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

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