb)
 628{
 629	switch (kcb->kprobe_status) {
 630	case KPROBE_HIT_SSDONE:
 631	case KPROBE_HIT_ACTIVE:
 632	case KPROBE_HIT_SS:
 633		kprobes_inc_nmissed_count(p);
 634		setup_singlestep(p, regs, kcb, 1);
 635		break;
 636	case KPROBE_REENTER:
 637		/* A probe has been hit in the codepath leading up to, or just
 638		 * after, single-stepping of a probed instruction. This entire
 639		 * codepath should strictly reside in .kprobes.text section.
 640		 * Raise a BUG or we'll continue in an endless reentering loop
 641		 * and eventually a stack overflow.
 642		 */
 643		printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
 644		       p->addr);
 645		dump_kprobe(p);
 646		BUG();
 647	default:
 648		/* impossible cases */
 649		WARN_ON(1);
 650		return 0;
 651	}
 652
 653	return 1;
 654}
 655NOKPROBE_SYMBOL(reenter_kprobe);
 656
 657/*
 658 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 659 * remain disabled throughout this function.
 660 */
 661int kprobe_int3_handler(struct pt_regs *regs)
 662{
 663	kprobe_opcode_t *addr;
 664	struct kprobe *p;
 665	struct kprobe_ctlblk *kcb;
 666
 667	if (user_mode(regs))
 668		return 0;
 669
 670	addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 671	/*
 672	 * We don't want to be preempted for the entire
 673	 * duration of kprobe processing. We conditionally
 674	 * re-enable preemption at the end of this function,
 675	 * and also in reenter_kprobe() and setup_singlestep().
 676	 */
 677	preempt_disable();
 678
 679	kcb = get_kprobe_ctlblk();
 680	p = get_kprobe(addr);
 681
 682	if (p) {
 683		if (kprobe_running()) {
 684			if (reenter_kprobe(p, regs, kcb))
 685				return 1;
 686		} else {
 687			set_current_kprobe(p, regs, kcb);
 688			kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 689
 690			/*
 691			 * If we have no pre-handler or it returned 0, we
 692			 * continue with normal processing.  If we have a
 693			 * pre-handler and it returned non-zero, it prepped
 694			 * for calling the break_handler below on re-entry
 695			 * for jprobe processing, so get out doing nothing
 696			 * more here.
 697			 */
 698			if (!p->pre_handler || !p->pre_handler(p, regs))
 699				setup_singlestep(p, regs, kcb, 0);
 
 
 700			return 1;
 701		}
 702	} else if (*addr != BREAKPOINT_INSTRUCTION) {
 703		/*
 704		 * The breakpoint instruction was removed right
 705		 * after we hit it.  Another cpu has removed
 706		 * either a probepoint or a debugger breakpoint
 707		 * at this address.  In either case, no further
 708		 * handling of this interrupt is appropriate.
 709		 * Back up over the (now missing) int3 and run
 710		 * the original instruction.
 711		 */
 712		regs->ip = (unsigned long)addr;
 713		preempt_enable_no_resched();
 714		return 1;
 715	} else if (kprobe_running()) {
 716		p = __this_cpu_read(current_kprobe);
 717		if (p->break_handler && p->break_handler(p, regs)) {
 718			if (!skip_singlestep(p, regs, kcb))
 719				setup_singlestep(p, regs, kcb, 0);
 720			return 1;
 721		}
 722	} /* else: not a kprobe fault; let the kernel handle it */
 723
 724	preempt_enable_no_resched();
 725	return 0;
 726}
 727NOKPROBE_SYMBOL(kprobe_int3_handler);
 728
 729/*
 730 * When a retprobed function returns, this code saves registers and
 731 * calls trampoline_handler() runs, which calls the kretprobe's handler.
 732 */
 733asm(
 
 734	".global kretprobe_trampoline\n"
 735	".type kretprobe_trampoline, @function\n"
 736	"kretprobe_trampoline:\n"
 
 737#ifdef CONFIG_X86_64
 738	/* We don't bother saving the ss register */
 739	"	pushq %rsp\n"
 740	"	pushfq\n"
 741	SAVE_REGS_STRING
 742	"	movq %rsp, %rdi\n"
 743	"	call trampoline_handler\n"
 744	/* Replace saved sp with true return address. */
 745	"	movq %rax, 152(%rsp)\n"
 746	RESTORE_REGS_STRING
 747	"	popfq\n"
 748#else
 749	"	pushf\n"
 
 750	SAVE_REGS_STRING
 751	"	movl %esp, %eax\n"
 752	"	call trampoline_handler\n"
 753	/* Move flags to cs */
 754	"	movl 56(%esp), %edx\n"
 755	"	movl %edx, 52(%esp)\n"
 756	/* Replace saved flags with true return address. */
 757	"	movl %eax, 56(%esp)\n"
 758	RESTORE_REGS_STRING
 759	"	popf\n"
 760#endif
 761	"	ret\n"
 762	".size kretprobe_trampoline, .-kretprobe_trampoline\n"
 763);
 764NOKPROBE_SYMBOL(kretprobe_trampoline);
 765STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
 766
 
 
 
 
 767/*
 768 * Called from kretprobe_trampoline
 769 */
 770__visible __used void *trampoline_handler(struct pt_regs *regs)
 771{
 
 772	struct kretprobe_instance *ri = NULL;
 773	struct hlist_head *head, empty_rp;
 774	struct hlist_node *tmp;
 775	unsigned long flags, orig_ret_address = 0;
 776	unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 777	kprobe_opcode_t *correct_ret_addr = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 778
 779	INIT_HLIST_HEAD(&empty_rp);
 780	kretprobe_hash_lock(current, &head, &flags);
 781	/* fixup registers */
 782#ifdef CONFIG_X86_64
 783	regs->cs = __KERNEL_CS;
 784#else
 785	regs->cs = __KERNEL_CS | get_kernel_rpl();
 786	regs->gs = 0;
 787#endif
 
 
 788	regs->ip = trampoline_address;
 789	regs->orig_ax = ~0UL;
 790
 791	/*
 792	 * It is possible to have multiple instances associated with a given
 793	 * task either because multiple functions in the call path have
 794	 * return probes installed on them, and/or more than one
 795	 * return probe was registered for a target function.
 796	 *
 797	 * We can handle this because:
 798	 *     - instances are always pushed into the head of the list
 799	 *     - when multiple return probes are registered for the same
 800	 *	 function, the (chronologically) first instance's ret_addr
 801	 *	 will be the real return address, and all the rest will
 802	 *	 point to kretprobe_trampoline.
 803	 */
 804	hlist_for_each_entry(ri, head, hlist) {
 805		if (ri->task != current)
 806			/* another task is sharing our hash bucket */
 807			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 808
 809		orig_ret_address = (unsigned long)ri->ret_addr;
 
 
 
 810
 811		if (orig_ret_address != trampoline_address)
 812			/*
 813			 * This is the real return address. Any other
 814			 * instances associated with this task are for
 815			 * other calls deeper on the call stack
 816			 */
 817			break;
 818	}
 819
 820	kretprobe_assert(ri, orig_ret_address, trampoline_address);
 821
 822	correct_ret_addr = ri->ret_addr;
 823	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
 824		if (ri->task != current)
 825			/* another task is sharing our hash bucket */
 826			continue;
 
 
 827
 828		orig_ret_address = (unsigned long)ri->ret_addr;
 829		if (ri->rp && ri->rp->handler) {
 830			__this_cpu_write(current_kprobe, &ri->rp->kp);
 831			get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 832			ri->ret_addr = correct_ret_addr;
 833			ri->rp->handler(ri, regs);
 834			__this_cpu_write(current_kprobe, NULL);
 835		}
 836
 837		recycle_rp_inst(ri, &empty_rp);
 838
 839		if (orig_ret_address != trampoline_address)
 840			/*
 841			 * This is the real return address. Any other
 842			 * instances associated with this task are for
 843			 * other calls deeper on the call stack
 844			 */
 845			break;
 846	}
 847
 848	kretprobe_hash_unlock(current, &flags);
 849
 
 
 
 850	hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
 851		hlist_del(&ri->hlist);
 852		kfree(ri);
 853	}
 854	return (void *)orig_ret_address;
 855}
 856NOKPROBE_SYMBOL(trampoline_handler);
 857
 858/*
 859 * Called after single-stepping.  p->addr is the address of the
 860 * instruction whose first byte has been replaced by the "int 3"
 861 * instruction.  To avoid the SMP problems that can occur when we
 862 * temporarily put back the original opcode to single-step, we
 863 * single-stepped a copy of the instruction.  The address of this
 864 * copy is p->ainsn.insn.
 865 *
 866 * This function prepares to return from the post-single-step
 867 * interrupt.  We have to fix up the stack as follows:
 868 *
 869 * 0) Except in the case of absolute or indirect jump or call instructions,
 870 * the new ip is relative to the copied instruction.  We need to make
 871 * it relative to the original instruction.
 872 *
 873 * 1) If the single-stepped instruction was pushfl, then the TF and IF
 874 * flags are set in the just-pushed flags, and may need to be cleared.
 875 *
 876 * 2) If the single-stepped instruction was a call, the return address
 877 * that is atop the stack is the address following the copied instruction.
 878 * We need to make it the address following the original instruction.
 879 *
 880 * If this is the first time we've single-stepped the instruction at
 881 * this probepoint, and the instruction is boostable, boost it: add a
 882 * jump instruction after the copied instruction, that jumps to the next
 883 * instruction after the probepoint.
 884 */
 885static void resume_execution(struct kprobe *p, struct pt_regs *regs,
 886			     struct kprobe_ctlblk *kcb)
 887{
 888	unsigned long *tos = stack_addr(regs);
 889	unsigned long copy_ip = (unsigned long)p->ainsn.insn;
 890	unsigned long orig_ip = (unsigned long)p->addr;
 891	kprobe_opcode_t *insn = p->ainsn.insn;
 892
 893	/* Skip prefixes */
 894	insn = skip_prefixes(insn);
 895
 896	regs->flags &= ~X86_EFLAGS_TF;
 897	switch (*insn) {
 898	case 0x9c:	/* pushfl */
 899		*tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
 900		*tos |= kcb->kprobe_old_flags;
 901		break;
 902	case 0xc2:	/* iret/ret/lret */
 903	case 0xc3:
 904	case 0xca:
 905	case 0xcb:
 906	case 0xcf:
 907	case 0xea:	/* jmp absolute -- ip is correct */
 908		/* ip is already adjusted, no more changes required */
 909		p->ainsn.boostable = true;
 910		goto no_change;
 911	case 0xe8:	/* call relative - Fix return addr */
 912		*tos = orig_ip + (*tos - copy_ip);
 913		break;
 914#ifdef CONFIG_X86_32
 915	case 0x9a:	/* call absolute -- same as call absolute, indirect */
 916		*tos = orig_ip + (*tos - copy_ip);
 917		goto no_change;
 918#endif
 919	case 0xff:
 920		if ((insn[1] & 0x30) == 0x10) {
 921			/*
 922			 * call absolute, indirect
 923			 * Fix return addr; ip is correct.
 924			 * But this is not boostable
 925			 */
 926			*tos = orig_ip + (*tos - copy_ip);
 927			goto no_change;
 928		} else if (((insn[1] & 0x31) == 0x20) ||
 929			   ((insn[1] & 0x31) == 0x21)) {
 930			/*
 931			 * jmp near and far, absolute indirect
 932			 * ip is correct. And this is boostable
 933			 */
 934			p->ainsn.boostable = true;
 935			goto no_change;
 936		}
 937	default:
 938		break;
 939	}
 940
 941	regs->ip += orig_ip - copy_ip;
 942
 943no_change:
 944	restore_btf();
 945}
 946NOKPROBE_SYMBOL(resume_execution);
 947
 948/*
 949 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 950 * remain disabled throughout this function.
 951 */
 952int kprobe_debug_handler(struct pt_regs *regs)
 953{
 954	struct kprobe *cur = kprobe_running();
 955	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 956
 957	if (!cur)
 958		return 0;
 959
 960	resume_execution(cur, regs, kcb);
 961	regs->flags |= kcb->kprobe_saved_flags;
 962
 963	if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 964		kcb->kprobe_status = KPROBE_HIT_SSDONE;
 965		cur->post_handler(cur, regs, 0);
 966	}
 967
 968	/* Restore back the original saved kprobes variables and continue. */
 969	if (kcb->kprobe_status == KPROBE_REENTER) {
 970		restore_previous_kprobe(kcb);
 971		goto out;
 972	}
 973	reset_current_kprobe();
 974out:
 975	preempt_enable_no_resched();
 976
 977	/*
 978	 * if somebody else is singlestepping across a probe point, flags
 979	 * will have TF set, in which case, continue the remaining processing
 980	 * of do_debug, as if this is not a probe hit.
 981	 */
 982	if (regs->flags & X86_EFLAGS_TF)
 983		return 0;
 984
 985	return 1;
 986}
 987NOKPROBE_SYMBOL(kprobe_debug_handler);
 988
 989int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 990{
 991	struct kprobe *cur = kprobe_running();
 992	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 993
 994	if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
 995		/* This must happen on single-stepping */
 996		WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
 997			kcb->kprobe_status != KPROBE_REENTER);
 998		/*
 999		 * We are here because the instruction being single
1000		 * stepped caused a page fault. We reset the current
1001		 * kprobe and the ip points back to the probe address
1002		 * and allow the page fault handler to continue as a
1003		 * normal page fault.
1004		 */
1005		regs->ip = (unsigned long)cur->addr;
1006		/*
1007		 * Trap flag (TF) has been set here because this fault
1008		 * happened where the single stepping will be done.
1009		 * So clear it by resetting the current kprobe:
1010		 */
1011		regs->flags &= ~X86_EFLAGS_TF;
1012
1013		/*
1014		 * If the TF flag was set before the kprobe hit,
1015		 * don't touch it:
1016		 */
1017		regs->flags |= kcb->kprobe_old_flags;
1018
1019		if (kcb->kprobe_status == KPROBE_REENTER)
1020			restore_previous_kprobe(kcb);
1021		else
1022			reset_current_kprobe();
1023		preempt_enable_no_resched();
1024	} else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
1025		   kcb->kprobe_status == KPROBE_HIT_SSDONE) {
1026		/*
1027		 * We increment the nmissed count for accounting,
1028		 * we can also use npre/npostfault count for accounting
1029		 * these specific fault cases.
1030		 */
1031		kprobes_inc_nmissed_count(cur);
1032
1033		/*
1034		 * We come here because instructions in the pre/post
1035		 * handler caused the page_fault, this could happen
1036		 * if handler tries to access user space by
1037		 * copy_from_user(), get_user() etc. Let the
1038		 * user-specified handler try to fix it first.
1039		 */
1040		if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1041			return 1;
1042
1043		/*
1044		 * In case the user-specified fault handler returned
1045		 * zero, try to fix up.
1046		 */
1047		if (fixup_exception(regs, trapnr))
1048			return 1;
1049
1050		/*
1051		 * fixup routine could not handle it,
1052		 * Let do_page_fault() fix it.
1053		 */
1054	}
1055
1056	return 0;
1057}
1058NOKPROBE_SYMBOL(kprobe_fault_handler);
1059
1060/*
1061 * Wrapper routine for handling exceptions.
1062 */
1063int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
1064			     void *data)
1065{
1066	struct die_args *args = data;
1067	int ret = NOTIFY_DONE;
1068
1069	if (args->regs && user_mode(args->regs))
 
 
1070		return ret;
1071
1072	if (val == DIE_GPF) {
1073		/*
1074		 * To be potentially processing a kprobe fault and to
1075		 * trust the result from kprobe_running(), we have
1076		 * be non-preemptible.
1077		 */
1078		if (!preemptible() && kprobe_running() &&
1079		    kprobe_fault_handler(args->regs, args->trapnr))
1080			ret = NOTIFY_STOP;
1081	}
1082	return ret;
1083}
1084NOKPROBE_SYMBOL(kprobe_exceptions_notify);
1085
1086int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1087{
1088	struct jprobe *jp = container_of(p, struct jprobe, kp);
1089	unsigned long addr;
1090	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1091
1092	kcb->jprobe_saved_regs = *regs;
1093	kcb->jprobe_saved_sp = stack_addr(regs);
1094	addr = (unsigned long)(kcb->jprobe_saved_sp);
1095
1096	/*
1097	 * As Linus pointed out, gcc assumes that the callee
1098	 * owns the argument space and could overwrite it, e.g.
1099	 * tailcall optimization. So, to be absolutely safe
1100	 * we also save and restore enough stack bytes to cover
1101	 * the argument area.
1102	 * Use __memcpy() to avoid KASAN stack out-of-bounds reports as we copy
1103	 * raw stack chunk with redzones:
1104	 */
1105	__memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr, MIN_STACK_SIZE(addr));
1106	regs->ip = (unsigned long)(jp->entry);
1107
1108	/*
1109	 * jprobes use jprobe_return() which skips the normal return
1110	 * path of the function, and this messes up the accounting of the
1111	 * function graph tracer to get messed up.
1112	 *
1113	 * Pause function graph tracing while performing the jprobe function.
1114	 */
1115	pause_graph_tracing();
1116	return 1;
1117}
1118NOKPROBE_SYMBOL(setjmp_pre_handler);
1119
1120void jprobe_return(void)
1121{
1122	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1123
1124	/* Unpoison stack redzones in the frames we are going to jump over. */
1125	kasan_unpoison_stack_above_sp_to(kcb->jprobe_saved_sp);
1126
1127	asm volatile (
1128#ifdef CONFIG_X86_64
1129			"       xchg   %%rbx,%%rsp	\n"
1130#else
1131			"       xchgl   %%ebx,%%esp	\n"
1132#endif
1133			"       int3			\n"
1134			"       .globl jprobe_return_end\n"
1135			"       jprobe_return_end:	\n"
1136			"       nop			\n"::"b"
1137			(kcb->jprobe_saved_sp):"memory");
1138}
1139NOKPROBE_SYMBOL(jprobe_return);
1140NOKPROBE_SYMBOL(jprobe_return_end);
1141
1142int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1143{
1144	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1145	u8 *addr = (u8 *) (regs->ip - 1);
1146	struct jprobe *jp = container_of(p, struct jprobe, kp);
1147	void *saved_sp = kcb->jprobe_saved_sp;
1148
1149	if ((addr > (u8 *) jprobe_return) &&
1150	    (addr < (u8 *) jprobe_return_end)) {
1151		if (stack_addr(regs) != saved_sp) {
1152			struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1153			printk(KERN_ERR
1154			       "current sp %p does not match saved sp %p\n",
1155			       stack_addr(regs), saved_sp);
1156			printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1157			show_regs(saved_regs);
1158			printk(KERN_ERR "Current registers\n");
1159			show_regs(regs);
1160			BUG();
1161		}
1162		/* It's OK to start function graph tracing again */
1163		unpause_graph_tracing();
1164		*regs = kcb->jprobe_saved_regs;
1165		__memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1166		preempt_enable_no_resched();
1167		return 1;
1168	}
1169	return 0;
1170}
1171NOKPROBE_SYMBOL(longjmp_break_handler);
1172
1173bool arch_within_kprobe_blacklist(unsigned long addr)
1174{
1175	bool is_in_entry_trampoline_section = false;
1176
1177#ifdef CONFIG_X86_64
1178	is_in_entry_trampoline_section =
1179		(addr >= (unsigned long)__entry_trampoline_start &&
1180		 addr < (unsigned long)__entry_trampoline_end);
1181#endif
1182	return  (addr >= (unsigned long)__kprobes_text_start &&
1183		 addr < (unsigned long)__kprobes_text_end) ||
1184		(addr >= (unsigned long)__entry_text_start &&
1185		 addr < (unsigned long)__entry_text_end) ||
1186		is_in_entry_trampoline_section;
1187}
1188
1189int __init arch_init_kprobes(void)
1190{
1191	return 0;
1192}
1193
1194int arch_trampoline_kprobe(struct kprobe *p)
1195{
1196	return 0;
1197}