1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 1995 Linus Torvalds
4 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
5 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
6 */
7#include <linux/sched.h> /* test_thread_flag(), ... */
8#include <linux/sched/task_stack.h> /* task_stack_*(), ... */
9#include <linux/kdebug.h> /* oops_begin/end, ... */
10#include <linux/extable.h> /* search_exception_tables */
11#include <linux/memblock.h> /* max_low_pfn */
12#include <linux/kfence.h> /* kfence_handle_page_fault */
13#include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
14#include <linux/mmiotrace.h> /* kmmio_handler, ... */
15#include <linux/perf_event.h> /* perf_sw_event */
16#include <linux/hugetlb.h> /* hstate_index_to_shift */
17#include <linux/prefetch.h> /* prefetchw */
18#include <linux/context_tracking.h> /* exception_enter(), ... */
19#include <linux/uaccess.h> /* faulthandler_disabled() */
20#include <linux/efi.h> /* efi_crash_gracefully_on_page_fault()*/
21#include <linux/mm_types.h>
22#include <linux/mm.h> /* find_and_lock_vma() */
23
24#include <asm/cpufeature.h> /* boot_cpu_has, ... */
25#include <asm/traps.h> /* dotraplinkage, ... */
26#include <asm/fixmap.h> /* VSYSCALL_ADDR */
27#include <asm/vsyscall.h> /* emulate_vsyscall */
28#include <asm/vm86.h> /* struct vm86 */
29#include <asm/mmu_context.h> /* vma_pkey() */
30#include <asm/efi.h> /* efi_crash_gracefully_on_page_fault()*/
31#include <asm/desc.h> /* store_idt(), ... */
32#include <asm/cpu_entry_area.h> /* exception stack */
33#include <asm/pgtable_areas.h> /* VMALLOC_START, ... */
34#include <asm/kvm_para.h> /* kvm_handle_async_pf */
35#include <asm/vdso.h> /* fixup_vdso_exception() */
36#include <asm/irq_stack.h>
37#include <asm/fred.h>
38#include <asm/sev.h> /* snp_dump_hva_rmpentry() */
39
40#define CREATE_TRACE_POINTS
41#include <asm/trace/exceptions.h>
42
43/*
44 * Returns 0 if mmiotrace is disabled, or if the fault is not
45 * handled by mmiotrace:
46 */
47static nokprobe_inline int
48kmmio_fault(struct pt_regs *regs, unsigned long addr)
49{
50 if (unlikely(is_kmmio_active()))
51 if (kmmio_handler(regs, addr) == 1)
52 return -1;
53 return 0;
54}
55
56/*
57 * Prefetch quirks:
58 *
59 * 32-bit mode:
60 *
61 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
62 * Check that here and ignore it. This is AMD erratum #91.
63 *
64 * 64-bit mode:
65 *
66 * Sometimes the CPU reports invalid exceptions on prefetch.
67 * Check that here and ignore it.
68 *
69 * Opcode checker based on code by Richard Brunner.
70 */
71static inline int
72check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
73 unsigned char opcode, int *prefetch)
74{
75 unsigned char instr_hi = opcode & 0xf0;
76 unsigned char instr_lo = opcode & 0x0f;
77
78 switch (instr_hi) {
79 case 0x20:
80 case 0x30:
81 /*
82 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
83 * In X86_64 long mode, the CPU will signal invalid
84 * opcode if some of these prefixes are present so
85 * X86_64 will never get here anyway
86 */
87 return ((instr_lo & 7) == 0x6);
88#ifdef CONFIG_X86_64
89 case 0x40:
90 /*
91 * In 64-bit mode 0x40..0x4F are valid REX prefixes
92 */
93 return (!user_mode(regs) || user_64bit_mode(regs));
94#endif
95 case 0x60:
96 /* 0x64 thru 0x67 are valid prefixes in all modes. */
97 return (instr_lo & 0xC) == 0x4;
98 case 0xF0:
99 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
100 return !instr_lo || (instr_lo>>1) == 1;
101 case 0x00:
102 /* Prefetch instruction is 0x0F0D or 0x0F18 */
103 if (get_kernel_nofault(opcode, instr))
104 return 0;
105
106 *prefetch = (instr_lo == 0xF) &&
107 (opcode == 0x0D || opcode == 0x18);
108 return 0;
109 default:
110 return 0;
111 }
112}
113
114static bool is_amd_k8_pre_npt(void)
115{
116 struct cpuinfo_x86 *c = &boot_cpu_data;
117
118 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
119 c->x86_vendor == X86_VENDOR_AMD &&
120 c->x86 == 0xf && c->x86_model < 0x40);
121}
122
123static int
124is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
125{
126 unsigned char *max_instr;
127 unsigned char *instr;
128 int prefetch = 0;
129
130 /* Erratum #91 affects AMD K8, pre-NPT CPUs */
131 if (!is_amd_k8_pre_npt())
132 return 0;
133
134 /*
135 * If it was a exec (instruction fetch) fault on NX page, then
136 * do not ignore the fault:
137 */
138 if (error_code & X86_PF_INSTR)
139 return 0;
140
141 instr = (void *)convert_ip_to_linear(current, regs);
142 max_instr = instr + 15;
143
144 /*
145 * This code has historically always bailed out if IP points to a
146 * not-present page (e.g. due to a race). No one has ever
147 * complained about this.
148 */
149 pagefault_disable();
150
151 while (instr < max_instr) {
152 unsigned char opcode;
153
154 if (user_mode(regs)) {
155 if (get_user(opcode, (unsigned char __user *) instr))
156 break;
157 } else {
158 if (get_kernel_nofault(opcode, instr))
159 break;
160 }
161
162 instr++;
163
164 if (!check_prefetch_opcode(regs, instr, opcode, prefetch: &prefetch))
165 break;
166 }
167
168 pagefault_enable();
169 return prefetch;
170}
171
172DEFINE_SPINLOCK(pgd_lock);
173LIST_HEAD(pgd_list);
174
175#ifdef CONFIG_X86_32
176static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
177{
178 unsigned index = pgd_index(address);
179 pgd_t *pgd_k;
180 p4d_t *p4d, *p4d_k;
181 pud_t *pud, *pud_k;
182 pmd_t *pmd, *pmd_k;
183
184 pgd += index;
185 pgd_k = init_mm.pgd + index;
186
187 if (!pgd_present(*pgd_k))
188 return NULL;
189
190 /*
191 * set_pgd(pgd, *pgd_k); here would be useless on PAE
192 * and redundant with the set_pmd() on non-PAE. As would
193 * set_p4d/set_pud.
194 */
195 p4d = p4d_offset(pgd, address);
196 p4d_k = p4d_offset(pgd_k, address);
197 if (!p4d_present(*p4d_k))
198 return NULL;
199
200 pud = pud_offset(p4d, address);
201 pud_k = pud_offset(p4d_k, address);
202 if (!pud_present(*pud_k))
203 return NULL;
204
205 pmd = pmd_offset(pud, address);
206 pmd_k = pmd_offset(pud_k, address);
207
208 if (pmd_present(*pmd) != pmd_present(*pmd_k))
209 set_pmd(pmd, *pmd_k);
210
211 if (!pmd_present(*pmd_k))
212 return NULL;
213 else
214 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
215
216 return pmd_k;
217}
218
219/*
220 * Handle a fault on the vmalloc or module mapping area
221 *
222 * This is needed because there is a race condition between the time
223 * when the vmalloc mapping code updates the PMD to the point in time
224 * where it synchronizes this update with the other page-tables in the
225 * system.
226 *
227 * In this race window another thread/CPU can map an area on the same
228 * PMD, finds it already present and does not synchronize it with the
229 * rest of the system yet. As a result v[mz]alloc might return areas
230 * which are not mapped in every page-table in the system, causing an
231 * unhandled page-fault when they are accessed.
232 */
233static noinline int vmalloc_fault(unsigned long address)
234{
235 unsigned long pgd_paddr;
236 pmd_t *pmd_k;
237 pte_t *pte_k;
238
239 /* Make sure we are in vmalloc area: */
240 if (!(address >= VMALLOC_START && address < VMALLOC_END))
241 return -1;
242
243 /*
244 * Synchronize this task's top level page-table
245 * with the 'reference' page table.
246 *
247 * Do _not_ use "current" here. We might be inside
248 * an interrupt in the middle of a task switch..
249 */
250 pgd_paddr = read_cr3_pa();
251 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
252 if (!pmd_k)
253 return -1;
254
255 if (pmd_leaf(*pmd_k))
256 return 0;
257
258 pte_k = pte_offset_kernel(pmd_k, address);
259 if (!pte_present(*pte_k))
260 return -1;
261
262 return 0;
263}
264NOKPROBE_SYMBOL(vmalloc_fault);
265
266void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
267{
268 unsigned long addr;
269
270 for (addr = start & PMD_MASK;
271 addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
272 addr += PMD_SIZE) {
273 struct page *page;
274
275 spin_lock(&pgd_lock);
276 list_for_each_entry(page, &pgd_list, lru) {
277 spinlock_t *pgt_lock;
278
279 /* the pgt_lock only for Xen */
280 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
281
282 spin_lock(pgt_lock);
283 vmalloc_sync_one(page_address(page), addr);
284 spin_unlock(pgt_lock);
285 }
286 spin_unlock(&pgd_lock);
287 }
288}
289
290static bool low_pfn(unsigned long pfn)
291{
292 return pfn < max_low_pfn;
293}
294
295static void dump_pagetable(unsigned long address)
296{
297 pgd_t *base = __va(read_cr3_pa());
298 pgd_t *pgd = &base[pgd_index(address)];
299 p4d_t *p4d;
300 pud_t *pud;
301 pmd_t *pmd;
302 pte_t *pte;
303
304#ifdef CONFIG_X86_PAE
305 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
306 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
307 goto out;
308#define pr_pde pr_cont
309#else
310#define pr_pde pr_info
311#endif
312 p4d = p4d_offset(pgd, address);
313 pud = pud_offset(p4d, address);
314 pmd = pmd_offset(pud, address);
315 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
316#undef pr_pde
317
318 /*
319 * We must not directly access the pte in the highpte
320 * case if the page table is located in highmem.
321 * And let's rather not kmap-atomic the pte, just in case
322 * it's allocated already:
323 */
324 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_leaf(*pmd))
325 goto out;
326
327 pte = pte_offset_kernel(pmd, address);
328 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
329out:
330 pr_cont("\n");
331}
332
333#else /* CONFIG_X86_64: */
334
335#ifdef CONFIG_CPU_SUP_AMD
336static const char errata93_warning[] =
337KERN_ERR
338"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
339"******* Working around it, but it may cause SEGVs or burn power.\n"
340"******* Please consider a BIOS update.\n"
341"******* Disabling USB legacy in the BIOS may also help.\n";
342#endif
343
344static int bad_address(void *p)
345{
346 unsigned long dummy;
347
348 return get_kernel_nofault(dummy, (unsigned long *)p);
349}
350
351static void dump_pagetable(unsigned long address)
352{
353 pgd_t *base = __va(read_cr3_pa());
354 pgd_t *pgd = base + pgd_index(address);
355 p4d_t *p4d;
356 pud_t *pud;
357 pmd_t *pmd;
358 pte_t *pte;
359
360 if (bad_address(p: pgd))
361 goto bad;
362
363 pr_info("PGD %lx ", pgd_val(*pgd));
364
365 if (!pgd_present(pgd: *pgd))
366 goto out;
367
368 p4d = p4d_offset(pgd, address);
369 if (bad_address(p: p4d))
370 goto bad;
371
372 pr_cont("P4D %lx ", p4d_val(*p4d));
373 if (!p4d_present(p4d: *p4d) || p4d_leaf(p4d: *p4d))
374 goto out;
375
376 pud = pud_offset(p4d, address);
377 if (bad_address(p: pud))
378 goto bad;
379
380 pr_cont("PUD %lx ", pud_val(*pud));
381 if (!pud_present(pud: *pud) || pud_leaf(pud: *pud))
382 goto out;
383
384 pmd = pmd_offset(pud, address);
385 if (bad_address(p: pmd))
386 goto bad;
387
388 pr_cont("PMD %lx ", pmd_val(*pmd));
389 if (!pmd_present(pmd: *pmd) || pmd_leaf(pte: *pmd))
390 goto out;
391
392 pte = pte_offset_kernel(pmd, address);
393 if (bad_address(p: pte))
394 goto bad;
395
396 pr_cont("PTE %lx", pte_val(*pte));
397out:
398 pr_cont("\n");
399 return;
400bad:
401 pr_info("BAD\n");
402}
403
404#endif /* CONFIG_X86_64 */
405
406/*
407 * Workaround for K8 erratum #93 & buggy BIOS.
408 *
409 * BIOS SMM functions are required to use a specific workaround
410 * to avoid corruption of the 64bit RIP register on C stepping K8.
411 *
412 * A lot of BIOS that didn't get tested properly miss this.
413 *
414 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
415 * Try to work around it here.
416 *
417 * Note we only handle faults in kernel here.
418 * Does nothing on 32-bit.
419 */
420static int is_errata93(struct pt_regs *regs, unsigned long address)
421{
422#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
423 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
424 || boot_cpu_data.x86 != 0xf)
425 return 0;
426
427 if (user_mode(regs))
428 return 0;
429
430 if (address != regs->ip)
431 return 0;
432
433 if ((address >> 32) != 0)
434 return 0;
435
436 address |= 0xffffffffUL << 32;
437 if ((address >= (u64)_stext && address <= (u64)_etext) ||
438 (address >= MODULES_VADDR && address <= MODULES_END)) {
439 printk_once(errata93_warning);
440 regs->ip = address;
441 return 1;
442 }
443#endif
444 return 0;
445}
446
447/*
448 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
449 * to illegal addresses >4GB.
450 *
451 * We catch this in the page fault handler because these addresses
452 * are not reachable. Just detect this case and return. Any code
453 * segment in LDT is compatibility mode.
454 */
455static int is_errata100(struct pt_regs *regs, unsigned long address)
456{
457#ifdef CONFIG_X86_64
458 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
459 return 1;
460#endif
461 return 0;
462}
463
464/* Pentium F0 0F C7 C8 bug workaround: */
465static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
466 unsigned long address)
467{
468#ifdef CONFIG_X86_F00F_BUG
469 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
470 idt_is_f00f_address(address)) {
471 handle_invalid_op(regs);
472 return 1;
473 }
474#endif
475 return 0;
476}
477
478static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
479{
480 u32 offset = (index >> 3) * sizeof(struct desc_struct);
481 unsigned long addr;
482 struct ldttss_desc desc;
483
484 if (index == 0) {
485 pr_alert("%s: NULL\n", name);
486 return;
487 }
488
489 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
490 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
491 return;
492 }
493
494 if (copy_from_kernel_nofault(dst: &desc, src: (void *)(gdt->address + offset),
495 size: sizeof(struct ldttss_desc))) {
496 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
497 name, index);
498 return;
499 }
500
501 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
502#ifdef CONFIG_X86_64
503 addr |= ((u64)desc.base3 << 32);
504#endif
505 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
506 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
507}
508
509static void
510show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
511{
512 if (!oops_may_print())
513 return;
514
515 if (error_code & X86_PF_INSTR) {
516 unsigned int level;
517 pgd_t *pgd;
518 pte_t *pte;
519
520 pgd = __va(read_cr3_pa());
521 pgd += pgd_index(address);
522
523 pte = lookup_address_in_pgd(pgd, address, level: &level);
524
525 if (pte && pte_present(a: *pte) && !pte_exec(pte: *pte))
526 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
527 from_kuid(&init_user_ns, current_uid()));
528 if (pte && pte_present(a: *pte) && pte_exec(pte: *pte) &&
529 (pgd_flags(pgd: *pgd) & _PAGE_USER) &&
530 (__read_cr4() & X86_CR4_SMEP))
531 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
532 from_kuid(&init_user_ns, current_uid()));
533 }
534
535 if (address < PAGE_SIZE && !user_mode(regs))
536 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
537 (void *)address);
538 else
539 pr_alert("BUG: unable to handle page fault for address: %px\n",
540 (void *)address);
541
542 pr_alert("#PF: %s %s in %s mode\n",
543 (error_code & X86_PF_USER) ? "user" : "supervisor",
544 (error_code & X86_PF_INSTR) ? "instruction fetch" :
545 (error_code & X86_PF_WRITE) ? "write access" :
546 "read access",
547 user_mode(regs) ? "user" : "kernel");
548 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
549 !(error_code & X86_PF_PROT) ? "not-present page" :
550 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
551 (error_code & X86_PF_PK) ? "protection keys violation" :
552 (error_code & X86_PF_RMP) ? "RMP violation" :
553 "permissions violation");
554
555 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
556 struct desc_ptr idt, gdt;
557 u16 ldtr, tr;
558
559 /*
560 * This can happen for quite a few reasons. The more obvious
561 * ones are faults accessing the GDT, or LDT. Perhaps
562 * surprisingly, if the CPU tries to deliver a benign or
563 * contributory exception from user code and gets a page fault
564 * during delivery, the page fault can be delivered as though
565 * it originated directly from user code. This could happen
566 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
567 * kernel or IST stack.
568 */
569 store_idt(dtr: &idt);
570
571 /* Usable even on Xen PV -- it's just slow. */
572 native_store_gdt(dtr: &gdt);
573
574 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
575 idt.address, idt.size, gdt.address, gdt.size);
576
577 store_ldt(ldtr);
578 show_ldttss(gdt: &gdt, name: "LDTR", index: ldtr);
579
580 store_tr(tr);
581 show_ldttss(gdt: &gdt, name: "TR", index: tr);
582 }
583
584 dump_pagetable(address);
585
586 if (error_code & X86_PF_RMP)
587 snp_dump_hva_rmpentry(address);
588}
589
590static noinline void
591pgtable_bad(struct pt_regs *regs, unsigned long error_code,
592 unsigned long address)
593{
594 struct task_struct *tsk;
595 unsigned long flags;
596 int sig;
597
598 flags = oops_begin();
599 tsk = current;
600 sig = SIGKILL;
601
602 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
603 tsk->comm, address);
604 dump_pagetable(address);
605
606 if (__die("Bad pagetable", regs, error_code))
607 sig = 0;
608
609 oops_end(flags, regs, signr: sig);
610}
611
612static void sanitize_error_code(unsigned long address,
613 unsigned long *error_code)
614{
615 /*
616 * To avoid leaking information about the kernel page
617 * table layout, pretend that user-mode accesses to
618 * kernel addresses are always protection faults.
619 *
620 * NB: This means that failed vsyscalls with vsyscall=none
621 * will have the PROT bit. This doesn't leak any
622 * information and does not appear to cause any problems.
623 */
624 if (address >= TASK_SIZE_MAX)
625 *error_code |= X86_PF_PROT;
626}
627
628static void set_signal_archinfo(unsigned long address,
629 unsigned long error_code)
630{
631 struct task_struct *tsk = current;
632
633 tsk->thread.trap_nr = X86_TRAP_PF;
634 tsk->thread.error_code = error_code | X86_PF_USER;
635 tsk->thread.cr2 = address;
636}
637
638static noinline void
639page_fault_oops(struct pt_regs *regs, unsigned long error_code,
640 unsigned long address)
641{
642#ifdef CONFIG_VMAP_STACK
643 struct stack_info info;
644#endif
645 unsigned long flags;
646 int sig;
647
648 if (user_mode(regs)) {
649 /*
650 * Implicit kernel access from user mode? Skip the stack
651 * overflow and EFI special cases.
652 */
653 goto oops;
654 }
655
656#ifdef CONFIG_VMAP_STACK
657 /*
658 * Stack overflow? During boot, we can fault near the initial
659 * stack in the direct map, but that's not an overflow -- check
660 * that we're in vmalloc space to avoid this.
661 */
662 if (is_vmalloc_addr(x: (void *)address) &&
663 get_stack_guard_info(stack: (void *)address, info: &info)) {
664 /*
665 * We're likely to be running with very little stack space
666 * left. It's plausible that we'd hit this condition but
667 * double-fault even before we get this far, in which case
668 * we're fine: the double-fault handler will deal with it.
669 *
670 * We don't want to make it all the way into the oops code
671 * and then double-fault, though, because we're likely to
672 * break the console driver and lose most of the stack dump.
673 */
674 call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
675 handle_stack_overflow,
676 ASM_CALL_ARG3,
677 , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
678
679 unreachable();
680 }
681#endif
682
683 /*
684 * Buggy firmware could access regions which might page fault. If
685 * this happens, EFI has a special OOPS path that will try to
686 * avoid hanging the system.
687 */
688 if (IS_ENABLED(CONFIG_EFI))
689 efi_crash_gracefully_on_page_fault(phys_addr: address);
690
691 /* Only not-present faults should be handled by KFENCE. */
692 if (!(error_code & X86_PF_PROT) &&
693 kfence_handle_page_fault(addr: address, is_write: error_code & X86_PF_WRITE, regs))
694 return;
695
696oops:
697 /*
698 * Oops. The kernel tried to access some bad page. We'll have to
699 * terminate things with extreme prejudice:
700 */
701 flags = oops_begin();
702
703 show_fault_oops(regs, error_code, address);
704
705 if (task_stack_end_corrupted(current))
706 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
707
708 sig = SIGKILL;
709 if (__die("Oops", regs, error_code))
710 sig = 0;
711
712 /* Executive summary in case the body of the oops scrolled away */
713 printk(KERN_DEFAULT "CR2: %016lx\n", address);
714
715 oops_end(flags, regs, signr: sig);
716}
717
718static noinline void
719kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
720 unsigned long address, int signal, int si_code,
721 u32 pkey)
722{
723 WARN_ON_ONCE(user_mode(regs));
724
725 /* Are we prepared to handle this kernel fault? */
726 if (fixup_exception(regs, X86_TRAP_PF, error_code, fault_addr: address)) {
727 /*
728 * Any interrupt that takes a fault gets the fixup. This makes
729 * the below recursive fault logic only apply to a faults from
730 * task context.
731 */
732 if (in_interrupt())
733 return;
734
735 /*
736 * Per the above we're !in_interrupt(), aka. task context.
737 *
738 * In this case we need to make sure we're not recursively
739 * faulting through the emulate_vsyscall() logic.
740 */
741 if (current->thread.sig_on_uaccess_err && signal) {
742 sanitize_error_code(address, error_code: &error_code);
743
744 set_signal_archinfo(address, error_code);
745
746 if (si_code == SEGV_PKUERR) {
747 force_sig_pkuerr(addr: (void __user *)address, pkey);
748 } else {
749 /* XXX: hwpoison faults will set the wrong code. */
750 force_sig_fault(sig: signal, code: si_code, addr: (void __user *)address);
751 }
752 }
753
754 /*
755 * Barring that, we can do the fixup and be happy.
756 */
757 return;
758 }
759
760 /*
761 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
762 * instruction.
763 */
764 if (is_prefetch(regs, error_code, addr: address))
765 return;
766
767 page_fault_oops(regs, error_code, address);
768}
769
770/*
771 * Print out info about fatal segfaults, if the show_unhandled_signals
772 * sysctl is set:
773 */
774static inline void
775show_signal_msg(struct pt_regs *regs, unsigned long error_code,
776 unsigned long address, struct task_struct *tsk)
777{
778 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
779 /* This is a racy snapshot, but it's better than nothing. */
780 int cpu = raw_smp_processor_id();
781
782 if (!unhandled_signal(tsk, SIGSEGV))
783 return;
784
785 if (!printk_ratelimit())
786 return;
787
788 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
789 loglvl, tsk->comm, task_pid_nr(tsk), address,
790 (void *)regs->ip, (void *)regs->sp, error_code);
791
792 print_vma_addr(KERN_CONT " in ", rip: regs->ip);
793
794 /*
795 * Dump the likely CPU where the fatal segfault happened.
796 * This can help identify faulty hardware.
797 */
798 printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu,
799 topology_core_id(cpu), topology_physical_package_id(cpu));
800
801
802 printk(KERN_CONT "\n");
803
804 show_opcodes(regs, loglvl);
805}
806
807static void
808__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
809 unsigned long address, u32 pkey, int si_code)
810{
811 struct task_struct *tsk = current;
812
813 if (!user_mode(regs)) {
814 kernelmode_fixup_or_oops(regs, error_code, address,
815 SIGSEGV, si_code, pkey);
816 return;
817 }
818
819 if (!(error_code & X86_PF_USER)) {
820 /* Implicit user access to kernel memory -- just oops */
821 page_fault_oops(regs, error_code, address);
822 return;
823 }
824
825 /*
826 * User mode accesses just cause a SIGSEGV.
827 * It's possible to have interrupts off here:
828 */
829 local_irq_enable();
830
831 /*
832 * Valid to do another page fault here because this one came
833 * from user space:
834 */
835 if (is_prefetch(regs, error_code, addr: address))
836 return;
837
838 if (is_errata100(regs, address))
839 return;
840
841 sanitize_error_code(address, error_code: &error_code);
842
843 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, fault_addr: address))
844 return;
845
846 if (likely(show_unhandled_signals))
847 show_signal_msg(regs, error_code, address, tsk);
848
849 set_signal_archinfo(address, error_code);
850
851 if (si_code == SEGV_PKUERR)
852 force_sig_pkuerr(addr: (void __user *)address, pkey);
853 else
854 force_sig_fault(SIGSEGV, code: si_code, addr: (void __user *)address);
855
856 local_irq_disable();
857}
858
859static noinline void
860bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
861 unsigned long address)
862{
863 __bad_area_nosemaphore(regs, error_code, address, pkey: 0, SEGV_MAPERR);
864}
865
866static void
867__bad_area(struct pt_regs *regs, unsigned long error_code,
868 unsigned long address, u32 pkey, int si_code)
869{
870 struct mm_struct *mm = current->mm;
871 /*
872 * Something tried to access memory that isn't in our memory map..
873 * Fix it, but check if it's kernel or user first..
874 */
875 mmap_read_unlock(mm);
876
877 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
878}
879
880static inline bool bad_area_access_from_pkeys(unsigned long error_code,
881 struct vm_area_struct *vma)
882{
883 /* This code is always called on the current mm */
884 bool foreign = false;
885
886 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
887 return false;
888 if (error_code & X86_PF_PK)
889 return true;
890 /* this checks permission keys on the VMA: */
891 if (!arch_vma_access_permitted(vma, write: (error_code & X86_PF_WRITE),
892 execute: (error_code & X86_PF_INSTR), foreign))
893 return true;
894 return false;
895}
896
897static noinline void
898bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
899 unsigned long address, struct vm_area_struct *vma)
900{
901 /*
902 * This OSPKE check is not strictly necessary at runtime.
903 * But, doing it this way allows compiler optimizations
904 * if pkeys are compiled out.
905 */
906 if (bad_area_access_from_pkeys(error_code, vma)) {
907 /*
908 * A protection key fault means that the PKRU value did not allow
909 * access to some PTE. Userspace can figure out what PKRU was
910 * from the XSAVE state. This function captures the pkey from
911 * the vma and passes it to userspace so userspace can discover
912 * which protection key was set on the PTE.
913 *
914 * If we get here, we know that the hardware signaled a X86_PF_PK
915 * fault and that there was a VMA once we got in the fault
916 * handler. It does *not* guarantee that the VMA we find here
917 * was the one that we faulted on.
918 *
919 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
920 * 2. T1 : set PKRU to deny access to pkey=4, touches page
921 * 3. T1 : faults...
922 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
923 * 5. T1 : enters fault handler, takes mmap_lock, etc...
924 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
925 * faulted on a pte with its pkey=4.
926 */
927 u32 pkey = vma_pkey(vma);
928
929 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
930 } else {
931 __bad_area(regs, error_code, address, pkey: 0, SEGV_ACCERR);
932 }
933}
934
935static void
936do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
937 vm_fault_t fault)
938{
939 /* Kernel mode? Handle exceptions or die: */
940 if (!user_mode(regs)) {
941 kernelmode_fixup_or_oops(regs, error_code, address,
942 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
943 return;
944 }
945
946 /* User-space => ok to do another page fault: */
947 if (is_prefetch(regs, error_code, addr: address))
948 return;
949
950 sanitize_error_code(address, error_code: &error_code);
951
952 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, fault_addr: address))
953 return;
954
955 set_signal_archinfo(address, error_code);
956
957#ifdef CONFIG_MEMORY_FAILURE
958 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
959 struct task_struct *tsk = current;
960 unsigned lsb = 0;
961
962 pr_err(
963 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
964 tsk->comm, tsk->pid, address);
965 if (fault & VM_FAULT_HWPOISON_LARGE)
966 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
967 if (fault & VM_FAULT_HWPOISON)
968 lsb = PAGE_SHIFT;
969 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
970 return;
971 }
972#endif
973 force_sig_fault(SIGBUS, BUS_ADRERR, addr: (void __user *)address);
974}
975
976static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
977{
978 if ((error_code & X86_PF_WRITE) && !pte_write(pte: *pte))
979 return 0;
980
981 if ((error_code & X86_PF_INSTR) && !pte_exec(pte: *pte))
982 return 0;
983
984 return 1;
985}
986
987/*
988 * Handle a spurious fault caused by a stale TLB entry.
989 *
990 * This allows us to lazily refresh the TLB when increasing the
991 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
992 * eagerly is very expensive since that implies doing a full
993 * cross-processor TLB flush, even if no stale TLB entries exist
994 * on other processors.
995 *
996 * Spurious faults may only occur if the TLB contains an entry with
997 * fewer permission than the page table entry. Non-present (P = 0)
998 * and reserved bit (R = 1) faults are never spurious.
999 *
1000 * There are no security implications to leaving a stale TLB when
1001 * increasing the permissions on a page.
1002 *
1003 * Returns non-zero if a spurious fault was handled, zero otherwise.
1004 *
1005 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1006 * (Optional Invalidation).
1007 */
1008static noinline int
1009spurious_kernel_fault(unsigned long error_code, unsigned long address)
1010{
1011 pgd_t *pgd;
1012 p4d_t *p4d;
1013 pud_t *pud;
1014 pmd_t *pmd;
1015 pte_t *pte;
1016 int ret;
1017
1018 /*
1019 * Only writes to RO or instruction fetches from NX may cause
1020 * spurious faults.
1021 *
1022 * These could be from user or supervisor accesses but the TLB
1023 * is only lazily flushed after a kernel mapping protection
1024 * change, so user accesses are not expected to cause spurious
1025 * faults.
1026 */
1027 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1028 error_code != (X86_PF_INSTR | X86_PF_PROT))
1029 return 0;
1030
1031 pgd = init_mm.pgd + pgd_index(address);
1032 if (!pgd_present(pgd: *pgd))
1033 return 0;
1034
1035 p4d = p4d_offset(pgd, address);
1036 if (!p4d_present(p4d: *p4d))
1037 return 0;
1038
1039 if (p4d_leaf(p4d: *p4d))
1040 return spurious_kernel_fault_check(error_code, pte: (pte_t *) p4d);
1041
1042 pud = pud_offset(p4d, address);
1043 if (!pud_present(pud: *pud))
1044 return 0;
1045
1046 if (pud_leaf(pud: *pud))
1047 return spurious_kernel_fault_check(error_code, pte: (pte_t *) pud);
1048
1049 pmd = pmd_offset(pud, address);
1050 if (!pmd_present(pmd: *pmd))
1051 return 0;
1052
1053 if (pmd_leaf(pte: *pmd))
1054 return spurious_kernel_fault_check(error_code, pte: (pte_t *) pmd);
1055
1056 pte = pte_offset_kernel(pmd, address);
1057 if (!pte_present(a: *pte))
1058 return 0;
1059
1060 ret = spurious_kernel_fault_check(error_code, pte);
1061 if (!ret)
1062 return 0;
1063
1064 /*
1065 * Make sure we have permissions in PMD.
1066 * If not, then there's a bug in the page tables:
1067 */
1068 ret = spurious_kernel_fault_check(error_code, pte: (pte_t *) pmd);
1069 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1070
1071 return ret;
1072}
1073NOKPROBE_SYMBOL(spurious_kernel_fault);
1074
1075int show_unhandled_signals = 1;
1076
1077static inline int
1078access_error(unsigned long error_code, struct vm_area_struct *vma)
1079{
1080 /* This is only called for the current mm, so: */
1081 bool foreign = false;
1082
1083 /*
1084 * Read or write was blocked by protection keys. This is
1085 * always an unconditional error and can never result in
1086 * a follow-up action to resolve the fault, like a COW.
1087 */
1088 if (error_code & X86_PF_PK)
1089 return 1;
1090
1091 /*
1092 * SGX hardware blocked the access. This usually happens
1093 * when the enclave memory contents have been destroyed, like
1094 * after a suspend/resume cycle. In any case, the kernel can't
1095 * fix the cause of the fault. Handle the fault as an access
1096 * error even in cases where no actual access violation
1097 * occurred. This allows userspace to rebuild the enclave in
1098 * response to the signal.
1099 */
1100 if (unlikely(error_code & X86_PF_SGX))
1101 return 1;
1102
1103 /*
1104 * Make sure to check the VMA so that we do not perform
1105 * faults just to hit a X86_PF_PK as soon as we fill in a
1106 * page.
1107 */
1108 if (!arch_vma_access_permitted(vma, write: (error_code & X86_PF_WRITE),
1109 execute: (error_code & X86_PF_INSTR), foreign))
1110 return 1;
1111
1112 /*
1113 * Shadow stack accesses (PF_SHSTK=1) are only permitted to
1114 * shadow stack VMAs. All other accesses result in an error.
1115 */
1116 if (error_code & X86_PF_SHSTK) {
1117 if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK)))
1118 return 1;
1119 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1120 return 1;
1121 return 0;
1122 }
1123
1124 if (error_code & X86_PF_WRITE) {
1125 /* write, present and write, not present: */
1126 if (unlikely(vma->vm_flags & VM_SHADOW_STACK))
1127 return 1;
1128 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1129 return 1;
1130 return 0;
1131 }
1132
1133 /* read, present: */
1134 if (unlikely(error_code & X86_PF_PROT))
1135 return 1;
1136
1137 /* read, not present: */
1138 if (unlikely(!vma_is_accessible(vma)))
1139 return 1;
1140
1141 return 0;
1142}
1143
1144bool fault_in_kernel_space(unsigned long address)
1145{
1146 /*
1147 * On 64-bit systems, the vsyscall page is at an address above
1148 * TASK_SIZE_MAX, but is not considered part of the kernel
1149 * address space.
1150 */
1151 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(vaddr: address))
1152 return false;
1153
1154 return address >= TASK_SIZE_MAX;
1155}
1156
1157/*
1158 * Called for all faults where 'address' is part of the kernel address
1159 * space. Might get called for faults that originate from *code* that
1160 * ran in userspace or the kernel.
1161 */
1162static void
1163do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1164 unsigned long address)
1165{
1166 /*
1167 * Protection keys exceptions only happen on user pages. We
1168 * have no user pages in the kernel portion of the address
1169 * space, so do not expect them here.
1170 */
1171 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1172
1173#ifdef CONFIG_X86_32
1174 /*
1175 * We can fault-in kernel-space virtual memory on-demand. The
1176 * 'reference' page table is init_mm.pgd.
1177 *
1178 * NOTE! We MUST NOT take any locks for this case. We may
1179 * be in an interrupt or a critical region, and should
1180 * only copy the information from the master page table,
1181 * nothing more.
1182 *
1183 * Before doing this on-demand faulting, ensure that the
1184 * fault is not any of the following:
1185 * 1. A fault on a PTE with a reserved bit set.
1186 * 2. A fault caused by a user-mode access. (Do not demand-
1187 * fault kernel memory due to user-mode accesses).
1188 * 3. A fault caused by a page-level protection violation.
1189 * (A demand fault would be on a non-present page which
1190 * would have X86_PF_PROT==0).
1191 *
1192 * This is only needed to close a race condition on x86-32 in
1193 * the vmalloc mapping/unmapping code. See the comment above
1194 * vmalloc_fault() for details. On x86-64 the race does not
1195 * exist as the vmalloc mappings don't need to be synchronized
1196 * there.
1197 */
1198 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1199 if (vmalloc_fault(address) >= 0)
1200 return;
1201 }
1202#endif
1203
1204 if (is_f00f_bug(regs, error_code: hw_error_code, address))
1205 return;
1206
1207 /* Was the fault spurious, caused by lazy TLB invalidation? */
1208 if (spurious_kernel_fault(error_code: hw_error_code, address))
1209 return;
1210
1211 /* kprobes don't want to hook the spurious faults: */
1212 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1213 return;
1214
1215 /*
1216 * Note, despite being a "bad area", there are quite a few
1217 * acceptable reasons to get here, such as erratum fixups
1218 * and handling kernel code that can fault, like get_user().
1219 *
1220 * Don't take the mm semaphore here. If we fixup a prefetch
1221 * fault we could otherwise deadlock:
1222 */
1223 bad_area_nosemaphore(regs, error_code: hw_error_code, address);
1224}
1225NOKPROBE_SYMBOL(do_kern_addr_fault);
1226
1227/*
1228 * Handle faults in the user portion of the address space. Nothing in here
1229 * should check X86_PF_USER without a specific justification: for almost
1230 * all purposes, we should treat a normal kernel access to user memory
1231 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1232 * The one exception is AC flag handling, which is, per the x86
1233 * architecture, special for WRUSS.
1234 */
1235static inline
1236void do_user_addr_fault(struct pt_regs *regs,
1237 unsigned long error_code,
1238 unsigned long address)
1239{
1240 struct vm_area_struct *vma;
1241 struct task_struct *tsk;
1242 struct mm_struct *mm;
1243 vm_fault_t fault;
1244 unsigned int flags = FAULT_FLAG_DEFAULT;
1245
1246 tsk = current;
1247 mm = tsk->mm;
1248
1249 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1250 /*
1251 * Whoops, this is kernel mode code trying to execute from
1252 * user memory. Unless this is AMD erratum #93, which
1253 * corrupts RIP such that it looks like a user address,
1254 * this is unrecoverable. Don't even try to look up the
1255 * VMA or look for extable entries.
1256 */
1257 if (is_errata93(regs, address))
1258 return;
1259
1260 page_fault_oops(regs, error_code, address);
1261 return;
1262 }
1263
1264 /* kprobes don't want to hook the spurious faults: */
1265 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1266 return;
1267
1268 /*
1269 * Reserved bits are never expected to be set on
1270 * entries in the user portion of the page tables.
1271 */
1272 if (unlikely(error_code & X86_PF_RSVD))
1273 pgtable_bad(regs, error_code, address);
1274
1275 /*
1276 * If SMAP is on, check for invalid kernel (supervisor) access to user
1277 * pages in the user address space. The odd case here is WRUSS,
1278 * which, according to the preliminary documentation, does not respect
1279 * SMAP and will have the USER bit set so, in all cases, SMAP
1280 * enforcement appears to be consistent with the USER bit.
1281 */
1282 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1283 !(error_code & X86_PF_USER) &&
1284 !(regs->flags & X86_EFLAGS_AC))) {
1285 /*
1286 * No extable entry here. This was a kernel access to an
1287 * invalid pointer. get_kernel_nofault() will not get here.
1288 */
1289 page_fault_oops(regs, error_code, address);
1290 return;
1291 }
1292
1293 /*
1294 * If we're in an interrupt, have no user context or are running
1295 * in a region with pagefaults disabled then we must not take the fault
1296 */
1297 if (unlikely(faulthandler_disabled() || !mm)) {
1298 bad_area_nosemaphore(regs, error_code, address);
1299 return;
1300 }
1301
1302 /* Legacy check - remove this after verifying that it doesn't trigger */
1303 if (WARN_ON_ONCE(!(regs->flags & X86_EFLAGS_IF))) {
1304 bad_area_nosemaphore(regs, error_code, address);
1305 return;
1306 }
1307
1308 local_irq_enable();
1309
1310 perf_sw_event(event_id: PERF_COUNT_SW_PAGE_FAULTS, nr: 1, regs, addr: address);
1311
1312 /*
1313 * Read-only permissions can not be expressed in shadow stack PTEs.
1314 * Treat all shadow stack accesses as WRITE faults. This ensures
1315 * that the MM will prepare everything (e.g., break COW) such that
1316 * maybe_mkwrite() can create a proper shadow stack PTE.
1317 */
1318 if (error_code & X86_PF_SHSTK)
1319 flags |= FAULT_FLAG_WRITE;
1320 if (error_code & X86_PF_WRITE)
1321 flags |= FAULT_FLAG_WRITE;
1322 if (error_code & X86_PF_INSTR)
1323 flags |= FAULT_FLAG_INSTRUCTION;
1324
1325 /*
1326 * We set FAULT_FLAG_USER based on the register state, not
1327 * based on X86_PF_USER. User space accesses that cause
1328 * system page faults are still user accesses.
1329 */
1330 if (user_mode(regs))
1331 flags |= FAULT_FLAG_USER;
1332
1333#ifdef CONFIG_X86_64
1334 /*
1335 * Faults in the vsyscall page might need emulation. The
1336 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1337 * considered to be part of the user address space.
1338 *
1339 * The vsyscall page does not have a "real" VMA, so do this
1340 * emulation before we go searching for VMAs.
1341 *
1342 * PKRU never rejects instruction fetches, so we don't need
1343 * to consider the PF_PK bit.
1344 */
1345 if (is_vsyscall_vaddr(vaddr: address)) {
1346 if (emulate_vsyscall(error_code, regs, address))
1347 return;
1348 }
1349#endif
1350
1351 if (!(flags & FAULT_FLAG_USER))
1352 goto lock_mmap;
1353
1354 vma = lock_vma_under_rcu(mm, address);
1355 if (!vma)
1356 goto lock_mmap;
1357
1358 if (unlikely(access_error(error_code, vma))) {
1359 vma_end_read(vma);
1360 goto lock_mmap;
1361 }
1362 fault = handle_mm_fault(vma, address, flags: flags | FAULT_FLAG_VMA_LOCK, regs);
1363 if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
1364 vma_end_read(vma);
1365
1366 if (!(fault & VM_FAULT_RETRY)) {
1367 count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1368 goto done;
1369 }
1370 count_vm_vma_lock_event(VMA_LOCK_RETRY);
1371 if (fault & VM_FAULT_MAJOR)
1372 flags |= FAULT_FLAG_TRIED;
1373
1374 /* Quick path to respond to signals */
1375 if (fault_signal_pending(fault_flags: fault, regs)) {
1376 if (!user_mode(regs))
1377 kernelmode_fixup_or_oops(regs, error_code, address,
1378 SIGBUS, BUS_ADRERR,
1379 ARCH_DEFAULT_PKEY);
1380 return;
1381 }
1382lock_mmap:
1383
1384retry:
1385 vma = lock_mm_and_find_vma(mm, address, regs);
1386 if (unlikely(!vma)) {
1387 bad_area_nosemaphore(regs, error_code, address);
1388 return;
1389 }
1390
1391 /*
1392 * Ok, we have a good vm_area for this memory access, so
1393 * we can handle it..
1394 */
1395 if (unlikely(access_error(error_code, vma))) {
1396 bad_area_access_error(regs, error_code, address, vma);
1397 return;
1398 }
1399
1400 /*
1401 * If for any reason at all we couldn't handle the fault,
1402 * make sure we exit gracefully rather than endlessly redo
1403 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1404 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1405 *
1406 * Note that handle_userfault() may also release and reacquire mmap_lock
1407 * (and not return with VM_FAULT_RETRY), when returning to userland to
1408 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1409 * (potentially after handling any pending signal during the return to
1410 * userland). The return to userland is identified whenever
1411 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1412 */
1413 fault = handle_mm_fault(vma, address, flags, regs);
1414
1415 if (fault_signal_pending(fault_flags: fault, regs)) {
1416 /*
1417 * Quick path to respond to signals. The core mm code
1418 * has unlocked the mm for us if we get here.
1419 */
1420 if (!user_mode(regs))
1421 kernelmode_fixup_or_oops(regs, error_code, address,
1422 SIGBUS, BUS_ADRERR,
1423 ARCH_DEFAULT_PKEY);
1424 return;
1425 }
1426
1427 /* The fault is fully completed (including releasing mmap lock) */
1428 if (fault & VM_FAULT_COMPLETED)
1429 return;
1430
1431 /*
1432 * If we need to retry the mmap_lock has already been released,
1433 * and if there is a fatal signal pending there is no guarantee
1434 * that we made any progress. Handle this case first.
1435 */
1436 if (unlikely(fault & VM_FAULT_RETRY)) {
1437 flags |= FAULT_FLAG_TRIED;
1438 goto retry;
1439 }
1440
1441 mmap_read_unlock(mm);
1442done:
1443 if (likely(!(fault & VM_FAULT_ERROR)))
1444 return;
1445
1446 if (fatal_signal_pending(current) && !user_mode(regs)) {
1447 kernelmode_fixup_or_oops(regs, error_code, address,
1448 signal: 0, si_code: 0, ARCH_DEFAULT_PKEY);
1449 return;
1450 }
1451
1452 if (fault & VM_FAULT_OOM) {
1453 /* Kernel mode? Handle exceptions or die: */
1454 if (!user_mode(regs)) {
1455 kernelmode_fixup_or_oops(regs, error_code, address,
1456 SIGSEGV, SEGV_MAPERR,
1457 ARCH_DEFAULT_PKEY);
1458 return;
1459 }
1460
1461 /*
1462 * We ran out of memory, call the OOM killer, and return the
1463 * userspace (which will retry the fault, or kill us if we got
1464 * oom-killed):
1465 */
1466 pagefault_out_of_memory();
1467 } else {
1468 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1469 VM_FAULT_HWPOISON_LARGE))
1470 do_sigbus(regs, error_code, address, fault);
1471 else if (fault & VM_FAULT_SIGSEGV)
1472 bad_area_nosemaphore(regs, error_code, address);
1473 else
1474 BUG();
1475 }
1476}
1477NOKPROBE_SYMBOL(do_user_addr_fault);
1478
1479static __always_inline void
1480trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1481 unsigned long address)
1482{
1483 if (!trace_pagefault_enabled())
1484 return;
1485
1486 if (user_mode(regs))
1487 trace_page_fault_user(address, regs, error_code);
1488 else
1489 trace_page_fault_kernel(address, regs, error_code);
1490}
1491
1492static __always_inline void
1493handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1494 unsigned long address)
1495{
1496 trace_page_fault_entries(regs, error_code, address);
1497
1498 if (unlikely(kmmio_fault(regs, address)))
1499 return;
1500
1501 /* Was the fault on kernel-controlled part of the address space? */
1502 if (unlikely(fault_in_kernel_space(address))) {
1503 do_kern_addr_fault(regs, hw_error_code: error_code, address);
1504 } else {
1505 do_user_addr_fault(regs, error_code, address);
1506 /*
1507 * User address page fault handling might have reenabled
1508 * interrupts. Fixing up all potential exit points of
1509 * do_user_addr_fault() and its leaf functions is just not
1510 * doable w/o creating an unholy mess or turning the code
1511 * upside down.
1512 */
1513 local_irq_disable();
1514 }
1515}
1516
1517DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1518{
1519 irqentry_state_t state;
1520 unsigned long address;
1521
1522 address = cpu_feature_enabled(X86_FEATURE_FRED) ? fred_event_data(regs) : read_cr2();
1523
1524 prefetchw(x: &current->mm->mmap_lock);
1525
1526 /*
1527 * KVM uses #PF vector to deliver 'page not present' events to guests
1528 * (asynchronous page fault mechanism). The event happens when a
1529 * userspace task is trying to access some valid (from guest's point of
1530 * view) memory which is not currently mapped by the host (e.g. the
1531 * memory is swapped out). Note, the corresponding "page ready" event
1532 * which is injected when the memory becomes available, is delivered via
1533 * an interrupt mechanism and not a #PF exception
1534 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1535 *
1536 * We are relying on the interrupted context being sane (valid RSP,
1537 * relevant locks not held, etc.), which is fine as long as the
1538 * interrupted context had IF=1. We are also relying on the KVM
1539 * async pf type field and CR2 being read consistently instead of
1540 * getting values from real and async page faults mixed up.
1541 *
1542 * Fingers crossed.
1543 *
1544 * The async #PF handling code takes care of idtentry handling
1545 * itself.
1546 */
1547 if (kvm_handle_async_pf(regs, token: (u32)address))
1548 return;
1549
1550 /*
1551 * Entry handling for valid #PF from kernel mode is slightly
1552 * different: RCU is already watching and ct_irq_enter() must not
1553 * be invoked because a kernel fault on a user space address might
1554 * sleep.
1555 *
1556 * In case the fault hit a RCU idle region the conditional entry
1557 * code reenabled RCU to avoid subsequent wreckage which helps
1558 * debuggability.
1559 */
1560 state = irqentry_enter(regs);
1561
1562 instrumentation_begin();
1563 handle_page_fault(regs, error_code, address);
1564 instrumentation_end();
1565
1566 irqentry_exit(regs, state);
1567}
1568

source code of linux/arch/x86/mm/fault.c