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/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
38#define CREATE_TRACE_POINTS
39#include <asm/trace/exceptions.h>
40
41/*
42 * Returns 0 if mmiotrace is disabled, or if the fault is not
43 * handled by mmiotrace:
44 */
45static nokprobe_inline int
46kmmio_fault(struct pt_regs *regs, unsigned long addr)
47{
48 if (unlikely(is_kmmio_active()))
49 if (kmmio_handler(regs, addr) == 1)
50 return -1;
51 return 0;
52}
53
54/*
55 * Prefetch quirks:
56 *
57 * 32-bit mode:
58 *
59 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
60 * Check that here and ignore it. This is AMD erratum #91.
61 *
62 * 64-bit mode:
63 *
64 * Sometimes the CPU reports invalid exceptions on prefetch.
65 * Check that here and ignore it.
66 *
67 * Opcode checker based on code by Richard Brunner.
68 */
69static inline int
70check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
71 unsigned char opcode, int *prefetch)
72{
73 unsigned char instr_hi = opcode & 0xf0;
74 unsigned char instr_lo = opcode & 0x0f;
75
76 switch (instr_hi) {
77 case 0x20:
78 case 0x30:
79 /*
80 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
81 * In X86_64 long mode, the CPU will signal invalid
82 * opcode if some of these prefixes are present so
83 * X86_64 will never get here anyway
84 */
85 return ((instr_lo & 7) == 0x6);
86#ifdef CONFIG_X86_64
87 case 0x40:
88 /*
89 * In 64-bit mode 0x40..0x4F are valid REX prefixes
90 */
91 return (!user_mode(regs) || user_64bit_mode(regs));
92#endif
93 case 0x60:
94 /* 0x64 thru 0x67 are valid prefixes in all modes. */
95 return (instr_lo & 0xC) == 0x4;
96 case 0xF0:
97 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
98 return !instr_lo || (instr_lo>>1) == 1;
99 case 0x00:
100 /* Prefetch instruction is 0x0F0D or 0x0F18 */
101 if (get_kernel_nofault(opcode, instr))
102 return 0;
103
104 *prefetch = (instr_lo == 0xF) &&
105 (opcode == 0x0D || opcode == 0x18);
106 return 0;
107 default:
108 return 0;
109 }
110}
111
112static bool is_amd_k8_pre_npt(void)
113{
114 struct cpuinfo_x86 *c = &boot_cpu_data;
115
116 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
117 c->x86_vendor == X86_VENDOR_AMD &&
118 c->x86 == 0xf && c->x86_model < 0x40);
119}
120
121static int
122is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
123{
124 unsigned char *max_instr;
125 unsigned char *instr;
126 int prefetch = 0;
127
128 /* Erratum #91 affects AMD K8, pre-NPT CPUs */
129 if (!is_amd_k8_pre_npt())
130 return 0;
131
132 /*
133 * If it was a exec (instruction fetch) fault on NX page, then
134 * do not ignore the fault:
135 */
136 if (error_code & X86_PF_INSTR)
137 return 0;
138
139 instr = (void *)convert_ip_to_linear(current, regs);
140 max_instr = instr + 15;
141
142 /*
143 * This code has historically always bailed out if IP points to a
144 * not-present page (e.g. due to a race). No one has ever
145 * complained about this.
146 */
147 pagefault_disable();
148
149 while (instr < max_instr) {
150 unsigned char opcode;
151
152 if (user_mode(regs)) {
153 if (get_user(opcode, (unsigned char __user *) instr))
154 break;
155 } else {
156 if (get_kernel_nofault(opcode, instr))
157 break;
158 }
159
160 instr++;
161
162 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
163 break;
164 }
165
166 pagefault_enable();
167 return prefetch;
168}
169
170DEFINE_SPINLOCK(pgd_lock);
171LIST_HEAD(pgd_list);
172
173#ifdef CONFIG_X86_32
174static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
175{
176 unsigned index = pgd_index(address);
177 pgd_t *pgd_k;
178 p4d_t *p4d, *p4d_k;
179 pud_t *pud, *pud_k;
180 pmd_t *pmd, *pmd_k;
181
182 pgd += index;
183 pgd_k = init_mm.pgd + index;
184
185 if (!pgd_present(*pgd_k))
186 return NULL;
187
188 /*
189 * set_pgd(pgd, *pgd_k); here would be useless on PAE
190 * and redundant with the set_pmd() on non-PAE. As would
191 * set_p4d/set_pud.
192 */
193 p4d = p4d_offset(pgd, address);
194 p4d_k = p4d_offset(pgd_k, address);
195 if (!p4d_present(*p4d_k))
196 return NULL;
197
198 pud = pud_offset(p4d, address);
199 pud_k = pud_offset(p4d_k, address);
200 if (!pud_present(*pud_k))
201 return NULL;
202
203 pmd = pmd_offset(pud, address);
204 pmd_k = pmd_offset(pud_k, address);
205
206 if (pmd_present(*pmd) != pmd_present(*pmd_k))
207 set_pmd(pmd, *pmd_k);
208
209 if (!pmd_present(*pmd_k))
210 return NULL;
211 else
212 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
213
214 return pmd_k;
215}
216
217/*
218 * Handle a fault on the vmalloc or module mapping area
219 *
220 * This is needed because there is a race condition between the time
221 * when the vmalloc mapping code updates the PMD to the point in time
222 * where it synchronizes this update with the other page-tables in the
223 * system.
224 *
225 * In this race window another thread/CPU can map an area on the same
226 * PMD, finds it already present and does not synchronize it with the
227 * rest of the system yet. As a result v[mz]alloc might return areas
228 * which are not mapped in every page-table in the system, causing an
229 * unhandled page-fault when they are accessed.
230 */
231static noinline int vmalloc_fault(unsigned long address)
232{
233 unsigned long pgd_paddr;
234 pmd_t *pmd_k;
235 pte_t *pte_k;
236
237 /* Make sure we are in vmalloc area: */
238 if (!(address >= VMALLOC_START && address < VMALLOC_END))
239 return -1;
240
241 /*
242 * Synchronize this task's top level page-table
243 * with the 'reference' page table.
244 *
245 * Do _not_ use "current" here. We might be inside
246 * an interrupt in the middle of a task switch..
247 */
248 pgd_paddr = read_cr3_pa();
249 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
250 if (!pmd_k)
251 return -1;
252
253 if (pmd_large(*pmd_k))
254 return 0;
255
256 pte_k = pte_offset_kernel(pmd_k, address);
257 if (!pte_present(*pte_k))
258 return -1;
259
260 return 0;
261}
262NOKPROBE_SYMBOL(vmalloc_fault);
263
264void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
265{
266 unsigned long addr;
267
268 for (addr = start & PMD_MASK;
269 addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
270 addr += PMD_SIZE) {
271 struct page *page;
272
273 spin_lock(&pgd_lock);
274 list_for_each_entry(page, &pgd_list, lru) {
275 spinlock_t *pgt_lock;
276
277 /* the pgt_lock only for Xen */
278 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
279
280 spin_lock(pgt_lock);
281 vmalloc_sync_one(page_address(page), addr);
282 spin_unlock(pgt_lock);
283 }
284 spin_unlock(&pgd_lock);
285 }
286}
287
288static bool low_pfn(unsigned long pfn)
289{
290 return pfn < max_low_pfn;
291}
292
293static void dump_pagetable(unsigned long address)
294{
295 pgd_t *base = __va(read_cr3_pa());
296 pgd_t *pgd = &base[pgd_index(address)];
297 p4d_t *p4d;
298 pud_t *pud;
299 pmd_t *pmd;
300 pte_t *pte;
301
302#ifdef CONFIG_X86_PAE
303 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
304 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
305 goto out;
306#define pr_pde pr_cont
307#else
308#define pr_pde pr_info
309#endif
310 p4d = p4d_offset(pgd, address);
311 pud = pud_offset(p4d, address);
312 pmd = pmd_offset(pud, address);
313 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
314#undef pr_pde
315
316 /*
317 * We must not directly access the pte in the highpte
318 * case if the page table is located in highmem.
319 * And let's rather not kmap-atomic the pte, just in case
320 * it's allocated already:
321 */
322 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
323 goto out;
324
325 pte = pte_offset_kernel(pmd, address);
326 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
327out:
328 pr_cont("\n");
329}
330
331#else /* CONFIG_X86_64: */
332
333#ifdef CONFIG_CPU_SUP_AMD
334static const char errata93_warning[] =
335KERN_ERR
336"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
337"******* Working around it, but it may cause SEGVs or burn power.\n"
338"******* Please consider a BIOS update.\n"
339"******* Disabling USB legacy in the BIOS may also help.\n";
340#endif
341
342static int bad_address(void *p)
343{
344 unsigned long dummy;
345
346 return get_kernel_nofault(dummy, (unsigned long *)p);
347}
348
349static void dump_pagetable(unsigned long address)
350{
351 pgd_t *base = __va(read_cr3_pa());
352 pgd_t *pgd = base + pgd_index(address);
353 p4d_t *p4d;
354 pud_t *pud;
355 pmd_t *pmd;
356 pte_t *pte;
357
358 if (bad_address(pgd))
359 goto bad;
360
361 pr_info("PGD %lx ", pgd_val(*pgd));
362
363 if (!pgd_present(*pgd))
364 goto out;
365
366 p4d = p4d_offset(pgd, address);
367 if (bad_address(p4d))
368 goto bad;
369
370 pr_cont("P4D %lx ", p4d_val(*p4d));
371 if (!p4d_present(*p4d) || p4d_large(*p4d))
372 goto out;
373
374 pud = pud_offset(p4d, address);
375 if (bad_address(pud))
376 goto bad;
377
378 pr_cont("PUD %lx ", pud_val(*pud));
379 if (!pud_present(*pud) || pud_large(*pud))
380 goto out;
381
382 pmd = pmd_offset(pud, address);
383 if (bad_address(pmd))
384 goto bad;
385
386 pr_cont("PMD %lx ", pmd_val(*pmd));
387 if (!pmd_present(*pmd) || pmd_large(*pmd))
388 goto out;
389
390 pte = pte_offset_kernel(pmd, address);
391 if (bad_address(pte))
392 goto bad;
393
394 pr_cont("PTE %lx", pte_val(*pte));
395out:
396 pr_cont("\n");
397 return;
398bad:
399 pr_info("BAD\n");
400}
401
402#endif /* CONFIG_X86_64 */
403
404/*
405 * Workaround for K8 erratum #93 & buggy BIOS.
406 *
407 * BIOS SMM functions are required to use a specific workaround
408 * to avoid corruption of the 64bit RIP register on C stepping K8.
409 *
410 * A lot of BIOS that didn't get tested properly miss this.
411 *
412 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
413 * Try to work around it here.
414 *
415 * Note we only handle faults in kernel here.
416 * Does nothing on 32-bit.
417 */
418static int is_errata93(struct pt_regs *regs, unsigned long address)
419{
420#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
421 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
422 || boot_cpu_data.x86 != 0xf)
423 return 0;
424
425 if (user_mode(regs))
426 return 0;
427
428 if (address != regs->ip)
429 return 0;
430
431 if ((address >> 32) != 0)
432 return 0;
433
434 address |= 0xffffffffUL << 32;
435 if ((address >= (u64)_stext && address <= (u64)_etext) ||
436 (address >= MODULES_VADDR && address <= MODULES_END)) {
437 printk_once(errata93_warning);
438 regs->ip = address;
439 return 1;
440 }
441#endif
442 return 0;
443}
444
445/*
446 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
447 * to illegal addresses >4GB.
448 *
449 * We catch this in the page fault handler because these addresses
450 * are not reachable. Just detect this case and return. Any code
451 * segment in LDT is compatibility mode.
452 */
453static int is_errata100(struct pt_regs *regs, unsigned long address)
454{
455#ifdef CONFIG_X86_64
456 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
457 return 1;
458#endif
459 return 0;
460}
461
462/* Pentium F0 0F C7 C8 bug workaround: */
463static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
464 unsigned long address)
465{
466#ifdef CONFIG_X86_F00F_BUG
467 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
468 idt_is_f00f_address(address)) {
469 handle_invalid_op(regs);
470 return 1;
471 }
472#endif
473 return 0;
474}
475
476static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
477{
478 u32 offset = (index >> 3) * sizeof(struct desc_struct);
479 unsigned long addr;
480 struct ldttss_desc desc;
481
482 if (index == 0) {
483 pr_alert("%s: NULL\n", name);
484 return;
485 }
486
487 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
488 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
489 return;
490 }
491
492 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
493 sizeof(struct ldttss_desc))) {
494 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
495 name, index);
496 return;
497 }
498
499 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
500#ifdef CONFIG_X86_64
501 addr |= ((u64)desc.base3 << 32);
502#endif
503 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
504 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
505}
506
507static void
508show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
509{
510 if (!oops_may_print())
511 return;
512
513 if (error_code & X86_PF_INSTR) {
514 unsigned int level;
515 pgd_t *pgd;
516 pte_t *pte;
517
518 pgd = __va(read_cr3_pa());
519 pgd += pgd_index(address);
520
521 pte = lookup_address_in_pgd(pgd, address, &level);
522
523 if (pte && pte_present(*pte) && !pte_exec(*pte))
524 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
525 from_kuid(&init_user_ns, current_uid()));
526 if (pte && pte_present(*pte) && pte_exec(*pte) &&
527 (pgd_flags(*pgd) & _PAGE_USER) &&
528 (__read_cr4() & X86_CR4_SMEP))
529 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
530 from_kuid(&init_user_ns, current_uid()));
531 }
532
533 if (address < PAGE_SIZE && !user_mode(regs))
534 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
535 (void *)address);
536 else
537 pr_alert("BUG: unable to handle page fault for address: %px\n",
538 (void *)address);
539
540 pr_alert("#PF: %s %s in %s mode\n",
541 (error_code & X86_PF_USER) ? "user" : "supervisor",
542 (error_code & X86_PF_INSTR) ? "instruction fetch" :
543 (error_code & X86_PF_WRITE) ? "write access" :
544 "read access",
545 user_mode(regs) ? "user" : "kernel");
546 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
547 !(error_code & X86_PF_PROT) ? "not-present page" :
548 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
549 (error_code & X86_PF_PK) ? "protection keys violation" :
550 "permissions violation");
551
552 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
553 struct desc_ptr idt, gdt;
554 u16 ldtr, tr;
555
556 /*
557 * This can happen for quite a few reasons. The more obvious
558 * ones are faults accessing the GDT, or LDT. Perhaps
559 * surprisingly, if the CPU tries to deliver a benign or
560 * contributory exception from user code and gets a page fault
561 * during delivery, the page fault can be delivered as though
562 * it originated directly from user code. This could happen
563 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
564 * kernel or IST stack.
565 */
566 store_idt(&idt);
567
568 /* Usable even on Xen PV -- it's just slow. */
569 native_store_gdt(&gdt);
570
571 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
572 idt.address, idt.size, gdt.address, gdt.size);
573
574 store_ldt(ldtr);
575 show_ldttss(&gdt, "LDTR", ldtr);
576
577 store_tr(tr);
578 show_ldttss(&gdt, "TR", tr);
579 }
580
581 dump_pagetable(address);
582}
583
584static noinline void
585pgtable_bad(struct pt_regs *regs, unsigned long error_code,
586 unsigned long address)
587{
588 struct task_struct *tsk;
589 unsigned long flags;
590 int sig;
591
592 flags = oops_begin();
593 tsk = current;
594 sig = SIGKILL;
595
596 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
597 tsk->comm, address);
598 dump_pagetable(address);
599
600 if (__die("Bad pagetable", regs, error_code))
601 sig = 0;
602
603 oops_end(flags, regs, sig);
604}
605
606static void sanitize_error_code(unsigned long address,
607 unsigned long *error_code)
608{
609 /*
610 * To avoid leaking information about the kernel page
611 * table layout, pretend that user-mode accesses to
612 * kernel addresses are always protection faults.
613 *
614 * NB: This means that failed vsyscalls with vsyscall=none
615 * will have the PROT bit. This doesn't leak any
616 * information and does not appear to cause any problems.
617 */
618 if (address >= TASK_SIZE_MAX)
619 *error_code |= X86_PF_PROT;
620}
621
622static void set_signal_archinfo(unsigned long address,
623 unsigned long error_code)
624{
625 struct task_struct *tsk = current;
626
627 tsk->thread.trap_nr = X86_TRAP_PF;
628 tsk->thread.error_code = error_code | X86_PF_USER;
629 tsk->thread.cr2 = address;
630}
631
632static noinline void
633page_fault_oops(struct pt_regs *regs, unsigned long error_code,
634 unsigned long address)
635{
636#ifdef CONFIG_VMAP_STACK
637 struct stack_info info;
638#endif
639 unsigned long flags;
640 int sig;
641
642 if (user_mode(regs)) {
643 /*
644 * Implicit kernel access from user mode? Skip the stack
645 * overflow and EFI special cases.
646 */
647 goto oops;
648 }
649
650#ifdef CONFIG_VMAP_STACK
651 /*
652 * Stack overflow? During boot, we can fault near the initial
653 * stack in the direct map, but that's not an overflow -- check
654 * that we're in vmalloc space to avoid this.
655 */
656 if (is_vmalloc_addr((void *)address) &&
657 get_stack_guard_info((void *)address, &info)) {
658 /*
659 * We're likely to be running with very little stack space
660 * left. It's plausible that we'd hit this condition but
661 * double-fault even before we get this far, in which case
662 * we're fine: the double-fault handler will deal with it.
663 *
664 * We don't want to make it all the way into the oops code
665 * and then double-fault, though, because we're likely to
666 * break the console driver and lose most of the stack dump.
667 */
668 call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
669 handle_stack_overflow,
670 ASM_CALL_ARG3,
671 , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
672
673 unreachable();
674 }
675#endif
676
677 /*
678 * Buggy firmware could access regions which might page fault. If
679 * this happens, EFI has a special OOPS path that will try to
680 * avoid hanging the system.
681 */
682 if (IS_ENABLED(CONFIG_EFI))
683 efi_crash_gracefully_on_page_fault(address);
684
685 /* Only not-present faults should be handled by KFENCE. */
686 if (!(error_code & X86_PF_PROT) &&
687 kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
688 return;
689
690oops:
691 /*
692 * Oops. The kernel tried to access some bad page. We'll have to
693 * terminate things with extreme prejudice:
694 */
695 flags = oops_begin();
696
697 show_fault_oops(regs, error_code, address);
698
699 if (task_stack_end_corrupted(current))
700 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
701
702 sig = SIGKILL;
703 if (__die("Oops", regs, error_code))
704 sig = 0;
705
706 /* Executive summary in case the body of the oops scrolled away */
707 printk(KERN_DEFAULT "CR2: %016lx\n", address);
708
709 oops_end(flags, regs, sig);
710}
711
712static noinline void
713kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
714 unsigned long address, int signal, int si_code,
715 u32 pkey)
716{
717 WARN_ON_ONCE(user_mode(regs));
718
719 /* Are we prepared to handle this kernel fault? */
720 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
721 /*
722 * Any interrupt that takes a fault gets the fixup. This makes
723 * the below recursive fault logic only apply to a faults from
724 * task context.
725 */
726 if (in_interrupt())
727 return;
728
729 /*
730 * Per the above we're !in_interrupt(), aka. task context.
731 *
732 * In this case we need to make sure we're not recursively
733 * faulting through the emulate_vsyscall() logic.
734 */
735 if (current->thread.sig_on_uaccess_err && signal) {
736 sanitize_error_code(address, &error_code);
737
738 set_signal_archinfo(address, error_code);
739
740 if (si_code == SEGV_PKUERR) {
741 force_sig_pkuerr((void __user *)address, pkey);
742 } else {
743 /* XXX: hwpoison faults will set the wrong code. */
744 force_sig_fault(signal, si_code, (void __user *)address);
745 }
746 }
747
748 /*
749 * Barring that, we can do the fixup and be happy.
750 */
751 return;
752 }
753
754 /*
755 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
756 * instruction.
757 */
758 if (is_prefetch(regs, error_code, address))
759 return;
760
761 page_fault_oops(regs, error_code, address);
762}
763
764/*
765 * Print out info about fatal segfaults, if the show_unhandled_signals
766 * sysctl is set:
767 */
768static inline void
769show_signal_msg(struct pt_regs *regs, unsigned long error_code,
770 unsigned long address, struct task_struct *tsk)
771{
772 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
773 /* This is a racy snapshot, but it's better than nothing. */
774 int cpu = raw_smp_processor_id();
775
776 if (!unhandled_signal(tsk, SIGSEGV))
777 return;
778
779 if (!printk_ratelimit())
780 return;
781
782 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
783 loglvl, tsk->comm, task_pid_nr(tsk), address,
784 (void *)regs->ip, (void *)regs->sp, error_code);
785
786 print_vma_addr(KERN_CONT " in ", regs->ip);
787
788 /*
789 * Dump the likely CPU where the fatal segfault happened.
790 * This can help identify faulty hardware.
791 */
792 printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu,
793 topology_core_id(cpu), topology_physical_package_id(cpu));
794
795
796 printk(KERN_CONT "\n");
797
798 show_opcodes(regs, loglvl);
799}
800
801static void
802__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
803 unsigned long address, u32 pkey, int si_code)
804{
805 struct task_struct *tsk = current;
806
807 if (!user_mode(regs)) {
808 kernelmode_fixup_or_oops(regs, error_code, address,
809 SIGSEGV, si_code, pkey);
810 return;
811 }
812
813 if (!(error_code & X86_PF_USER)) {
814 /* Implicit user access to kernel memory -- just oops */
815 page_fault_oops(regs, error_code, address);
816 return;
817 }
818
819 /*
820 * User mode accesses just cause a SIGSEGV.
821 * It's possible to have interrupts off here:
822 */
823 local_irq_enable();
824
825 /*
826 * Valid to do another page fault here because this one came
827 * from user space:
828 */
829 if (is_prefetch(regs, error_code, address))
830 return;
831
832 if (is_errata100(regs, address))
833 return;
834
835 sanitize_error_code(address, &error_code);
836
837 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
838 return;
839
840 if (likely(show_unhandled_signals))
841 show_signal_msg(regs, error_code, address, tsk);
842
843 set_signal_archinfo(address, error_code);
844
845 if (si_code == SEGV_PKUERR)
846 force_sig_pkuerr((void __user *)address, pkey);
847 else
848 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
849
850 local_irq_disable();
851}
852
853static noinline void
854bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
855 unsigned long address)
856{
857 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
858}
859
860static void
861__bad_area(struct pt_regs *regs, unsigned long error_code,
862 unsigned long address, u32 pkey, int si_code)
863{
864 struct mm_struct *mm = current->mm;
865 /*
866 * Something tried to access memory that isn't in our memory map..
867 * Fix it, but check if it's kernel or user first..
868 */
869 mmap_read_unlock(mm);
870
871 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
872}
873
874static inline bool bad_area_access_from_pkeys(unsigned long error_code,
875 struct vm_area_struct *vma)
876{
877 /* This code is always called on the current mm */
878 bool foreign = false;
879
880 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
881 return false;
882 if (error_code & X86_PF_PK)
883 return true;
884 /* this checks permission keys on the VMA: */
885 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
886 (error_code & X86_PF_INSTR), foreign))
887 return true;
888 return false;
889}
890
891static noinline void
892bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
893 unsigned long address, struct vm_area_struct *vma)
894{
895 /*
896 * This OSPKE check is not strictly necessary at runtime.
897 * But, doing it this way allows compiler optimizations
898 * if pkeys are compiled out.
899 */
900 if (bad_area_access_from_pkeys(error_code, vma)) {
901 /*
902 * A protection key fault means that the PKRU value did not allow
903 * access to some PTE. Userspace can figure out what PKRU was
904 * from the XSAVE state. This function captures the pkey from
905 * the vma and passes it to userspace so userspace can discover
906 * which protection key was set on the PTE.
907 *
908 * If we get here, we know that the hardware signaled a X86_PF_PK
909 * fault and that there was a VMA once we got in the fault
910 * handler. It does *not* guarantee that the VMA we find here
911 * was the one that we faulted on.
912 *
913 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
914 * 2. T1 : set PKRU to deny access to pkey=4, touches page
915 * 3. T1 : faults...
916 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
917 * 5. T1 : enters fault handler, takes mmap_lock, etc...
918 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
919 * faulted on a pte with its pkey=4.
920 */
921 u32 pkey = vma_pkey(vma);
922
923 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
924 } else {
925 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
926 }
927}
928
929static void
930do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
931 vm_fault_t fault)
932{
933 /* Kernel mode? Handle exceptions or die: */
934 if (!user_mode(regs)) {
935 kernelmode_fixup_or_oops(regs, error_code, address,
936 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
937 return;
938 }
939
940 /* User-space => ok to do another page fault: */
941 if (is_prefetch(regs, error_code, address))
942 return;
943
944 sanitize_error_code(address, &error_code);
945
946 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
947 return;
948
949 set_signal_archinfo(address, error_code);
950
951#ifdef CONFIG_MEMORY_FAILURE
952 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
953 struct task_struct *tsk = current;
954 unsigned lsb = 0;
955
956 pr_err(
957 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
958 tsk->comm, tsk->pid, address);
959 if (fault & VM_FAULT_HWPOISON_LARGE)
960 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
961 if (fault & VM_FAULT_HWPOISON)
962 lsb = PAGE_SHIFT;
963 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
964 return;
965 }
966#endif
967 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
968}
969
970static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
971{
972 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
973 return 0;
974
975 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
976 return 0;
977
978 return 1;
979}
980
981/*
982 * Handle a spurious fault caused by a stale TLB entry.
983 *
984 * This allows us to lazily refresh the TLB when increasing the
985 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
986 * eagerly is very expensive since that implies doing a full
987 * cross-processor TLB flush, even if no stale TLB entries exist
988 * on other processors.
989 *
990 * Spurious faults may only occur if the TLB contains an entry with
991 * fewer permission than the page table entry. Non-present (P = 0)
992 * and reserved bit (R = 1) faults are never spurious.
993 *
994 * There are no security implications to leaving a stale TLB when
995 * increasing the permissions on a page.
996 *
997 * Returns non-zero if a spurious fault was handled, zero otherwise.
998 *
999 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1000 * (Optional Invalidation).
1001 */
1002static noinline int
1003spurious_kernel_fault(unsigned long error_code, unsigned long address)
1004{
1005 pgd_t *pgd;
1006 p4d_t *p4d;
1007 pud_t *pud;
1008 pmd_t *pmd;
1009 pte_t *pte;
1010 int ret;
1011
1012 /*
1013 * Only writes to RO or instruction fetches from NX may cause
1014 * spurious faults.
1015 *
1016 * These could be from user or supervisor accesses but the TLB
1017 * is only lazily flushed after a kernel mapping protection
1018 * change, so user accesses are not expected to cause spurious
1019 * faults.
1020 */
1021 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1022 error_code != (X86_PF_INSTR | X86_PF_PROT))
1023 return 0;
1024
1025 pgd = init_mm.pgd + pgd_index(address);
1026 if (!pgd_present(*pgd))
1027 return 0;
1028
1029 p4d = p4d_offset(pgd, address);
1030 if (!p4d_present(*p4d))
1031 return 0;
1032
1033 if (p4d_large(*p4d))
1034 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1035
1036 pud = pud_offset(p4d, address);
1037 if (!pud_present(*pud))
1038 return 0;
1039
1040 if (pud_large(*pud))
1041 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1042
1043 pmd = pmd_offset(pud, address);
1044 if (!pmd_present(*pmd))
1045 return 0;
1046
1047 if (pmd_large(*pmd))
1048 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1049
1050 pte = pte_offset_kernel(pmd, address);
1051 if (!pte_present(*pte))
1052 return 0;
1053
1054 ret = spurious_kernel_fault_check(error_code, pte);
1055 if (!ret)
1056 return 0;
1057
1058 /*
1059 * Make sure we have permissions in PMD.
1060 * If not, then there's a bug in the page tables:
1061 */
1062 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1063 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1064
1065 return ret;
1066}
1067NOKPROBE_SYMBOL(spurious_kernel_fault);
1068
1069int show_unhandled_signals = 1;
1070
1071static inline int
1072access_error(unsigned long error_code, struct vm_area_struct *vma)
1073{
1074 /* This is only called for the current mm, so: */
1075 bool foreign = false;
1076
1077 /*
1078 * Read or write was blocked by protection keys. This is
1079 * always an unconditional error and can never result in
1080 * a follow-up action to resolve the fault, like a COW.
1081 */
1082 if (error_code & X86_PF_PK)
1083 return 1;
1084
1085 /*
1086 * SGX hardware blocked the access. This usually happens
1087 * when the enclave memory contents have been destroyed, like
1088 * after a suspend/resume cycle. In any case, the kernel can't
1089 * fix the cause of the fault. Handle the fault as an access
1090 * error even in cases where no actual access violation
1091 * occurred. This allows userspace to rebuild the enclave in
1092 * response to the signal.
1093 */
1094 if (unlikely(error_code & X86_PF_SGX))
1095 return 1;
1096
1097 /*
1098 * Make sure to check the VMA so that we do not perform
1099 * faults just to hit a X86_PF_PK as soon as we fill in a
1100 * page.
1101 */
1102 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1103 (error_code & X86_PF_INSTR), foreign))
1104 return 1;
1105
1106 /*
1107 * Shadow stack accesses (PF_SHSTK=1) are only permitted to
1108 * shadow stack VMAs. All other accesses result in an error.
1109 */
1110 if (error_code & X86_PF_SHSTK) {
1111 if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK)))
1112 return 1;
1113 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1114 return 1;
1115 return 0;
1116 }
1117
1118 if (error_code & X86_PF_WRITE) {
1119 /* write, present and write, not present: */
1120 if (unlikely(vma->vm_flags & VM_SHADOW_STACK))
1121 return 1;
1122 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1123 return 1;
1124 return 0;
1125 }
1126
1127 /* read, present: */
1128 if (unlikely(error_code & X86_PF_PROT))
1129 return 1;
1130
1131 /* read, not present: */
1132 if (unlikely(!vma_is_accessible(vma)))
1133 return 1;
1134
1135 return 0;
1136}
1137
1138bool fault_in_kernel_space(unsigned long address)
1139{
1140 /*
1141 * On 64-bit systems, the vsyscall page is at an address above
1142 * TASK_SIZE_MAX, but is not considered part of the kernel
1143 * address space.
1144 */
1145 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1146 return false;
1147
1148 return address >= TASK_SIZE_MAX;
1149}
1150
1151/*
1152 * Called for all faults where 'address' is part of the kernel address
1153 * space. Might get called for faults that originate from *code* that
1154 * ran in userspace or the kernel.
1155 */
1156static void
1157do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1158 unsigned long address)
1159{
1160 /*
1161 * Protection keys exceptions only happen on user pages. We
1162 * have no user pages in the kernel portion of the address
1163 * space, so do not expect them here.
1164 */
1165 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1166
1167#ifdef CONFIG_X86_32
1168 /*
1169 * We can fault-in kernel-space virtual memory on-demand. The
1170 * 'reference' page table is init_mm.pgd.
1171 *
1172 * NOTE! We MUST NOT take any locks for this case. We may
1173 * be in an interrupt or a critical region, and should
1174 * only copy the information from the master page table,
1175 * nothing more.
1176 *
1177 * Before doing this on-demand faulting, ensure that the
1178 * fault is not any of the following:
1179 * 1. A fault on a PTE with a reserved bit set.
1180 * 2. A fault caused by a user-mode access. (Do not demand-
1181 * fault kernel memory due to user-mode accesses).
1182 * 3. A fault caused by a page-level protection violation.
1183 * (A demand fault would be on a non-present page which
1184 * would have X86_PF_PROT==0).
1185 *
1186 * This is only needed to close a race condition on x86-32 in
1187 * the vmalloc mapping/unmapping code. See the comment above
1188 * vmalloc_fault() for details. On x86-64 the race does not
1189 * exist as the vmalloc mappings don't need to be synchronized
1190 * there.
1191 */
1192 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1193 if (vmalloc_fault(address) >= 0)
1194 return;
1195 }
1196#endif
1197
1198 if (is_f00f_bug(regs, hw_error_code, address))
1199 return;
1200
1201 /* Was the fault spurious, caused by lazy TLB invalidation? */
1202 if (spurious_kernel_fault(hw_error_code, address))
1203 return;
1204
1205 /* kprobes don't want to hook the spurious faults: */
1206 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1207 return;
1208
1209 /*
1210 * Note, despite being a "bad area", there are quite a few
1211 * acceptable reasons to get here, such as erratum fixups
1212 * and handling kernel code that can fault, like get_user().
1213 *
1214 * Don't take the mm semaphore here. If we fixup a prefetch
1215 * fault we could otherwise deadlock:
1216 */
1217 bad_area_nosemaphore(regs, hw_error_code, address);
1218}
1219NOKPROBE_SYMBOL(do_kern_addr_fault);
1220
1221/*
1222 * Handle faults in the user portion of the address space. Nothing in here
1223 * should check X86_PF_USER without a specific justification: for almost
1224 * all purposes, we should treat a normal kernel access to user memory
1225 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1226 * The one exception is AC flag handling, which is, per the x86
1227 * architecture, special for WRUSS.
1228 */
1229static inline
1230void do_user_addr_fault(struct pt_regs *regs,
1231 unsigned long error_code,
1232 unsigned long address)
1233{
1234 struct vm_area_struct *vma;
1235 struct task_struct *tsk;
1236 struct mm_struct *mm;
1237 vm_fault_t fault;
1238 unsigned int flags = FAULT_FLAG_DEFAULT;
1239
1240 tsk = current;
1241 mm = tsk->mm;
1242
1243 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1244 /*
1245 * Whoops, this is kernel mode code trying to execute from
1246 * user memory. Unless this is AMD erratum #93, which
1247 * corrupts RIP such that it looks like a user address,
1248 * this is unrecoverable. Don't even try to look up the
1249 * VMA or look for extable entries.
1250 */
1251 if (is_errata93(regs, address))
1252 return;
1253
1254 page_fault_oops(regs, error_code, address);
1255 return;
1256 }
1257
1258 /* kprobes don't want to hook the spurious faults: */
1259 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1260 return;
1261
1262 /*
1263 * Reserved bits are never expected to be set on
1264 * entries in the user portion of the page tables.
1265 */
1266 if (unlikely(error_code & X86_PF_RSVD))
1267 pgtable_bad(regs, error_code, address);
1268
1269 /*
1270 * If SMAP is on, check for invalid kernel (supervisor) access to user
1271 * pages in the user address space. The odd case here is WRUSS,
1272 * which, according to the preliminary documentation, does not respect
1273 * SMAP and will have the USER bit set so, in all cases, SMAP
1274 * enforcement appears to be consistent with the USER bit.
1275 */
1276 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1277 !(error_code & X86_PF_USER) &&
1278 !(regs->flags & X86_EFLAGS_AC))) {
1279 /*
1280 * No extable entry here. This was a kernel access to an
1281 * invalid pointer. get_kernel_nofault() will not get here.
1282 */
1283 page_fault_oops(regs, error_code, address);
1284 return;
1285 }
1286
1287 /*
1288 * If we're in an interrupt, have no user context or are running
1289 * in a region with pagefaults disabled then we must not take the fault
1290 */
1291 if (unlikely(faulthandler_disabled() || !mm)) {
1292 bad_area_nosemaphore(regs, error_code, address);
1293 return;
1294 }
1295
1296 /*
1297 * It's safe to allow irq's after cr2 has been saved and the
1298 * vmalloc fault has been handled.
1299 *
1300 * User-mode registers count as a user access even for any
1301 * potential system fault or CPU buglet:
1302 */
1303 if (user_mode(regs)) {
1304 local_irq_enable();
1305 flags |= FAULT_FLAG_USER;
1306 } else {
1307 if (regs->flags & X86_EFLAGS_IF)
1308 local_irq_enable();
1309 }
1310
1311 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1312
1313 /*
1314 * Read-only permissions can not be expressed in shadow stack PTEs.
1315 * Treat all shadow stack accesses as WRITE faults. This ensures
1316 * that the MM will prepare everything (e.g., break COW) such that
1317 * maybe_mkwrite() can create a proper shadow stack PTE.
1318 */
1319 if (error_code & X86_PF_SHSTK)
1320 flags |= FAULT_FLAG_WRITE;
1321 if (error_code & X86_PF_WRITE)
1322 flags |= FAULT_FLAG_WRITE;
1323 if (error_code & X86_PF_INSTR)
1324 flags |= FAULT_FLAG_INSTRUCTION;
1325
1326#ifdef CONFIG_X86_64
1327 /*
1328 * Faults in the vsyscall page might need emulation. The
1329 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1330 * considered to be part of the user address space.
1331 *
1332 * The vsyscall page does not have a "real" VMA, so do this
1333 * emulation before we go searching for VMAs.
1334 *
1335 * PKRU never rejects instruction fetches, so we don't need
1336 * to consider the PF_PK bit.
1337 */
1338 if (is_vsyscall_vaddr(address)) {
1339 if (emulate_vsyscall(error_code, regs, address))
1340 return;
1341 }
1342#endif
1343
1344 if (!(flags & FAULT_FLAG_USER))
1345 goto lock_mmap;
1346
1347 vma = lock_vma_under_rcu(mm, address);
1348 if (!vma)
1349 goto lock_mmap;
1350
1351 if (unlikely(access_error(error_code, vma))) {
1352 vma_end_read(vma);
1353 goto lock_mmap;
1354 }
1355 fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
1356 if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
1357 vma_end_read(vma);
1358
1359 if (!(fault & VM_FAULT_RETRY)) {
1360 count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1361 goto done;
1362 }
1363 count_vm_vma_lock_event(VMA_LOCK_RETRY);
1364 if (fault & VM_FAULT_MAJOR)
1365 flags |= FAULT_FLAG_TRIED;
1366
1367 /* Quick path to respond to signals */
1368 if (fault_signal_pending(fault, regs)) {
1369 if (!user_mode(regs))
1370 kernelmode_fixup_or_oops(regs, error_code, address,
1371 SIGBUS, BUS_ADRERR,
1372 ARCH_DEFAULT_PKEY);
1373 return;
1374 }
1375lock_mmap:
1376
1377retry:
1378 vma = lock_mm_and_find_vma(mm, address, regs);
1379 if (unlikely(!vma)) {
1380 bad_area_nosemaphore(regs, error_code, address);
1381 return;
1382 }
1383
1384 /*
1385 * Ok, we have a good vm_area for this memory access, so
1386 * we can handle it..
1387 */
1388 if (unlikely(access_error(error_code, vma))) {
1389 bad_area_access_error(regs, error_code, address, vma);
1390 return;
1391 }
1392
1393 /*
1394 * If for any reason at all we couldn't handle the fault,
1395 * make sure we exit gracefully rather than endlessly redo
1396 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1397 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1398 *
1399 * Note that handle_userfault() may also release and reacquire mmap_lock
1400 * (and not return with VM_FAULT_RETRY), when returning to userland to
1401 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1402 * (potentially after handling any pending signal during the return to
1403 * userland). The return to userland is identified whenever
1404 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1405 */
1406 fault = handle_mm_fault(vma, address, flags, regs);
1407
1408 if (fault_signal_pending(fault, regs)) {
1409 /*
1410 * Quick path to respond to signals. The core mm code
1411 * has unlocked the mm for us if we get here.
1412 */
1413 if (!user_mode(regs))
1414 kernelmode_fixup_or_oops(regs, error_code, address,
1415 SIGBUS, BUS_ADRERR,
1416 ARCH_DEFAULT_PKEY);
1417 return;
1418 }
1419
1420 /* The fault is fully completed (including releasing mmap lock) */
1421 if (fault & VM_FAULT_COMPLETED)
1422 return;
1423
1424 /*
1425 * If we need to retry the mmap_lock has already been released,
1426 * and if there is a fatal signal pending there is no guarantee
1427 * that we made any progress. Handle this case first.
1428 */
1429 if (unlikely(fault & VM_FAULT_RETRY)) {
1430 flags |= FAULT_FLAG_TRIED;
1431 goto retry;
1432 }
1433
1434 mmap_read_unlock(mm);
1435done:
1436 if (likely(!(fault & VM_FAULT_ERROR)))
1437 return;
1438
1439 if (fatal_signal_pending(current) && !user_mode(regs)) {
1440 kernelmode_fixup_or_oops(regs, error_code, address,
1441 0, 0, ARCH_DEFAULT_PKEY);
1442 return;
1443 }
1444
1445 if (fault & VM_FAULT_OOM) {
1446 /* Kernel mode? Handle exceptions or die: */
1447 if (!user_mode(regs)) {
1448 kernelmode_fixup_or_oops(regs, error_code, address,
1449 SIGSEGV, SEGV_MAPERR,
1450 ARCH_DEFAULT_PKEY);
1451 return;
1452 }
1453
1454 /*
1455 * We ran out of memory, call the OOM killer, and return the
1456 * userspace (which will retry the fault, or kill us if we got
1457 * oom-killed):
1458 */
1459 pagefault_out_of_memory();
1460 } else {
1461 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1462 VM_FAULT_HWPOISON_LARGE))
1463 do_sigbus(regs, error_code, address, fault);
1464 else if (fault & VM_FAULT_SIGSEGV)
1465 bad_area_nosemaphore(regs, error_code, address);
1466 else
1467 BUG();
1468 }
1469}
1470NOKPROBE_SYMBOL(do_user_addr_fault);
1471
1472static __always_inline void
1473trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1474 unsigned long address)
1475{
1476 if (!trace_pagefault_enabled())
1477 return;
1478
1479 if (user_mode(regs))
1480 trace_page_fault_user(address, regs, error_code);
1481 else
1482 trace_page_fault_kernel(address, regs, error_code);
1483}
1484
1485static __always_inline void
1486handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1487 unsigned long address)
1488{
1489 trace_page_fault_entries(regs, error_code, address);
1490
1491 if (unlikely(kmmio_fault(regs, address)))
1492 return;
1493
1494 /* Was the fault on kernel-controlled part of the address space? */
1495 if (unlikely(fault_in_kernel_space(address))) {
1496 do_kern_addr_fault(regs, error_code, address);
1497 } else {
1498 do_user_addr_fault(regs, error_code, address);
1499 /*
1500 * User address page fault handling might have reenabled
1501 * interrupts. Fixing up all potential exit points of
1502 * do_user_addr_fault() and its leaf functions is just not
1503 * doable w/o creating an unholy mess or turning the code
1504 * upside down.
1505 */
1506 local_irq_disable();
1507 }
1508}
1509
1510DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1511{
1512 unsigned long address = read_cr2();
1513 irqentry_state_t state;
1514
1515 prefetchw(¤t->mm->mmap_lock);
1516
1517 /*
1518 * KVM uses #PF vector to deliver 'page not present' events to guests
1519 * (asynchronous page fault mechanism). The event happens when a
1520 * userspace task is trying to access some valid (from guest's point of
1521 * view) memory which is not currently mapped by the host (e.g. the
1522 * memory is swapped out). Note, the corresponding "page ready" event
1523 * which is injected when the memory becomes available, is delivered via
1524 * an interrupt mechanism and not a #PF exception
1525 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1526 *
1527 * We are relying on the interrupted context being sane (valid RSP,
1528 * relevant locks not held, etc.), which is fine as long as the
1529 * interrupted context had IF=1. We are also relying on the KVM
1530 * async pf type field and CR2 being read consistently instead of
1531 * getting values from real and async page faults mixed up.
1532 *
1533 * Fingers crossed.
1534 *
1535 * The async #PF handling code takes care of idtentry handling
1536 * itself.
1537 */
1538 if (kvm_handle_async_pf(regs, (u32)address))
1539 return;
1540
1541 /*
1542 * Entry handling for valid #PF from kernel mode is slightly
1543 * different: RCU is already watching and ct_irq_enter() must not
1544 * be invoked because a kernel fault on a user space address might
1545 * sleep.
1546 *
1547 * In case the fault hit a RCU idle region the conditional entry
1548 * code reenabled RCU to avoid subsequent wreckage which helps
1549 * debuggability.
1550 */
1551 state = irqentry_enter(regs);
1552
1553 instrumentation_begin();
1554 handle_page_fault(regs, error_code, address);
1555 instrumentation_end();
1556
1557 irqentry_exit(regs, state);
1558}
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/bootmem.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
20#include <asm/cpufeature.h> /* boot_cpu_has, ... */
21#include <asm/traps.h> /* dotraplinkage, ... */
22#include <asm/pgalloc.h> /* pgd_*(), ... */
23#include <asm/fixmap.h> /* VSYSCALL_ADDR */
24#include <asm/vsyscall.h> /* emulate_vsyscall */
25#include <asm/vm86.h> /* struct vm86 */
26#include <asm/mmu_context.h> /* vma_pkey() */
27
28#define CREATE_TRACE_POINTS
29#include <asm/trace/exceptions.h>
30
31/*
32 * Returns 0 if mmiotrace is disabled, or if the fault is not
33 * handled by mmiotrace:
34 */
35static nokprobe_inline int
36kmmio_fault(struct pt_regs *regs, unsigned long addr)
37{
38 if (unlikely(is_kmmio_active()))
39 if (kmmio_handler(regs, addr) == 1)
40 return -1;
41 return 0;
42}
43
44static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
45{
46 int ret = 0;
47
48 /* kprobe_running() needs smp_processor_id() */
49 if (kprobes_built_in() && !user_mode(regs)) {
50 preempt_disable();
51 if (kprobe_running() && kprobe_fault_handler(regs, 14))
52 ret = 1;
53 preempt_enable();
54 }
55
56 return ret;
57}
58
59/*
60 * Prefetch quirks:
61 *
62 * 32-bit mode:
63 *
64 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
65 * Check that here and ignore it.
66 *
67 * 64-bit mode:
68 *
69 * Sometimes the CPU reports invalid exceptions on prefetch.
70 * Check that here and ignore it.
71 *
72 * Opcode checker based on code by Richard Brunner.
73 */
74static inline int
75check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
76 unsigned char opcode, int *prefetch)
77{
78 unsigned char instr_hi = opcode & 0xf0;
79 unsigned char instr_lo = opcode & 0x0f;
80
81 switch (instr_hi) {
82 case 0x20:
83 case 0x30:
84 /*
85 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
86 * In X86_64 long mode, the CPU will signal invalid
87 * opcode if some of these prefixes are present so
88 * X86_64 will never get here anyway
89 */
90 return ((instr_lo & 7) == 0x6);
91#ifdef CONFIG_X86_64
92 case 0x40:
93 /*
94 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
95 * Need to figure out under what instruction mode the
96 * instruction was issued. Could check the LDT for lm,
97 * but for now it's good enough to assume that long
98 * mode only uses well known segments or kernel.
99 */
100 return (!user_mode(regs) || user_64bit_mode(regs));
101#endif
102 case 0x60:
103 /* 0x64 thru 0x67 are valid prefixes in all modes. */
104 return (instr_lo & 0xC) == 0x4;
105 case 0xF0:
106 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
107 return !instr_lo || (instr_lo>>1) == 1;
108 case 0x00:
109 /* Prefetch instruction is 0x0F0D or 0x0F18 */
110 if (probe_kernel_address(instr, opcode))
111 return 0;
112
113 *prefetch = (instr_lo == 0xF) &&
114 (opcode == 0x0D || opcode == 0x18);
115 return 0;
116 default:
117 return 0;
118 }
119}
120
121static int
122is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
123{
124 unsigned char *max_instr;
125 unsigned char *instr;
126 int prefetch = 0;
127
128 /*
129 * If it was a exec (instruction fetch) fault on NX page, then
130 * do not ignore the fault:
131 */
132 if (error_code & X86_PF_INSTR)
133 return 0;
134
135 instr = (void *)convert_ip_to_linear(current, regs);
136 max_instr = instr + 15;
137
138 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
139 return 0;
140
141 while (instr < max_instr) {
142 unsigned char opcode;
143
144 if (probe_kernel_address(instr, opcode))
145 break;
146
147 instr++;
148
149 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
150 break;
151 }
152 return prefetch;
153}
154
155/*
156 * A protection key fault means that the PKRU value did not allow
157 * access to some PTE. Userspace can figure out what PKRU was
158 * from the XSAVE state, and this function fills out a field in
159 * siginfo so userspace can discover which protection key was set
160 * on the PTE.
161 *
162 * If we get here, we know that the hardware signaled a X86_PF_PK
163 * fault and that there was a VMA once we got in the fault
164 * handler. It does *not* guarantee that the VMA we find here
165 * was the one that we faulted on.
166 *
167 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
168 * 2. T1 : set PKRU to deny access to pkey=4, touches page
169 * 3. T1 : faults...
170 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
171 * 5. T1 : enters fault handler, takes mmap_sem, etc...
172 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
173 * faulted on a pte with its pkey=4.
174 */
175static void fill_sig_info_pkey(int si_signo, int si_code, siginfo_t *info,
176 u32 *pkey)
177{
178 /* This is effectively an #ifdef */
179 if (!boot_cpu_has(X86_FEATURE_OSPKE))
180 return;
181
182 /* Fault not from Protection Keys: nothing to do */
183 if ((si_code != SEGV_PKUERR) || (si_signo != SIGSEGV))
184 return;
185 /*
186 * force_sig_info_fault() is called from a number of
187 * contexts, some of which have a VMA and some of which
188 * do not. The X86_PF_PK handing happens after we have a
189 * valid VMA, so we should never reach this without a
190 * valid VMA.
191 */
192 if (!pkey) {
193 WARN_ONCE(1, "PKU fault with no VMA passed in");
194 info->si_pkey = 0;
195 return;
196 }
197 /*
198 * si_pkey should be thought of as a strong hint, but not
199 * absolutely guranteed to be 100% accurate because of
200 * the race explained above.
201 */
202 info->si_pkey = *pkey;
203}
204
205static void
206force_sig_info_fault(int si_signo, int si_code, unsigned long address,
207 struct task_struct *tsk, u32 *pkey, int fault)
208{
209 unsigned lsb = 0;
210 siginfo_t info;
211
212 info.si_signo = si_signo;
213 info.si_errno = 0;
214 info.si_code = si_code;
215 info.si_addr = (void __user *)address;
216 if (fault & VM_FAULT_HWPOISON_LARGE)
217 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
218 if (fault & VM_FAULT_HWPOISON)
219 lsb = PAGE_SHIFT;
220 info.si_addr_lsb = lsb;
221
222 fill_sig_info_pkey(si_signo, si_code, &info, pkey);
223
224 force_sig_info(si_signo, &info, tsk);
225}
226
227DEFINE_SPINLOCK(pgd_lock);
228LIST_HEAD(pgd_list);
229
230#ifdef CONFIG_X86_32
231static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
232{
233 unsigned index = pgd_index(address);
234 pgd_t *pgd_k;
235 p4d_t *p4d, *p4d_k;
236 pud_t *pud, *pud_k;
237 pmd_t *pmd, *pmd_k;
238
239 pgd += index;
240 pgd_k = init_mm.pgd + index;
241
242 if (!pgd_present(*pgd_k))
243 return NULL;
244
245 /*
246 * set_pgd(pgd, *pgd_k); here would be useless on PAE
247 * and redundant with the set_pmd() on non-PAE. As would
248 * set_p4d/set_pud.
249 */
250 p4d = p4d_offset(pgd, address);
251 p4d_k = p4d_offset(pgd_k, address);
252 if (!p4d_present(*p4d_k))
253 return NULL;
254
255 pud = pud_offset(p4d, address);
256 pud_k = pud_offset(p4d_k, address);
257 if (!pud_present(*pud_k))
258 return NULL;
259
260 pmd = pmd_offset(pud, address);
261 pmd_k = pmd_offset(pud_k, address);
262 if (!pmd_present(*pmd_k))
263 return NULL;
264
265 if (!pmd_present(*pmd))
266 set_pmd(pmd, *pmd_k);
267 else
268 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
269
270 return pmd_k;
271}
272
273void vmalloc_sync_all(void)
274{
275 unsigned long address;
276
277 if (SHARED_KERNEL_PMD)
278 return;
279
280 for (address = VMALLOC_START & PMD_MASK;
281 address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
282 address += PMD_SIZE) {
283 struct page *page;
284
285 spin_lock(&pgd_lock);
286 list_for_each_entry(page, &pgd_list, lru) {
287 spinlock_t *pgt_lock;
288 pmd_t *ret;
289
290 /* the pgt_lock only for Xen */
291 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
292
293 spin_lock(pgt_lock);
294 ret = vmalloc_sync_one(page_address(page), address);
295 spin_unlock(pgt_lock);
296
297 if (!ret)
298 break;
299 }
300 spin_unlock(&pgd_lock);
301 }
302}
303
304/*
305 * 32-bit:
306 *
307 * Handle a fault on the vmalloc or module mapping area
308 */
309static noinline int vmalloc_fault(unsigned long address)
310{
311 unsigned long pgd_paddr;
312 pmd_t *pmd_k;
313 pte_t *pte_k;
314
315 /* Make sure we are in vmalloc area: */
316 if (!(address >= VMALLOC_START && address < VMALLOC_END))
317 return -1;
318
319 WARN_ON_ONCE(in_nmi());
320
321 /*
322 * Synchronize this task's top level page-table
323 * with the 'reference' page table.
324 *
325 * Do _not_ use "current" here. We might be inside
326 * an interrupt in the middle of a task switch..
327 */
328 pgd_paddr = read_cr3_pa();
329 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
330 if (!pmd_k)
331 return -1;
332
333 if (pmd_large(*pmd_k))
334 return 0;
335
336 pte_k = pte_offset_kernel(pmd_k, address);
337 if (!pte_present(*pte_k))
338 return -1;
339
340 return 0;
341}
342NOKPROBE_SYMBOL(vmalloc_fault);
343
344/*
345 * Did it hit the DOS screen memory VA from vm86 mode?
346 */
347static inline void
348check_v8086_mode(struct pt_regs *regs, unsigned long address,
349 struct task_struct *tsk)
350{
351#ifdef CONFIG_VM86
352 unsigned long bit;
353
354 if (!v8086_mode(regs) || !tsk->thread.vm86)
355 return;
356
357 bit = (address - 0xA0000) >> PAGE_SHIFT;
358 if (bit < 32)
359 tsk->thread.vm86->screen_bitmap |= 1 << bit;
360#endif
361}
362
363static bool low_pfn(unsigned long pfn)
364{
365 return pfn < max_low_pfn;
366}
367
368static void dump_pagetable(unsigned long address)
369{
370 pgd_t *base = __va(read_cr3_pa());
371 pgd_t *pgd = &base[pgd_index(address)];
372 p4d_t *p4d;
373 pud_t *pud;
374 pmd_t *pmd;
375 pte_t *pte;
376
377#ifdef CONFIG_X86_PAE
378 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
379 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
380 goto out;
381#define pr_pde pr_cont
382#else
383#define pr_pde pr_info
384#endif
385 p4d = p4d_offset(pgd, address);
386 pud = pud_offset(p4d, address);
387 pmd = pmd_offset(pud, address);
388 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
389#undef pr_pde
390
391 /*
392 * We must not directly access the pte in the highpte
393 * case if the page table is located in highmem.
394 * And let's rather not kmap-atomic the pte, just in case
395 * it's allocated already:
396 */
397 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
398 goto out;
399
400 pte = pte_offset_kernel(pmd, address);
401 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
402out:
403 pr_cont("\n");
404}
405
406#else /* CONFIG_X86_64: */
407
408void vmalloc_sync_all(void)
409{
410 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
411}
412
413/*
414 * 64-bit:
415 *
416 * Handle a fault on the vmalloc area
417 */
418static noinline int vmalloc_fault(unsigned long address)
419{
420 pgd_t *pgd, *pgd_k;
421 p4d_t *p4d, *p4d_k;
422 pud_t *pud;
423 pmd_t *pmd;
424 pte_t *pte;
425
426 /* Make sure we are in vmalloc area: */
427 if (!(address >= VMALLOC_START && address < VMALLOC_END))
428 return -1;
429
430 WARN_ON_ONCE(in_nmi());
431
432 /*
433 * Copy kernel mappings over when needed. This can also
434 * happen within a race in page table update. In the later
435 * case just flush:
436 */
437 pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address);
438 pgd_k = pgd_offset_k(address);
439 if (pgd_none(*pgd_k))
440 return -1;
441
442 if (pgtable_l5_enabled) {
443 if (pgd_none(*pgd)) {
444 set_pgd(pgd, *pgd_k);
445 arch_flush_lazy_mmu_mode();
446 } else {
447 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_k));
448 }
449 }
450
451 /* With 4-level paging, copying happens on the p4d level. */
452 p4d = p4d_offset(pgd, address);
453 p4d_k = p4d_offset(pgd_k, address);
454 if (p4d_none(*p4d_k))
455 return -1;
456
457 if (p4d_none(*p4d) && !pgtable_l5_enabled) {
458 set_p4d(p4d, *p4d_k);
459 arch_flush_lazy_mmu_mode();
460 } else {
461 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_k));
462 }
463
464 BUILD_BUG_ON(CONFIG_PGTABLE_LEVELS < 4);
465
466 pud = pud_offset(p4d, address);
467 if (pud_none(*pud))
468 return -1;
469
470 if (pud_large(*pud))
471 return 0;
472
473 pmd = pmd_offset(pud, address);
474 if (pmd_none(*pmd))
475 return -1;
476
477 if (pmd_large(*pmd))
478 return 0;
479
480 pte = pte_offset_kernel(pmd, address);
481 if (!pte_present(*pte))
482 return -1;
483
484 return 0;
485}
486NOKPROBE_SYMBOL(vmalloc_fault);
487
488#ifdef CONFIG_CPU_SUP_AMD
489static const char errata93_warning[] =
490KERN_ERR
491"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
492"******* Working around it, but it may cause SEGVs or burn power.\n"
493"******* Please consider a BIOS update.\n"
494"******* Disabling USB legacy in the BIOS may also help.\n";
495#endif
496
497/*
498 * No vm86 mode in 64-bit mode:
499 */
500static inline void
501check_v8086_mode(struct pt_regs *regs, unsigned long address,
502 struct task_struct *tsk)
503{
504}
505
506static int bad_address(void *p)
507{
508 unsigned long dummy;
509
510 return probe_kernel_address((unsigned long *)p, dummy);
511}
512
513static void dump_pagetable(unsigned long address)
514{
515 pgd_t *base = __va(read_cr3_pa());
516 pgd_t *pgd = base + pgd_index(address);
517 p4d_t *p4d;
518 pud_t *pud;
519 pmd_t *pmd;
520 pte_t *pte;
521
522 if (bad_address(pgd))
523 goto bad;
524
525 pr_info("PGD %lx ", pgd_val(*pgd));
526
527 if (!pgd_present(*pgd))
528 goto out;
529
530 p4d = p4d_offset(pgd, address);
531 if (bad_address(p4d))
532 goto bad;
533
534 pr_cont("P4D %lx ", p4d_val(*p4d));
535 if (!p4d_present(*p4d) || p4d_large(*p4d))
536 goto out;
537
538 pud = pud_offset(p4d, address);
539 if (bad_address(pud))
540 goto bad;
541
542 pr_cont("PUD %lx ", pud_val(*pud));
543 if (!pud_present(*pud) || pud_large(*pud))
544 goto out;
545
546 pmd = pmd_offset(pud, address);
547 if (bad_address(pmd))
548 goto bad;
549
550 pr_cont("PMD %lx ", pmd_val(*pmd));
551 if (!pmd_present(*pmd) || pmd_large(*pmd))
552 goto out;
553
554 pte = pte_offset_kernel(pmd, address);
555 if (bad_address(pte))
556 goto bad;
557
558 pr_cont("PTE %lx", pte_val(*pte));
559out:
560 pr_cont("\n");
561 return;
562bad:
563 pr_info("BAD\n");
564}
565
566#endif /* CONFIG_X86_64 */
567
568/*
569 * Workaround for K8 erratum #93 & buggy BIOS.
570 *
571 * BIOS SMM functions are required to use a specific workaround
572 * to avoid corruption of the 64bit RIP register on C stepping K8.
573 *
574 * A lot of BIOS that didn't get tested properly miss this.
575 *
576 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
577 * Try to work around it here.
578 *
579 * Note we only handle faults in kernel here.
580 * Does nothing on 32-bit.
581 */
582static int is_errata93(struct pt_regs *regs, unsigned long address)
583{
584#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
585 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
586 || boot_cpu_data.x86 != 0xf)
587 return 0;
588
589 if (address != regs->ip)
590 return 0;
591
592 if ((address >> 32) != 0)
593 return 0;
594
595 address |= 0xffffffffUL << 32;
596 if ((address >= (u64)_stext && address <= (u64)_etext) ||
597 (address >= MODULES_VADDR && address <= MODULES_END)) {
598 printk_once(errata93_warning);
599 regs->ip = address;
600 return 1;
601 }
602#endif
603 return 0;
604}
605
606/*
607 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
608 * to illegal addresses >4GB.
609 *
610 * We catch this in the page fault handler because these addresses
611 * are not reachable. Just detect this case and return. Any code
612 * segment in LDT is compatibility mode.
613 */
614static int is_errata100(struct pt_regs *regs, unsigned long address)
615{
616#ifdef CONFIG_X86_64
617 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
618 return 1;
619#endif
620 return 0;
621}
622
623static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
624{
625#ifdef CONFIG_X86_F00F_BUG
626 unsigned long nr;
627
628 /*
629 * Pentium F0 0F C7 C8 bug workaround:
630 */
631 if (boot_cpu_has_bug(X86_BUG_F00F)) {
632 nr = (address - idt_descr.address) >> 3;
633
634 if (nr == 6) {
635 do_invalid_op(regs, 0);
636 return 1;
637 }
638 }
639#endif
640 return 0;
641}
642
643static const char nx_warning[] = KERN_CRIT
644"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
645static const char smep_warning[] = KERN_CRIT
646"unable to execute userspace code (SMEP?) (uid: %d)\n";
647
648static void
649show_fault_oops(struct pt_regs *regs, unsigned long error_code,
650 unsigned long address)
651{
652 if (!oops_may_print())
653 return;
654
655 if (error_code & X86_PF_INSTR) {
656 unsigned int level;
657 pgd_t *pgd;
658 pte_t *pte;
659
660 pgd = __va(read_cr3_pa());
661 pgd += pgd_index(address);
662
663 pte = lookup_address_in_pgd(pgd, address, &level);
664
665 if (pte && pte_present(*pte) && !pte_exec(*pte))
666 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
667 if (pte && pte_present(*pte) && pte_exec(*pte) &&
668 (pgd_flags(*pgd) & _PAGE_USER) &&
669 (__read_cr4() & X86_CR4_SMEP))
670 printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
671 }
672
673 printk(KERN_ALERT "BUG: unable to handle kernel ");
674 if (address < PAGE_SIZE)
675 printk(KERN_CONT "NULL pointer dereference");
676 else
677 printk(KERN_CONT "paging request");
678
679 printk(KERN_CONT " at %px\n", (void *) address);
680
681 dump_pagetable(address);
682}
683
684static noinline void
685pgtable_bad(struct pt_regs *regs, unsigned long error_code,
686 unsigned long address)
687{
688 struct task_struct *tsk;
689 unsigned long flags;
690 int sig;
691
692 flags = oops_begin();
693 tsk = current;
694 sig = SIGKILL;
695
696 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
697 tsk->comm, address);
698 dump_pagetable(address);
699
700 tsk->thread.cr2 = address;
701 tsk->thread.trap_nr = X86_TRAP_PF;
702 tsk->thread.error_code = error_code;
703
704 if (__die("Bad pagetable", regs, error_code))
705 sig = 0;
706
707 oops_end(flags, regs, sig);
708}
709
710static noinline void
711no_context(struct pt_regs *regs, unsigned long error_code,
712 unsigned long address, int signal, int si_code)
713{
714 struct task_struct *tsk = current;
715 unsigned long flags;
716 int sig;
717
718 /* Are we prepared to handle this kernel fault? */
719 if (fixup_exception(regs, X86_TRAP_PF)) {
720 /*
721 * Any interrupt that takes a fault gets the fixup. This makes
722 * the below recursive fault logic only apply to a faults from
723 * task context.
724 */
725 if (in_interrupt())
726 return;
727
728 /*
729 * Per the above we're !in_interrupt(), aka. task context.
730 *
731 * In this case we need to make sure we're not recursively
732 * faulting through the emulate_vsyscall() logic.
733 */
734 if (current->thread.sig_on_uaccess_err && signal) {
735 tsk->thread.trap_nr = X86_TRAP_PF;
736 tsk->thread.error_code = error_code | X86_PF_USER;
737 tsk->thread.cr2 = address;
738
739 /* XXX: hwpoison faults will set the wrong code. */
740 force_sig_info_fault(signal, si_code, address,
741 tsk, NULL, 0);
742 }
743
744 /*
745 * Barring that, we can do the fixup and be happy.
746 */
747 return;
748 }
749
750#ifdef CONFIG_VMAP_STACK
751 /*
752 * Stack overflow? During boot, we can fault near the initial
753 * stack in the direct map, but that's not an overflow -- check
754 * that we're in vmalloc space to avoid this.
755 */
756 if (is_vmalloc_addr((void *)address) &&
757 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
758 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
759 unsigned long stack = this_cpu_read(orig_ist.ist[DOUBLEFAULT_STACK]) - sizeof(void *);
760 /*
761 * We're likely to be running with very little stack space
762 * left. It's plausible that we'd hit this condition but
763 * double-fault even before we get this far, in which case
764 * we're fine: the double-fault handler will deal with it.
765 *
766 * We don't want to make it all the way into the oops code
767 * and then double-fault, though, because we're likely to
768 * break the console driver and lose most of the stack dump.
769 */
770 asm volatile ("movq %[stack], %%rsp\n\t"
771 "call handle_stack_overflow\n\t"
772 "1: jmp 1b"
773 : ASM_CALL_CONSTRAINT
774 : "D" ("kernel stack overflow (page fault)"),
775 "S" (regs), "d" (address),
776 [stack] "rm" (stack));
777 unreachable();
778 }
779#endif
780
781 /*
782 * 32-bit:
783 *
784 * Valid to do another page fault here, because if this fault
785 * had been triggered by is_prefetch fixup_exception would have
786 * handled it.
787 *
788 * 64-bit:
789 *
790 * Hall of shame of CPU/BIOS bugs.
791 */
792 if (is_prefetch(regs, error_code, address))
793 return;
794
795 if (is_errata93(regs, address))
796 return;
797
798 /*
799 * Oops. The kernel tried to access some bad page. We'll have to
800 * terminate things with extreme prejudice:
801 */
802 flags = oops_begin();
803
804 show_fault_oops(regs, error_code, address);
805
806 if (task_stack_end_corrupted(tsk))
807 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
808
809 tsk->thread.cr2 = address;
810 tsk->thread.trap_nr = X86_TRAP_PF;
811 tsk->thread.error_code = error_code;
812
813 sig = SIGKILL;
814 if (__die("Oops", regs, error_code))
815 sig = 0;
816
817 /* Executive summary in case the body of the oops scrolled away */
818 printk(KERN_DEFAULT "CR2: %016lx\n", address);
819
820 oops_end(flags, regs, sig);
821}
822
823/*
824 * Print out info about fatal segfaults, if the show_unhandled_signals
825 * sysctl is set:
826 */
827static inline void
828show_signal_msg(struct pt_regs *regs, unsigned long error_code,
829 unsigned long address, struct task_struct *tsk)
830{
831 if (!unhandled_signal(tsk, SIGSEGV))
832 return;
833
834 if (!printk_ratelimit())
835 return;
836
837 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
838 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
839 tsk->comm, task_pid_nr(tsk), address,
840 (void *)regs->ip, (void *)regs->sp, error_code);
841
842 print_vma_addr(KERN_CONT " in ", regs->ip);
843
844 printk(KERN_CONT "\n");
845}
846
847static void
848__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
849 unsigned long address, u32 *pkey, int si_code)
850{
851 struct task_struct *tsk = current;
852
853 /* User mode accesses just cause a SIGSEGV */
854 if (error_code & X86_PF_USER) {
855 /*
856 * It's possible to have interrupts off here:
857 */
858 local_irq_enable();
859
860 /*
861 * Valid to do another page fault here because this one came
862 * from user space:
863 */
864 if (is_prefetch(regs, error_code, address))
865 return;
866
867 if (is_errata100(regs, address))
868 return;
869
870#ifdef CONFIG_X86_64
871 /*
872 * Instruction fetch faults in the vsyscall page might need
873 * emulation.
874 */
875 if (unlikely((error_code & X86_PF_INSTR) &&
876 ((address & ~0xfff) == VSYSCALL_ADDR))) {
877 if (emulate_vsyscall(regs, address))
878 return;
879 }
880#endif
881
882 /*
883 * To avoid leaking information about the kernel page table
884 * layout, pretend that user-mode accesses to kernel addresses
885 * are always protection faults.
886 */
887 if (address >= TASK_SIZE_MAX)
888 error_code |= X86_PF_PROT;
889
890 if (likely(show_unhandled_signals))
891 show_signal_msg(regs, error_code, address, tsk);
892
893 tsk->thread.cr2 = address;
894 tsk->thread.error_code = error_code;
895 tsk->thread.trap_nr = X86_TRAP_PF;
896
897 force_sig_info_fault(SIGSEGV, si_code, address, tsk, pkey, 0);
898
899 return;
900 }
901
902 if (is_f00f_bug(regs, address))
903 return;
904
905 no_context(regs, error_code, address, SIGSEGV, si_code);
906}
907
908static noinline void
909bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
910 unsigned long address, u32 *pkey)
911{
912 __bad_area_nosemaphore(regs, error_code, address, pkey, SEGV_MAPERR);
913}
914
915static void
916__bad_area(struct pt_regs *regs, unsigned long error_code,
917 unsigned long address, struct vm_area_struct *vma, int si_code)
918{
919 struct mm_struct *mm = current->mm;
920 u32 pkey;
921
922 if (vma)
923 pkey = vma_pkey(vma);
924
925 /*
926 * Something tried to access memory that isn't in our memory map..
927 * Fix it, but check if it's kernel or user first..
928 */
929 up_read(&mm->mmap_sem);
930
931 __bad_area_nosemaphore(regs, error_code, address,
932 (vma) ? &pkey : NULL, si_code);
933}
934
935static noinline void
936bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
937{
938 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
939}
940
941static inline bool bad_area_access_from_pkeys(unsigned long error_code,
942 struct vm_area_struct *vma)
943{
944 /* This code is always called on the current mm */
945 bool foreign = false;
946
947 if (!boot_cpu_has(X86_FEATURE_OSPKE))
948 return false;
949 if (error_code & X86_PF_PK)
950 return true;
951 /* this checks permission keys on the VMA: */
952 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
953 (error_code & X86_PF_INSTR), foreign))
954 return true;
955 return false;
956}
957
958static noinline void
959bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
960 unsigned long address, struct vm_area_struct *vma)
961{
962 /*
963 * This OSPKE check is not strictly necessary at runtime.
964 * But, doing it this way allows compiler optimizations
965 * if pkeys are compiled out.
966 */
967 if (bad_area_access_from_pkeys(error_code, vma))
968 __bad_area(regs, error_code, address, vma, SEGV_PKUERR);
969 else
970 __bad_area(regs, error_code, address, vma, SEGV_ACCERR);
971}
972
973static void
974do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
975 u32 *pkey, unsigned int fault)
976{
977 struct task_struct *tsk = current;
978 int code = BUS_ADRERR;
979
980 /* Kernel mode? Handle exceptions or die: */
981 if (!(error_code & X86_PF_USER)) {
982 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
983 return;
984 }
985
986 /* User-space => ok to do another page fault: */
987 if (is_prefetch(regs, error_code, address))
988 return;
989
990 tsk->thread.cr2 = address;
991 tsk->thread.error_code = error_code;
992 tsk->thread.trap_nr = X86_TRAP_PF;
993
994#ifdef CONFIG_MEMORY_FAILURE
995 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
996 printk(KERN_ERR
997 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
998 tsk->comm, tsk->pid, address);
999 code = BUS_MCEERR_AR;
1000 }
1001#endif
1002 force_sig_info_fault(SIGBUS, code, address, tsk, pkey, fault);
1003}
1004
1005static noinline void
1006mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1007 unsigned long address, u32 *pkey, unsigned int fault)
1008{
1009 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
1010 no_context(regs, error_code, address, 0, 0);
1011 return;
1012 }
1013
1014 if (fault & VM_FAULT_OOM) {
1015 /* Kernel mode? Handle exceptions or die: */
1016 if (!(error_code & X86_PF_USER)) {
1017 no_context(regs, error_code, address,
1018 SIGSEGV, SEGV_MAPERR);
1019 return;
1020 }
1021
1022 /*
1023 * We ran out of memory, call the OOM killer, and return the
1024 * userspace (which will retry the fault, or kill us if we got
1025 * oom-killed):
1026 */
1027 pagefault_out_of_memory();
1028 } else {
1029 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1030 VM_FAULT_HWPOISON_LARGE))
1031 do_sigbus(regs, error_code, address, pkey, fault);
1032 else if (fault & VM_FAULT_SIGSEGV)
1033 bad_area_nosemaphore(regs, error_code, address, pkey);
1034 else
1035 BUG();
1036 }
1037}
1038
1039static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1040{
1041 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1042 return 0;
1043
1044 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1045 return 0;
1046 /*
1047 * Note: We do not do lazy flushing on protection key
1048 * changes, so no spurious fault will ever set X86_PF_PK.
1049 */
1050 if ((error_code & X86_PF_PK))
1051 return 1;
1052
1053 return 1;
1054}
1055
1056/*
1057 * Handle a spurious fault caused by a stale TLB entry.
1058 *
1059 * This allows us to lazily refresh the TLB when increasing the
1060 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1061 * eagerly is very expensive since that implies doing a full
1062 * cross-processor TLB flush, even if no stale TLB entries exist
1063 * on other processors.
1064 *
1065 * Spurious faults may only occur if the TLB contains an entry with
1066 * fewer permission than the page table entry. Non-present (P = 0)
1067 * and reserved bit (R = 1) faults are never spurious.
1068 *
1069 * There are no security implications to leaving a stale TLB when
1070 * increasing the permissions on a page.
1071 *
1072 * Returns non-zero if a spurious fault was handled, zero otherwise.
1073 *
1074 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1075 * (Optional Invalidation).
1076 */
1077static noinline int
1078spurious_fault(unsigned long error_code, unsigned long address)
1079{
1080 pgd_t *pgd;
1081 p4d_t *p4d;
1082 pud_t *pud;
1083 pmd_t *pmd;
1084 pte_t *pte;
1085 int ret;
1086
1087 /*
1088 * Only writes to RO or instruction fetches from NX may cause
1089 * spurious faults.
1090 *
1091 * These could be from user or supervisor accesses but the TLB
1092 * is only lazily flushed after a kernel mapping protection
1093 * change, so user accesses are not expected to cause spurious
1094 * faults.
1095 */
1096 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1097 error_code != (X86_PF_INSTR | X86_PF_PROT))
1098 return 0;
1099
1100 pgd = init_mm.pgd + pgd_index(address);
1101 if (!pgd_present(*pgd))
1102 return 0;
1103
1104 p4d = p4d_offset(pgd, address);
1105 if (!p4d_present(*p4d))
1106 return 0;
1107
1108 if (p4d_large(*p4d))
1109 return spurious_fault_check(error_code, (pte_t *) p4d);
1110
1111 pud = pud_offset(p4d, address);
1112 if (!pud_present(*pud))
1113 return 0;
1114
1115 if (pud_large(*pud))
1116 return spurious_fault_check(error_code, (pte_t *) pud);
1117
1118 pmd = pmd_offset(pud, address);
1119 if (!pmd_present(*pmd))
1120 return 0;
1121
1122 if (pmd_large(*pmd))
1123 return spurious_fault_check(error_code, (pte_t *) pmd);
1124
1125 pte = pte_offset_kernel(pmd, address);
1126 if (!pte_present(*pte))
1127 return 0;
1128
1129 ret = spurious_fault_check(error_code, pte);
1130 if (!ret)
1131 return 0;
1132
1133 /*
1134 * Make sure we have permissions in PMD.
1135 * If not, then there's a bug in the page tables:
1136 */
1137 ret = spurious_fault_check(error_code, (pte_t *) pmd);
1138 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1139
1140 return ret;
1141}
1142NOKPROBE_SYMBOL(spurious_fault);
1143
1144int show_unhandled_signals = 1;
1145
1146static inline int
1147access_error(unsigned long error_code, struct vm_area_struct *vma)
1148{
1149 /* This is only called for the current mm, so: */
1150 bool foreign = false;
1151
1152 /*
1153 * Read or write was blocked by protection keys. This is
1154 * always an unconditional error and can never result in
1155 * a follow-up action to resolve the fault, like a COW.
1156 */
1157 if (error_code & X86_PF_PK)
1158 return 1;
1159
1160 /*
1161 * Make sure to check the VMA so that we do not perform
1162 * faults just to hit a X86_PF_PK as soon as we fill in a
1163 * page.
1164 */
1165 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1166 (error_code & X86_PF_INSTR), foreign))
1167 return 1;
1168
1169 if (error_code & X86_PF_WRITE) {
1170 /* write, present and write, not present: */
1171 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1172 return 1;
1173 return 0;
1174 }
1175
1176 /* read, present: */
1177 if (unlikely(error_code & X86_PF_PROT))
1178 return 1;
1179
1180 /* read, not present: */
1181 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1182 return 1;
1183
1184 return 0;
1185}
1186
1187static int fault_in_kernel_space(unsigned long address)
1188{
1189 return address >= TASK_SIZE_MAX;
1190}
1191
1192static inline bool smap_violation(int error_code, struct pt_regs *regs)
1193{
1194 if (!IS_ENABLED(CONFIG_X86_SMAP))
1195 return false;
1196
1197 if (!static_cpu_has(X86_FEATURE_SMAP))
1198 return false;
1199
1200 if (error_code & X86_PF_USER)
1201 return false;
1202
1203 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1204 return false;
1205
1206 return true;
1207}
1208
1209/*
1210 * This routine handles page faults. It determines the address,
1211 * and the problem, and then passes it off to one of the appropriate
1212 * routines.
1213 */
1214static noinline void
1215__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1216 unsigned long address)
1217{
1218 struct vm_area_struct *vma;
1219 struct task_struct *tsk;
1220 struct mm_struct *mm;
1221 int fault, major = 0;
1222 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1223 u32 pkey;
1224
1225 tsk = current;
1226 mm = tsk->mm;
1227
1228 prefetchw(&mm->mmap_sem);
1229
1230 if (unlikely(kmmio_fault(regs, address)))
1231 return;
1232
1233 /*
1234 * We fault-in kernel-space virtual memory on-demand. The
1235 * 'reference' page table is init_mm.pgd.
1236 *
1237 * NOTE! We MUST NOT take any locks for this case. We may
1238 * be in an interrupt or a critical region, and should
1239 * only copy the information from the master page table,
1240 * nothing more.
1241 *
1242 * This verifies that the fault happens in kernel space
1243 * (error_code & 4) == 0, and that the fault was not a
1244 * protection error (error_code & 9) == 0.
1245 */
1246 if (unlikely(fault_in_kernel_space(address))) {
1247 if (!(error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1248 if (vmalloc_fault(address) >= 0)
1249 return;
1250 }
1251
1252 /* Can handle a stale RO->RW TLB: */
1253 if (spurious_fault(error_code, address))
1254 return;
1255
1256 /* kprobes don't want to hook the spurious faults: */
1257 if (kprobes_fault(regs))
1258 return;
1259 /*
1260 * Don't take the mm semaphore here. If we fixup a prefetch
1261 * fault we could otherwise deadlock:
1262 */
1263 bad_area_nosemaphore(regs, error_code, address, NULL);
1264
1265 return;
1266 }
1267
1268 /* kprobes don't want to hook the spurious faults: */
1269 if (unlikely(kprobes_fault(regs)))
1270 return;
1271
1272 if (unlikely(error_code & X86_PF_RSVD))
1273 pgtable_bad(regs, error_code, address);
1274
1275 if (unlikely(smap_violation(error_code, regs))) {
1276 bad_area_nosemaphore(regs, error_code, address, NULL);
1277 return;
1278 }
1279
1280 /*
1281 * If we're in an interrupt, have no user context or are running
1282 * in a region with pagefaults disabled then we must not take the fault
1283 */
1284 if (unlikely(faulthandler_disabled() || !mm)) {
1285 bad_area_nosemaphore(regs, error_code, address, NULL);
1286 return;
1287 }
1288
1289 /*
1290 * It's safe to allow irq's after cr2 has been saved and the
1291 * vmalloc fault has been handled.
1292 *
1293 * User-mode registers count as a user access even for any
1294 * potential system fault or CPU buglet:
1295 */
1296 if (user_mode(regs)) {
1297 local_irq_enable();
1298 error_code |= X86_PF_USER;
1299 flags |= FAULT_FLAG_USER;
1300 } else {
1301 if (regs->flags & X86_EFLAGS_IF)
1302 local_irq_enable();
1303 }
1304
1305 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1306
1307 if (error_code & X86_PF_WRITE)
1308 flags |= FAULT_FLAG_WRITE;
1309 if (error_code & X86_PF_INSTR)
1310 flags |= FAULT_FLAG_INSTRUCTION;
1311
1312 /*
1313 * When running in the kernel we expect faults to occur only to
1314 * addresses in user space. All other faults represent errors in
1315 * the kernel and should generate an OOPS. Unfortunately, in the
1316 * case of an erroneous fault occurring in a code path which already
1317 * holds mmap_sem we will deadlock attempting to validate the fault
1318 * against the address space. Luckily the kernel only validly
1319 * references user space from well defined areas of code, which are
1320 * listed in the exceptions table.
1321 *
1322 * As the vast majority of faults will be valid we will only perform
1323 * the source reference check when there is a possibility of a
1324 * deadlock. Attempt to lock the address space, if we cannot we then
1325 * validate the source. If this is invalid we can skip the address
1326 * space check, thus avoiding the deadlock:
1327 */
1328 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1329 if (!(error_code & X86_PF_USER) &&
1330 !search_exception_tables(regs->ip)) {
1331 bad_area_nosemaphore(regs, error_code, address, NULL);
1332 return;
1333 }
1334retry:
1335 down_read(&mm->mmap_sem);
1336 } else {
1337 /*
1338 * The above down_read_trylock() might have succeeded in
1339 * which case we'll have missed the might_sleep() from
1340 * down_read():
1341 */
1342 might_sleep();
1343 }
1344
1345 vma = find_vma(mm, address);
1346 if (unlikely(!vma)) {
1347 bad_area(regs, error_code, address);
1348 return;
1349 }
1350 if (likely(vma->vm_start <= address))
1351 goto good_area;
1352 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1353 bad_area(regs, error_code, address);
1354 return;
1355 }
1356 if (error_code & X86_PF_USER) {
1357 /*
1358 * Accessing the stack below %sp is always a bug.
1359 * The large cushion allows instructions like enter
1360 * and pusha to work. ("enter $65535, $31" pushes
1361 * 32 pointers and then decrements %sp by 65535.)
1362 */
1363 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1364 bad_area(regs, error_code, address);
1365 return;
1366 }
1367 }
1368 if (unlikely(expand_stack(vma, address))) {
1369 bad_area(regs, error_code, address);
1370 return;
1371 }
1372
1373 /*
1374 * Ok, we have a good vm_area for this memory access, so
1375 * we can handle it..
1376 */
1377good_area:
1378 if (unlikely(access_error(error_code, vma))) {
1379 bad_area_access_error(regs, error_code, address, vma);
1380 return;
1381 }
1382
1383 /*
1384 * If for any reason at all we couldn't handle the fault,
1385 * make sure we exit gracefully rather than endlessly redo
1386 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1387 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1388 *
1389 * Note that handle_userfault() may also release and reacquire mmap_sem
1390 * (and not return with VM_FAULT_RETRY), when returning to userland to
1391 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1392 * (potentially after handling any pending signal during the return to
1393 * userland). The return to userland is identified whenever
1394 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1395 * Thus we have to be careful about not touching vma after handling the
1396 * fault, so we read the pkey beforehand.
1397 */
1398 pkey = vma_pkey(vma);
1399 fault = handle_mm_fault(vma, address, flags);
1400 major |= fault & VM_FAULT_MAJOR;
1401
1402 /*
1403 * If we need to retry the mmap_sem has already been released,
1404 * and if there is a fatal signal pending there is no guarantee
1405 * that we made any progress. Handle this case first.
1406 */
1407 if (unlikely(fault & VM_FAULT_RETRY)) {
1408 /* Retry at most once */
1409 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1410 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1411 flags |= FAULT_FLAG_TRIED;
1412 if (!fatal_signal_pending(tsk))
1413 goto retry;
1414 }
1415
1416 /* User mode? Just return to handle the fatal exception */
1417 if (flags & FAULT_FLAG_USER)
1418 return;
1419
1420 /* Not returning to user mode? Handle exceptions or die: */
1421 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1422 return;
1423 }
1424
1425 up_read(&mm->mmap_sem);
1426 if (unlikely(fault & VM_FAULT_ERROR)) {
1427 mm_fault_error(regs, error_code, address, &pkey, fault);
1428 return;
1429 }
1430
1431 /*
1432 * Major/minor page fault accounting. If any of the events
1433 * returned VM_FAULT_MAJOR, we account it as a major fault.
1434 */
1435 if (major) {
1436 tsk->maj_flt++;
1437 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1438 } else {
1439 tsk->min_flt++;
1440 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1441 }
1442
1443 check_v8086_mode(regs, address, tsk);
1444}
1445NOKPROBE_SYMBOL(__do_page_fault);
1446
1447static nokprobe_inline void
1448trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1449 unsigned long error_code)
1450{
1451 if (user_mode(regs))
1452 trace_page_fault_user(address, regs, error_code);
1453 else
1454 trace_page_fault_kernel(address, regs, error_code);
1455}
1456
1457/*
1458 * We must have this function blacklisted from kprobes, tagged with notrace
1459 * and call read_cr2() before calling anything else. To avoid calling any
1460 * kind of tracing machinery before we've observed the CR2 value.
1461 *
1462 * exception_{enter,exit}() contains all sorts of tracepoints.
1463 */
1464dotraplinkage void notrace
1465do_page_fault(struct pt_regs *regs, unsigned long error_code)
1466{
1467 unsigned long address = read_cr2(); /* Get the faulting address */
1468 enum ctx_state prev_state;
1469
1470 prev_state = exception_enter();
1471 if (trace_pagefault_enabled())
1472 trace_page_fault_entries(address, regs, error_code);
1473
1474 __do_page_fault(regs, error_code, address);
1475 exception_exit(prev_state);
1476}
1477NOKPROBE_SYMBOL(do_page_fault);