1 // SPDX-License-Identifier: GPL-2.0-only
2 #define pr_fmt(fmt) "SMP alternatives: " fmt
4 #include <linux/module.h>
5 #include <linux/sched.h>
6 #include <linux/perf_event.h>
7 #include <linux/mutex.h>
8 #include <linux/list.h>
9 #include <linux/stringify.h>
10 #include <linux/highmem.h>
12 #include <linux/vmalloc.h>
13 #include <linux/memory.h>
14 #include <linux/stop_machine.h>
15 #include <linux/slab.h>
16 #include <linux/kdebug.h>
17 #include <linux/kprobes.h>
18 #include <linux/mmu_context.h>
19 #include <linux/bsearch.h>
20 #include <linux/sync_core.h>
21 #include <asm/text-patching.h>
22 #include <asm/alternative.h>
23 #include <asm/sections.h>
26 #include <asm/cacheflush.h>
27 #include <asm/tlbflush.h>
30 #include <asm/fixmap.h>
32 int __read_mostly alternatives_patched;
34 EXPORT_SYMBOL_GPL(alternatives_patched);
36 #define MAX_PATCH_LEN (255-1)
38 static int __initdata_or_module debug_alternative;
40 static int __init debug_alt(char *str)
42 debug_alternative = 1;
45 __setup("debug-alternative", debug_alt);
47 static int noreplace_smp;
49 static int __init setup_noreplace_smp(char *str)
54 __setup("noreplace-smp", setup_noreplace_smp);
56 #define DPRINTK(fmt, args...) \
58 if (debug_alternative) \
59 printk(KERN_DEBUG pr_fmt(fmt) "\n", ##args); \
62 #define DUMP_BYTES(buf, len, fmt, args...) \
64 if (unlikely(debug_alternative)) { \
70 printk(KERN_DEBUG pr_fmt(fmt), ##args); \
71 for (j = 0; j < (len) - 1; j++) \
72 printk(KERN_CONT "%02hhx ", buf[j]); \
73 printk(KERN_CONT "%02hhx\n", buf[j]); \
78 * Each GENERIC_NOPX is of X bytes, and defined as an array of bytes
79 * that correspond to that nop. Getting from one nop to the next, we
80 * add to the array the offset that is equal to the sum of all sizes of
81 * nops preceding the one we are after.
83 * Note: The GENERIC_NOP5_ATOMIC is at the end, as it breaks the
84 * nice symmetry of sizes of the previous nops.
86 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
87 static const unsigned char intelnops[] =
99 static const unsigned char * const intel_nops[ASM_NOP_MAX+2] =
105 intelnops + 1 + 2 + 3,
106 intelnops + 1 + 2 + 3 + 4,
107 intelnops + 1 + 2 + 3 + 4 + 5,
108 intelnops + 1 + 2 + 3 + 4 + 5 + 6,
109 intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
110 intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
115 static const unsigned char k8nops[] =
127 static const unsigned char * const k8_nops[ASM_NOP_MAX+2] =
134 k8nops + 1 + 2 + 3 + 4,
135 k8nops + 1 + 2 + 3 + 4 + 5,
136 k8nops + 1 + 2 + 3 + 4 + 5 + 6,
137 k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
138 k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
142 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
143 static const unsigned char k7nops[] =
155 static const unsigned char * const k7_nops[ASM_NOP_MAX+2] =
162 k7nops + 1 + 2 + 3 + 4,
163 k7nops + 1 + 2 + 3 + 4 + 5,
164 k7nops + 1 + 2 + 3 + 4 + 5 + 6,
165 k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
166 k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
171 static const unsigned char p6nops[] =
183 static const unsigned char * const p6_nops[ASM_NOP_MAX+2] =
190 p6nops + 1 + 2 + 3 + 4,
191 p6nops + 1 + 2 + 3 + 4 + 5,
192 p6nops + 1 + 2 + 3 + 4 + 5 + 6,
193 p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
194 p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
198 /* Initialize these to a safe default */
200 const unsigned char * const *ideal_nops = p6_nops;
202 const unsigned char * const *ideal_nops = intel_nops;
205 void __init arch_init_ideal_nops(void)
207 switch (boot_cpu_data.x86_vendor) {
208 case X86_VENDOR_INTEL:
210 * Due to a decoder implementation quirk, some
211 * specific Intel CPUs actually perform better with
212 * the "k8_nops" than with the SDM-recommended NOPs.
214 if (boot_cpu_data.x86 == 6 &&
215 boot_cpu_data.x86_model >= 0x0f &&
216 boot_cpu_data.x86_model != 0x1c &&
217 boot_cpu_data.x86_model != 0x26 &&
218 boot_cpu_data.x86_model != 0x27 &&
219 boot_cpu_data.x86_model < 0x30) {
220 ideal_nops = k8_nops;
221 } else if (boot_cpu_has(X86_FEATURE_NOPL)) {
222 ideal_nops = p6_nops;
225 ideal_nops = k8_nops;
227 ideal_nops = intel_nops;
232 case X86_VENDOR_HYGON:
233 ideal_nops = p6_nops;
237 if (boot_cpu_data.x86 > 0xf) {
238 ideal_nops = p6_nops;
246 ideal_nops = k8_nops;
248 if (boot_cpu_has(X86_FEATURE_K8))
249 ideal_nops = k8_nops;
250 else if (boot_cpu_has(X86_FEATURE_K7))
251 ideal_nops = k7_nops;
253 ideal_nops = intel_nops;
258 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
259 static void __init_or_module add_nops(void *insns, unsigned int len)
262 unsigned int noplen = len;
263 if (noplen > ASM_NOP_MAX)
264 noplen = ASM_NOP_MAX;
265 memcpy(insns, ideal_nops[noplen], noplen);
271 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
272 extern s32 __smp_locks[], __smp_locks_end[];
273 void text_poke_early(void *addr, const void *opcode, size_t len);
276 * Are we looking at a near JMP with a 1 or 4-byte displacement.
278 static inline bool is_jmp(const u8 opcode)
280 return opcode == 0xeb || opcode == 0xe9;
283 static void __init_or_module
284 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insn_buff)
286 u8 *next_rip, *tgt_rip;
290 if (a->replacementlen != 5)
293 o_dspl = *(s32 *)(insn_buff + 1);
295 /* next_rip of the replacement JMP */
296 next_rip = repl_insn + a->replacementlen;
297 /* target rip of the replacement JMP */
298 tgt_rip = next_rip + o_dspl;
299 n_dspl = tgt_rip - orig_insn;
301 DPRINTK("target RIP: %px, new_displ: 0x%x", tgt_rip, n_dspl);
303 if (tgt_rip - orig_insn >= 0) {
304 if (n_dspl - 2 <= 127)
308 /* negative offset */
310 if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
320 insn_buff[1] = (s8)n_dspl;
321 add_nops(insn_buff + 2, 3);
330 *(s32 *)&insn_buff[1] = n_dspl;
336 DPRINTK("final displ: 0x%08x, JMP 0x%lx",
337 n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
341 * "noinline" to cause control flow change and thus invalidate I$ and
342 * cause refetch after modification.
344 static void __init_or_module noinline optimize_nops(struct alt_instr *a, u8 *instr)
349 for (i = 0; i < a->padlen; i++) {
350 if (instr[i] != 0x90)
354 local_irq_save(flags);
355 add_nops(instr + (a->instrlen - a->padlen), a->padlen);
356 local_irq_restore(flags);
358 DUMP_BYTES(instr, a->instrlen, "%px: [%d:%d) optimized NOPs: ",
359 instr, a->instrlen - a->padlen, a->padlen);
363 * Replace instructions with better alternatives for this CPU type. This runs
364 * before SMP is initialized to avoid SMP problems with self modifying code.
365 * This implies that asymmetric systems where APs have less capabilities than
366 * the boot processor are not handled. Tough. Make sure you disable such
369 * Marked "noinline" to cause control flow change and thus insn cache
370 * to refetch changed I$ lines.
372 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
373 struct alt_instr *end)
376 u8 *instr, *replacement;
377 u8 insn_buff[MAX_PATCH_LEN];
379 DPRINTK("alt table %px, -> %px", start, end);
381 * The scan order should be from start to end. A later scanned
382 * alternative code can overwrite previously scanned alternative code.
383 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
386 * So be careful if you want to change the scan order to any other
389 for (a = start; a < end; a++) {
390 int insn_buff_sz = 0;
392 instr = (u8 *)&a->instr_offset + a->instr_offset;
393 replacement = (u8 *)&a->repl_offset + a->repl_offset;
394 BUG_ON(a->instrlen > sizeof(insn_buff));
395 BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
396 if (!boot_cpu_has(a->cpuid)) {
398 optimize_nops(a, instr);
403 DPRINTK("feat: %d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d), pad: %d",
406 instr, instr, a->instrlen,
407 replacement, a->replacementlen, a->padlen);
409 DUMP_BYTES(instr, a->instrlen, "%px: old_insn: ", instr);
410 DUMP_BYTES(replacement, a->replacementlen, "%px: rpl_insn: ", replacement);
412 memcpy(insn_buff, replacement, a->replacementlen);
413 insn_buff_sz = a->replacementlen;
416 * 0xe8 is a relative jump; fix the offset.
418 * Instruction length is checked before the opcode to avoid
419 * accessing uninitialized bytes for zero-length replacements.
421 if (a->replacementlen == 5 && *insn_buff == 0xe8) {
422 *(s32 *)(insn_buff + 1) += replacement - instr;
423 DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
424 *(s32 *)(insn_buff + 1),
425 (unsigned long)instr + *(s32 *)(insn_buff + 1) + 5);
428 if (a->replacementlen && is_jmp(replacement[0]))
429 recompute_jump(a, instr, replacement, insn_buff);
431 if (a->instrlen > a->replacementlen) {
432 add_nops(insn_buff + a->replacementlen,
433 a->instrlen - a->replacementlen);
434 insn_buff_sz += a->instrlen - a->replacementlen;
436 DUMP_BYTES(insn_buff, insn_buff_sz, "%px: final_insn: ", instr);
438 text_poke_early(instr, insn_buff, insn_buff_sz);
443 static void alternatives_smp_lock(const s32 *start, const s32 *end,
444 u8 *text, u8 *text_end)
448 for (poff = start; poff < end; poff++) {
449 u8 *ptr = (u8 *)poff + *poff;
451 if (!*poff || ptr < text || ptr >= text_end)
453 /* turn DS segment override prefix into lock prefix */
455 text_poke(ptr, ((unsigned char []){0xf0}), 1);
459 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
460 u8 *text, u8 *text_end)
464 for (poff = start; poff < end; poff++) {
465 u8 *ptr = (u8 *)poff + *poff;
467 if (!*poff || ptr < text || ptr >= text_end)
469 /* turn lock prefix into DS segment override prefix */
471 text_poke(ptr, ((unsigned char []){0x3E}), 1);
475 struct smp_alt_module {
476 /* what is this ??? */
480 /* ptrs to lock prefixes */
482 const s32 *locks_end;
484 /* .text segment, needed to avoid patching init code ;) */
488 struct list_head next;
490 static LIST_HEAD(smp_alt_modules);
491 static bool uniproc_patched = false; /* protected by text_mutex */
493 void __init_or_module alternatives_smp_module_add(struct module *mod,
495 void *locks, void *locks_end,
496 void *text, void *text_end)
498 struct smp_alt_module *smp;
500 mutex_lock(&text_mutex);
501 if (!uniproc_patched)
504 if (num_possible_cpus() == 1)
505 /* Don't bother remembering, we'll never have to undo it. */
508 smp = kzalloc(sizeof(*smp), GFP_KERNEL);
510 /* we'll run the (safe but slow) SMP code then ... */
516 smp->locks_end = locks_end;
518 smp->text_end = text_end;
519 DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
520 smp->locks, smp->locks_end,
521 smp->text, smp->text_end, smp->name);
523 list_add_tail(&smp->next, &smp_alt_modules);
525 alternatives_smp_unlock(locks, locks_end, text, text_end);
527 mutex_unlock(&text_mutex);
530 void __init_or_module alternatives_smp_module_del(struct module *mod)
532 struct smp_alt_module *item;
534 mutex_lock(&text_mutex);
535 list_for_each_entry(item, &smp_alt_modules, next) {
536 if (mod != item->mod)
538 list_del(&item->next);
542 mutex_unlock(&text_mutex);
545 void alternatives_enable_smp(void)
547 struct smp_alt_module *mod;
549 /* Why bother if there are no other CPUs? */
550 BUG_ON(num_possible_cpus() == 1);
552 mutex_lock(&text_mutex);
554 if (uniproc_patched) {
555 pr_info("switching to SMP code\n");
556 BUG_ON(num_online_cpus() != 1);
557 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
558 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
559 list_for_each_entry(mod, &smp_alt_modules, next)
560 alternatives_smp_lock(mod->locks, mod->locks_end,
561 mod->text, mod->text_end);
562 uniproc_patched = false;
564 mutex_unlock(&text_mutex);
568 * Return 1 if the address range is reserved for SMP-alternatives.
569 * Must hold text_mutex.
571 int alternatives_text_reserved(void *start, void *end)
573 struct smp_alt_module *mod;
575 u8 *text_start = start;
578 lockdep_assert_held(&text_mutex);
580 list_for_each_entry(mod, &smp_alt_modules, next) {
581 if (mod->text > text_end || mod->text_end < text_start)
583 for (poff = mod->locks; poff < mod->locks_end; poff++) {
584 const u8 *ptr = (const u8 *)poff + *poff;
586 if (text_start <= ptr && text_end > ptr)
593 #endif /* CONFIG_SMP */
595 #ifdef CONFIG_PARAVIRT
596 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
597 struct paravirt_patch_site *end)
599 struct paravirt_patch_site *p;
600 char insn_buff[MAX_PATCH_LEN];
602 for (p = start; p < end; p++) {
605 BUG_ON(p->len > MAX_PATCH_LEN);
606 /* prep the buffer with the original instructions */
607 memcpy(insn_buff, p->instr, p->len);
608 used = pv_ops.init.patch(p->type, insn_buff, (unsigned long)p->instr, p->len);
610 BUG_ON(used > p->len);
612 /* Pad the rest with nops */
613 add_nops(insn_buff + used, p->len - used);
614 text_poke_early(p->instr, insn_buff, p->len);
617 extern struct paravirt_patch_site __start_parainstructions[],
618 __stop_parainstructions[];
619 #endif /* CONFIG_PARAVIRT */
622 * Self-test for the INT3 based CALL emulation code.
624 * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up
625 * properly and that there is a stack gap between the INT3 frame and the
626 * previous context. Without this gap doing a virtual PUSH on the interrupted
627 * stack would corrupt the INT3 IRET frame.
629 * See entry_{32,64}.S for more details.
633 * We define the int3_magic() function in assembly to control the calling
634 * convention such that we can 'call' it from assembly.
637 extern void int3_magic(unsigned int *ptr); /* defined in asm */
640 " .pushsection .init.text, \"ax\", @progbits\n"
641 " .type int3_magic, @function\n"
643 " movl $1, (%" _ASM_ARG1 ")\n"
645 " .size int3_magic, .-int3_magic\n"
649 extern __initdata unsigned long int3_selftest_ip; /* defined in asm below */
652 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data)
654 struct die_args *args = data;
655 struct pt_regs *regs = args->regs;
657 if (!regs || user_mode(regs))
663 if (regs->ip - INT3_INSN_SIZE != int3_selftest_ip)
666 int3_emulate_call(regs, (unsigned long)&int3_magic);
670 static void __init int3_selftest(void)
672 static __initdata struct notifier_block int3_exception_nb = {
673 .notifier_call = int3_exception_notify,
674 .priority = INT_MAX-1, /* last */
676 unsigned int val = 0;
678 BUG_ON(register_die_notifier(&int3_exception_nb));
681 * Basically: int3_magic(&val); but really complicated :-)
683 * Stick the address of the INT3 instruction into int3_selftest_ip,
684 * then trigger the INT3, padded with NOPs to match a CALL instruction
687 asm volatile ("1: int3; nop; nop; nop; nop\n\t"
688 ".pushsection .init.data,\"aw\"\n\t"
689 ".align " __ASM_SEL(4, 8) "\n\t"
690 ".type int3_selftest_ip, @object\n\t"
691 ".size int3_selftest_ip, " __ASM_SEL(4, 8) "\n\t"
692 "int3_selftest_ip:\n\t"
693 __ASM_SEL(.long, .quad) " 1b\n\t"
695 : ASM_CALL_CONSTRAINT
696 : __ASM_SEL_RAW(a, D) (&val)
701 unregister_die_notifier(&int3_exception_nb);
704 void __init alternative_instructions(void)
709 * The patching is not fully atomic, so try to avoid local
710 * interruptions that might execute the to be patched code.
711 * Other CPUs are not running.
716 * Don't stop machine check exceptions while patching.
717 * MCEs only happen when something got corrupted and in this
718 * case we must do something about the corruption.
719 * Ignoring it is worse than an unlikely patching race.
720 * Also machine checks tend to be broadcast and if one CPU
721 * goes into machine check the others follow quickly, so we don't
722 * expect a machine check to cause undue problems during to code
726 apply_alternatives(__alt_instructions, __alt_instructions_end);
729 /* Patch to UP if other cpus not imminent. */
730 if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
731 uniproc_patched = true;
732 alternatives_smp_module_add(NULL, "core kernel",
733 __smp_locks, __smp_locks_end,
737 if (!uniproc_patched || num_possible_cpus() == 1) {
738 free_init_pages("SMP alternatives",
739 (unsigned long)__smp_locks,
740 (unsigned long)__smp_locks_end);
744 apply_paravirt(__parainstructions, __parainstructions_end);
747 alternatives_patched = 1;
751 * text_poke_early - Update instructions on a live kernel at boot time
752 * @addr: address to modify
753 * @opcode: source of the copy
754 * @len: length to copy
756 * When you use this code to patch more than one byte of an instruction
757 * you need to make sure that other CPUs cannot execute this code in parallel.
758 * Also no thread must be currently preempted in the middle of these
759 * instructions. And on the local CPU you need to be protected against NMI or
760 * MCE handlers seeing an inconsistent instruction while you patch.
762 void __init_or_module text_poke_early(void *addr, const void *opcode,
767 if (boot_cpu_has(X86_FEATURE_NX) &&
768 is_module_text_address((unsigned long)addr)) {
770 * Modules text is marked initially as non-executable, so the
771 * code cannot be running and speculative code-fetches are
772 * prevented. Just change the code.
774 memcpy(addr, opcode, len);
776 local_irq_save(flags);
777 memcpy(addr, opcode, len);
778 local_irq_restore(flags);
782 * Could also do a CLFLUSH here to speed up CPU recovery; but
783 * that causes hangs on some VIA CPUs.
789 struct mm_struct *mm;
793 * Using a temporary mm allows to set temporary mappings that are not accessible
794 * by other CPUs. Such mappings are needed to perform sensitive memory writes
795 * that override the kernel memory protections (e.g., W^X), without exposing the
796 * temporary page-table mappings that are required for these write operations to
797 * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
798 * mapping is torn down.
800 * Context: The temporary mm needs to be used exclusively by a single core. To
801 * harden security IRQs must be disabled while the temporary mm is
802 * loaded, thereby preventing interrupt handler bugs from overriding
803 * the kernel memory protection.
805 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
807 temp_mm_state_t temp_state;
809 lockdep_assert_irqs_disabled();
810 temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
811 switch_mm_irqs_off(NULL, mm, current);
814 * If breakpoints are enabled, disable them while the temporary mm is
815 * used. Userspace might set up watchpoints on addresses that are used
816 * in the temporary mm, which would lead to wrong signals being sent or
819 * Note that breakpoints are not disabled selectively, which also causes
820 * kernel breakpoints (e.g., perf's) to be disabled. This might be
821 * undesirable, but still seems reasonable as the code that runs in the
822 * temporary mm should be short.
824 if (hw_breakpoint_active())
825 hw_breakpoint_disable();
830 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
832 lockdep_assert_irqs_disabled();
833 switch_mm_irqs_off(NULL, prev_state.mm, current);
836 * Restore the breakpoints if they were disabled before the temporary mm
839 if (hw_breakpoint_active())
840 hw_breakpoint_restore();
843 __ro_after_init struct mm_struct *poking_mm;
844 __ro_after_init unsigned long poking_addr;
846 static void *__text_poke(void *addr, const void *opcode, size_t len)
848 bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE;
849 struct page *pages[2] = {NULL};
850 temp_mm_state_t prev;
857 * While boot memory allocator is running we cannot use struct pages as
858 * they are not yet initialized. There is no way to recover.
860 BUG_ON(!after_bootmem);
862 if (!core_kernel_text((unsigned long)addr)) {
863 pages[0] = vmalloc_to_page(addr);
864 if (cross_page_boundary)
865 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
867 pages[0] = virt_to_page(addr);
868 WARN_ON(!PageReserved(pages[0]));
869 if (cross_page_boundary)
870 pages[1] = virt_to_page(addr + PAGE_SIZE);
873 * If something went wrong, crash and burn since recovery paths are not
876 BUG_ON(!pages[0] || (cross_page_boundary && !pages[1]));
879 * Map the page without the global bit, as TLB flushing is done with
880 * flush_tlb_mm_range(), which is intended for non-global PTEs.
882 pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL);
885 * The lock is not really needed, but this allows to avoid open-coding.
887 ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
890 * This must not fail; preallocated in poking_init().
894 local_irq_save(flags);
896 pte = mk_pte(pages[0], pgprot);
897 set_pte_at(poking_mm, poking_addr, ptep, pte);
899 if (cross_page_boundary) {
900 pte = mk_pte(pages[1], pgprot);
901 set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte);
905 * Loading the temporary mm behaves as a compiler barrier, which
906 * guarantees that the PTE will be set at the time memcpy() is done.
908 prev = use_temporary_mm(poking_mm);
910 kasan_disable_current();
911 memcpy((u8 *)poking_addr + offset_in_page(addr), opcode, len);
912 kasan_enable_current();
915 * Ensure that the PTE is only cleared after the instructions of memcpy
916 * were issued by using a compiler barrier.
920 pte_clear(poking_mm, poking_addr, ptep);
921 if (cross_page_boundary)
922 pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1);
925 * Loading the previous page-table hierarchy requires a serializing
926 * instruction that already allows the core to see the updated version.
927 * Xen-PV is assumed to serialize execution in a similar manner.
929 unuse_temporary_mm(prev);
932 * Flushing the TLB might involve IPIs, which would require enabled
933 * IRQs, but not if the mm is not used, as it is in this point.
935 flush_tlb_mm_range(poking_mm, poking_addr, poking_addr +
936 (cross_page_boundary ? 2 : 1) * PAGE_SIZE,
940 * If the text does not match what we just wrote then something is
941 * fundamentally screwy; there's nothing we can really do about that.
943 BUG_ON(memcmp(addr, opcode, len));
945 local_irq_restore(flags);
946 pte_unmap_unlock(ptep, ptl);
951 * text_poke - Update instructions on a live kernel
952 * @addr: address to modify
953 * @opcode: source of the copy
954 * @len: length to copy
956 * Only atomic text poke/set should be allowed when not doing early patching.
957 * It means the size must be writable atomically and the address must be aligned
958 * in a way that permits an atomic write. It also makes sure we fit on a single
961 * Note that the caller must ensure that if the modified code is part of a
962 * module, the module would not be removed during poking. This can be achieved
963 * by registering a module notifier, and ordering module removal and patching
966 void *text_poke(void *addr, const void *opcode, size_t len)
968 lockdep_assert_held(&text_mutex);
970 return __text_poke(addr, opcode, len);
974 * text_poke_kgdb - Update instructions on a live kernel by kgdb
975 * @addr: address to modify
976 * @opcode: source of the copy
977 * @len: length to copy
979 * Only atomic text poke/set should be allowed when not doing early patching.
980 * It means the size must be writable atomically and the address must be aligned
981 * in a way that permits an atomic write. It also makes sure we fit on a single
984 * Context: should only be used by kgdb, which ensures no other core is running,
985 * despite the fact it does not hold the text_mutex.
987 void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
989 return __text_poke(addr, opcode, len);
992 static void do_sync_core(void *info)
997 void text_poke_sync(void)
999 on_each_cpu(do_sync_core, NULL, 1);
1002 struct text_poke_loc {
1003 s32 rel_addr; /* addr := _stext + rel_addr */
1006 const u8 text[POKE_MAX_OPCODE_SIZE];
1010 struct bp_patching_desc {
1011 struct text_poke_loc *vec;
1016 static struct bp_patching_desc *bp_desc;
1018 static __always_inline
1019 struct bp_patching_desc *try_get_desc(struct bp_patching_desc **descp)
1021 struct bp_patching_desc *desc = __READ_ONCE(*descp); /* rcu_dereference */
1023 if (!desc || !arch_atomic_inc_not_zero(&desc->refs))
1029 static __always_inline void put_desc(struct bp_patching_desc *desc)
1031 smp_mb__before_atomic();
1032 arch_atomic_dec(&desc->refs);
1035 static __always_inline void *text_poke_addr(struct text_poke_loc *tp)
1037 return _stext + tp->rel_addr;
1040 static __always_inline int patch_cmp(const void *key, const void *elt)
1042 struct text_poke_loc *tp = (struct text_poke_loc *) elt;
1044 if (key < text_poke_addr(tp))
1046 if (key > text_poke_addr(tp))
1051 noinstr int poke_int3_handler(struct pt_regs *regs)
1053 struct bp_patching_desc *desc;
1054 struct text_poke_loc *tp;
1058 if (user_mode(regs))
1062 * Having observed our INT3 instruction, we now must observe
1065 * bp_desc = desc INT3
1067 * write INT3 if (desc)
1071 desc = try_get_desc(&bp_desc);
1076 * Discount the INT3. See text_poke_bp_batch().
1078 ip = (void *) regs->ip - INT3_INSN_SIZE;
1081 * Skip the binary search if there is a single member in the vector.
1083 if (unlikely(desc->nr_entries > 1)) {
1084 tp = __inline_bsearch(ip, desc->vec, desc->nr_entries,
1085 sizeof(struct text_poke_loc),
1091 if (text_poke_addr(tp) != ip)
1095 len = text_opcode_size(tp->opcode);
1098 switch (tp->opcode) {
1099 case INT3_INSN_OPCODE:
1101 * Someone poked an explicit INT3, they'll want to handle it,
1106 case RET_INSN_OPCODE:
1107 int3_emulate_ret(regs);
1110 case CALL_INSN_OPCODE:
1111 int3_emulate_call(regs, (long)ip + tp->rel32);
1114 case JMP32_INSN_OPCODE:
1115 case JMP8_INSN_OPCODE:
1116 int3_emulate_jmp(regs, (long)ip + tp->rel32);
1130 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc))
1131 static struct text_poke_loc tp_vec[TP_VEC_MAX];
1132 static int tp_vec_nr;
1135 * text_poke_bp_batch() -- update instructions on live kernel on SMP
1136 * @tp: vector of instructions to patch
1137 * @nr_entries: number of entries in the vector
1139 * Modify multi-byte instruction by using int3 breakpoint on SMP.
1140 * We completely avoid stop_machine() here, and achieve the
1141 * synchronization using int3 breakpoint.
1143 * The way it is done:
1144 * - For each entry in the vector:
1145 * - add a int3 trap to the address that will be patched
1147 * - For each entry in the vector:
1148 * - update all but the first byte of the patched range
1150 * - For each entry in the vector:
1151 * - replace the first byte (int3) by the first byte of
1155 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
1157 struct bp_patching_desc desc = {
1159 .nr_entries = nr_entries,
1160 .refs = ATOMIC_INIT(1),
1162 unsigned char int3 = INT3_INSN_OPCODE;
1166 lockdep_assert_held(&text_mutex);
1168 smp_store_release(&bp_desc, &desc); /* rcu_assign_pointer */
1171 * Corresponding read barrier in int3 notifier for making sure the
1172 * nr_entries and handler are correctly ordered wrt. patching.
1177 * First step: add a int3 trap to the address that will be patched.
1179 for (i = 0; i < nr_entries; i++) {
1180 tp[i].old = *(u8 *)text_poke_addr(&tp[i]);
1181 text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE);
1187 * Second step: update all but the first byte of the patched range.
1189 for (do_sync = 0, i = 0; i < nr_entries; i++) {
1190 u8 old[POKE_MAX_OPCODE_SIZE] = { tp[i].old, };
1191 int len = text_opcode_size(tp[i].opcode);
1193 if (len - INT3_INSN_SIZE > 0) {
1194 memcpy(old + INT3_INSN_SIZE,
1195 text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1196 len - INT3_INSN_SIZE);
1197 text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1198 (const char *)tp[i].text + INT3_INSN_SIZE,
1199 len - INT3_INSN_SIZE);
1204 * Emit a perf event to record the text poke, primarily to
1205 * support Intel PT decoding which must walk the executable code
1206 * to reconstruct the trace. The flow up to here is:
1209 * - write instruction tail
1210 * At this point the actual control flow will be through the
1211 * INT3 and handler and not hit the old or new instruction.
1212 * Intel PT outputs FUP/TIP packets for the INT3, so the flow
1213 * can still be decoded. Subsequently:
1214 * - emit RECORD_TEXT_POKE with the new instruction
1216 * - write first byte
1218 * So before the text poke event timestamp, the decoder will see
1219 * either the old instruction flow or FUP/TIP of INT3. After the
1220 * text poke event timestamp, the decoder will see either the
1221 * new instruction flow or FUP/TIP of INT3. Thus decoders can
1222 * use the timestamp as the point at which to modify the
1224 * The old instruction is recorded so that the event can be
1225 * processed forwards or backwards.
1227 perf_event_text_poke(text_poke_addr(&tp[i]), old, len,
1233 * According to Intel, this core syncing is very likely
1234 * not necessary and we'd be safe even without it. But
1235 * better safe than sorry (plus there's not only Intel).
1241 * Third step: replace the first byte (int3) by the first byte of
1244 for (do_sync = 0, i = 0; i < nr_entries; i++) {
1245 if (tp[i].text[0] == INT3_INSN_OPCODE)
1248 text_poke(text_poke_addr(&tp[i]), tp[i].text, INT3_INSN_SIZE);
1256 * Remove and synchronize_rcu(), except we have a very primitive
1257 * refcount based completion.
1259 WRITE_ONCE(bp_desc, NULL); /* RCU_INIT_POINTER */
1260 if (!atomic_dec_and_test(&desc.refs))
1261 atomic_cond_read_acquire(&desc.refs, !VAL);
1264 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
1265 const void *opcode, size_t len, const void *emulate)
1269 memcpy((void *)tp->text, opcode, len);
1273 kernel_insn_init(&insn, emulate, MAX_INSN_SIZE);
1274 insn_get_length(&insn);
1276 BUG_ON(!insn_complete(&insn));
1277 BUG_ON(len != insn.length);
1279 tp->rel_addr = addr - (void *)_stext;
1280 tp->opcode = insn.opcode.bytes[0];
1282 switch (tp->opcode) {
1283 case INT3_INSN_OPCODE:
1284 case RET_INSN_OPCODE:
1287 case CALL_INSN_OPCODE:
1288 case JMP32_INSN_OPCODE:
1289 case JMP8_INSN_OPCODE:
1290 tp->rel32 = insn.immediate.value;
1293 default: /* assume NOP */
1295 case 2: /* NOP2 -- emulate as JMP8+0 */
1296 BUG_ON(memcmp(emulate, ideal_nops[len], len));
1297 tp->opcode = JMP8_INSN_OPCODE;
1301 case 5: /* NOP5 -- emulate as JMP32+0 */
1302 BUG_ON(memcmp(emulate, ideal_nops[NOP_ATOMIC5], len));
1303 tp->opcode = JMP32_INSN_OPCODE;
1307 default: /* unknown instruction */
1315 * We hard rely on the tp_vec being ordered; ensure this is so by flushing
1318 static bool tp_order_fail(void *addr)
1320 struct text_poke_loc *tp;
1325 if (!addr) /* force */
1328 tp = &tp_vec[tp_vec_nr - 1];
1329 if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr)
1335 static void text_poke_flush(void *addr)
1337 if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) {
1338 text_poke_bp_batch(tp_vec, tp_vec_nr);
1343 void text_poke_finish(void)
1345 text_poke_flush(NULL);
1348 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate)
1350 struct text_poke_loc *tp;
1352 if (unlikely(system_state == SYSTEM_BOOTING)) {
1353 text_poke_early(addr, opcode, len);
1357 text_poke_flush(addr);
1359 tp = &tp_vec[tp_vec_nr++];
1360 text_poke_loc_init(tp, addr, opcode, len, emulate);
1364 * text_poke_bp() -- update instructions on live kernel on SMP
1365 * @addr: address to patch
1366 * @opcode: opcode of new instruction
1367 * @len: length to copy
1368 * @handler: address to jump to when the temporary breakpoint is hit
1370 * Update a single instruction with the vector in the stack, avoiding
1371 * dynamically allocated memory. This function should be used when it is
1372 * not possible to allocate memory.
1374 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate)
1376 struct text_poke_loc tp;
1378 if (unlikely(system_state == SYSTEM_BOOTING)) {
1379 text_poke_early(addr, opcode, len);
1383 text_poke_loc_init(&tp, addr, opcode, len, emulate);
1384 text_poke_bp_batch(&tp, 1);