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}
v4.6
 
   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/module.h>
 
  49#include <linux/kdebug.h>
  50#include <linux/kallsyms.h>
  51#include <linux/ftrace.h>
  52#include <linux/frame.h>
 
 
  53
 
  54#include <asm/cacheflush.h>
  55#include <asm/desc.h>
  56#include <asm/pgtable.h>
  57#include <asm/uaccess.h>
  58#include <asm/alternative.h>
  59#include <asm/insn.h>
  60#include <asm/debugreg.h>
 
  61
  62#include "common.h"
  63
  64void jprobe_return_end(void);
  65
  66DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  67DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  68
  69#define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
  70
  71#define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
  72	(((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  73	  (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  74	  (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  75	  (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  76	 << (row % 32))
  77	/*
  78	 * Undefined/reserved opcodes, conditional jump, Opcode Extension
  79	 * Groups, and some special opcodes can not boost.
  80	 * This is non-const and volatile to keep gcc from statically
  81	 * optimizing it out, as variable_test_bit makes gcc think only
  82	 * *(unsigned long*) is used.
  83	 */
  84static volatile u32 twobyte_is_boostable[256 / 32] = {
  85	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  86	/*      ----------------------------------------------          */
  87	W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
  88	W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
  89	W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
  90	W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
  91	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
  92	W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
  93	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
  94	W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
  95	W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
  96	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
  97	W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
  98	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
  99	W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
 100	W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
 101	W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
 102	W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
 103	/*      -----------------------------------------------         */
 104	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
 105};
 106#undef W
 107
 108struct kretprobe_blackpoint kretprobe_blacklist[] = {
 109	{"__switch_to", }, /* This function switches only current task, but
 110			      doesn't switch kernel stack.*/
 111	{NULL, NULL}	/* Terminator */
 112};
 113
 114const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
 115
 116static nokprobe_inline void
 117__synthesize_relative_insn(void *from, void *to, u8 op)
 118{
 119	struct __arch_relative_insn {
 120		u8 op;
 121		s32 raddr;
 122	} __packed *insn;
 123
 124	insn = (struct __arch_relative_insn *)from;
 125	insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
 126	insn->op = op;
 127}
 128
 129/* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
 130void synthesize_reljump(void *from, void *to)
 131{
 132	__synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
 133}
 134NOKPROBE_SYMBOL(synthesize_reljump);
 135
 136/* Insert a call instruction at address 'from', which calls address 'to'.*/
 137void synthesize_relcall(void *from, void *to)
 138{
 139	__synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
 140}
 141NOKPROBE_SYMBOL(synthesize_relcall);
 142
 143/*
 144 * Skip the prefixes of the instruction.
 145 */
 146static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
 147{
 148	insn_attr_t attr;
 149
 150	attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 151	while (inat_is_legacy_prefix(attr)) {
 152		insn++;
 153		attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 154	}
 155#ifdef CONFIG_X86_64
 156	if (inat_is_rex_prefix(attr))
 157		insn++;
 158#endif
 159	return insn;
 160}
 161NOKPROBE_SYMBOL(skip_prefixes);
 162
 163/*
 164 * Returns non-zero if opcode is boostable.
 165 * RIP relative instructions are adjusted at copying time in 64 bits mode
 166 */
 167int can_boost(kprobe_opcode_t *opcodes)
 168{
 169	kprobe_opcode_t opcode;
 170	kprobe_opcode_t *orig_opcodes = opcodes;
 171
 172	if (search_exception_tables((unsigned long)opcodes))
 173		return 0;	/* Page fault may occur on this address. */
 174
 175retry:
 176	if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 
 
 
 
 
 
 
 
 177		return 0;
 178	opcode = *(opcodes++);
 179
 180	/* 2nd-byte opcode */
 181	if (opcode == 0x0f) {
 182		if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 183			return 0;
 184		return test_bit(*opcodes,
 185				(unsigned long *)twobyte_is_boostable);
 186	}
 187
 188	switch (opcode & 0xf0) {
 189#ifdef CONFIG_X86_64
 190	case 0x40:
 191		goto retry; /* REX prefix is boostable */
 192#endif
 193	case 0x60:
 194		if (0x63 < opcode && opcode < 0x67)
 195			goto retry; /* prefixes */
 196		/* can't boost Address-size override and bound */
 197		return (opcode != 0x62 && opcode != 0x67);
 198	case 0x70:
 199		return 0; /* can't boost conditional jump */
 
 
 200	case 0xc0:
 201		/* can't boost software-interruptions */
 202		return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 203	case 0xd0:
 204		/* can boost AA* and XLAT */
 205		return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 206	case 0xe0:
 207		/* can boost in/out and absolute jmps */
 208		return ((opcode & 0x04) || opcode == 0xea);
 209	case 0xf0:
 210		if ((opcode & 0x0c) == 0 && opcode != 0xf1)
 211			goto retry; /* lock/rep(ne) prefix */
 212		/* clear and set flags are boostable */
 213		return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 214	default:
 215		/* segment override prefixes are boostable */
 216		if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
 217			goto retry; /* prefixes */
 218		/* CS override prefix and call are not boostable */
 219		return (opcode != 0x2e && opcode != 0x9a);
 220	}
 221}
 222
 223static unsigned long
 224__recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
 225{
 226	struct kprobe *kp;
 227	unsigned long faddr;
 228
 229	kp = get_kprobe((void *)addr);
 230	faddr = ftrace_location(addr);
 231	/*
 232	 * Addresses inside the ftrace location are refused by
 233	 * arch_check_ftrace_location(). Something went terribly wrong
 234	 * if such an address is checked here.
 235	 */
 236	if (WARN_ON(faddr && faddr != addr))
 237		return 0UL;
 238	/*
 239	 * Use the current code if it is not modified by Kprobe
 240	 * and it cannot be modified by ftrace.
 241	 */
 242	if (!kp && !faddr)
 243		return addr;
 244
 245	/*
 246	 * Basically, kp->ainsn.insn has an original instruction.
 247	 * However, RIP-relative instruction can not do single-stepping
 248	 * at different place, __copy_instruction() tweaks the displacement of
 249	 * that instruction. In that case, we can't recover the instruction
 250	 * from the kp->ainsn.insn.
 251	 *
 252	 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
 253	 * of the first byte of the probed instruction, which is overwritten
 254	 * by int3. And the instruction at kp->addr is not modified by kprobes
 255	 * except for the first byte, we can recover the original instruction
 256	 * from it and kp->opcode.
 257	 *
 258	 * In case of Kprobes using ftrace, we do not have a copy of
 259	 * the original instruction. In fact, the ftrace location might
 260	 * be modified at anytime and even could be in an inconsistent state.
 261	 * Fortunately, we know that the original code is the ideal 5-byte
 262	 * long NOP.
 263	 */
 264	memcpy(buf, (void *)addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 
 
 
 265	if (faddr)
 266		memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
 267	else
 268		buf[0] = kp->opcode;
 269	return (unsigned long)buf;
 270}
 271
 272/*
 273 * Recover the probed instruction at addr for further analysis.
 274 * Caller must lock kprobes by kprobe_mutex, or disable preemption
 275 * for preventing to release referencing kprobes.
 276 * Returns zero if the instruction can not get recovered.
 277 */
 278unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
 279{
 280	unsigned long __addr;
 281
 282	__addr = __recover_optprobed_insn(buf, addr);
 283	if (__addr != addr)
 284		return __addr;
 285
 286	return __recover_probed_insn(buf, addr);
 287}
 288
 289/* Check if paddr is at an instruction boundary */
 290static int can_probe(unsigned long paddr)
 291{
 292	unsigned long addr, __addr, offset = 0;
 293	struct insn insn;
 294	kprobe_opcode_t buf[MAX_INSN_SIZE];
 295
 296	if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
 297		return 0;
 298
 299	/* Decode instructions */
 300	addr = paddr - offset;
 301	while (addr < paddr) {
 302		/*
 303		 * Check if the instruction has been modified by another
 304		 * kprobe, in which case we replace the breakpoint by the
 305		 * original instruction in our buffer.
 306		 * Also, jump optimization will change the breakpoint to
 307		 * relative-jump. Since the relative-jump itself is
 308		 * normally used, we just go through if there is no kprobe.
 309		 */
 310		__addr = recover_probed_instruction(buf, addr);
 311		if (!__addr)
 312			return 0;
 313		kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
 314		insn_get_length(&insn);
 315
 316		/*
 317		 * Another debugging subsystem might insert this breakpoint.
 318		 * In that case, we can't recover it.
 319		 */
 320		if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 321			return 0;
 322		addr += insn.length;
 323	}
 324
 325	return (addr == paddr);
 326}
 327
 328/*
 329 * Returns non-zero if opcode modifies the interrupt flag.
 330 */
 331static int is_IF_modifier(kprobe_opcode_t *insn)
 332{
 333	/* Skip prefixes */
 334	insn = skip_prefixes(insn);
 335
 336	switch (*insn) {
 337	case 0xfa:		/* cli */
 338	case 0xfb:		/* sti */
 339	case 0xcf:		/* iret/iretd */
 340	case 0x9d:		/* popf/popfd */
 341		return 1;
 342	}
 343
 344	return 0;
 345}
 346
 347/*
 348 * Copy an instruction and adjust the displacement if the instruction
 349 * uses the %rip-relative addressing mode.
 350 * If it does, Return the address of the 32-bit displacement word.
 351 * If not, return null.
 352 * Only applicable to 64-bit x86.
 353 */
 354int __copy_instruction(u8 *dest, u8 *src)
 355{
 356	struct insn insn;
 357	kprobe_opcode_t buf[MAX_INSN_SIZE];
 358	int length;
 359	unsigned long recovered_insn =
 360		recover_probed_instruction(buf, (unsigned long)src);
 361
 362	if (!recovered_insn)
 363		return 0;
 364	kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
 365	insn_get_length(&insn);
 366	length = insn.length;
 
 
 
 
 367
 368	/* Another subsystem puts a breakpoint, failed to recover */
 369	if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 
 
 
 
 370		return 0;
 371	memcpy(dest, insn.kaddr, length);
 372
 373#ifdef CONFIG_X86_64
 374	if (insn_rip_relative(&insn)) {
 
 375		s64 newdisp;
 376		u8 *disp;
 377		kernel_insn_init(&insn, dest, length);
 378		insn_get_displacement(&insn);
 379		/*
 380		 * The copied instruction uses the %rip-relative addressing
 381		 * mode.  Adjust the displacement for the difference between
 382		 * the original location of this instruction and the location
 383		 * of the copy that will actually be run.  The tricky bit here
 384		 * is making sure that the sign extension happens correctly in
 385		 * this calculation, since we need a signed 32-bit result to
 386		 * be sign-extended to 64 bits when it's added to the %rip
 387		 * value and yield the same 64-bit result that the sign-
 388		 * extension of the original signed 32-bit displacement would
 389		 * have given.
 390		 */
 391		newdisp = (u8 *) src + (s64) insn.displacement.value - (u8 *) dest;
 
 392		if ((s64) (s32) newdisp != newdisp) {
 393			pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
 394			pr_err("\tSrc: %p, Dest: %p, old disp: %x\n", src, dest, insn.displacement.value);
 395			return 0;
 396		}
 397		disp = (u8 *) dest + insn_offset_displacement(&insn);
 398		*(s32 *) disp = (s32) newdisp;
 399	}
 400#endif
 401	return length;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 402}
 403
 404static int arch_copy_kprobe(struct kprobe *p)
 405{
 406	int ret;
 
 
 407
 408	/* Copy an instruction with recovering if other optprobe modifies it.*/
 409	ret = __copy_instruction(p->ainsn.insn, p->addr);
 410	if (!ret)
 411		return -EINVAL;
 412
 413	/*
 414	 * __copy_instruction can modify the displacement of the instruction,
 415	 * but it doesn't affect boostable check.
 416	 */
 417	if (can_boost(p->ainsn.insn))
 418		p->ainsn.boostable = 0;
 419	else
 420		p->ainsn.boostable = -1;
 421
 422	/* Check whether the instruction modifies Interrupt Flag or not */
 423	p->ainsn.if_modifier = is_IF_modifier(p->ainsn.insn);
 424
 425	/* Also, displacement change doesn't affect the first byte */
 426	p->opcode = p->ainsn.insn[0];
 
 
 
 427
 428	return 0;
 429}
 430
 431int arch_prepare_kprobe(struct kprobe *p)
 432{
 
 
 433	if (alternatives_text_reserved(p->addr, p->addr))
 434		return -EINVAL;
 435
 436	if (!can_probe((unsigned long)p->addr))
 437		return -EILSEQ;
 438	/* insn: must be on special executable page on x86. */
 439	p->ainsn.insn = get_insn_slot();
 440	if (!p->ainsn.insn)
 441		return -ENOMEM;
 442
 443	return arch_copy_kprobe(p);
 
 
 
 
 
 
 444}
 445
 446void arch_arm_kprobe(struct kprobe *p)
 447{
 448	text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 449}
 450
 451void arch_disarm_kprobe(struct kprobe *p)
 452{
 453	text_poke(p->addr, &p->opcode, 1);
 454}
 455
 456void arch_remove_kprobe(struct kprobe *p)
 457{
 458	if (p->ainsn.insn) {
 459		free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
 460		p->ainsn.insn = NULL;
 461	}
 462}
 463
 464static nokprobe_inline void
 465save_previous_kprobe(struct kprobe_ctlblk *kcb)
 466{
 467	kcb->prev_kprobe.kp = kprobe_running();
 468	kcb->prev_kprobe.status = kcb->kprobe_status;
 469	kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
 470	kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
 471}
 472
 473static nokprobe_inline void
 474restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 475{
 476	__this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
 477	kcb->kprobe_status = kcb->prev_kprobe.status;
 478	kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
 479	kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
 480}
 481
 482static nokprobe_inline void
 483set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 484		   struct kprobe_ctlblk *kcb)
 485{
 486	__this_cpu_write(current_kprobe, p);
 487	kcb->kprobe_saved_flags = kcb->kprobe_old_flags
 488		= (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 489	if (p->ainsn.if_modifier)
 490		kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
 491}
 492
 493static nokprobe_inline void clear_btf(void)
 494{
 495	if (test_thread_flag(TIF_BLOCKSTEP)) {
 496		unsigned long debugctl = get_debugctlmsr();
 497
 498		debugctl &= ~DEBUGCTLMSR_BTF;
 499		update_debugctlmsr(debugctl);
 500	}
 501}
 502
 503static nokprobe_inline void restore_btf(void)
 504{
 505	if (test_thread_flag(TIF_BLOCKSTEP)) {
 506		unsigned long debugctl = get_debugctlmsr();
 507
 508		debugctl |= DEBUGCTLMSR_BTF;
 509		update_debugctlmsr(debugctl);
 510	}
 511}
 512
 513void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
 514{
 515	unsigned long *sara = stack_addr(regs);
 516
 517	ri->ret_addr = (kprobe_opcode_t *) *sara;
 
 518
 519	/* Replace the return addr with trampoline addr */
 520	*sara = (unsigned long) &kretprobe_trampoline;
 521}
 522NOKPROBE_SYMBOL(arch_prepare_kretprobe);
 523
 524static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
 525			     struct kprobe_ctlblk *kcb, int reenter)
 526{
 527	if (setup_detour_execution(p, regs, reenter))
 528		return;
 529
 530#if !defined(CONFIG_PREEMPT)
 531	if (p->ainsn.boostable == 1 && !p->post_handler) {
 532		/* Boost up -- we can execute copied instructions directly */
 533		if (!reenter)
 534			reset_current_kprobe();
 535		/*
 536		 * Reentering boosted probe doesn't reset current_kprobe,
 537		 * nor set current_kprobe, because it doesn't use single
 538		 * stepping.
 539		 */
 540		regs->ip = (unsigned long)p->ainsn.insn;
 541		preempt_enable_no_resched();
 542		return;
 543	}
 544#endif
 545	if (reenter) {
 546		save_previous_kprobe(kcb);
 547		set_current_kprobe(p, regs, kcb);
 548		kcb->kprobe_status = KPROBE_REENTER;
 549	} else
 550		kcb->kprobe_status = KPROBE_HIT_SS;
 551	/* Prepare real single stepping */
 552	clear_btf();
 553	regs->flags |= X86_EFLAGS_TF;
 554	regs->flags &= ~X86_EFLAGS_IF;
 555	/* single step inline if the instruction is an int3 */
 556	if (p->opcode == BREAKPOINT_INSTRUCTION)
 557		regs->ip = (unsigned long)p->addr;
 558	else
 559		regs->ip = (unsigned long)p->ainsn.insn;
 560}
 561NOKPROBE_SYMBOL(setup_singlestep);
 562
 563/*
 564 * We have reentered the kprobe_handler(), since another probe was hit while
 565 * within the handler. We save the original kprobes variables and just single
 566 * step on the instruction of the new probe without calling any user handlers.
 567 */
 568static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
 569			  struct kprobe_ctlblk *kcb)
 570{
 571	switch (kcb->kprobe_status) {
 572	case KPROBE_HIT_SSDONE:
 573	case KPROBE_HIT_ACTIVE:
 574	case KPROBE_HIT_SS:
 575		kprobes_inc_nmissed_count(p);
 576		setup_singlestep(p, regs, kcb, 1);
 577		break;
 578	case KPROBE_REENTER:
 579		/* A probe has been hit in the codepath leading up to, or just
 580		 * after, single-stepping of a probed instruction. This entire
 581		 * codepath should strictly reside in .kprobes.text section.
 582		 * Raise a BUG or we'll continue in an endless reentering loop
 583		 * and eventually a stack overflow.
 584		 */
 585		printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
 586		       p->addr);
 587		dump_kprobe(p);
 588		BUG();
 589	default:
 590		/* impossible cases */
 591		WARN_ON(1);
 592		return 0;
 593	}
 594
 595	return 1;
 596}
 597NOKPROBE_SYMBOL(reenter_kprobe);
 598
 599/*
 600 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 601 * remain disabled throughout this function.
 602 */
 603int kprobe_int3_handler(struct pt_regs *regs)
 604{
 605	kprobe_opcode_t *addr;
 606	struct kprobe *p;
 607	struct kprobe_ctlblk *kcb;
 608
 609	if (user_mode(regs))
 610		return 0;
 611
 612	addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 613	/*
 614	 * We don't want to be preempted for the entire
 615	 * duration of kprobe processing. We conditionally
 616	 * re-enable preemption at the end of this function,
 617	 * and also in reenter_kprobe() and setup_singlestep().
 618	 */
 619	preempt_disable();
 620
 621	kcb = get_kprobe_ctlblk();
 622	p = get_kprobe(addr);
 623
 624	if (p) {
 625		if (kprobe_running()) {
 626			if (reenter_kprobe(p, regs, kcb))
 627				return 1;
 628		} else {
 629			set_current_kprobe(p, regs, kcb);
 630			kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 631
 632			/*
 633			 * If we have no pre-handler or it returned 0, we
 634			 * continue with normal processing.  If we have a
 635			 * pre-handler and it returned non-zero, it prepped
 636			 * for calling the break_handler below on re-entry
 637			 * for jprobe processing, so get out doing nothing
 638			 * more here.
 639			 */
 640			if (!p->pre_handler || !p->pre_handler(p, regs))
 641				setup_singlestep(p, regs, kcb, 0);
 
 
 642			return 1;
 643		}
 644	} else if (*addr != BREAKPOINT_INSTRUCTION) {
 645		/*
 646		 * The breakpoint instruction was removed right
 647		 * after we hit it.  Another cpu has removed
 648		 * either a probepoint or a debugger breakpoint
 649		 * at this address.  In either case, no further
 650		 * handling of this interrupt is appropriate.
 651		 * Back up over the (now missing) int3 and run
 652		 * the original instruction.
 653		 */
 654		regs->ip = (unsigned long)addr;
 655		preempt_enable_no_resched();
 656		return 1;
 657	} else if (kprobe_running()) {
 658		p = __this_cpu_read(current_kprobe);
 659		if (p->break_handler && p->break_handler(p, regs)) {
 660			if (!skip_singlestep(p, regs, kcb))
 661				setup_singlestep(p, regs, kcb, 0);
 662			return 1;
 663		}
 664	} /* else: not a kprobe fault; let the kernel handle it */
 665
 666	preempt_enable_no_resched();
 667	return 0;
 668}
 669NOKPROBE_SYMBOL(kprobe_int3_handler);
 670
 671/*
 672 * When a retprobed function returns, this code saves registers and
 673 * calls trampoline_handler() runs, which calls the kretprobe's handler.
 674 */
 675asm(
 
 676	".global kretprobe_trampoline\n"
 677	".type kretprobe_trampoline, @function\n"
 678	"kretprobe_trampoline:\n"
 
 679#ifdef CONFIG_X86_64
 680	/* We don't bother saving the ss register */
 681	"	pushq %rsp\n"
 682	"	pushfq\n"
 683	SAVE_REGS_STRING
 684	"	movq %rsp, %rdi\n"
 685	"	call trampoline_handler\n"
 686	/* Replace saved sp with true return address. */
 687	"	movq %rax, 152(%rsp)\n"
 688	RESTORE_REGS_STRING
 689	"	popfq\n"
 690#else
 691	"	pushf\n"
 
 692	SAVE_REGS_STRING
 693	"	movl %esp, %eax\n"
 694	"	call trampoline_handler\n"
 695	/* Move flags to cs */
 696	"	movl 56(%esp), %edx\n"
 697	"	movl %edx, 52(%esp)\n"
 698	/* Replace saved flags with true return address. */
 699	"	movl %eax, 56(%esp)\n"
 700	RESTORE_REGS_STRING
 701	"	popf\n"
 702#endif
 703	"	ret\n"
 704	".size kretprobe_trampoline, .-kretprobe_trampoline\n"
 705);
 706NOKPROBE_SYMBOL(kretprobe_trampoline);
 707STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
 708
 
 
 
 
 709/*
 710 * Called from kretprobe_trampoline
 711 */
 712__visible __used void *trampoline_handler(struct pt_regs *regs)
 713{
 
 714	struct kretprobe_instance *ri = NULL;
 715	struct hlist_head *head, empty_rp;
 716	struct hlist_node *tmp;
 717	unsigned long flags, orig_ret_address = 0;
 718	unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 719	kprobe_opcode_t *correct_ret_addr = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 720
 721	INIT_HLIST_HEAD(&empty_rp);
 722	kretprobe_hash_lock(current, &head, &flags);
 723	/* fixup registers */
 724#ifdef CONFIG_X86_64
 725	regs->cs = __KERNEL_CS;
 726#else
 727	regs->cs = __KERNEL_CS | get_kernel_rpl();
 728	regs->gs = 0;
 729#endif
 
 
 730	regs->ip = trampoline_address;
 731	regs->orig_ax = ~0UL;
 732
 733	/*
 734	 * It is possible to have multiple instances associated with a given
 735	 * task either because multiple functions in the call path have
 736	 * return probes installed on them, and/or more than one
 737	 * return probe was registered for a target function.
 738	 *
 739	 * We can handle this because:
 740	 *     - instances are always pushed into the head of the list
 741	 *     - when multiple return probes are registered for the same
 742	 *	 function, the (chronologically) first instance's ret_addr
 743	 *	 will be the real return address, and all the rest will
 744	 *	 point to kretprobe_trampoline.
 745	 */
 746	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
 747		if (ri->task != current)
 748			/* another task is sharing our hash bucket */
 749			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 750
 751		orig_ret_address = (unsigned long)ri->ret_addr;
 
 
 
 752
 753		if (orig_ret_address != trampoline_address)
 754			/*
 755			 * This is the real return address. Any other
 756			 * instances associated with this task are for
 757			 * other calls deeper on the call stack
 758			 */
 759			break;
 760	}
 761
 762	kretprobe_assert(ri, orig_ret_address, trampoline_address);
 763
 764	correct_ret_addr = ri->ret_addr;
 765	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
 766		if (ri->task != current)
 767			/* another task is sharing our hash bucket */
 768			continue;
 
 
 769
 770		orig_ret_address = (unsigned long)ri->ret_addr;
 771		if (ri->rp && ri->rp->handler) {
 772			__this_cpu_write(current_kprobe, &ri->rp->kp);
 773			get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 774			ri->ret_addr = correct_ret_addr;
 775			ri->rp->handler(ri, regs);
 776			__this_cpu_write(current_kprobe, NULL);
 777		}
 778
 779		recycle_rp_inst(ri, &empty_rp);
 780
 781		if (orig_ret_address != trampoline_address)
 782			/*
 783			 * This is the real return address. Any other
 784			 * instances associated with this task are for
 785			 * other calls deeper on the call stack
 786			 */
 787			break;
 788	}
 789
 790	kretprobe_hash_unlock(current, &flags);
 791
 
 
 
 792	hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
 793		hlist_del(&ri->hlist);
 794		kfree(ri);
 795	}
 796	return (void *)orig_ret_address;
 797}
 798NOKPROBE_SYMBOL(trampoline_handler);
 799
 800/*
 801 * Called after single-stepping.  p->addr is the address of the
 802 * instruction whose first byte has been replaced by the "int 3"
 803 * instruction.  To avoid the SMP problems that can occur when we
 804 * temporarily put back the original opcode to single-step, we
 805 * single-stepped a copy of the instruction.  The address of this
 806 * copy is p->ainsn.insn.
 807 *
 808 * This function prepares to return from the post-single-step
 809 * interrupt.  We have to fix up the stack as follows:
 810 *
 811 * 0) Except in the case of absolute or indirect jump or call instructions,
 812 * the new ip is relative to the copied instruction.  We need to make
 813 * it relative to the original instruction.
 814 *
 815 * 1) If the single-stepped instruction was pushfl, then the TF and IF
 816 * flags are set in the just-pushed flags, and may need to be cleared.
 817 *
 818 * 2) If the single-stepped instruction was a call, the return address
 819 * that is atop the stack is the address following the copied instruction.
 820 * We need to make it the address following the original instruction.
 821 *
 822 * If this is the first time we've single-stepped the instruction at
 823 * this probepoint, and the instruction is boostable, boost it: add a
 824 * jump instruction after the copied instruction, that jumps to the next
 825 * instruction after the probepoint.
 826 */
 827static void resume_execution(struct kprobe *p, struct pt_regs *regs,
 828			     struct kprobe_ctlblk *kcb)
 829{
 830	unsigned long *tos = stack_addr(regs);
 831	unsigned long copy_ip = (unsigned long)p->ainsn.insn;
 832	unsigned long orig_ip = (unsigned long)p->addr;
 833	kprobe_opcode_t *insn = p->ainsn.insn;
 834
 835	/* Skip prefixes */
 836	insn = skip_prefixes(insn);
 837
 838	regs->flags &= ~X86_EFLAGS_TF;
 839	switch (*insn) {
 840	case 0x9c:	/* pushfl */
 841		*tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
 842		*tos |= kcb->kprobe_old_flags;
 843		break;
 844	case 0xc2:	/* iret/ret/lret */
 845	case 0xc3:
 846	case 0xca:
 847	case 0xcb:
 848	case 0xcf:
 849	case 0xea:	/* jmp absolute -- ip is correct */
 850		/* ip is already adjusted, no more changes required */
 851		p->ainsn.boostable = 1;
 852		goto no_change;
 853	case 0xe8:	/* call relative - Fix return addr */
 854		*tos = orig_ip + (*tos - copy_ip);
 855		break;
 856#ifdef CONFIG_X86_32
 857	case 0x9a:	/* call absolute -- same as call absolute, indirect */
 858		*tos = orig_ip + (*tos - copy_ip);
 859		goto no_change;
 860#endif
 861	case 0xff:
 862		if ((insn[1] & 0x30) == 0x10) {
 863			/*
 864			 * call absolute, indirect
 865			 * Fix return addr; ip is correct.
 866			 * But this is not boostable
 867			 */
 868			*tos = orig_ip + (*tos - copy_ip);
 869			goto no_change;
 870		} else if (((insn[1] & 0x31) == 0x20) ||
 871			   ((insn[1] & 0x31) == 0x21)) {
 872			/*
 873			 * jmp near and far, absolute indirect
 874			 * ip is correct. And this is boostable
 875			 */
 876			p->ainsn.boostable = 1;
 877			goto no_change;
 878		}
 879	default:
 880		break;
 881	}
 882
 883	if (p->ainsn.boostable == 0) {
 884		if ((regs->ip > copy_ip) &&
 885		    (regs->ip - copy_ip) + 5 < MAX_INSN_SIZE) {
 886			/*
 887			 * These instructions can be executed directly if it
 888			 * jumps back to correct address.
 889			 */
 890			synthesize_reljump((void *)regs->ip,
 891				(void *)orig_ip + (regs->ip - copy_ip));
 892			p->ainsn.boostable = 1;
 893		} else {
 894			p->ainsn.boostable = -1;
 895		}
 896	}
 897
 898	regs->ip += orig_ip - copy_ip;
 899
 900no_change:
 901	restore_btf();
 902}
 903NOKPROBE_SYMBOL(resume_execution);
 904
 905/*
 906 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 907 * remain disabled throughout this function.
 908 */
 909int kprobe_debug_handler(struct pt_regs *regs)
 910{
 911	struct kprobe *cur = kprobe_running();
 912	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 913
 914	if (!cur)
 915		return 0;
 916
 917	resume_execution(cur, regs, kcb);
 918	regs->flags |= kcb->kprobe_saved_flags;
 919
 920	if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 921		kcb->kprobe_status = KPROBE_HIT_SSDONE;
 922		cur->post_handler(cur, regs, 0);
 923	}
 924
 925	/* Restore back the original saved kprobes variables and continue. */
 926	if (kcb->kprobe_status == KPROBE_REENTER) {
 927		restore_previous_kprobe(kcb);
 928		goto out;
 929	}
 930	reset_current_kprobe();
 931out:
 932	preempt_enable_no_resched();
 933
 934	/*
 935	 * if somebody else is singlestepping across a probe point, flags
 936	 * will have TF set, in which case, continue the remaining processing
 937	 * of do_debug, as if this is not a probe hit.
 938	 */
 939	if (regs->flags & X86_EFLAGS_TF)
 940		return 0;
 941
 942	return 1;
 943}
 944NOKPROBE_SYMBOL(kprobe_debug_handler);
 945
 946int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 947{
 948	struct kprobe *cur = kprobe_running();
 949	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 950
 951	if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
 952		/* This must happen on single-stepping */
 953		WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
 954			kcb->kprobe_status != KPROBE_REENTER);
 955		/*
 956		 * We are here because the instruction being single
 957		 * stepped caused a page fault. We reset the current
 958		 * kprobe and the ip points back to the probe address
 959		 * and allow the page fault handler to continue as a
 960		 * normal page fault.
 961		 */
 962		regs->ip = (unsigned long)cur->addr;
 
 
 
 
 
 
 
 
 
 
 
 963		regs->flags |= kcb->kprobe_old_flags;
 
 964		if (kcb->kprobe_status == KPROBE_REENTER)
 965			restore_previous_kprobe(kcb);
 966		else
 967			reset_current_kprobe();
 968		preempt_enable_no_resched();
 969	} else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
 970		   kcb->kprobe_status == KPROBE_HIT_SSDONE) {
 971		/*
 972		 * We increment the nmissed count for accounting,
 973		 * we can also use npre/npostfault count for accounting
 974		 * these specific fault cases.
 975		 */
 976		kprobes_inc_nmissed_count(cur);
 977
 978		/*
 979		 * We come here because instructions in the pre/post
 980		 * handler caused the page_fault, this could happen
 981		 * if handler tries to access user space by
 982		 * copy_from_user(), get_user() etc. Let the
 983		 * user-specified handler try to fix it first.
 984		 */
 985		if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 986			return 1;
 987
 988		/*
 989		 * In case the user-specified fault handler returned
 990		 * zero, try to fix up.
 991		 */
 992		if (fixup_exception(regs, trapnr))
 993			return 1;
 994
 995		/*
 996		 * fixup routine could not handle it,
 997		 * Let do_page_fault() fix it.
 998		 */
 999	}
1000
1001	return 0;
1002}
1003NOKPROBE_SYMBOL(kprobe_fault_handler);
1004
1005/*
1006 * Wrapper routine for handling exceptions.
1007 */
1008int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
1009			     void *data)
1010{
1011	struct die_args *args = data;
1012	int ret = NOTIFY_DONE;
1013
1014	if (args->regs && user_mode(args->regs))
 
 
1015		return ret;
1016
1017	if (val == DIE_GPF) {
1018		/*
1019		 * To be potentially processing a kprobe fault and to
1020		 * trust the result from kprobe_running(), we have
1021		 * be non-preemptible.
1022		 */
1023		if (!preemptible() && kprobe_running() &&
1024		    kprobe_fault_handler(args->regs, args->trapnr))
1025			ret = NOTIFY_STOP;
1026	}
1027	return ret;
1028}
1029NOKPROBE_SYMBOL(kprobe_exceptions_notify);
1030
1031int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1032{
1033	struct jprobe *jp = container_of(p, struct jprobe, kp);
1034	unsigned long addr;
1035	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1036
1037	kcb->jprobe_saved_regs = *regs;
1038	kcb->jprobe_saved_sp = stack_addr(regs);
1039	addr = (unsigned long)(kcb->jprobe_saved_sp);
1040
1041	/*
1042	 * As Linus pointed out, gcc assumes that the callee
1043	 * owns the argument space and could overwrite it, e.g.
1044	 * tailcall optimization. So, to be absolutely safe
1045	 * we also save and restore enough stack bytes to cover
1046	 * the argument area.
1047	 */
1048	memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
1049	       MIN_STACK_SIZE(addr));
1050	regs->flags &= ~X86_EFLAGS_IF;
1051	trace_hardirqs_off();
1052	regs->ip = (unsigned long)(jp->entry);
1053
1054	/*
1055	 * jprobes use jprobe_return() which skips the normal return
1056	 * path of the function, and this messes up the accounting of the
1057	 * function graph tracer to get messed up.
1058	 *
1059	 * Pause function graph tracing while performing the jprobe function.
1060	 */
1061	pause_graph_tracing();
1062	return 1;
1063}
1064NOKPROBE_SYMBOL(setjmp_pre_handler);
1065
1066void jprobe_return(void)
1067{
1068	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1069
1070	asm volatile (
1071#ifdef CONFIG_X86_64
1072			"       xchg   %%rbx,%%rsp	\n"
1073#else
1074			"       xchgl   %%ebx,%%esp	\n"
1075#endif
1076			"       int3			\n"
1077			"       .globl jprobe_return_end\n"
1078			"       jprobe_return_end:	\n"
1079			"       nop			\n"::"b"
1080			(kcb->jprobe_saved_sp):"memory");
1081}
1082NOKPROBE_SYMBOL(jprobe_return);
1083NOKPROBE_SYMBOL(jprobe_return_end);
1084
1085int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1086{
1087	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1088	u8 *addr = (u8 *) (regs->ip - 1);
1089	struct jprobe *jp = container_of(p, struct jprobe, kp);
1090	void *saved_sp = kcb->jprobe_saved_sp;
1091
1092	if ((addr > (u8 *) jprobe_return) &&
1093	    (addr < (u8 *) jprobe_return_end)) {
1094		if (stack_addr(regs) != saved_sp) {
1095			struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1096			printk(KERN_ERR
1097			       "current sp %p does not match saved sp %p\n",
1098			       stack_addr(regs), saved_sp);
1099			printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1100			show_regs(saved_regs);
1101			printk(KERN_ERR "Current registers\n");
1102			show_regs(regs);
1103			BUG();
1104		}
1105		/* It's OK to start function graph tracing again */
1106		unpause_graph_tracing();
1107		*regs = kcb->jprobe_saved_regs;
1108		memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1109		preempt_enable_no_resched();
1110		return 1;
1111	}
1112	return 0;
1113}
1114NOKPROBE_SYMBOL(longjmp_break_handler);
1115
1116bool arch_within_kprobe_blacklist(unsigned long addr)
1117{
1118	return  (addr >= (unsigned long)__kprobes_text_start &&
1119		 addr < (unsigned long)__kprobes_text_end) ||
1120		(addr >= (unsigned long)__entry_text_start &&
1121		 addr < (unsigned long)__entry_text_end);
1122}
1123
1124int __init arch_init_kprobes(void)
1125{
1126	return 0;
1127}
1128
1129int arch_trampoline_kprobe(struct kprobe *p)
1130{
1131	return 0;
1132}