Loading...
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/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/prefetch.h> /* prefetchw */
17#include <linux/context_tracking.h> /* exception_enter(), ... */
18#include <linux/uaccess.h> /* faulthandler_disabled() */
19#include <linux/efi.h> /* efi_recover_from_page_fault()*/
20#include <linux/mm_types.h>
21
22#include <asm/cpufeature.h> /* boot_cpu_has, ... */
23#include <asm/traps.h> /* dotraplinkage, ... */
24#include <asm/pgalloc.h> /* pgd_*(), ... */
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_recover_from_page_fault()*/
30#include <asm/desc.h> /* store_idt(), ... */
31#include <asm/cpu_entry_area.h> /* exception stack */
32
33#define CREATE_TRACE_POINTS
34#include <asm/trace/exceptions.h>
35
36/*
37 * Returns 0 if mmiotrace is disabled, or if the fault is not
38 * handled by mmiotrace:
39 */
40static nokprobe_inline int
41kmmio_fault(struct pt_regs *regs, unsigned long addr)
42{
43 if (unlikely(is_kmmio_active()))
44 if (kmmio_handler(regs, addr) == 1)
45 return -1;
46 return 0;
47}
48
49/*
50 * Prefetch quirks:
51 *
52 * 32-bit mode:
53 *
54 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
55 * Check that here and ignore it.
56 *
57 * 64-bit mode:
58 *
59 * Sometimes the CPU reports invalid exceptions on prefetch.
60 * Check that here and ignore it.
61 *
62 * Opcode checker based on code by Richard Brunner.
63 */
64static inline int
65check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
66 unsigned char opcode, int *prefetch)
67{
68 unsigned char instr_hi = opcode & 0xf0;
69 unsigned char instr_lo = opcode & 0x0f;
70
71 switch (instr_hi) {
72 case 0x20:
73 case 0x30:
74 /*
75 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
76 * In X86_64 long mode, the CPU will signal invalid
77 * opcode if some of these prefixes are present so
78 * X86_64 will never get here anyway
79 */
80 return ((instr_lo & 7) == 0x6);
81#ifdef CONFIG_X86_64
82 case 0x40:
83 /*
84 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
85 * Need to figure out under what instruction mode the
86 * instruction was issued. Could check the LDT for lm,
87 * but for now it's good enough to assume that long
88 * mode only uses well known segments or kernel.
89 */
90 return (!user_mode(regs) || user_64bit_mode(regs));
91#endif
92 case 0x60:
93 /* 0x64 thru 0x67 are valid prefixes in all modes. */
94 return (instr_lo & 0xC) == 0x4;
95 case 0xF0:
96 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
97 return !instr_lo || (instr_lo>>1) == 1;
98 case 0x00:
99 /* Prefetch instruction is 0x0F0D or 0x0F18 */
100 if (probe_kernel_address(instr, opcode))
101 return 0;
102
103 *prefetch = (instr_lo == 0xF) &&
104 (opcode == 0x0D || opcode == 0x18);
105 return 0;
106 default:
107 return 0;
108 }
109}
110
111static int
112is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
113{
114 unsigned char *max_instr;
115 unsigned char *instr;
116 int prefetch = 0;
117
118 /*
119 * If it was a exec (instruction fetch) fault on NX page, then
120 * do not ignore the fault:
121 */
122 if (error_code & X86_PF_INSTR)
123 return 0;
124
125 instr = (void *)convert_ip_to_linear(current, regs);
126 max_instr = instr + 15;
127
128 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
129 return 0;
130
131 while (instr < max_instr) {
132 unsigned char opcode;
133
134 if (probe_kernel_address(instr, opcode))
135 break;
136
137 instr++;
138
139 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
140 break;
141 }
142 return prefetch;
143}
144
145DEFINE_SPINLOCK(pgd_lock);
146LIST_HEAD(pgd_list);
147
148#ifdef CONFIG_X86_32
149static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
150{
151 unsigned index = pgd_index(address);
152 pgd_t *pgd_k;
153 p4d_t *p4d, *p4d_k;
154 pud_t *pud, *pud_k;
155 pmd_t *pmd, *pmd_k;
156
157 pgd += index;
158 pgd_k = init_mm.pgd + index;
159
160 if (!pgd_present(*pgd_k))
161 return NULL;
162
163 /*
164 * set_pgd(pgd, *pgd_k); here would be useless on PAE
165 * and redundant with the set_pmd() on non-PAE. As would
166 * set_p4d/set_pud.
167 */
168 p4d = p4d_offset(pgd, address);
169 p4d_k = p4d_offset(pgd_k, address);
170 if (!p4d_present(*p4d_k))
171 return NULL;
172
173 pud = pud_offset(p4d, address);
174 pud_k = pud_offset(p4d_k, address);
175 if (!pud_present(*pud_k))
176 return NULL;
177
178 pmd = pmd_offset(pud, address);
179 pmd_k = pmd_offset(pud_k, address);
180
181 if (pmd_present(*pmd) != pmd_present(*pmd_k))
182 set_pmd(pmd, *pmd_k);
183
184 if (!pmd_present(*pmd_k))
185 return NULL;
186 else
187 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
188
189 return pmd_k;
190}
191
192void vmalloc_sync_all(void)
193{
194 unsigned long address;
195
196 if (SHARED_KERNEL_PMD)
197 return;
198
199 for (address = VMALLOC_START & PMD_MASK;
200 address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
201 address += PMD_SIZE) {
202 struct page *page;
203
204 spin_lock(&pgd_lock);
205 list_for_each_entry(page, &pgd_list, lru) {
206 spinlock_t *pgt_lock;
207
208 /* the pgt_lock only for Xen */
209 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
210
211 spin_lock(pgt_lock);
212 vmalloc_sync_one(page_address(page), address);
213 spin_unlock(pgt_lock);
214 }
215 spin_unlock(&pgd_lock);
216 }
217}
218
219/*
220 * 32-bit:
221 *
222 * Handle a fault on the vmalloc or module mapping area
223 */
224static noinline int vmalloc_fault(unsigned long address)
225{
226 unsigned long pgd_paddr;
227 pmd_t *pmd_k;
228 pte_t *pte_k;
229
230 /* Make sure we are in vmalloc area: */
231 if (!(address >= VMALLOC_START && address < VMALLOC_END))
232 return -1;
233
234 /*
235 * Synchronize this task's top level page-table
236 * with the 'reference' page table.
237 *
238 * Do _not_ use "current" here. We might be inside
239 * an interrupt in the middle of a task switch..
240 */
241 pgd_paddr = read_cr3_pa();
242 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
243 if (!pmd_k)
244 return -1;
245
246 if (pmd_large(*pmd_k))
247 return 0;
248
249 pte_k = pte_offset_kernel(pmd_k, address);
250 if (!pte_present(*pte_k))
251 return -1;
252
253 return 0;
254}
255NOKPROBE_SYMBOL(vmalloc_fault);
256
257/*
258 * Did it hit the DOS screen memory VA from vm86 mode?
259 */
260static inline void
261check_v8086_mode(struct pt_regs *regs, unsigned long address,
262 struct task_struct *tsk)
263{
264#ifdef CONFIG_VM86
265 unsigned long bit;
266
267 if (!v8086_mode(regs) || !tsk->thread.vm86)
268 return;
269
270 bit = (address - 0xA0000) >> PAGE_SHIFT;
271 if (bit < 32)
272 tsk->thread.vm86->screen_bitmap |= 1 << bit;
273#endif
274}
275
276static bool low_pfn(unsigned long pfn)
277{
278 return pfn < max_low_pfn;
279}
280
281static void dump_pagetable(unsigned long address)
282{
283 pgd_t *base = __va(read_cr3_pa());
284 pgd_t *pgd = &base[pgd_index(address)];
285 p4d_t *p4d;
286 pud_t *pud;
287 pmd_t *pmd;
288 pte_t *pte;
289
290#ifdef CONFIG_X86_PAE
291 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
292 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
293 goto out;
294#define pr_pde pr_cont
295#else
296#define pr_pde pr_info
297#endif
298 p4d = p4d_offset(pgd, address);
299 pud = pud_offset(p4d, address);
300 pmd = pmd_offset(pud, address);
301 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
302#undef pr_pde
303
304 /*
305 * We must not directly access the pte in the highpte
306 * case if the page table is located in highmem.
307 * And let's rather not kmap-atomic the pte, just in case
308 * it's allocated already:
309 */
310 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
311 goto out;
312
313 pte = pte_offset_kernel(pmd, address);
314 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
315out:
316 pr_cont("\n");
317}
318
319#else /* CONFIG_X86_64: */
320
321void vmalloc_sync_all(void)
322{
323 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
324}
325
326/*
327 * 64-bit:
328 *
329 * Handle a fault on the vmalloc area
330 */
331static noinline int vmalloc_fault(unsigned long address)
332{
333 pgd_t *pgd, *pgd_k;
334 p4d_t *p4d, *p4d_k;
335 pud_t *pud;
336 pmd_t *pmd;
337 pte_t *pte;
338
339 /* Make sure we are in vmalloc area: */
340 if (!(address >= VMALLOC_START && address < VMALLOC_END))
341 return -1;
342
343 /*
344 * Copy kernel mappings over when needed. This can also
345 * happen within a race in page table update. In the later
346 * case just flush:
347 */
348 pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address);
349 pgd_k = pgd_offset_k(address);
350 if (pgd_none(*pgd_k))
351 return -1;
352
353 if (pgtable_l5_enabled()) {
354 if (pgd_none(*pgd)) {
355 set_pgd(pgd, *pgd_k);
356 arch_flush_lazy_mmu_mode();
357 } else {
358 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_k));
359 }
360 }
361
362 /* With 4-level paging, copying happens on the p4d level. */
363 p4d = p4d_offset(pgd, address);
364 p4d_k = p4d_offset(pgd_k, address);
365 if (p4d_none(*p4d_k))
366 return -1;
367
368 if (p4d_none(*p4d) && !pgtable_l5_enabled()) {
369 set_p4d(p4d, *p4d_k);
370 arch_flush_lazy_mmu_mode();
371 } else {
372 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_k));
373 }
374
375 BUILD_BUG_ON(CONFIG_PGTABLE_LEVELS < 4);
376
377 pud = pud_offset(p4d, address);
378 if (pud_none(*pud))
379 return -1;
380
381 if (pud_large(*pud))
382 return 0;
383
384 pmd = pmd_offset(pud, address);
385 if (pmd_none(*pmd))
386 return -1;
387
388 if (pmd_large(*pmd))
389 return 0;
390
391 pte = pte_offset_kernel(pmd, address);
392 if (!pte_present(*pte))
393 return -1;
394
395 return 0;
396}
397NOKPROBE_SYMBOL(vmalloc_fault);
398
399#ifdef CONFIG_CPU_SUP_AMD
400static const char errata93_warning[] =
401KERN_ERR
402"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
403"******* Working around it, but it may cause SEGVs or burn power.\n"
404"******* Please consider a BIOS update.\n"
405"******* Disabling USB legacy in the BIOS may also help.\n";
406#endif
407
408/*
409 * No vm86 mode in 64-bit mode:
410 */
411static inline void
412check_v8086_mode(struct pt_regs *regs, unsigned long address,
413 struct task_struct *tsk)
414{
415}
416
417static int bad_address(void *p)
418{
419 unsigned long dummy;
420
421 return probe_kernel_address((unsigned long *)p, dummy);
422}
423
424static void dump_pagetable(unsigned long address)
425{
426 pgd_t *base = __va(read_cr3_pa());
427 pgd_t *pgd = base + pgd_index(address);
428 p4d_t *p4d;
429 pud_t *pud;
430 pmd_t *pmd;
431 pte_t *pte;
432
433 if (bad_address(pgd))
434 goto bad;
435
436 pr_info("PGD %lx ", pgd_val(*pgd));
437
438 if (!pgd_present(*pgd))
439 goto out;
440
441 p4d = p4d_offset(pgd, address);
442 if (bad_address(p4d))
443 goto bad;
444
445 pr_cont("P4D %lx ", p4d_val(*p4d));
446 if (!p4d_present(*p4d) || p4d_large(*p4d))
447 goto out;
448
449 pud = pud_offset(p4d, address);
450 if (bad_address(pud))
451 goto bad;
452
453 pr_cont("PUD %lx ", pud_val(*pud));
454 if (!pud_present(*pud) || pud_large(*pud))
455 goto out;
456
457 pmd = pmd_offset(pud, address);
458 if (bad_address(pmd))
459 goto bad;
460
461 pr_cont("PMD %lx ", pmd_val(*pmd));
462 if (!pmd_present(*pmd) || pmd_large(*pmd))
463 goto out;
464
465 pte = pte_offset_kernel(pmd, address);
466 if (bad_address(pte))
467 goto bad;
468
469 pr_cont("PTE %lx", pte_val(*pte));
470out:
471 pr_cont("\n");
472 return;
473bad:
474 pr_info("BAD\n");
475}
476
477#endif /* CONFIG_X86_64 */
478
479/*
480 * Workaround for K8 erratum #93 & buggy BIOS.
481 *
482 * BIOS SMM functions are required to use a specific workaround
483 * to avoid corruption of the 64bit RIP register on C stepping K8.
484 *
485 * A lot of BIOS that didn't get tested properly miss this.
486 *
487 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
488 * Try to work around it here.
489 *
490 * Note we only handle faults in kernel here.
491 * Does nothing on 32-bit.
492 */
493static int is_errata93(struct pt_regs *regs, unsigned long address)
494{
495#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
496 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
497 || boot_cpu_data.x86 != 0xf)
498 return 0;
499
500 if (address != regs->ip)
501 return 0;
502
503 if ((address >> 32) != 0)
504 return 0;
505
506 address |= 0xffffffffUL << 32;
507 if ((address >= (u64)_stext && address <= (u64)_etext) ||
508 (address >= MODULES_VADDR && address <= MODULES_END)) {
509 printk_once(errata93_warning);
510 regs->ip = address;
511 return 1;
512 }
513#endif
514 return 0;
515}
516
517/*
518 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
519 * to illegal addresses >4GB.
520 *
521 * We catch this in the page fault handler because these addresses
522 * are not reachable. Just detect this case and return. Any code
523 * segment in LDT is compatibility mode.
524 */
525static int is_errata100(struct pt_regs *regs, unsigned long address)
526{
527#ifdef CONFIG_X86_64
528 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
529 return 1;
530#endif
531 return 0;
532}
533
534static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
535{
536#ifdef CONFIG_X86_F00F_BUG
537 unsigned long nr;
538
539 /*
540 * Pentium F0 0F C7 C8 bug workaround:
541 */
542 if (boot_cpu_has_bug(X86_BUG_F00F)) {
543 nr = (address - idt_descr.address) >> 3;
544
545 if (nr == 6) {
546 do_invalid_op(regs, 0);
547 return 1;
548 }
549 }
550#endif
551 return 0;
552}
553
554static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
555{
556 u32 offset = (index >> 3) * sizeof(struct desc_struct);
557 unsigned long addr;
558 struct ldttss_desc desc;
559
560 if (index == 0) {
561 pr_alert("%s: NULL\n", name);
562 return;
563 }
564
565 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
566 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
567 return;
568 }
569
570 if (probe_kernel_read(&desc, (void *)(gdt->address + offset),
571 sizeof(struct ldttss_desc))) {
572 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
573 name, index);
574 return;
575 }
576
577 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
578#ifdef CONFIG_X86_64
579 addr |= ((u64)desc.base3 << 32);
580#endif
581 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
582 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
583}
584
585static void
586show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
587{
588 if (!oops_may_print())
589 return;
590
591 if (error_code & X86_PF_INSTR) {
592 unsigned int level;
593 pgd_t *pgd;
594 pte_t *pte;
595
596 pgd = __va(read_cr3_pa());
597 pgd += pgd_index(address);
598
599 pte = lookup_address_in_pgd(pgd, address, &level);
600
601 if (pte && pte_present(*pte) && !pte_exec(*pte))
602 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
603 from_kuid(&init_user_ns, current_uid()));
604 if (pte && pte_present(*pte) && pte_exec(*pte) &&
605 (pgd_flags(*pgd) & _PAGE_USER) &&
606 (__read_cr4() & X86_CR4_SMEP))
607 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
608 from_kuid(&init_user_ns, current_uid()));
609 }
610
611 if (address < PAGE_SIZE && !user_mode(regs))
612 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
613 (void *)address);
614 else
615 pr_alert("BUG: unable to handle page fault for address: %px\n",
616 (void *)address);
617
618 pr_alert("#PF: %s %s in %s mode\n",
619 (error_code & X86_PF_USER) ? "user" : "supervisor",
620 (error_code & X86_PF_INSTR) ? "instruction fetch" :
621 (error_code & X86_PF_WRITE) ? "write access" :
622 "read access",
623 user_mode(regs) ? "user" : "kernel");
624 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
625 !(error_code & X86_PF_PROT) ? "not-present page" :
626 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
627 (error_code & X86_PF_PK) ? "protection keys violation" :
628 "permissions violation");
629
630 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
631 struct desc_ptr idt, gdt;
632 u16 ldtr, tr;
633
634 /*
635 * This can happen for quite a few reasons. The more obvious
636 * ones are faults accessing the GDT, or LDT. Perhaps
637 * surprisingly, if the CPU tries to deliver a benign or
638 * contributory exception from user code and gets a page fault
639 * during delivery, the page fault can be delivered as though
640 * it originated directly from user code. This could happen
641 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
642 * kernel or IST stack.
643 */
644 store_idt(&idt);
645
646 /* Usable even on Xen PV -- it's just slow. */
647 native_store_gdt(&gdt);
648
649 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
650 idt.address, idt.size, gdt.address, gdt.size);
651
652 store_ldt(ldtr);
653 show_ldttss(&gdt, "LDTR", ldtr);
654
655 store_tr(tr);
656 show_ldttss(&gdt, "TR", tr);
657 }
658
659 dump_pagetable(address);
660}
661
662static noinline void
663pgtable_bad(struct pt_regs *regs, unsigned long error_code,
664 unsigned long address)
665{
666 struct task_struct *tsk;
667 unsigned long flags;
668 int sig;
669
670 flags = oops_begin();
671 tsk = current;
672 sig = SIGKILL;
673
674 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
675 tsk->comm, address);
676 dump_pagetable(address);
677
678 if (__die("Bad pagetable", regs, error_code))
679 sig = 0;
680
681 oops_end(flags, regs, sig);
682}
683
684static void set_signal_archinfo(unsigned long address,
685 unsigned long error_code)
686{
687 struct task_struct *tsk = current;
688
689 /*
690 * To avoid leaking information about the kernel page
691 * table layout, pretend that user-mode accesses to
692 * kernel addresses are always protection faults.
693 *
694 * NB: This means that failed vsyscalls with vsyscall=none
695 * will have the PROT bit. This doesn't leak any
696 * information and does not appear to cause any problems.
697 */
698 if (address >= TASK_SIZE_MAX)
699 error_code |= X86_PF_PROT;
700
701 tsk->thread.trap_nr = X86_TRAP_PF;
702 tsk->thread.error_code = error_code | X86_PF_USER;
703 tsk->thread.cr2 = address;
704}
705
706static noinline void
707no_context(struct pt_regs *regs, unsigned long error_code,
708 unsigned long address, int signal, int si_code)
709{
710 struct task_struct *tsk = current;
711 unsigned long flags;
712 int sig;
713
714 if (user_mode(regs)) {
715 /*
716 * This is an implicit supervisor-mode access from user
717 * mode. Bypass all the kernel-mode recovery code and just
718 * OOPS.
719 */
720 goto oops;
721 }
722
723 /* Are we prepared to handle this kernel fault? */
724 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
725 /*
726 * Any interrupt that takes a fault gets the fixup. This makes
727 * the below recursive fault logic only apply to a faults from
728 * task context.
729 */
730 if (in_interrupt())
731 return;
732
733 /*
734 * Per the above we're !in_interrupt(), aka. task context.
735 *
736 * In this case we need to make sure we're not recursively
737 * faulting through the emulate_vsyscall() logic.
738 */
739 if (current->thread.sig_on_uaccess_err && signal) {
740 set_signal_archinfo(address, error_code);
741
742 /* XXX: hwpoison faults will set the wrong code. */
743 force_sig_fault(signal, si_code, (void __user *)address);
744 }
745
746 /*
747 * Barring that, we can do the fixup and be happy.
748 */
749 return;
750 }
751
752#ifdef CONFIG_VMAP_STACK
753 /*
754 * Stack overflow? During boot, we can fault near the initial
755 * stack in the direct map, but that's not an overflow -- check
756 * that we're in vmalloc space to avoid this.
757 */
758 if (is_vmalloc_addr((void *)address) &&
759 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
760 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
761 unsigned long stack = __this_cpu_ist_top_va(DF) - sizeof(void *);
762 /*
763 * We're likely to be running with very little stack space
764 * left. It's plausible that we'd hit this condition but
765 * double-fault even before we get this far, in which case
766 * we're fine: the double-fault handler will deal with it.
767 *
768 * We don't want to make it all the way into the oops code
769 * and then double-fault, though, because we're likely to
770 * break the console driver and lose most of the stack dump.
771 */
772 asm volatile ("movq %[stack], %%rsp\n\t"
773 "call handle_stack_overflow\n\t"
774 "1: jmp 1b"
775 : ASM_CALL_CONSTRAINT
776 : "D" ("kernel stack overflow (page fault)"),
777 "S" (regs), "d" (address),
778 [stack] "rm" (stack));
779 unreachable();
780 }
781#endif
782
783 /*
784 * 32-bit:
785 *
786 * Valid to do another page fault here, because if this fault
787 * had been triggered by is_prefetch fixup_exception would have
788 * handled it.
789 *
790 * 64-bit:
791 *
792 * Hall of shame of CPU/BIOS bugs.
793 */
794 if (is_prefetch(regs, error_code, address))
795 return;
796
797 if (is_errata93(regs, address))
798 return;
799
800 /*
801 * Buggy firmware could access regions which might page fault, try to
802 * recover from such faults.
803 */
804 if (IS_ENABLED(CONFIG_EFI))
805 efi_recover_from_page_fault(address);
806
807oops:
808 /*
809 * Oops. The kernel tried to access some bad page. We'll have to
810 * terminate things with extreme prejudice:
811 */
812 flags = oops_begin();
813
814 show_fault_oops(regs, error_code, address);
815
816 if (task_stack_end_corrupted(tsk))
817 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
818
819 sig = SIGKILL;
820 if (__die("Oops", regs, error_code))
821 sig = 0;
822
823 /* Executive summary in case the body of the oops scrolled away */
824 printk(KERN_DEFAULT "CR2: %016lx\n", address);
825
826 oops_end(flags, regs, sig);
827}
828
829/*
830 * Print out info about fatal segfaults, if the show_unhandled_signals
831 * sysctl is set:
832 */
833static inline void
834show_signal_msg(struct pt_regs *regs, unsigned long error_code,
835 unsigned long address, struct task_struct *tsk)
836{
837 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
838
839 if (!unhandled_signal(tsk, SIGSEGV))
840 return;
841
842 if (!printk_ratelimit())
843 return;
844
845 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
846 loglvl, tsk->comm, task_pid_nr(tsk), address,
847 (void *)regs->ip, (void *)regs->sp, error_code);
848
849 print_vma_addr(KERN_CONT " in ", regs->ip);
850
851 printk(KERN_CONT "\n");
852
853 show_opcodes(regs, loglvl);
854}
855
856/*
857 * The (legacy) vsyscall page is the long page in the kernel portion
858 * of the address space that has user-accessible permissions.
859 */
860static bool is_vsyscall_vaddr(unsigned long vaddr)
861{
862 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
863}
864
865static void
866__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
867 unsigned long address, u32 pkey, int si_code)
868{
869 struct task_struct *tsk = current;
870
871 /* User mode accesses just cause a SIGSEGV */
872 if (user_mode(regs) && (error_code & X86_PF_USER)) {
873 /*
874 * It's possible to have interrupts off here:
875 */
876 local_irq_enable();
877
878 /*
879 * Valid to do another page fault here because this one came
880 * from user space:
881 */
882 if (is_prefetch(regs, error_code, address))
883 return;
884
885 if (is_errata100(regs, address))
886 return;
887
888 /*
889 * To avoid leaking information about the kernel page table
890 * layout, pretend that user-mode accesses to kernel addresses
891 * are always protection faults.
892 */
893 if (address >= TASK_SIZE_MAX)
894 error_code |= X86_PF_PROT;
895
896 if (likely(show_unhandled_signals))
897 show_signal_msg(regs, error_code, address, tsk);
898
899 set_signal_archinfo(address, error_code);
900
901 if (si_code == SEGV_PKUERR)
902 force_sig_pkuerr((void __user *)address, pkey);
903
904 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
905
906 return;
907 }
908
909 if (is_f00f_bug(regs, address))
910 return;
911
912 no_context(regs, error_code, address, SIGSEGV, si_code);
913}
914
915static noinline void
916bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
917 unsigned long address)
918{
919 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
920}
921
922static void
923__bad_area(struct pt_regs *regs, unsigned long error_code,
924 unsigned long address, u32 pkey, int si_code)
925{
926 struct mm_struct *mm = current->mm;
927 /*
928 * Something tried to access memory that isn't in our memory map..
929 * Fix it, but check if it's kernel or user first..
930 */
931 up_read(&mm->mmap_sem);
932
933 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
934}
935
936static noinline void
937bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
938{
939 __bad_area(regs, error_code, address, 0, SEGV_MAPERR);
940}
941
942static inline bool bad_area_access_from_pkeys(unsigned long error_code,
943 struct vm_area_struct *vma)
944{
945 /* This code is always called on the current mm */
946 bool foreign = false;
947
948 if (!boot_cpu_has(X86_FEATURE_OSPKE))
949 return false;
950 if (error_code & X86_PF_PK)
951 return true;
952 /* this checks permission keys on the VMA: */
953 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
954 (error_code & X86_PF_INSTR), foreign))
955 return true;
956 return false;
957}
958
959static noinline void
960bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
961 unsigned long address, struct vm_area_struct *vma)
962{
963 /*
964 * This OSPKE check is not strictly necessary at runtime.
965 * But, doing it this way allows compiler optimizations
966 * if pkeys are compiled out.
967 */
968 if (bad_area_access_from_pkeys(error_code, vma)) {
969 /*
970 * A protection key fault means that the PKRU value did not allow
971 * access to some PTE. Userspace can figure out what PKRU was
972 * from the XSAVE state. This function captures the pkey from
973 * the vma and passes it to userspace so userspace can discover
974 * which protection key was set on the PTE.
975 *
976 * If we get here, we know that the hardware signaled a X86_PF_PK
977 * fault and that there was a VMA once we got in the fault
978 * handler. It does *not* guarantee that the VMA we find here
979 * was the one that we faulted on.
980 *
981 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
982 * 2. T1 : set PKRU to deny access to pkey=4, touches page
983 * 3. T1 : faults...
984 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
985 * 5. T1 : enters fault handler, takes mmap_sem, etc...
986 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
987 * faulted on a pte with its pkey=4.
988 */
989 u32 pkey = vma_pkey(vma);
990
991 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
992 } else {
993 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
994 }
995}
996
997static void
998do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
999 vm_fault_t fault)
1000{
1001 /* Kernel mode? Handle exceptions or die: */
1002 if (!(error_code & X86_PF_USER)) {
1003 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1004 return;
1005 }
1006
1007 /* User-space => ok to do another page fault: */
1008 if (is_prefetch(regs, error_code, address))
1009 return;
1010
1011 set_signal_archinfo(address, error_code);
1012
1013#ifdef CONFIG_MEMORY_FAILURE
1014 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
1015 struct task_struct *tsk = current;
1016 unsigned lsb = 0;
1017
1018 pr_err(
1019 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
1020 tsk->comm, tsk->pid, address);
1021 if (fault & VM_FAULT_HWPOISON_LARGE)
1022 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
1023 if (fault & VM_FAULT_HWPOISON)
1024 lsb = PAGE_SHIFT;
1025 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
1026 return;
1027 }
1028#endif
1029 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
1030}
1031
1032static noinline void
1033mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1034 unsigned long address, vm_fault_t fault)
1035{
1036 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
1037 no_context(regs, error_code, address, 0, 0);
1038 return;
1039 }
1040
1041 if (fault & VM_FAULT_OOM) {
1042 /* Kernel mode? Handle exceptions or die: */
1043 if (!(error_code & X86_PF_USER)) {
1044 no_context(regs, error_code, address,
1045 SIGSEGV, SEGV_MAPERR);
1046 return;
1047 }
1048
1049 /*
1050 * We ran out of memory, call the OOM killer, and return the
1051 * userspace (which will retry the fault, or kill us if we got
1052 * oom-killed):
1053 */
1054 pagefault_out_of_memory();
1055 } else {
1056 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1057 VM_FAULT_HWPOISON_LARGE))
1058 do_sigbus(regs, error_code, address, fault);
1059 else if (fault & VM_FAULT_SIGSEGV)
1060 bad_area_nosemaphore(regs, error_code, address);
1061 else
1062 BUG();
1063 }
1064}
1065
1066static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
1067{
1068 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1069 return 0;
1070
1071 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1072 return 0;
1073
1074 return 1;
1075}
1076
1077/*
1078 * Handle a spurious fault caused by a stale TLB entry.
1079 *
1080 * This allows us to lazily refresh the TLB when increasing the
1081 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1082 * eagerly is very expensive since that implies doing a full
1083 * cross-processor TLB flush, even if no stale TLB entries exist
1084 * on other processors.
1085 *
1086 * Spurious faults may only occur if the TLB contains an entry with
1087 * fewer permission than the page table entry. Non-present (P = 0)
1088 * and reserved bit (R = 1) faults are never spurious.
1089 *
1090 * There are no security implications to leaving a stale TLB when
1091 * increasing the permissions on a page.
1092 *
1093 * Returns non-zero if a spurious fault was handled, zero otherwise.
1094 *
1095 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1096 * (Optional Invalidation).
1097 */
1098static noinline int
1099spurious_kernel_fault(unsigned long error_code, unsigned long address)
1100{
1101 pgd_t *pgd;
1102 p4d_t *p4d;
1103 pud_t *pud;
1104 pmd_t *pmd;
1105 pte_t *pte;
1106 int ret;
1107
1108 /*
1109 * Only writes to RO or instruction fetches from NX may cause
1110 * spurious faults.
1111 *
1112 * These could be from user or supervisor accesses but the TLB
1113 * is only lazily flushed after a kernel mapping protection
1114 * change, so user accesses are not expected to cause spurious
1115 * faults.
1116 */
1117 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1118 error_code != (X86_PF_INSTR | X86_PF_PROT))
1119 return 0;
1120
1121 pgd = init_mm.pgd + pgd_index(address);
1122 if (!pgd_present(*pgd))
1123 return 0;
1124
1125 p4d = p4d_offset(pgd, address);
1126 if (!p4d_present(*p4d))
1127 return 0;
1128
1129 if (p4d_large(*p4d))
1130 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1131
1132 pud = pud_offset(p4d, address);
1133 if (!pud_present(*pud))
1134 return 0;
1135
1136 if (pud_large(*pud))
1137 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1138
1139 pmd = pmd_offset(pud, address);
1140 if (!pmd_present(*pmd))
1141 return 0;
1142
1143 if (pmd_large(*pmd))
1144 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1145
1146 pte = pte_offset_kernel(pmd, address);
1147 if (!pte_present(*pte))
1148 return 0;
1149
1150 ret = spurious_kernel_fault_check(error_code, pte);
1151 if (!ret)
1152 return 0;
1153
1154 /*
1155 * Make sure we have permissions in PMD.
1156 * If not, then there's a bug in the page tables:
1157 */
1158 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1159 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1160
1161 return ret;
1162}
1163NOKPROBE_SYMBOL(spurious_kernel_fault);
1164
1165int show_unhandled_signals = 1;
1166
1167static inline int
1168access_error(unsigned long error_code, struct vm_area_struct *vma)
1169{
1170 /* This is only called for the current mm, so: */
1171 bool foreign = false;
1172
1173 /*
1174 * Read or write was blocked by protection keys. This is
1175 * always an unconditional error and can never result in
1176 * a follow-up action to resolve the fault, like a COW.
1177 */
1178 if (error_code & X86_PF_PK)
1179 return 1;
1180
1181 /*
1182 * Make sure to check the VMA so that we do not perform
1183 * faults just to hit a X86_PF_PK as soon as we fill in a
1184 * page.
1185 */
1186 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1187 (error_code & X86_PF_INSTR), foreign))
1188 return 1;
1189
1190 if (error_code & X86_PF_WRITE) {
1191 /* write, present and write, not present: */
1192 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1193 return 1;
1194 return 0;
1195 }
1196
1197 /* read, present: */
1198 if (unlikely(error_code & X86_PF_PROT))
1199 return 1;
1200
1201 /* read, not present: */
1202 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1203 return 1;
1204
1205 return 0;
1206}
1207
1208static int fault_in_kernel_space(unsigned long address)
1209{
1210 /*
1211 * On 64-bit systems, the vsyscall page is at an address above
1212 * TASK_SIZE_MAX, but is not considered part of the kernel
1213 * address space.
1214 */
1215 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1216 return false;
1217
1218 return address >= TASK_SIZE_MAX;
1219}
1220
1221/*
1222 * Called for all faults where 'address' is part of the kernel address
1223 * space. Might get called for faults that originate from *code* that
1224 * ran in userspace or the kernel.
1225 */
1226static void
1227do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1228 unsigned long address)
1229{
1230 /*
1231 * Protection keys exceptions only happen on user pages. We
1232 * have no user pages in the kernel portion of the address
1233 * space, so do not expect them here.
1234 */
1235 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1236
1237 /*
1238 * We can fault-in kernel-space virtual memory on-demand. The
1239 * 'reference' page table is init_mm.pgd.
1240 *
1241 * NOTE! We MUST NOT take any locks for this case. We may
1242 * be in an interrupt or a critical region, and should
1243 * only copy the information from the master page table,
1244 * nothing more.
1245 *
1246 * Before doing this on-demand faulting, ensure that the
1247 * fault is not any of the following:
1248 * 1. A fault on a PTE with a reserved bit set.
1249 * 2. A fault caused by a user-mode access. (Do not demand-
1250 * fault kernel memory due to user-mode accesses).
1251 * 3. A fault caused by a page-level protection violation.
1252 * (A demand fault would be on a non-present page which
1253 * would have X86_PF_PROT==0).
1254 */
1255 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1256 if (vmalloc_fault(address) >= 0)
1257 return;
1258 }
1259
1260 /* Was the fault spurious, caused by lazy TLB invalidation? */
1261 if (spurious_kernel_fault(hw_error_code, address))
1262 return;
1263
1264 /* kprobes don't want to hook the spurious faults: */
1265 if (kprobe_page_fault(regs, X86_TRAP_PF))
1266 return;
1267
1268 /*
1269 * Note, despite being a "bad area", there are quite a few
1270 * acceptable reasons to get here, such as erratum fixups
1271 * and handling kernel code that can fault, like get_user().
1272 *
1273 * Don't take the mm semaphore here. If we fixup a prefetch
1274 * fault we could otherwise deadlock:
1275 */
1276 bad_area_nosemaphore(regs, hw_error_code, address);
1277}
1278NOKPROBE_SYMBOL(do_kern_addr_fault);
1279
1280/* Handle faults in the user portion of the address space */
1281static inline
1282void do_user_addr_fault(struct pt_regs *regs,
1283 unsigned long hw_error_code,
1284 unsigned long address)
1285{
1286 struct vm_area_struct *vma;
1287 struct task_struct *tsk;
1288 struct mm_struct *mm;
1289 vm_fault_t fault, major = 0;
1290 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1291
1292 tsk = current;
1293 mm = tsk->mm;
1294
1295 /* kprobes don't want to hook the spurious faults: */
1296 if (unlikely(kprobe_page_fault(regs, X86_TRAP_PF)))
1297 return;
1298
1299 /*
1300 * Reserved bits are never expected to be set on
1301 * entries in the user portion of the page tables.
1302 */
1303 if (unlikely(hw_error_code & X86_PF_RSVD))
1304 pgtable_bad(regs, hw_error_code, address);
1305
1306 /*
1307 * If SMAP is on, check for invalid kernel (supervisor) access to user
1308 * pages in the user address space. The odd case here is WRUSS,
1309 * which, according to the preliminary documentation, does not respect
1310 * SMAP and will have the USER bit set so, in all cases, SMAP
1311 * enforcement appears to be consistent with the USER bit.
1312 */
1313 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1314 !(hw_error_code & X86_PF_USER) &&
1315 !(regs->flags & X86_EFLAGS_AC)))
1316 {
1317 bad_area_nosemaphore(regs, hw_error_code, address);
1318 return;
1319 }
1320
1321 /*
1322 * If we're in an interrupt, have no user context or are running
1323 * in a region with pagefaults disabled then we must not take the fault
1324 */
1325 if (unlikely(faulthandler_disabled() || !mm)) {
1326 bad_area_nosemaphore(regs, hw_error_code, address);
1327 return;
1328 }
1329
1330 /*
1331 * It's safe to allow irq's after cr2 has been saved and the
1332 * vmalloc fault has been handled.
1333 *
1334 * User-mode registers count as a user access even for any
1335 * potential system fault or CPU buglet:
1336 */
1337 if (user_mode(regs)) {
1338 local_irq_enable();
1339 flags |= FAULT_FLAG_USER;
1340 } else {
1341 if (regs->flags & X86_EFLAGS_IF)
1342 local_irq_enable();
1343 }
1344
1345 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1346
1347 if (hw_error_code & X86_PF_WRITE)
1348 flags |= FAULT_FLAG_WRITE;
1349 if (hw_error_code & X86_PF_INSTR)
1350 flags |= FAULT_FLAG_INSTRUCTION;
1351
1352#ifdef CONFIG_X86_64
1353 /*
1354 * Faults in the vsyscall page might need emulation. The
1355 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1356 * considered to be part of the user address space.
1357 *
1358 * The vsyscall page does not have a "real" VMA, so do this
1359 * emulation before we go searching for VMAs.
1360 *
1361 * PKRU never rejects instruction fetches, so we don't need
1362 * to consider the PF_PK bit.
1363 */
1364 if (is_vsyscall_vaddr(address)) {
1365 if (emulate_vsyscall(hw_error_code, regs, address))
1366 return;
1367 }
1368#endif
1369
1370 /*
1371 * Kernel-mode access to the user address space should only occur
1372 * on well-defined single instructions listed in the exception
1373 * tables. But, an erroneous kernel fault occurring outside one of
1374 * those areas which also holds mmap_sem might deadlock attempting
1375 * to validate the fault against the address space.
1376 *
1377 * Only do the expensive exception table search when we might be at
1378 * risk of a deadlock. This happens if we
1379 * 1. Failed to acquire mmap_sem, and
1380 * 2. The access did not originate in userspace.
1381 */
1382 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1383 if (!user_mode(regs) && !search_exception_tables(regs->ip)) {
1384 /*
1385 * Fault from code in kernel from
1386 * which we do not expect faults.
1387 */
1388 bad_area_nosemaphore(regs, hw_error_code, address);
1389 return;
1390 }
1391retry:
1392 down_read(&mm->mmap_sem);
1393 } else {
1394 /*
1395 * The above down_read_trylock() might have succeeded in
1396 * which case we'll have missed the might_sleep() from
1397 * down_read():
1398 */
1399 might_sleep();
1400 }
1401
1402 vma = find_vma(mm, address);
1403 if (unlikely(!vma)) {
1404 bad_area(regs, hw_error_code, address);
1405 return;
1406 }
1407 if (likely(vma->vm_start <= address))
1408 goto good_area;
1409 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1410 bad_area(regs, hw_error_code, address);
1411 return;
1412 }
1413 if (unlikely(expand_stack(vma, address))) {
1414 bad_area(regs, hw_error_code, address);
1415 return;
1416 }
1417
1418 /*
1419 * Ok, we have a good vm_area for this memory access, so
1420 * we can handle it..
1421 */
1422good_area:
1423 if (unlikely(access_error(hw_error_code, vma))) {
1424 bad_area_access_error(regs, hw_error_code, address, vma);
1425 return;
1426 }
1427
1428 /*
1429 * If for any reason at all we couldn't handle the fault,
1430 * make sure we exit gracefully rather than endlessly redo
1431 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1432 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1433 *
1434 * Note that handle_userfault() may also release and reacquire mmap_sem
1435 * (and not return with VM_FAULT_RETRY), when returning to userland to
1436 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1437 * (potentially after handling any pending signal during the return to
1438 * userland). The return to userland is identified whenever
1439 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1440 */
1441 fault = handle_mm_fault(vma, address, flags);
1442 major |= fault & VM_FAULT_MAJOR;
1443
1444 /*
1445 * If we need to retry the mmap_sem has already been released,
1446 * and if there is a fatal signal pending there is no guarantee
1447 * that we made any progress. Handle this case first.
1448 */
1449 if (unlikely(fault & VM_FAULT_RETRY)) {
1450 /* Retry at most once */
1451 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1452 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1453 flags |= FAULT_FLAG_TRIED;
1454 if (!fatal_signal_pending(tsk))
1455 goto retry;
1456 }
1457
1458 /* User mode? Just return to handle the fatal exception */
1459 if (flags & FAULT_FLAG_USER)
1460 return;
1461
1462 /* Not returning to user mode? Handle exceptions or die: */
1463 no_context(regs, hw_error_code, address, SIGBUS, BUS_ADRERR);
1464 return;
1465 }
1466
1467 up_read(&mm->mmap_sem);
1468 if (unlikely(fault & VM_FAULT_ERROR)) {
1469 mm_fault_error(regs, hw_error_code, address, fault);
1470 return;
1471 }
1472
1473 /*
1474 * Major/minor page fault accounting. If any of the events
1475 * returned VM_FAULT_MAJOR, we account it as a major fault.
1476 */
1477 if (major) {
1478 tsk->maj_flt++;
1479 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1480 } else {
1481 tsk->min_flt++;
1482 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1483 }
1484
1485 check_v8086_mode(regs, address, tsk);
1486}
1487NOKPROBE_SYMBOL(do_user_addr_fault);
1488
1489/*
1490 * Explicitly marked noinline such that the function tracer sees this as the
1491 * page_fault entry point.
1492 */
1493static noinline void
1494__do_page_fault(struct pt_regs *regs, unsigned long hw_error_code,
1495 unsigned long address)
1496{
1497 prefetchw(¤t->mm->mmap_sem);
1498
1499 if (unlikely(kmmio_fault(regs, address)))
1500 return;
1501
1502 /* Was the fault on kernel-controlled part of the address space? */
1503 if (unlikely(fault_in_kernel_space(address)))
1504 do_kern_addr_fault(regs, hw_error_code, address);
1505 else
1506 do_user_addr_fault(regs, hw_error_code, address);
1507}
1508NOKPROBE_SYMBOL(__do_page_fault);
1509
1510static __always_inline void
1511trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1512 unsigned long address)
1513{
1514 if (!trace_pagefault_enabled())
1515 return;
1516
1517 if (user_mode(regs))
1518 trace_page_fault_user(address, regs, error_code);
1519 else
1520 trace_page_fault_kernel(address, regs, error_code);
1521}
1522
1523dotraplinkage void
1524do_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
1525{
1526 enum ctx_state prev_state;
1527
1528 prev_state = exception_enter();
1529 trace_page_fault_entries(regs, error_code, address);
1530 __do_page_fault(regs, error_code, address);
1531 exception_exit(prev_state);
1532}
1533NOKPROBE_SYMBOL(do_page_fault);
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))
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(pgd))
361 goto bad;
362
363 pr_info("PGD %lx ", pgd_val(*pgd));
364
365 if (!pgd_present(*pgd))
366 goto out;
367
368 p4d = p4d_offset(pgd, address);
369 if (bad_address(p4d))
370 goto bad;
371
372 pr_cont("P4D %lx ", p4d_val(*p4d));
373 if (!p4d_present(*p4d) || p4d_leaf(*p4d))
374 goto out;
375
376 pud = pud_offset(p4d, address);
377 if (bad_address(pud))
378 goto bad;
379
380 pr_cont("PUD %lx ", pud_val(*pud));
381 if (!pud_present(*pud) || pud_leaf(*pud))
382 goto out;
383
384 pmd = pmd_offset(pud, address);
385 if (bad_address(pmd))
386 goto bad;
387
388 pr_cont("PMD %lx ", pmd_val(*pmd));
389 if (!pmd_present(*pmd) || pmd_leaf(*pmd))
390 goto out;
391
392 pte = pte_offset_kernel(pmd, address);
393 if (bad_address(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(&desc, (void *)(gdt->address + offset),
495 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);
524
525 if (pte && pte_present(*pte) && !pte_exec(*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(*pte) && pte_exec(*pte) &&
529 (pgd_flags(*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(&idt);
570
571 /* Usable even on Xen PV -- it's just slow. */
572 native_store_gdt(&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, "LDTR", ldtr);
579
580 store_tr(tr);
581 show_ldttss(&gdt, "TR", 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, 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((void *)address) &&
663 get_stack_guard_info((void *)address, &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(address);
690
691 /* Only not-present faults should be handled by KFENCE. */
692 if (!(error_code & X86_PF_PROT) &&
693 kfence_handle_page_fault(address, 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, 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, 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, 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 ", 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, address))
805 return;
806
807 if (is_errata100(regs, address))
808 return;
809
810 sanitize_error_code(address, &error_code);
811
812 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, 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((void __user *)address, pkey);
822 else
823 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
824
825 local_irq_disable();
826}
827
828static noinline void
829bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
830 unsigned long address)
831{
832 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
833}
834
835static void
836__bad_area(struct pt_regs *regs, unsigned long error_code,
837 unsigned long address, u32 pkey, int si_code)
838{
839 struct mm_struct *mm = current->mm;
840 /*
841 * Something tried to access memory that isn't in our memory map..
842 * Fix it, but check if it's kernel or user first..
843 */
844 mmap_read_unlock(mm);
845
846 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
847}
848
849static inline bool bad_area_access_from_pkeys(unsigned long error_code,
850 struct vm_area_struct *vma)
851{
852 /* This code is always called on the current mm */
853 bool foreign = false;
854
855 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
856 return false;
857 if (error_code & X86_PF_PK)
858 return true;
859 /* this checks permission keys on the VMA: */
860 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
861 (error_code & X86_PF_INSTR), foreign))
862 return true;
863 return false;
864}
865
866static noinline void
867bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
868 unsigned long address, struct vm_area_struct *vma)
869{
870 /*
871 * This OSPKE check is not strictly necessary at runtime.
872 * But, doing it this way allows compiler optimizations
873 * if pkeys are compiled out.
874 */
875 if (bad_area_access_from_pkeys(error_code, vma)) {
876 /*
877 * A protection key fault means that the PKRU value did not allow
878 * access to some PTE. Userspace can figure out what PKRU was
879 * from the XSAVE state. This function captures the pkey from
880 * the vma and passes it to userspace so userspace can discover
881 * which protection key was set on the PTE.
882 *
883 * If we get here, we know that the hardware signaled a X86_PF_PK
884 * fault and that there was a VMA once we got in the fault
885 * handler. It does *not* guarantee that the VMA we find here
886 * was the one that we faulted on.
887 *
888 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
889 * 2. T1 : set PKRU to deny access to pkey=4, touches page
890 * 3. T1 : faults...
891 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
892 * 5. T1 : enters fault handler, takes mmap_lock, etc...
893 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
894 * faulted on a pte with its pkey=4.
895 */
896 u32 pkey = vma_pkey(vma);
897
898 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
899 } else {
900 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
901 }
902}
903
904static void
905do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
906 vm_fault_t fault)
907{
908 /* Kernel mode? Handle exceptions or die: */
909 if (!user_mode(regs)) {
910 kernelmode_fixup_or_oops(regs, error_code, address,
911 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
912 return;
913 }
914
915 /* User-space => ok to do another page fault: */
916 if (is_prefetch(regs, error_code, address))
917 return;
918
919 sanitize_error_code(address, &error_code);
920
921 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
922 return;
923
924 set_signal_archinfo(address, error_code);
925
926#ifdef CONFIG_MEMORY_FAILURE
927 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
928 struct task_struct *tsk = current;
929 unsigned lsb = 0;
930
931 pr_err(
932 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
933 tsk->comm, tsk->pid, address);
934 if (fault & VM_FAULT_HWPOISON_LARGE)
935 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
936 if (fault & VM_FAULT_HWPOISON)
937 lsb = PAGE_SHIFT;
938 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
939 return;
940 }
941#endif
942 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
943}
944
945static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
946{
947 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
948 return 0;
949
950 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
951 return 0;
952
953 return 1;
954}
955
956/*
957 * Handle a spurious fault caused by a stale TLB entry.
958 *
959 * This allows us to lazily refresh the TLB when increasing the
960 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
961 * eagerly is very expensive since that implies doing a full
962 * cross-processor TLB flush, even if no stale TLB entries exist
963 * on other processors.
964 *
965 * Spurious faults may only occur if the TLB contains an entry with
966 * fewer permission than the page table entry. Non-present (P = 0)
967 * and reserved bit (R = 1) faults are never spurious.
968 *
969 * There are no security implications to leaving a stale TLB when
970 * increasing the permissions on a page.
971 *
972 * Returns non-zero if a spurious fault was handled, zero otherwise.
973 *
974 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
975 * (Optional Invalidation).
976 */
977static noinline int
978spurious_kernel_fault(unsigned long error_code, unsigned long address)
979{
980 pgd_t *pgd;
981 p4d_t *p4d;
982 pud_t *pud;
983 pmd_t *pmd;
984 pte_t *pte;
985 int ret;
986
987 /*
988 * Only writes to RO or instruction fetches from NX may cause
989 * spurious faults.
990 *
991 * These could be from user or supervisor accesses but the TLB
992 * is only lazily flushed after a kernel mapping protection
993 * change, so user accesses are not expected to cause spurious
994 * faults.
995 */
996 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
997 error_code != (X86_PF_INSTR | X86_PF_PROT))
998 return 0;
999
1000 pgd = init_mm.pgd + pgd_index(address);
1001 if (!pgd_present(*pgd))
1002 return 0;
1003
1004 p4d = p4d_offset(pgd, address);
1005 if (!p4d_present(*p4d))
1006 return 0;
1007
1008 if (p4d_leaf(*p4d))
1009 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1010
1011 pud = pud_offset(p4d, address);
1012 if (!pud_present(*pud))
1013 return 0;
1014
1015 if (pud_leaf(*pud))
1016 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1017
1018 pmd = pmd_offset(pud, address);
1019 if (!pmd_present(*pmd))
1020 return 0;
1021
1022 if (pmd_leaf(*pmd))
1023 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1024
1025 pte = pte_offset_kernel(pmd, address);
1026 if (!pte_present(*pte))
1027 return 0;
1028
1029 ret = spurious_kernel_fault_check(error_code, pte);
1030 if (!ret)
1031 return 0;
1032
1033 /*
1034 * Make sure we have permissions in PMD.
1035 * If not, then there's a bug in the page tables:
1036 */
1037 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1038 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1039
1040 return ret;
1041}
1042NOKPROBE_SYMBOL(spurious_kernel_fault);
1043
1044int show_unhandled_signals = 1;
1045
1046static inline int
1047access_error(unsigned long error_code, struct vm_area_struct *vma)
1048{
1049 /* This is only called for the current mm, so: */
1050 bool foreign = false;
1051
1052 /*
1053 * Read or write was blocked by protection keys. This is
1054 * always an unconditional error and can never result in
1055 * a follow-up action to resolve the fault, like a COW.
1056 */
1057 if (error_code & X86_PF_PK)
1058 return 1;
1059
1060 /*
1061 * SGX hardware blocked the access. This usually happens
1062 * when the enclave memory contents have been destroyed, like
1063 * after a suspend/resume cycle. In any case, the kernel can't
1064 * fix the cause of the fault. Handle the fault as an access
1065 * error even in cases where no actual access violation
1066 * occurred. This allows userspace to rebuild the enclave in
1067 * response to the signal.
1068 */
1069 if (unlikely(error_code & X86_PF_SGX))
1070 return 1;
1071
1072 /*
1073 * Make sure to check the VMA so that we do not perform
1074 * faults just to hit a X86_PF_PK as soon as we fill in a
1075 * page.
1076 */
1077 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1078 (error_code & X86_PF_INSTR), foreign))
1079 return 1;
1080
1081 /*
1082 * Shadow stack accesses (PF_SHSTK=1) are only permitted to
1083 * shadow stack VMAs. All other accesses result in an error.
1084 */
1085 if (error_code & X86_PF_SHSTK) {
1086 if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK)))
1087 return 1;
1088 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1089 return 1;
1090 return 0;
1091 }
1092
1093 if (error_code & X86_PF_WRITE) {
1094 /* write, present and write, not present: */
1095 if (unlikely(vma->vm_flags & VM_SHADOW_STACK))
1096 return 1;
1097 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1098 return 1;
1099 return 0;
1100 }
1101
1102 /* read, present: */
1103 if (unlikely(error_code & X86_PF_PROT))
1104 return 1;
1105
1106 /* read, not present: */
1107 if (unlikely(!vma_is_accessible(vma)))
1108 return 1;
1109
1110 return 0;
1111}
1112
1113bool fault_in_kernel_space(unsigned long address)
1114{
1115 /*
1116 * On 64-bit systems, the vsyscall page is at an address above
1117 * TASK_SIZE_MAX, but is not considered part of the kernel
1118 * address space.
1119 */
1120 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1121 return false;
1122
1123 return address >= TASK_SIZE_MAX;
1124}
1125
1126/*
1127 * Called for all faults where 'address' is part of the kernel address
1128 * space. Might get called for faults that originate from *code* that
1129 * ran in userspace or the kernel.
1130 */
1131static void
1132do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1133 unsigned long address)
1134{
1135 /*
1136 * Protection keys exceptions only happen on user pages. We
1137 * have no user pages in the kernel portion of the address
1138 * space, so do not expect them here.
1139 */
1140 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1141
1142#ifdef CONFIG_X86_32
1143 /*
1144 * We can fault-in kernel-space virtual memory on-demand. The
1145 * 'reference' page table is init_mm.pgd.
1146 *
1147 * NOTE! We MUST NOT take any locks for this case. We may
1148 * be in an interrupt or a critical region, and should
1149 * only copy the information from the master page table,
1150 * nothing more.
1151 *
1152 * Before doing this on-demand faulting, ensure that the
1153 * fault is not any of the following:
1154 * 1. A fault on a PTE with a reserved bit set.
1155 * 2. A fault caused by a user-mode access. (Do not demand-
1156 * fault kernel memory due to user-mode accesses).
1157 * 3. A fault caused by a page-level protection violation.
1158 * (A demand fault would be on a non-present page which
1159 * would have X86_PF_PROT==0).
1160 *
1161 * This is only needed to close a race condition on x86-32 in
1162 * the vmalloc mapping/unmapping code. See the comment above
1163 * vmalloc_fault() for details. On x86-64 the race does not
1164 * exist as the vmalloc mappings don't need to be synchronized
1165 * there.
1166 */
1167 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1168 if (vmalloc_fault(address) >= 0)
1169 return;
1170 }
1171#endif
1172
1173 if (is_f00f_bug(regs, hw_error_code, address))
1174 return;
1175
1176 /* Was the fault spurious, caused by lazy TLB invalidation? */
1177 if (spurious_kernel_fault(hw_error_code, address))
1178 return;
1179
1180 /* kprobes don't want to hook the spurious faults: */
1181 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1182 return;
1183
1184 /*
1185 * Note, despite being a "bad area", there are quite a few
1186 * acceptable reasons to get here, such as erratum fixups
1187 * and handling kernel code that can fault, like get_user().
1188 *
1189 * Don't take the mm semaphore here. If we fixup a prefetch
1190 * fault we could otherwise deadlock:
1191 */
1192 bad_area_nosemaphore(regs, hw_error_code, address);
1193}
1194NOKPROBE_SYMBOL(do_kern_addr_fault);
1195
1196/*
1197 * Handle faults in the user portion of the address space. Nothing in here
1198 * should check X86_PF_USER without a specific justification: for almost
1199 * all purposes, we should treat a normal kernel access to user memory
1200 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1201 * The one exception is AC flag handling, which is, per the x86
1202 * architecture, special for WRUSS.
1203 */
1204static inline
1205void do_user_addr_fault(struct pt_regs *regs,
1206 unsigned long error_code,
1207 unsigned long address)
1208{
1209 struct vm_area_struct *vma;
1210 struct task_struct *tsk;
1211 struct mm_struct *mm;
1212 vm_fault_t fault;
1213 unsigned int flags = FAULT_FLAG_DEFAULT;
1214
1215 tsk = current;
1216 mm = tsk->mm;
1217
1218 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1219 /*
1220 * Whoops, this is kernel mode code trying to execute from
1221 * user memory. Unless this is AMD erratum #93, which
1222 * corrupts RIP such that it looks like a user address,
1223 * this is unrecoverable. Don't even try to look up the
1224 * VMA or look for extable entries.
1225 */
1226 if (is_errata93(regs, address))
1227 return;
1228
1229 page_fault_oops(regs, error_code, address);
1230 return;
1231 }
1232
1233 /* kprobes don't want to hook the spurious faults: */
1234 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1235 return;
1236
1237 /*
1238 * Reserved bits are never expected to be set on
1239 * entries in the user portion of the page tables.
1240 */
1241 if (unlikely(error_code & X86_PF_RSVD))
1242 pgtable_bad(regs, error_code, address);
1243
1244 /*
1245 * If SMAP is on, check for invalid kernel (supervisor) access to user
1246 * pages in the user address space. The odd case here is WRUSS,
1247 * which, according to the preliminary documentation, does not respect
1248 * SMAP and will have the USER bit set so, in all cases, SMAP
1249 * enforcement appears to be consistent with the USER bit.
1250 */
1251 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1252 !(error_code & X86_PF_USER) &&
1253 !(regs->flags & X86_EFLAGS_AC))) {
1254 /*
1255 * No extable entry here. This was a kernel access to an
1256 * invalid pointer. get_kernel_nofault() will not get here.
1257 */
1258 page_fault_oops(regs, error_code, address);
1259 return;
1260 }
1261
1262 /*
1263 * If we're in an interrupt, have no user context or are running
1264 * in a region with pagefaults disabled then we must not take the fault
1265 */
1266 if (unlikely(faulthandler_disabled() || !mm)) {
1267 bad_area_nosemaphore(regs, error_code, address);
1268 return;
1269 }
1270
1271 /* Legacy check - remove this after verifying that it doesn't trigger */
1272 if (WARN_ON_ONCE(!(regs->flags & X86_EFLAGS_IF))) {
1273 bad_area_nosemaphore(regs, error_code, address);
1274 return;
1275 }
1276
1277 local_irq_enable();
1278
1279 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1280
1281 /*
1282 * Read-only permissions can not be expressed in shadow stack PTEs.
1283 * Treat all shadow stack accesses as WRITE faults. This ensures
1284 * that the MM will prepare everything (e.g., break COW) such that
1285 * maybe_mkwrite() can create a proper shadow stack PTE.
1286 */
1287 if (error_code & X86_PF_SHSTK)
1288 flags |= FAULT_FLAG_WRITE;
1289 if (error_code & X86_PF_WRITE)
1290 flags |= FAULT_FLAG_WRITE;
1291 if (error_code & X86_PF_INSTR)
1292 flags |= FAULT_FLAG_INSTRUCTION;
1293
1294 /*
1295 * We set FAULT_FLAG_USER based on the register state, not
1296 * based on X86_PF_USER. User space accesses that cause
1297 * system page faults are still user accesses.
1298 */
1299 if (user_mode(regs))
1300 flags |= FAULT_FLAG_USER;
1301
1302#ifdef CONFIG_X86_64
1303 /*
1304 * Faults in the vsyscall page might need emulation. The
1305 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1306 * considered to be part of the user address space.
1307 *
1308 * The vsyscall page does not have a "real" VMA, so do this
1309 * emulation before we go searching for VMAs.
1310 *
1311 * PKRU never rejects instruction fetches, so we don't need
1312 * to consider the PF_PK bit.
1313 */
1314 if (is_vsyscall_vaddr(address)) {
1315 if (emulate_vsyscall(error_code, regs, address))
1316 return;
1317 }
1318#endif
1319
1320 if (!(flags & FAULT_FLAG_USER))
1321 goto lock_mmap;
1322
1323 vma = lock_vma_under_rcu(mm, address);
1324 if (!vma)
1325 goto lock_mmap;
1326
1327 if (unlikely(access_error(error_code, vma))) {
1328 vma_end_read(vma);
1329 goto lock_mmap;
1330 }
1331 fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
1332 if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
1333 vma_end_read(vma);
1334
1335 if (!(fault & VM_FAULT_RETRY)) {
1336 count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1337 goto done;
1338 }
1339 count_vm_vma_lock_event(VMA_LOCK_RETRY);
1340 if (fault & VM_FAULT_MAJOR)
1341 flags |= FAULT_FLAG_TRIED;
1342
1343 /* Quick path to respond to signals */
1344 if (fault_signal_pending(fault, regs)) {
1345 if (!user_mode(regs))
1346 kernelmode_fixup_or_oops(regs, error_code, address,
1347 SIGBUS, BUS_ADRERR,
1348 ARCH_DEFAULT_PKEY);
1349 return;
1350 }
1351lock_mmap:
1352
1353retry:
1354 vma = lock_mm_and_find_vma(mm, address, regs);
1355 if (unlikely(!vma)) {
1356 bad_area_nosemaphore(regs, error_code, address);
1357 return;
1358 }
1359
1360 /*
1361 * Ok, we have a good vm_area for this memory access, so
1362 * we can handle it..
1363 */
1364 if (unlikely(access_error(error_code, vma))) {
1365 bad_area_access_error(regs, error_code, address, vma);
1366 return;
1367 }
1368
1369 /*
1370 * If for any reason at all we couldn't handle the fault,
1371 * make sure we exit gracefully rather than endlessly redo
1372 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1373 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1374 *
1375 * Note that handle_userfault() may also release and reacquire mmap_lock
1376 * (and not return with VM_FAULT_RETRY), when returning to userland to
1377 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1378 * (potentially after handling any pending signal during the return to
1379 * userland). The return to userland is identified whenever
1380 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1381 */
1382 fault = handle_mm_fault(vma, address, flags, regs);
1383
1384 if (fault_signal_pending(fault, regs)) {
1385 /*
1386 * Quick path to respond to signals. The core mm code
1387 * has unlocked the mm for us if we get here.
1388 */
1389 if (!user_mode(regs))
1390 kernelmode_fixup_or_oops(regs, error_code, address,
1391 SIGBUS, BUS_ADRERR,
1392 ARCH_DEFAULT_PKEY);
1393 return;
1394 }
1395
1396 /* The fault is fully completed (including releasing mmap lock) */
1397 if (fault & VM_FAULT_COMPLETED)
1398 return;
1399
1400 /*
1401 * If we need to retry the mmap_lock has already been released,
1402 * and if there is a fatal signal pending there is no guarantee
1403 * that we made any progress. Handle this case first.
1404 */
1405 if (unlikely(fault & VM_FAULT_RETRY)) {
1406 flags |= FAULT_FLAG_TRIED;
1407 goto retry;
1408 }
1409
1410 mmap_read_unlock(mm);
1411done:
1412 if (likely(!(fault & VM_FAULT_ERROR)))
1413 return;
1414
1415 if (fatal_signal_pending(current) && !user_mode(regs)) {
1416 kernelmode_fixup_or_oops(regs, error_code, address,
1417 0, 0, ARCH_DEFAULT_PKEY);
1418 return;
1419 }
1420
1421 if (fault & VM_FAULT_OOM) {
1422 /* Kernel mode? Handle exceptions or die: */
1423 if (!user_mode(regs)) {
1424 kernelmode_fixup_or_oops(regs, error_code, address,
1425 SIGSEGV, SEGV_MAPERR,
1426 ARCH_DEFAULT_PKEY);
1427 return;
1428 }
1429
1430 /*
1431 * We ran out of memory, call the OOM killer, and return the
1432 * userspace (which will retry the fault, or kill us if we got
1433 * oom-killed):
1434 */
1435 pagefault_out_of_memory();
1436 } else {
1437 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1438 VM_FAULT_HWPOISON_LARGE))
1439 do_sigbus(regs, error_code, address, fault);
1440 else if (fault & VM_FAULT_SIGSEGV)
1441 bad_area_nosemaphore(regs, error_code, address);
1442 else
1443 BUG();
1444 }
1445}
1446NOKPROBE_SYMBOL(do_user_addr_fault);
1447
1448static __always_inline void
1449trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1450 unsigned long address)
1451{
1452 if (!trace_pagefault_enabled())
1453 return;
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, error_code, address);
1473 } else {
1474 do_user_addr_fault(regs, error_code, address);
1475 /*
1476 * User address page fault handling might have reenabled
1477 * interrupts. Fixing up all potential exit points of
1478 * do_user_addr_fault() and its leaf functions is just not
1479 * doable w/o creating an unholy mess or turning the code
1480 * upside down.
1481 */
1482 local_irq_disable();
1483 }
1484}
1485
1486DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1487{
1488 irqentry_state_t state;
1489 unsigned long address;
1490
1491 address = cpu_feature_enabled(X86_FEATURE_FRED) ? fred_event_data(regs) : read_cr2();
1492
1493 prefetchw(¤t->mm->mmap_lock);
1494
1495 /*
1496 * KVM uses #PF vector to deliver 'page not present' events to guests
1497 * (asynchronous page fault mechanism). The event happens when a
1498 * userspace task is trying to access some valid (from guest's point of
1499 * view) memory which is not currently mapped by the host (e.g. the
1500 * memory is swapped out). Note, the corresponding "page ready" event
1501 * which is injected when the memory becomes available, is delivered via
1502 * an interrupt mechanism and not a #PF exception
1503 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1504 *
1505 * We are relying on the interrupted context being sane (valid RSP,
1506 * relevant locks not held, etc.), which is fine as long as the
1507 * interrupted context had IF=1. We are also relying on the KVM
1508 * async pf type field and CR2 being read consistently instead of
1509 * getting values from real and async page faults mixed up.
1510 *
1511 * Fingers crossed.
1512 *
1513 * The async #PF handling code takes care of idtentry handling
1514 * itself.
1515 */
1516 if (kvm_handle_async_pf(regs, (u32)address))
1517 return;
1518
1519 /*
1520 * Entry handling for valid #PF from kernel mode is slightly
1521 * different: RCU is already watching and ct_irq_enter() must not
1522 * be invoked because a kernel fault on a user space address might
1523 * sleep.
1524 *
1525 * In case the fault hit a RCU idle region the conditional entry
1526 * code reenabled RCU to avoid subsequent wreckage which helps
1527 * debuggability.
1528 */
1529 state = irqentry_enter(regs);
1530
1531 instrumentation_begin();
1532 handle_page_fault(regs, error_code, address);
1533 instrumentation_end();
1534
1535 irqentry_exit(regs, state);
1536}