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/*
2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
4 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
5 */
6#include <linux/sched.h> /* test_thread_flag(), ... */
7#include <linux/kdebug.h> /* oops_begin/end, ... */
8#include <linux/module.h> /* search_exception_table */
9#include <linux/bootmem.h> /* max_low_pfn */
10#include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
11#include <linux/mmiotrace.h> /* kmmio_handler, ... */
12#include <linux/perf_event.h> /* perf_sw_event */
13#include <linux/hugetlb.h> /* hstate_index_to_shift */
14#include <linux/prefetch.h> /* prefetchw */
15#include <linux/context_tracking.h> /* exception_enter(), ... */
16#include <linux/uaccess.h> /* faulthandler_disabled() */
17
18#include <asm/cpufeature.h> /* boot_cpu_has, ... */
19#include <asm/traps.h> /* dotraplinkage, ... */
20#include <asm/pgalloc.h> /* pgd_*(), ... */
21#include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
22#include <asm/fixmap.h> /* VSYSCALL_ADDR */
23#include <asm/vsyscall.h> /* emulate_vsyscall */
24#include <asm/vm86.h> /* struct vm86 */
25#include <asm/mmu_context.h> /* vma_pkey() */
26
27#define CREATE_TRACE_POINTS
28#include <asm/trace/exceptions.h>
29
30/*
31 * Page fault error code bits:
32 *
33 * bit 0 == 0: no page found 1: protection fault
34 * bit 1 == 0: read access 1: write access
35 * bit 2 == 0: kernel-mode access 1: user-mode access
36 * bit 3 == 1: use of reserved bit detected
37 * bit 4 == 1: fault was an instruction fetch
38 * bit 5 == 1: protection keys block access
39 */
40enum x86_pf_error_code {
41
42 PF_PROT = 1 << 0,
43 PF_WRITE = 1 << 1,
44 PF_USER = 1 << 2,
45 PF_RSVD = 1 << 3,
46 PF_INSTR = 1 << 4,
47 PF_PK = 1 << 5,
48};
49
50/*
51 * Returns 0 if mmiotrace is disabled, or if the fault is not
52 * handled by mmiotrace:
53 */
54static nokprobe_inline int
55kmmio_fault(struct pt_regs *regs, unsigned long addr)
56{
57 if (unlikely(is_kmmio_active()))
58 if (kmmio_handler(regs, addr) == 1)
59 return -1;
60 return 0;
61}
62
63static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
64{
65 int ret = 0;
66
67 /* kprobe_running() needs smp_processor_id() */
68 if (kprobes_built_in() && !user_mode(regs)) {
69 preempt_disable();
70 if (kprobe_running() && kprobe_fault_handler(regs, 14))
71 ret = 1;
72 preempt_enable();
73 }
74
75 return ret;
76}
77
78/*
79 * Prefetch quirks:
80 *
81 * 32-bit mode:
82 *
83 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
84 * Check that here and ignore it.
85 *
86 * 64-bit mode:
87 *
88 * Sometimes the CPU reports invalid exceptions on prefetch.
89 * Check that here and ignore it.
90 *
91 * Opcode checker based on code by Richard Brunner.
92 */
93static inline int
94check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
95 unsigned char opcode, int *prefetch)
96{
97 unsigned char instr_hi = opcode & 0xf0;
98 unsigned char instr_lo = opcode & 0x0f;
99
100 switch (instr_hi) {
101 case 0x20:
102 case 0x30:
103 /*
104 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
105 * In X86_64 long mode, the CPU will signal invalid
106 * opcode if some of these prefixes are present so
107 * X86_64 will never get here anyway
108 */
109 return ((instr_lo & 7) == 0x6);
110#ifdef CONFIG_X86_64
111 case 0x40:
112 /*
113 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
114 * Need to figure out under what instruction mode the
115 * instruction was issued. Could check the LDT for lm,
116 * but for now it's good enough to assume that long
117 * mode only uses well known segments or kernel.
118 */
119 return (!user_mode(regs) || user_64bit_mode(regs));
120#endif
121 case 0x60:
122 /* 0x64 thru 0x67 are valid prefixes in all modes. */
123 return (instr_lo & 0xC) == 0x4;
124 case 0xF0:
125 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
126 return !instr_lo || (instr_lo>>1) == 1;
127 case 0x00:
128 /* Prefetch instruction is 0x0F0D or 0x0F18 */
129 if (probe_kernel_address(instr, opcode))
130 return 0;
131
132 *prefetch = (instr_lo == 0xF) &&
133 (opcode == 0x0D || opcode == 0x18);
134 return 0;
135 default:
136 return 0;
137 }
138}
139
140static int
141is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
142{
143 unsigned char *max_instr;
144 unsigned char *instr;
145 int prefetch = 0;
146
147 /*
148 * If it was a exec (instruction fetch) fault on NX page, then
149 * do not ignore the fault:
150 */
151 if (error_code & PF_INSTR)
152 return 0;
153
154 instr = (void *)convert_ip_to_linear(current, regs);
155 max_instr = instr + 15;
156
157 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
158 return 0;
159
160 while (instr < max_instr) {
161 unsigned char opcode;
162
163 if (probe_kernel_address(instr, opcode))
164 break;
165
166 instr++;
167
168 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
169 break;
170 }
171 return prefetch;
172}
173
174/*
175 * A protection key fault means that the PKRU value did not allow
176 * access to some PTE. Userspace can figure out what PKRU was
177 * from the XSAVE state, and this function fills out a field in
178 * siginfo so userspace can discover which protection key was set
179 * on the PTE.
180 *
181 * If we get here, we know that the hardware signaled a PF_PK
182 * fault and that there was a VMA once we got in the fault
183 * handler. It does *not* guarantee that the VMA we find here
184 * was the one that we faulted on.
185 *
186 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
187 * 2. T1 : set PKRU to deny access to pkey=4, touches page
188 * 3. T1 : faults...
189 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
190 * 5. T1 : enters fault handler, takes mmap_sem, etc...
191 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
192 * faulted on a pte with its pkey=4.
193 */
194static void fill_sig_info_pkey(int si_code, siginfo_t *info,
195 struct vm_area_struct *vma)
196{
197 /* This is effectively an #ifdef */
198 if (!boot_cpu_has(X86_FEATURE_OSPKE))
199 return;
200
201 /* Fault not from Protection Keys: nothing to do */
202 if (si_code != SEGV_PKUERR)
203 return;
204 /*
205 * force_sig_info_fault() is called from a number of
206 * contexts, some of which have a VMA and some of which
207 * do not. The PF_PK handing happens after we have a
208 * valid VMA, so we should never reach this without a
209 * valid VMA.
210 */
211 if (!vma) {
212 WARN_ONCE(1, "PKU fault with no VMA passed in");
213 info->si_pkey = 0;
214 return;
215 }
216 /*
217 * si_pkey should be thought of as a strong hint, but not
218 * absolutely guranteed to be 100% accurate because of
219 * the race explained above.
220 */
221 info->si_pkey = vma_pkey(vma);
222}
223
224static void
225force_sig_info_fault(int si_signo, int si_code, unsigned long address,
226 struct task_struct *tsk, struct vm_area_struct *vma,
227 int fault)
228{
229 unsigned lsb = 0;
230 siginfo_t info;
231
232 info.si_signo = si_signo;
233 info.si_errno = 0;
234 info.si_code = si_code;
235 info.si_addr = (void __user *)address;
236 if (fault & VM_FAULT_HWPOISON_LARGE)
237 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
238 if (fault & VM_FAULT_HWPOISON)
239 lsb = PAGE_SHIFT;
240 info.si_addr_lsb = lsb;
241
242 fill_sig_info_pkey(si_code, &info, vma);
243
244 force_sig_info(si_signo, &info, tsk);
245}
246
247DEFINE_SPINLOCK(pgd_lock);
248LIST_HEAD(pgd_list);
249
250#ifdef CONFIG_X86_32
251static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
252{
253 unsigned index = pgd_index(address);
254 pgd_t *pgd_k;
255 pud_t *pud, *pud_k;
256 pmd_t *pmd, *pmd_k;
257
258 pgd += index;
259 pgd_k = init_mm.pgd + index;
260
261 if (!pgd_present(*pgd_k))
262 return NULL;
263
264 /*
265 * set_pgd(pgd, *pgd_k); here would be useless on PAE
266 * and redundant with the set_pmd() on non-PAE. As would
267 * set_pud.
268 */
269 pud = pud_offset(pgd, address);
270 pud_k = pud_offset(pgd_k, address);
271 if (!pud_present(*pud_k))
272 return NULL;
273
274 pmd = pmd_offset(pud, address);
275 pmd_k = pmd_offset(pud_k, address);
276 if (!pmd_present(*pmd_k))
277 return NULL;
278
279 if (!pmd_present(*pmd))
280 set_pmd(pmd, *pmd_k);
281 else
282 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
283
284 return pmd_k;
285}
286
287void vmalloc_sync_all(void)
288{
289 unsigned long address;
290
291 if (SHARED_KERNEL_PMD)
292 return;
293
294 for (address = VMALLOC_START & PMD_MASK;
295 address >= TASK_SIZE && address < FIXADDR_TOP;
296 address += PMD_SIZE) {
297 struct page *page;
298
299 spin_lock(&pgd_lock);
300 list_for_each_entry(page, &pgd_list, lru) {
301 spinlock_t *pgt_lock;
302 pmd_t *ret;
303
304 /* the pgt_lock only for Xen */
305 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
306
307 spin_lock(pgt_lock);
308 ret = vmalloc_sync_one(page_address(page), address);
309 spin_unlock(pgt_lock);
310
311 if (!ret)
312 break;
313 }
314 spin_unlock(&pgd_lock);
315 }
316}
317
318/*
319 * 32-bit:
320 *
321 * Handle a fault on the vmalloc or module mapping area
322 */
323static noinline int vmalloc_fault(unsigned long address)
324{
325 unsigned long pgd_paddr;
326 pmd_t *pmd_k;
327 pte_t *pte_k;
328
329 /* Make sure we are in vmalloc area: */
330 if (!(address >= VMALLOC_START && address < VMALLOC_END))
331 return -1;
332
333 WARN_ON_ONCE(in_nmi());
334
335 /*
336 * Synchronize this task's top level page-table
337 * with the 'reference' page table.
338 *
339 * Do _not_ use "current" here. We might be inside
340 * an interrupt in the middle of a task switch..
341 */
342 pgd_paddr = read_cr3();
343 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
344 if (!pmd_k)
345 return -1;
346
347 if (pmd_huge(*pmd_k))
348 return 0;
349
350 pte_k = pte_offset_kernel(pmd_k, address);
351 if (!pte_present(*pte_k))
352 return -1;
353
354 return 0;
355}
356NOKPROBE_SYMBOL(vmalloc_fault);
357
358/*
359 * Did it hit the DOS screen memory VA from vm86 mode?
360 */
361static inline void
362check_v8086_mode(struct pt_regs *regs, unsigned long address,
363 struct task_struct *tsk)
364{
365#ifdef CONFIG_VM86
366 unsigned long bit;
367
368 if (!v8086_mode(regs) || !tsk->thread.vm86)
369 return;
370
371 bit = (address - 0xA0000) >> PAGE_SHIFT;
372 if (bit < 32)
373 tsk->thread.vm86->screen_bitmap |= 1 << bit;
374#endif
375}
376
377static bool low_pfn(unsigned long pfn)
378{
379 return pfn < max_low_pfn;
380}
381
382static void dump_pagetable(unsigned long address)
383{
384 pgd_t *base = __va(read_cr3());
385 pgd_t *pgd = &base[pgd_index(address)];
386 pmd_t *pmd;
387 pte_t *pte;
388
389#ifdef CONFIG_X86_PAE
390 printk("*pdpt = %016Lx ", pgd_val(*pgd));
391 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
392 goto out;
393#endif
394 pmd = pmd_offset(pud_offset(pgd, address), address);
395 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
396
397 /*
398 * We must not directly access the pte in the highpte
399 * case if the page table is located in highmem.
400 * And let's rather not kmap-atomic the pte, just in case
401 * it's allocated already:
402 */
403 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
404 goto out;
405
406 pte = pte_offset_kernel(pmd, address);
407 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
408out:
409 printk("\n");
410}
411
412#else /* CONFIG_X86_64: */
413
414void vmalloc_sync_all(void)
415{
416 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0);
417}
418
419/*
420 * 64-bit:
421 *
422 * Handle a fault on the vmalloc area
423 */
424static noinline int vmalloc_fault(unsigned long address)
425{
426 pgd_t *pgd, *pgd_ref;
427 pud_t *pud, *pud_ref;
428 pmd_t *pmd, *pmd_ref;
429 pte_t *pte, *pte_ref;
430
431 /* Make sure we are in vmalloc area: */
432 if (!(address >= VMALLOC_START && address < VMALLOC_END))
433 return -1;
434
435 WARN_ON_ONCE(in_nmi());
436
437 /*
438 * Copy kernel mappings over when needed. This can also
439 * happen within a race in page table update. In the later
440 * case just flush:
441 */
442 pgd = pgd_offset(current->active_mm, address);
443 pgd_ref = pgd_offset_k(address);
444 if (pgd_none(*pgd_ref))
445 return -1;
446
447 if (pgd_none(*pgd)) {
448 set_pgd(pgd, *pgd_ref);
449 arch_flush_lazy_mmu_mode();
450 } else {
451 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
452 }
453
454 /*
455 * Below here mismatches are bugs because these lower tables
456 * are shared:
457 */
458
459 pud = pud_offset(pgd, address);
460 pud_ref = pud_offset(pgd_ref, address);
461 if (pud_none(*pud_ref))
462 return -1;
463
464 if (pud_none(*pud) || pud_pfn(*pud) != pud_pfn(*pud_ref))
465 BUG();
466
467 if (pud_huge(*pud))
468 return 0;
469
470 pmd = pmd_offset(pud, address);
471 pmd_ref = pmd_offset(pud_ref, address);
472 if (pmd_none(*pmd_ref))
473 return -1;
474
475 if (pmd_none(*pmd) || pmd_pfn(*pmd) != pmd_pfn(*pmd_ref))
476 BUG();
477
478 if (pmd_huge(*pmd))
479 return 0;
480
481 pte_ref = pte_offset_kernel(pmd_ref, address);
482 if (!pte_present(*pte_ref))
483 return -1;
484
485 pte = pte_offset_kernel(pmd, address);
486
487 /*
488 * Don't use pte_page here, because the mappings can point
489 * outside mem_map, and the NUMA hash lookup cannot handle
490 * that:
491 */
492 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
493 BUG();
494
495 return 0;
496}
497NOKPROBE_SYMBOL(vmalloc_fault);
498
499#ifdef CONFIG_CPU_SUP_AMD
500static const char errata93_warning[] =
501KERN_ERR
502"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
503"******* Working around it, but it may cause SEGVs or burn power.\n"
504"******* Please consider a BIOS update.\n"
505"******* Disabling USB legacy in the BIOS may also help.\n";
506#endif
507
508/*
509 * No vm86 mode in 64-bit mode:
510 */
511static inline void
512check_v8086_mode(struct pt_regs *regs, unsigned long address,
513 struct task_struct *tsk)
514{
515}
516
517static int bad_address(void *p)
518{
519 unsigned long dummy;
520
521 return probe_kernel_address((unsigned long *)p, dummy);
522}
523
524static void dump_pagetable(unsigned long address)
525{
526 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
527 pgd_t *pgd = base + pgd_index(address);
528 pud_t *pud;
529 pmd_t *pmd;
530 pte_t *pte;
531
532 if (bad_address(pgd))
533 goto bad;
534
535 printk("PGD %lx ", pgd_val(*pgd));
536
537 if (!pgd_present(*pgd))
538 goto out;
539
540 pud = pud_offset(pgd, address);
541 if (bad_address(pud))
542 goto bad;
543
544 printk("PUD %lx ", pud_val(*pud));
545 if (!pud_present(*pud) || pud_large(*pud))
546 goto out;
547
548 pmd = pmd_offset(pud, address);
549 if (bad_address(pmd))
550 goto bad;
551
552 printk("PMD %lx ", pmd_val(*pmd));
553 if (!pmd_present(*pmd) || pmd_large(*pmd))
554 goto out;
555
556 pte = pte_offset_kernel(pmd, address);
557 if (bad_address(pte))
558 goto bad;
559
560 printk("PTE %lx", pte_val(*pte));
561out:
562 printk("\n");
563 return;
564bad:
565 printk("BAD\n");
566}
567
568#endif /* CONFIG_X86_64 */
569
570/*
571 * Workaround for K8 erratum #93 & buggy BIOS.
572 *
573 * BIOS SMM functions are required to use a specific workaround
574 * to avoid corruption of the 64bit RIP register on C stepping K8.
575 *
576 * A lot of BIOS that didn't get tested properly miss this.
577 *
578 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
579 * Try to work around it here.
580 *
581 * Note we only handle faults in kernel here.
582 * Does nothing on 32-bit.
583 */
584static int is_errata93(struct pt_regs *regs, unsigned long address)
585{
586#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
587 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
588 || boot_cpu_data.x86 != 0xf)
589 return 0;
590
591 if (address != regs->ip)
592 return 0;
593
594 if ((address >> 32) != 0)
595 return 0;
596
597 address |= 0xffffffffUL << 32;
598 if ((address >= (u64)_stext && address <= (u64)_etext) ||
599 (address >= MODULES_VADDR && address <= MODULES_END)) {
600 printk_once(errata93_warning);
601 regs->ip = address;
602 return 1;
603 }
604#endif
605 return 0;
606}
607
608/*
609 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
610 * to illegal addresses >4GB.
611 *
612 * We catch this in the page fault handler because these addresses
613 * are not reachable. Just detect this case and return. Any code
614 * segment in LDT is compatibility mode.
615 */
616static int is_errata100(struct pt_regs *regs, unsigned long address)
617{
618#ifdef CONFIG_X86_64
619 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
620 return 1;
621#endif
622 return 0;
623}
624
625static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
626{
627#ifdef CONFIG_X86_F00F_BUG
628 unsigned long nr;
629
630 /*
631 * Pentium F0 0F C7 C8 bug workaround:
632 */
633 if (boot_cpu_has_bug(X86_BUG_F00F)) {
634 nr = (address - idt_descr.address) >> 3;
635
636 if (nr == 6) {
637 do_invalid_op(regs, 0);
638 return 1;
639 }
640 }
641#endif
642 return 0;
643}
644
645static const char nx_warning[] = KERN_CRIT
646"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
647static const char smep_warning[] = KERN_CRIT
648"unable to execute userspace code (SMEP?) (uid: %d)\n";
649
650static void
651show_fault_oops(struct pt_regs *regs, unsigned long error_code,
652 unsigned long address)
653{
654 if (!oops_may_print())
655 return;
656
657 if (error_code & PF_INSTR) {
658 unsigned int level;
659 pgd_t *pgd;
660 pte_t *pte;
661
662 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
663 pgd += pgd_index(address);
664
665 pte = lookup_address_in_pgd(pgd, address, &level);
666
667 if (pte && pte_present(*pte) && !pte_exec(*pte))
668 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
669 if (pte && pte_present(*pte) && pte_exec(*pte) &&
670 (pgd_flags(*pgd) & _PAGE_USER) &&
671 (__read_cr4() & X86_CR4_SMEP))
672 printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
673 }
674
675 printk(KERN_ALERT "BUG: unable to handle kernel ");
676 if (address < PAGE_SIZE)
677 printk(KERN_CONT "NULL pointer dereference");
678 else
679 printk(KERN_CONT "paging request");
680
681 printk(KERN_CONT " at %p\n", (void *) address);
682 printk(KERN_ALERT "IP:");
683 printk_address(regs->ip);
684
685 dump_pagetable(address);
686}
687
688static noinline void
689pgtable_bad(struct pt_regs *regs, unsigned long error_code,
690 unsigned long address)
691{
692 struct task_struct *tsk;
693 unsigned long flags;
694 int sig;
695
696 flags = oops_begin();
697 tsk = current;
698 sig = SIGKILL;
699
700 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
701 tsk->comm, address);
702 dump_pagetable(address);
703
704 tsk->thread.cr2 = address;
705 tsk->thread.trap_nr = X86_TRAP_PF;
706 tsk->thread.error_code = error_code;
707
708 if (__die("Bad pagetable", regs, error_code))
709 sig = 0;
710
711 oops_end(flags, regs, sig);
712}
713
714static noinline void
715no_context(struct pt_regs *regs, unsigned long error_code,
716 unsigned long address, int signal, int si_code)
717{
718 struct task_struct *tsk = current;
719 unsigned long flags;
720 int sig;
721 /* No context means no VMA to pass down */
722 struct vm_area_struct *vma = NULL;
723
724 /* Are we prepared to handle this kernel fault? */
725 if (fixup_exception(regs, X86_TRAP_PF)) {
726 /*
727 * Any interrupt that takes a fault gets the fixup. This makes
728 * the below recursive fault logic only apply to a faults from
729 * task context.
730 */
731 if (in_interrupt())
732 return;
733
734 /*
735 * Per the above we're !in_interrupt(), aka. task context.
736 *
737 * In this case we need to make sure we're not recursively
738 * faulting through the emulate_vsyscall() logic.
739 */
740 if (current_thread_info()->sig_on_uaccess_error && signal) {
741 tsk->thread.trap_nr = X86_TRAP_PF;
742 tsk->thread.error_code = error_code | PF_USER;
743 tsk->thread.cr2 = address;
744
745 /* XXX: hwpoison faults will set the wrong code. */
746 force_sig_info_fault(signal, si_code, address,
747 tsk, vma, 0);
748 }
749
750 /*
751 * Barring that, we can do the fixup and be happy.
752 */
753 return;
754 }
755
756 /*
757 * 32-bit:
758 *
759 * Valid to do another page fault here, because if this fault
760 * had been triggered by is_prefetch fixup_exception would have
761 * handled it.
762 *
763 * 64-bit:
764 *
765 * Hall of shame of CPU/BIOS bugs.
766 */
767 if (is_prefetch(regs, error_code, address))
768 return;
769
770 if (is_errata93(regs, address))
771 return;
772
773 /*
774 * Oops. The kernel tried to access some bad page. We'll have to
775 * terminate things with extreme prejudice:
776 */
777 flags = oops_begin();
778
779 show_fault_oops(regs, error_code, address);
780
781 if (task_stack_end_corrupted(tsk))
782 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
783
784 tsk->thread.cr2 = address;
785 tsk->thread.trap_nr = X86_TRAP_PF;
786 tsk->thread.error_code = error_code;
787
788 sig = SIGKILL;
789 if (__die("Oops", regs, error_code))
790 sig = 0;
791
792 /* Executive summary in case the body of the oops scrolled away */
793 printk(KERN_DEFAULT "CR2: %016lx\n", address);
794
795 oops_end(flags, regs, sig);
796}
797
798/*
799 * Print out info about fatal segfaults, if the show_unhandled_signals
800 * sysctl is set:
801 */
802static inline void
803show_signal_msg(struct pt_regs *regs, unsigned long error_code,
804 unsigned long address, struct task_struct *tsk)
805{
806 if (!unhandled_signal(tsk, SIGSEGV))
807 return;
808
809 if (!printk_ratelimit())
810 return;
811
812 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
813 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
814 tsk->comm, task_pid_nr(tsk), address,
815 (void *)regs->ip, (void *)regs->sp, error_code);
816
817 print_vma_addr(KERN_CONT " in ", regs->ip);
818
819 printk(KERN_CONT "\n");
820}
821
822static void
823__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
824 unsigned long address, struct vm_area_struct *vma,
825 int si_code)
826{
827 struct task_struct *tsk = current;
828
829 /* User mode accesses just cause a SIGSEGV */
830 if (error_code & PF_USER) {
831 /*
832 * It's possible to have interrupts off here:
833 */
834 local_irq_enable();
835
836 /*
837 * Valid to do another page fault here because this one came
838 * from user space:
839 */
840 if (is_prefetch(regs, error_code, address))
841 return;
842
843 if (is_errata100(regs, address))
844 return;
845
846#ifdef CONFIG_X86_64
847 /*
848 * Instruction fetch faults in the vsyscall page might need
849 * emulation.
850 */
851 if (unlikely((error_code & PF_INSTR) &&
852 ((address & ~0xfff) == VSYSCALL_ADDR))) {
853 if (emulate_vsyscall(regs, address))
854 return;
855 }
856#endif
857 /* Kernel addresses are always protection faults: */
858 if (address >= TASK_SIZE)
859 error_code |= PF_PROT;
860
861 if (likely(show_unhandled_signals))
862 show_signal_msg(regs, error_code, address, tsk);
863
864 tsk->thread.cr2 = address;
865 tsk->thread.error_code = error_code;
866 tsk->thread.trap_nr = X86_TRAP_PF;
867
868 force_sig_info_fault(SIGSEGV, si_code, address, tsk, vma, 0);
869
870 return;
871 }
872
873 if (is_f00f_bug(regs, address))
874 return;
875
876 no_context(regs, error_code, address, SIGSEGV, si_code);
877}
878
879static noinline void
880bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
881 unsigned long address, struct vm_area_struct *vma)
882{
883 __bad_area_nosemaphore(regs, error_code, address, vma, SEGV_MAPERR);
884}
885
886static void
887__bad_area(struct pt_regs *regs, unsigned long error_code,
888 unsigned long address, struct vm_area_struct *vma, int si_code)
889{
890 struct mm_struct *mm = current->mm;
891
892 /*
893 * Something tried to access memory that isn't in our memory map..
894 * Fix it, but check if it's kernel or user first..
895 */
896 up_read(&mm->mmap_sem);
897
898 __bad_area_nosemaphore(regs, error_code, address, vma, si_code);
899}
900
901static noinline void
902bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
903{
904 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
905}
906
907static inline bool bad_area_access_from_pkeys(unsigned long error_code,
908 struct vm_area_struct *vma)
909{
910 /* This code is always called on the current mm */
911 bool foreign = false;
912
913 if (!boot_cpu_has(X86_FEATURE_OSPKE))
914 return false;
915 if (error_code & PF_PK)
916 return true;
917 /* this checks permission keys on the VMA: */
918 if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
919 (error_code & PF_INSTR), foreign))
920 return true;
921 return false;
922}
923
924static noinline void
925bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
926 unsigned long address, struct vm_area_struct *vma)
927{
928 /*
929 * This OSPKE check is not strictly necessary at runtime.
930 * But, doing it this way allows compiler optimizations
931 * if pkeys are compiled out.
932 */
933 if (bad_area_access_from_pkeys(error_code, vma))
934 __bad_area(regs, error_code, address, vma, SEGV_PKUERR);
935 else
936 __bad_area(regs, error_code, address, vma, SEGV_ACCERR);
937}
938
939static void
940do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
941 struct vm_area_struct *vma, unsigned int fault)
942{
943 struct task_struct *tsk = current;
944 int code = BUS_ADRERR;
945
946 /* Kernel mode? Handle exceptions or die: */
947 if (!(error_code & PF_USER)) {
948 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
949 return;
950 }
951
952 /* User-space => ok to do another page fault: */
953 if (is_prefetch(regs, error_code, address))
954 return;
955
956 tsk->thread.cr2 = address;
957 tsk->thread.error_code = error_code;
958 tsk->thread.trap_nr = X86_TRAP_PF;
959
960#ifdef CONFIG_MEMORY_FAILURE
961 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
962 printk(KERN_ERR
963 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
964 tsk->comm, tsk->pid, address);
965 code = BUS_MCEERR_AR;
966 }
967#endif
968 force_sig_info_fault(SIGBUS, code, address, tsk, vma, fault);
969}
970
971static noinline void
972mm_fault_error(struct pt_regs *regs, unsigned long error_code,
973 unsigned long address, struct vm_area_struct *vma,
974 unsigned int fault)
975{
976 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
977 no_context(regs, error_code, address, 0, 0);
978 return;
979 }
980
981 if (fault & VM_FAULT_OOM) {
982 /* Kernel mode? Handle exceptions or die: */
983 if (!(error_code & PF_USER)) {
984 no_context(regs, error_code, address,
985 SIGSEGV, SEGV_MAPERR);
986 return;
987 }
988
989 /*
990 * We ran out of memory, call the OOM killer, and return the
991 * userspace (which will retry the fault, or kill us if we got
992 * oom-killed):
993 */
994 pagefault_out_of_memory();
995 } else {
996 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
997 VM_FAULT_HWPOISON_LARGE))
998 do_sigbus(regs, error_code, address, vma, fault);
999 else if (fault & VM_FAULT_SIGSEGV)
1000 bad_area_nosemaphore(regs, error_code, address, vma);
1001 else
1002 BUG();
1003 }
1004}
1005
1006static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1007{
1008 if ((error_code & PF_WRITE) && !pte_write(*pte))
1009 return 0;
1010
1011 if ((error_code & PF_INSTR) && !pte_exec(*pte))
1012 return 0;
1013 /*
1014 * Note: We do not do lazy flushing on protection key
1015 * changes, so no spurious fault will ever set PF_PK.
1016 */
1017 if ((error_code & PF_PK))
1018 return 1;
1019
1020 return 1;
1021}
1022
1023/*
1024 * Handle a spurious fault caused by a stale TLB entry.
1025 *
1026 * This allows us to lazily refresh the TLB when increasing the
1027 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1028 * eagerly is very expensive since that implies doing a full
1029 * cross-processor TLB flush, even if no stale TLB entries exist
1030 * on other processors.
1031 *
1032 * Spurious faults may only occur if the TLB contains an entry with
1033 * fewer permission than the page table entry. Non-present (P = 0)
1034 * and reserved bit (R = 1) faults are never spurious.
1035 *
1036 * There are no security implications to leaving a stale TLB when
1037 * increasing the permissions on a page.
1038 *
1039 * Returns non-zero if a spurious fault was handled, zero otherwise.
1040 *
1041 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1042 * (Optional Invalidation).
1043 */
1044static noinline int
1045spurious_fault(unsigned long error_code, unsigned long address)
1046{
1047 pgd_t *pgd;
1048 pud_t *pud;
1049 pmd_t *pmd;
1050 pte_t *pte;
1051 int ret;
1052
1053 /*
1054 * Only writes to RO or instruction fetches from NX may cause
1055 * spurious faults.
1056 *
1057 * These could be from user or supervisor accesses but the TLB
1058 * is only lazily flushed after a kernel mapping protection
1059 * change, so user accesses are not expected to cause spurious
1060 * faults.
1061 */
1062 if (error_code != (PF_WRITE | PF_PROT)
1063 && error_code != (PF_INSTR | PF_PROT))
1064 return 0;
1065
1066 pgd = init_mm.pgd + pgd_index(address);
1067 if (!pgd_present(*pgd))
1068 return 0;
1069
1070 pud = pud_offset(pgd, address);
1071 if (!pud_present(*pud))
1072 return 0;
1073
1074 if (pud_large(*pud))
1075 return spurious_fault_check(error_code, (pte_t *) pud);
1076
1077 pmd = pmd_offset(pud, address);
1078 if (!pmd_present(*pmd))
1079 return 0;
1080
1081 if (pmd_large(*pmd))
1082 return spurious_fault_check(error_code, (pte_t *) pmd);
1083
1084 pte = pte_offset_kernel(pmd, address);
1085 if (!pte_present(*pte))
1086 return 0;
1087
1088 ret = spurious_fault_check(error_code, pte);
1089 if (!ret)
1090 return 0;
1091
1092 /*
1093 * Make sure we have permissions in PMD.
1094 * If not, then there's a bug in the page tables:
1095 */
1096 ret = spurious_fault_check(error_code, (pte_t *) pmd);
1097 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1098
1099 return ret;
1100}
1101NOKPROBE_SYMBOL(spurious_fault);
1102
1103int show_unhandled_signals = 1;
1104
1105static inline int
1106access_error(unsigned long error_code, struct vm_area_struct *vma)
1107{
1108 /* This is only called for the current mm, so: */
1109 bool foreign = false;
1110 /*
1111 * Make sure to check the VMA so that we do not perform
1112 * faults just to hit a PF_PK as soon as we fill in a
1113 * page.
1114 */
1115 if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
1116 (error_code & PF_INSTR), foreign))
1117 return 1;
1118
1119 if (error_code & PF_WRITE) {
1120 /* write, present and write, not present: */
1121 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1122 return 1;
1123 return 0;
1124 }
1125
1126 /* read, present: */
1127 if (unlikely(error_code & PF_PROT))
1128 return 1;
1129
1130 /* read, not present: */
1131 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1132 return 1;
1133
1134 return 0;
1135}
1136
1137static int fault_in_kernel_space(unsigned long address)
1138{
1139 return address >= TASK_SIZE_MAX;
1140}
1141
1142static inline bool smap_violation(int error_code, struct pt_regs *regs)
1143{
1144 if (!IS_ENABLED(CONFIG_X86_SMAP))
1145 return false;
1146
1147 if (!static_cpu_has(X86_FEATURE_SMAP))
1148 return false;
1149
1150 if (error_code & PF_USER)
1151 return false;
1152
1153 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1154 return false;
1155
1156 return true;
1157}
1158
1159/*
1160 * This routine handles page faults. It determines the address,
1161 * and the problem, and then passes it off to one of the appropriate
1162 * routines.
1163 *
1164 * This function must have noinline because both callers
1165 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1166 * guarantees there's a function trace entry.
1167 */
1168static noinline void
1169__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1170 unsigned long address)
1171{
1172 struct vm_area_struct *vma;
1173 struct task_struct *tsk;
1174 struct mm_struct *mm;
1175 int fault, major = 0;
1176 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1177
1178 tsk = current;
1179 mm = tsk->mm;
1180
1181 /*
1182 * Detect and handle instructions that would cause a page fault for
1183 * both a tracked kernel page and a userspace page.
1184 */
1185 if (kmemcheck_active(regs))
1186 kmemcheck_hide(regs);
1187 prefetchw(&mm->mmap_sem);
1188
1189 if (unlikely(kmmio_fault(regs, address)))
1190 return;
1191
1192 /*
1193 * We fault-in kernel-space virtual memory on-demand. The
1194 * 'reference' page table is init_mm.pgd.
1195 *
1196 * NOTE! We MUST NOT take any locks for this case. We may
1197 * be in an interrupt or a critical region, and should
1198 * only copy the information from the master page table,
1199 * nothing more.
1200 *
1201 * This verifies that the fault happens in kernel space
1202 * (error_code & 4) == 0, and that the fault was not a
1203 * protection error (error_code & 9) == 0.
1204 */
1205 if (unlikely(fault_in_kernel_space(address))) {
1206 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1207 if (vmalloc_fault(address) >= 0)
1208 return;
1209
1210 if (kmemcheck_fault(regs, address, error_code))
1211 return;
1212 }
1213
1214 /* Can handle a stale RO->RW TLB: */
1215 if (spurious_fault(error_code, address))
1216 return;
1217
1218 /* kprobes don't want to hook the spurious faults: */
1219 if (kprobes_fault(regs))
1220 return;
1221 /*
1222 * Don't take the mm semaphore here. If we fixup a prefetch
1223 * fault we could otherwise deadlock:
1224 */
1225 bad_area_nosemaphore(regs, error_code, address, NULL);
1226
1227 return;
1228 }
1229
1230 /* kprobes don't want to hook the spurious faults: */
1231 if (unlikely(kprobes_fault(regs)))
1232 return;
1233
1234 if (unlikely(error_code & PF_RSVD))
1235 pgtable_bad(regs, error_code, address);
1236
1237 if (unlikely(smap_violation(error_code, regs))) {
1238 bad_area_nosemaphore(regs, error_code, address, NULL);
1239 return;
1240 }
1241
1242 /*
1243 * If we're in an interrupt, have no user context or are running
1244 * in a region with pagefaults disabled then we must not take the fault
1245 */
1246 if (unlikely(faulthandler_disabled() || !mm)) {
1247 bad_area_nosemaphore(regs, error_code, address, NULL);
1248 return;
1249 }
1250
1251 /*
1252 * It's safe to allow irq's after cr2 has been saved and the
1253 * vmalloc fault has been handled.
1254 *
1255 * User-mode registers count as a user access even for any
1256 * potential system fault or CPU buglet:
1257 */
1258 if (user_mode(regs)) {
1259 local_irq_enable();
1260 error_code |= PF_USER;
1261 flags |= FAULT_FLAG_USER;
1262 } else {
1263 if (regs->flags & X86_EFLAGS_IF)
1264 local_irq_enable();
1265 }
1266
1267 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1268
1269 if (error_code & PF_WRITE)
1270 flags |= FAULT_FLAG_WRITE;
1271 if (error_code & PF_INSTR)
1272 flags |= FAULT_FLAG_INSTRUCTION;
1273
1274 /*
1275 * When running in the kernel we expect faults to occur only to
1276 * addresses in user space. All other faults represent errors in
1277 * the kernel and should generate an OOPS. Unfortunately, in the
1278 * case of an erroneous fault occurring in a code path which already
1279 * holds mmap_sem we will deadlock attempting to validate the fault
1280 * against the address space. Luckily the kernel only validly
1281 * references user space from well defined areas of code, which are
1282 * listed in the exceptions table.
1283 *
1284 * As the vast majority of faults will be valid we will only perform
1285 * the source reference check when there is a possibility of a
1286 * deadlock. Attempt to lock the address space, if we cannot we then
1287 * validate the source. If this is invalid we can skip the address
1288 * space check, thus avoiding the deadlock:
1289 */
1290 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1291 if ((error_code & PF_USER) == 0 &&
1292 !search_exception_tables(regs->ip)) {
1293 bad_area_nosemaphore(regs, error_code, address, NULL);
1294 return;
1295 }
1296retry:
1297 down_read(&mm->mmap_sem);
1298 } else {
1299 /*
1300 * The above down_read_trylock() might have succeeded in
1301 * which case we'll have missed the might_sleep() from
1302 * down_read():
1303 */
1304 might_sleep();
1305 }
1306
1307 vma = find_vma(mm, address);
1308 if (unlikely(!vma)) {
1309 bad_area(regs, error_code, address);
1310 return;
1311 }
1312 if (likely(vma->vm_start <= address))
1313 goto good_area;
1314 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1315 bad_area(regs, error_code, address);
1316 return;
1317 }
1318 if (error_code & PF_USER) {
1319 /*
1320 * Accessing the stack below %sp is always a bug.
1321 * The large cushion allows instructions like enter
1322 * and pusha to work. ("enter $65535, $31" pushes
1323 * 32 pointers and then decrements %sp by 65535.)
1324 */
1325 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1326 bad_area(regs, error_code, address);
1327 return;
1328 }
1329 }
1330 if (unlikely(expand_stack(vma, address))) {
1331 bad_area(regs, error_code, address);
1332 return;
1333 }
1334
1335 /*
1336 * Ok, we have a good vm_area for this memory access, so
1337 * we can handle it..
1338 */
1339good_area:
1340 if (unlikely(access_error(error_code, vma))) {
1341 bad_area_access_error(regs, error_code, address, vma);
1342 return;
1343 }
1344
1345 /*
1346 * If for any reason at all we couldn't handle the fault,
1347 * make sure we exit gracefully rather than endlessly redo
1348 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1349 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1350 */
1351 fault = handle_mm_fault(mm, vma, address, flags);
1352 major |= fault & VM_FAULT_MAJOR;
1353
1354 /*
1355 * If we need to retry the mmap_sem has already been released,
1356 * and if there is a fatal signal pending there is no guarantee
1357 * that we made any progress. Handle this case first.
1358 */
1359 if (unlikely(fault & VM_FAULT_RETRY)) {
1360 /* Retry at most once */
1361 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1362 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1363 flags |= FAULT_FLAG_TRIED;
1364 if (!fatal_signal_pending(tsk))
1365 goto retry;
1366 }
1367
1368 /* User mode? Just return to handle the fatal exception */
1369 if (flags & FAULT_FLAG_USER)
1370 return;
1371
1372 /* Not returning to user mode? Handle exceptions or die: */
1373 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1374 return;
1375 }
1376
1377 up_read(&mm->mmap_sem);
1378 if (unlikely(fault & VM_FAULT_ERROR)) {
1379 mm_fault_error(regs, error_code, address, vma, fault);
1380 return;
1381 }
1382
1383 /*
1384 * Major/minor page fault accounting. If any of the events
1385 * returned VM_FAULT_MAJOR, we account it as a major fault.
1386 */
1387 if (major) {
1388 tsk->maj_flt++;
1389 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1390 } else {
1391 tsk->min_flt++;
1392 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1393 }
1394
1395 check_v8086_mode(regs, address, tsk);
1396}
1397NOKPROBE_SYMBOL(__do_page_fault);
1398
1399dotraplinkage void notrace
1400do_page_fault(struct pt_regs *regs, unsigned long error_code)
1401{
1402 unsigned long address = read_cr2(); /* Get the faulting address */
1403 enum ctx_state prev_state;
1404
1405 /*
1406 * We must have this function tagged with __kprobes, notrace and call
1407 * read_cr2() before calling anything else. To avoid calling any kind
1408 * of tracing machinery before we've observed the CR2 value.
1409 *
1410 * exception_{enter,exit}() contain all sorts of tracepoints.
1411 */
1412
1413 prev_state = exception_enter();
1414 __do_page_fault(regs, error_code, address);
1415 exception_exit(prev_state);
1416}
1417NOKPROBE_SYMBOL(do_page_fault);
1418
1419#ifdef CONFIG_TRACING
1420static nokprobe_inline void
1421trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1422 unsigned long error_code)
1423{
1424 if (user_mode(regs))
1425 trace_page_fault_user(address, regs, error_code);
1426 else
1427 trace_page_fault_kernel(address, regs, error_code);
1428}
1429
1430dotraplinkage void notrace
1431trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1432{
1433 /*
1434 * The exception_enter and tracepoint processing could
1435 * trigger another page faults (user space callchain
1436 * reading) and destroy the original cr2 value, so read
1437 * the faulting address now.
1438 */
1439 unsigned long address = read_cr2();
1440 enum ctx_state prev_state;
1441
1442 prev_state = exception_enter();
1443 trace_page_fault_entries(address, regs, error_code);
1444 __do_page_fault(regs, error_code, address);
1445 exception_exit(prev_state);
1446}
1447NOKPROBE_SYMBOL(trace_do_page_fault);
1448#endif /* CONFIG_TRACING */