Loading...
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/magic.h> /* STACK_END_MAGIC */
7#include <linux/sched.h> /* test_thread_flag(), ... */
8#include <linux/kdebug.h> /* oops_begin/end, ... */
9#include <linux/module.h> /* search_exception_table */
10#include <linux/bootmem.h> /* max_low_pfn */
11#include <linux/kprobes.h> /* __kprobes, ... */
12#include <linux/mmiotrace.h> /* kmmio_handler, ... */
13#include <linux/perf_event.h> /* perf_sw_event */
14#include <linux/hugetlb.h> /* hstate_index_to_shift */
15#include <linux/prefetch.h> /* prefetchw */
16
17#include <asm/traps.h> /* dotraplinkage, ... */
18#include <asm/pgalloc.h> /* pgd_*(), ... */
19#include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
20#include <asm/vsyscall.h>
21
22/*
23 * Page fault error code bits:
24 *
25 * bit 0 == 0: no page found 1: protection fault
26 * bit 1 == 0: read access 1: write access
27 * bit 2 == 0: kernel-mode access 1: user-mode access
28 * bit 3 == 1: use of reserved bit detected
29 * bit 4 == 1: fault was an instruction fetch
30 */
31enum x86_pf_error_code {
32
33 PF_PROT = 1 << 0,
34 PF_WRITE = 1 << 1,
35 PF_USER = 1 << 2,
36 PF_RSVD = 1 << 3,
37 PF_INSTR = 1 << 4,
38};
39
40/*
41 * Returns 0 if mmiotrace is disabled, or if the fault is not
42 * handled by mmiotrace:
43 */
44static inline int __kprobes
45kmmio_fault(struct pt_regs *regs, unsigned long addr)
46{
47 if (unlikely(is_kmmio_active()))
48 if (kmmio_handler(regs, addr) == 1)
49 return -1;
50 return 0;
51}
52
53static inline int __kprobes notify_page_fault(struct pt_regs *regs)
54{
55 int ret = 0;
56
57 /* kprobe_running() needs smp_processor_id() */
58 if (kprobes_built_in() && !user_mode_vm(regs)) {
59 preempt_disable();
60 if (kprobe_running() && kprobe_fault_handler(regs, 14))
61 ret = 1;
62 preempt_enable();
63 }
64
65 return ret;
66}
67
68/*
69 * Prefetch quirks:
70 *
71 * 32-bit mode:
72 *
73 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
74 * Check that here and ignore it.
75 *
76 * 64-bit mode:
77 *
78 * Sometimes the CPU reports invalid exceptions on prefetch.
79 * Check that here and ignore it.
80 *
81 * Opcode checker based on code by Richard Brunner.
82 */
83static inline int
84check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
85 unsigned char opcode, int *prefetch)
86{
87 unsigned char instr_hi = opcode & 0xf0;
88 unsigned char instr_lo = opcode & 0x0f;
89
90 switch (instr_hi) {
91 case 0x20:
92 case 0x30:
93 /*
94 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
95 * In X86_64 long mode, the CPU will signal invalid
96 * opcode if some of these prefixes are present so
97 * X86_64 will never get here anyway
98 */
99 return ((instr_lo & 7) == 0x6);
100#ifdef CONFIG_X86_64
101 case 0x40:
102 /*
103 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
104 * Need to figure out under what instruction mode the
105 * instruction was issued. Could check the LDT for lm,
106 * but for now it's good enough to assume that long
107 * mode only uses well known segments or kernel.
108 */
109 return (!user_mode(regs) || user_64bit_mode(regs));
110#endif
111 case 0x60:
112 /* 0x64 thru 0x67 are valid prefixes in all modes. */
113 return (instr_lo & 0xC) == 0x4;
114 case 0xF0:
115 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
116 return !instr_lo || (instr_lo>>1) == 1;
117 case 0x00:
118 /* Prefetch instruction is 0x0F0D or 0x0F18 */
119 if (probe_kernel_address(instr, opcode))
120 return 0;
121
122 *prefetch = (instr_lo == 0xF) &&
123 (opcode == 0x0D || opcode == 0x18);
124 return 0;
125 default:
126 return 0;
127 }
128}
129
130static int
131is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
132{
133 unsigned char *max_instr;
134 unsigned char *instr;
135 int prefetch = 0;
136
137 /*
138 * If it was a exec (instruction fetch) fault on NX page, then
139 * do not ignore the fault:
140 */
141 if (error_code & PF_INSTR)
142 return 0;
143
144 instr = (void *)convert_ip_to_linear(current, regs);
145 max_instr = instr + 15;
146
147 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
148 return 0;
149
150 while (instr < max_instr) {
151 unsigned char opcode;
152
153 if (probe_kernel_address(instr, opcode))
154 break;
155
156 instr++;
157
158 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
159 break;
160 }
161 return prefetch;
162}
163
164static void
165force_sig_info_fault(int si_signo, int si_code, unsigned long address,
166 struct task_struct *tsk, int fault)
167{
168 unsigned lsb = 0;
169 siginfo_t info;
170
171 info.si_signo = si_signo;
172 info.si_errno = 0;
173 info.si_code = si_code;
174 info.si_addr = (void __user *)address;
175 if (fault & VM_FAULT_HWPOISON_LARGE)
176 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
177 if (fault & VM_FAULT_HWPOISON)
178 lsb = PAGE_SHIFT;
179 info.si_addr_lsb = lsb;
180
181 force_sig_info(si_signo, &info, tsk);
182}
183
184DEFINE_SPINLOCK(pgd_lock);
185LIST_HEAD(pgd_list);
186
187#ifdef CONFIG_X86_32
188static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
189{
190 unsigned index = pgd_index(address);
191 pgd_t *pgd_k;
192 pud_t *pud, *pud_k;
193 pmd_t *pmd, *pmd_k;
194
195 pgd += index;
196 pgd_k = init_mm.pgd + index;
197
198 if (!pgd_present(*pgd_k))
199 return NULL;
200
201 /*
202 * set_pgd(pgd, *pgd_k); here would be useless on PAE
203 * and redundant with the set_pmd() on non-PAE. As would
204 * set_pud.
205 */
206 pud = pud_offset(pgd, address);
207 pud_k = pud_offset(pgd_k, address);
208 if (!pud_present(*pud_k))
209 return NULL;
210
211 pmd = pmd_offset(pud, address);
212 pmd_k = pmd_offset(pud_k, address);
213 if (!pmd_present(*pmd_k))
214 return NULL;
215
216 if (!pmd_present(*pmd))
217 set_pmd(pmd, *pmd_k);
218 else
219 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
220
221 return pmd_k;
222}
223
224void vmalloc_sync_all(void)
225{
226 unsigned long address;
227
228 if (SHARED_KERNEL_PMD)
229 return;
230
231 for (address = VMALLOC_START & PMD_MASK;
232 address >= TASK_SIZE && address < FIXADDR_TOP;
233 address += PMD_SIZE) {
234 struct page *page;
235
236 spin_lock(&pgd_lock);
237 list_for_each_entry(page, &pgd_list, lru) {
238 spinlock_t *pgt_lock;
239 pmd_t *ret;
240
241 /* the pgt_lock only for Xen */
242 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
243
244 spin_lock(pgt_lock);
245 ret = vmalloc_sync_one(page_address(page), address);
246 spin_unlock(pgt_lock);
247
248 if (!ret)
249 break;
250 }
251 spin_unlock(&pgd_lock);
252 }
253}
254
255/*
256 * 32-bit:
257 *
258 * Handle a fault on the vmalloc or module mapping area
259 */
260static noinline __kprobes int vmalloc_fault(unsigned long address)
261{
262 unsigned long pgd_paddr;
263 pmd_t *pmd_k;
264 pte_t *pte_k;
265
266 /* Make sure we are in vmalloc area: */
267 if (!(address >= VMALLOC_START && address < VMALLOC_END))
268 return -1;
269
270 WARN_ON_ONCE(in_nmi());
271
272 /*
273 * Synchronize this task's top level page-table
274 * with the 'reference' page table.
275 *
276 * Do _not_ use "current" here. We might be inside
277 * an interrupt in the middle of a task switch..
278 */
279 pgd_paddr = read_cr3();
280 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
281 if (!pmd_k)
282 return -1;
283
284 pte_k = pte_offset_kernel(pmd_k, address);
285 if (!pte_present(*pte_k))
286 return -1;
287
288 return 0;
289}
290
291/*
292 * Did it hit the DOS screen memory VA from vm86 mode?
293 */
294static inline void
295check_v8086_mode(struct pt_regs *regs, unsigned long address,
296 struct task_struct *tsk)
297{
298 unsigned long bit;
299
300 if (!v8086_mode(regs))
301 return;
302
303 bit = (address - 0xA0000) >> PAGE_SHIFT;
304 if (bit < 32)
305 tsk->thread.screen_bitmap |= 1 << bit;
306}
307
308static bool low_pfn(unsigned long pfn)
309{
310 return pfn < max_low_pfn;
311}
312
313static void dump_pagetable(unsigned long address)
314{
315 pgd_t *base = __va(read_cr3());
316 pgd_t *pgd = &base[pgd_index(address)];
317 pmd_t *pmd;
318 pte_t *pte;
319
320#ifdef CONFIG_X86_PAE
321 printk("*pdpt = %016Lx ", pgd_val(*pgd));
322 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
323 goto out;
324#endif
325 pmd = pmd_offset(pud_offset(pgd, address), address);
326 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
327
328 /*
329 * We must not directly access the pte in the highpte
330 * case if the page table is located in highmem.
331 * And let's rather not kmap-atomic the pte, just in case
332 * it's allocated already:
333 */
334 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
335 goto out;
336
337 pte = pte_offset_kernel(pmd, address);
338 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
339out:
340 printk("\n");
341}
342
343#else /* CONFIG_X86_64: */
344
345void vmalloc_sync_all(void)
346{
347 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
348}
349
350/*
351 * 64-bit:
352 *
353 * Handle a fault on the vmalloc area
354 *
355 * This assumes no large pages in there.
356 */
357static noinline __kprobes int vmalloc_fault(unsigned long address)
358{
359 pgd_t *pgd, *pgd_ref;
360 pud_t *pud, *pud_ref;
361 pmd_t *pmd, *pmd_ref;
362 pte_t *pte, *pte_ref;
363
364 /* Make sure we are in vmalloc area: */
365 if (!(address >= VMALLOC_START && address < VMALLOC_END))
366 return -1;
367
368 WARN_ON_ONCE(in_nmi());
369
370 /*
371 * Copy kernel mappings over when needed. This can also
372 * happen within a race in page table update. In the later
373 * case just flush:
374 */
375 pgd = pgd_offset(current->active_mm, address);
376 pgd_ref = pgd_offset_k(address);
377 if (pgd_none(*pgd_ref))
378 return -1;
379
380 if (pgd_none(*pgd))
381 set_pgd(pgd, *pgd_ref);
382 else
383 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
384
385 /*
386 * Below here mismatches are bugs because these lower tables
387 * are shared:
388 */
389
390 pud = pud_offset(pgd, address);
391 pud_ref = pud_offset(pgd_ref, address);
392 if (pud_none(*pud_ref))
393 return -1;
394
395 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
396 BUG();
397
398 pmd = pmd_offset(pud, address);
399 pmd_ref = pmd_offset(pud_ref, address);
400 if (pmd_none(*pmd_ref))
401 return -1;
402
403 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
404 BUG();
405
406 pte_ref = pte_offset_kernel(pmd_ref, address);
407 if (!pte_present(*pte_ref))
408 return -1;
409
410 pte = pte_offset_kernel(pmd, address);
411
412 /*
413 * Don't use pte_page here, because the mappings can point
414 * outside mem_map, and the NUMA hash lookup cannot handle
415 * that:
416 */
417 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
418 BUG();
419
420 return 0;
421}
422
423static const char errata93_warning[] =
424KERN_ERR
425"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
426"******* Working around it, but it may cause SEGVs or burn power.\n"
427"******* Please consider a BIOS update.\n"
428"******* Disabling USB legacy in the BIOS may also help.\n";
429
430/*
431 * No vm86 mode in 64-bit mode:
432 */
433static inline void
434check_v8086_mode(struct pt_regs *regs, unsigned long address,
435 struct task_struct *tsk)
436{
437}
438
439static int bad_address(void *p)
440{
441 unsigned long dummy;
442
443 return probe_kernel_address((unsigned long *)p, dummy);
444}
445
446static void dump_pagetable(unsigned long address)
447{
448 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
449 pgd_t *pgd = base + pgd_index(address);
450 pud_t *pud;
451 pmd_t *pmd;
452 pte_t *pte;
453
454 if (bad_address(pgd))
455 goto bad;
456
457 printk("PGD %lx ", pgd_val(*pgd));
458
459 if (!pgd_present(*pgd))
460 goto out;
461
462 pud = pud_offset(pgd, address);
463 if (bad_address(pud))
464 goto bad;
465
466 printk("PUD %lx ", pud_val(*pud));
467 if (!pud_present(*pud) || pud_large(*pud))
468 goto out;
469
470 pmd = pmd_offset(pud, address);
471 if (bad_address(pmd))
472 goto bad;
473
474 printk("PMD %lx ", pmd_val(*pmd));
475 if (!pmd_present(*pmd) || pmd_large(*pmd))
476 goto out;
477
478 pte = pte_offset_kernel(pmd, address);
479 if (bad_address(pte))
480 goto bad;
481
482 printk("PTE %lx", pte_val(*pte));
483out:
484 printk("\n");
485 return;
486bad:
487 printk("BAD\n");
488}
489
490#endif /* CONFIG_X86_64 */
491
492/*
493 * Workaround for K8 erratum #93 & buggy BIOS.
494 *
495 * BIOS SMM functions are required to use a specific workaround
496 * to avoid corruption of the 64bit RIP register on C stepping K8.
497 *
498 * A lot of BIOS that didn't get tested properly miss this.
499 *
500 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
501 * Try to work around it here.
502 *
503 * Note we only handle faults in kernel here.
504 * Does nothing on 32-bit.
505 */
506static int is_errata93(struct pt_regs *regs, unsigned long address)
507{
508#ifdef CONFIG_X86_64
509 if (address != regs->ip)
510 return 0;
511
512 if ((address >> 32) != 0)
513 return 0;
514
515 address |= 0xffffffffUL << 32;
516 if ((address >= (u64)_stext && address <= (u64)_etext) ||
517 (address >= MODULES_VADDR && address <= MODULES_END)) {
518 printk_once(errata93_warning);
519 regs->ip = address;
520 return 1;
521 }
522#endif
523 return 0;
524}
525
526/*
527 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
528 * to illegal addresses >4GB.
529 *
530 * We catch this in the page fault handler because these addresses
531 * are not reachable. Just detect this case and return. Any code
532 * segment in LDT is compatibility mode.
533 */
534static int is_errata100(struct pt_regs *regs, unsigned long address)
535{
536#ifdef CONFIG_X86_64
537 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
538 return 1;
539#endif
540 return 0;
541}
542
543static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
544{
545#ifdef CONFIG_X86_F00F_BUG
546 unsigned long nr;
547
548 /*
549 * Pentium F0 0F C7 C8 bug workaround:
550 */
551 if (boot_cpu_data.f00f_bug) {
552 nr = (address - idt_descr.address) >> 3;
553
554 if (nr == 6) {
555 do_invalid_op(regs, 0);
556 return 1;
557 }
558 }
559#endif
560 return 0;
561}
562
563static const char nx_warning[] = KERN_CRIT
564"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
565
566static void
567show_fault_oops(struct pt_regs *regs, unsigned long error_code,
568 unsigned long address)
569{
570 if (!oops_may_print())
571 return;
572
573 if (error_code & PF_INSTR) {
574 unsigned int level;
575
576 pte_t *pte = lookup_address(address, &level);
577
578 if (pte && pte_present(*pte) && !pte_exec(*pte))
579 printk(nx_warning, current_uid());
580 }
581
582 printk(KERN_ALERT "BUG: unable to handle kernel ");
583 if (address < PAGE_SIZE)
584 printk(KERN_CONT "NULL pointer dereference");
585 else
586 printk(KERN_CONT "paging request");
587
588 printk(KERN_CONT " at %p\n", (void *) address);
589 printk(KERN_ALERT "IP:");
590 printk_address(regs->ip, 1);
591
592 dump_pagetable(address);
593}
594
595static noinline void
596pgtable_bad(struct pt_regs *regs, unsigned long error_code,
597 unsigned long address)
598{
599 struct task_struct *tsk;
600 unsigned long flags;
601 int sig;
602
603 flags = oops_begin();
604 tsk = current;
605 sig = SIGKILL;
606
607 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
608 tsk->comm, address);
609 dump_pagetable(address);
610
611 tsk->thread.cr2 = address;
612 tsk->thread.trap_no = 14;
613 tsk->thread.error_code = error_code;
614
615 if (__die("Bad pagetable", regs, error_code))
616 sig = 0;
617
618 oops_end(flags, regs, sig);
619}
620
621static noinline void
622no_context(struct pt_regs *regs, unsigned long error_code,
623 unsigned long address)
624{
625 struct task_struct *tsk = current;
626 unsigned long *stackend;
627 unsigned long flags;
628 int sig;
629
630 /* Are we prepared to handle this kernel fault? */
631 if (fixup_exception(regs))
632 return;
633
634 /*
635 * 32-bit:
636 *
637 * Valid to do another page fault here, because if this fault
638 * had been triggered by is_prefetch fixup_exception would have
639 * handled it.
640 *
641 * 64-bit:
642 *
643 * Hall of shame of CPU/BIOS bugs.
644 */
645 if (is_prefetch(regs, error_code, address))
646 return;
647
648 if (is_errata93(regs, address))
649 return;
650
651 /*
652 * Oops. The kernel tried to access some bad page. We'll have to
653 * terminate things with extreme prejudice:
654 */
655 flags = oops_begin();
656
657 show_fault_oops(regs, error_code, address);
658
659 stackend = end_of_stack(tsk);
660 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
661 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
662
663 tsk->thread.cr2 = address;
664 tsk->thread.trap_no = 14;
665 tsk->thread.error_code = error_code;
666
667 sig = SIGKILL;
668 if (__die("Oops", regs, error_code))
669 sig = 0;
670
671 /* Executive summary in case the body of the oops scrolled away */
672 printk(KERN_EMERG "CR2: %016lx\n", address);
673
674 oops_end(flags, regs, sig);
675}
676
677/*
678 * Print out info about fatal segfaults, if the show_unhandled_signals
679 * sysctl is set:
680 */
681static inline void
682show_signal_msg(struct pt_regs *regs, unsigned long error_code,
683 unsigned long address, struct task_struct *tsk)
684{
685 if (!unhandled_signal(tsk, SIGSEGV))
686 return;
687
688 if (!printk_ratelimit())
689 return;
690
691 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
692 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
693 tsk->comm, task_pid_nr(tsk), address,
694 (void *)regs->ip, (void *)regs->sp, error_code);
695
696 print_vma_addr(KERN_CONT " in ", regs->ip);
697
698 printk(KERN_CONT "\n");
699}
700
701static void
702__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
703 unsigned long address, int si_code)
704{
705 struct task_struct *tsk = current;
706
707 /* User mode accesses just cause a SIGSEGV */
708 if (error_code & PF_USER) {
709 /*
710 * It's possible to have interrupts off here:
711 */
712 local_irq_enable();
713
714 /*
715 * Valid to do another page fault here because this one came
716 * from user space:
717 */
718 if (is_prefetch(regs, error_code, address))
719 return;
720
721 if (is_errata100(regs, address))
722 return;
723
724#ifdef CONFIG_X86_64
725 /*
726 * Instruction fetch faults in the vsyscall page might need
727 * emulation.
728 */
729 if (unlikely((error_code & PF_INSTR) &&
730 ((address & ~0xfff) == VSYSCALL_START))) {
731 if (emulate_vsyscall(regs, address))
732 return;
733 }
734#endif
735
736 if (unlikely(show_unhandled_signals))
737 show_signal_msg(regs, error_code, address, tsk);
738
739 /* Kernel addresses are always protection faults: */
740 tsk->thread.cr2 = address;
741 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
742 tsk->thread.trap_no = 14;
743
744 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
745
746 return;
747 }
748
749 if (is_f00f_bug(regs, address))
750 return;
751
752 no_context(regs, error_code, address);
753}
754
755static noinline void
756bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
757 unsigned long address)
758{
759 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
760}
761
762static void
763__bad_area(struct pt_regs *regs, unsigned long error_code,
764 unsigned long address, int si_code)
765{
766 struct mm_struct *mm = current->mm;
767
768 /*
769 * Something tried to access memory that isn't in our memory map..
770 * Fix it, but check if it's kernel or user first..
771 */
772 up_read(&mm->mmap_sem);
773
774 __bad_area_nosemaphore(regs, error_code, address, si_code);
775}
776
777static noinline void
778bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
779{
780 __bad_area(regs, error_code, address, SEGV_MAPERR);
781}
782
783static noinline void
784bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
785 unsigned long address)
786{
787 __bad_area(regs, error_code, address, SEGV_ACCERR);
788}
789
790/* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
791static void
792out_of_memory(struct pt_regs *regs, unsigned long error_code,
793 unsigned long address)
794{
795 /*
796 * We ran out of memory, call the OOM killer, and return the userspace
797 * (which will retry the fault, or kill us if we got oom-killed):
798 */
799 up_read(¤t->mm->mmap_sem);
800
801 pagefault_out_of_memory();
802}
803
804static void
805do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
806 unsigned int fault)
807{
808 struct task_struct *tsk = current;
809 struct mm_struct *mm = tsk->mm;
810 int code = BUS_ADRERR;
811
812 up_read(&mm->mmap_sem);
813
814 /* Kernel mode? Handle exceptions or die: */
815 if (!(error_code & PF_USER)) {
816 no_context(regs, error_code, address);
817 return;
818 }
819
820 /* User-space => ok to do another page fault: */
821 if (is_prefetch(regs, error_code, address))
822 return;
823
824 tsk->thread.cr2 = address;
825 tsk->thread.error_code = error_code;
826 tsk->thread.trap_no = 14;
827
828#ifdef CONFIG_MEMORY_FAILURE
829 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
830 printk(KERN_ERR
831 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
832 tsk->comm, tsk->pid, address);
833 code = BUS_MCEERR_AR;
834 }
835#endif
836 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
837}
838
839static noinline int
840mm_fault_error(struct pt_regs *regs, unsigned long error_code,
841 unsigned long address, unsigned int fault)
842{
843 /*
844 * Pagefault was interrupted by SIGKILL. We have no reason to
845 * continue pagefault.
846 */
847 if (fatal_signal_pending(current)) {
848 if (!(fault & VM_FAULT_RETRY))
849 up_read(¤t->mm->mmap_sem);
850 if (!(error_code & PF_USER))
851 no_context(regs, error_code, address);
852 return 1;
853 }
854 if (!(fault & VM_FAULT_ERROR))
855 return 0;
856
857 if (fault & VM_FAULT_OOM) {
858 /* Kernel mode? Handle exceptions or die: */
859 if (!(error_code & PF_USER)) {
860 up_read(¤t->mm->mmap_sem);
861 no_context(regs, error_code, address);
862 return 1;
863 }
864
865 out_of_memory(regs, error_code, address);
866 } else {
867 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
868 VM_FAULT_HWPOISON_LARGE))
869 do_sigbus(regs, error_code, address, fault);
870 else
871 BUG();
872 }
873 return 1;
874}
875
876static int spurious_fault_check(unsigned long error_code, pte_t *pte)
877{
878 if ((error_code & PF_WRITE) && !pte_write(*pte))
879 return 0;
880
881 if ((error_code & PF_INSTR) && !pte_exec(*pte))
882 return 0;
883
884 return 1;
885}
886
887/*
888 * Handle a spurious fault caused by a stale TLB entry.
889 *
890 * This allows us to lazily refresh the TLB when increasing the
891 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
892 * eagerly is very expensive since that implies doing a full
893 * cross-processor TLB flush, even if no stale TLB entries exist
894 * on other processors.
895 *
896 * There are no security implications to leaving a stale TLB when
897 * increasing the permissions on a page.
898 */
899static noinline __kprobes int
900spurious_fault(unsigned long error_code, unsigned long address)
901{
902 pgd_t *pgd;
903 pud_t *pud;
904 pmd_t *pmd;
905 pte_t *pte;
906 int ret;
907
908 /* Reserved-bit violation or user access to kernel space? */
909 if (error_code & (PF_USER | PF_RSVD))
910 return 0;
911
912 pgd = init_mm.pgd + pgd_index(address);
913 if (!pgd_present(*pgd))
914 return 0;
915
916 pud = pud_offset(pgd, address);
917 if (!pud_present(*pud))
918 return 0;
919
920 if (pud_large(*pud))
921 return spurious_fault_check(error_code, (pte_t *) pud);
922
923 pmd = pmd_offset(pud, address);
924 if (!pmd_present(*pmd))
925 return 0;
926
927 if (pmd_large(*pmd))
928 return spurious_fault_check(error_code, (pte_t *) pmd);
929
930 /*
931 * Note: don't use pte_present() here, since it returns true
932 * if the _PAGE_PROTNONE bit is set. However, this aliases the
933 * _PAGE_GLOBAL bit, which for kernel pages give false positives
934 * when CONFIG_DEBUG_PAGEALLOC is used.
935 */
936 pte = pte_offset_kernel(pmd, address);
937 if (!(pte_flags(*pte) & _PAGE_PRESENT))
938 return 0;
939
940 ret = spurious_fault_check(error_code, pte);
941 if (!ret)
942 return 0;
943
944 /*
945 * Make sure we have permissions in PMD.
946 * If not, then there's a bug in the page tables:
947 */
948 ret = spurious_fault_check(error_code, (pte_t *) pmd);
949 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
950
951 return ret;
952}
953
954int show_unhandled_signals = 1;
955
956static inline int
957access_error(unsigned long error_code, struct vm_area_struct *vma)
958{
959 if (error_code & PF_WRITE) {
960 /* write, present and write, not present: */
961 if (unlikely(!(vma->vm_flags & VM_WRITE)))
962 return 1;
963 return 0;
964 }
965
966 /* read, present: */
967 if (unlikely(error_code & PF_PROT))
968 return 1;
969
970 /* read, not present: */
971 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
972 return 1;
973
974 return 0;
975}
976
977static int fault_in_kernel_space(unsigned long address)
978{
979 return address >= TASK_SIZE_MAX;
980}
981
982/*
983 * This routine handles page faults. It determines the address,
984 * and the problem, and then passes it off to one of the appropriate
985 * routines.
986 */
987dotraplinkage void __kprobes
988do_page_fault(struct pt_regs *regs, unsigned long error_code)
989{
990 struct vm_area_struct *vma;
991 struct task_struct *tsk;
992 unsigned long address;
993 struct mm_struct *mm;
994 int fault;
995 int write = error_code & PF_WRITE;
996 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
997 (write ? FAULT_FLAG_WRITE : 0);
998
999 tsk = current;
1000 mm = tsk->mm;
1001
1002 /* Get the faulting address: */
1003 address = read_cr2();
1004
1005 /*
1006 * Detect and handle instructions that would cause a page fault for
1007 * both a tracked kernel page and a userspace page.
1008 */
1009 if (kmemcheck_active(regs))
1010 kmemcheck_hide(regs);
1011 prefetchw(&mm->mmap_sem);
1012
1013 if (unlikely(kmmio_fault(regs, address)))
1014 return;
1015
1016 /*
1017 * We fault-in kernel-space virtual memory on-demand. The
1018 * 'reference' page table is init_mm.pgd.
1019 *
1020 * NOTE! We MUST NOT take any locks for this case. We may
1021 * be in an interrupt or a critical region, and should
1022 * only copy the information from the master page table,
1023 * nothing more.
1024 *
1025 * This verifies that the fault happens in kernel space
1026 * (error_code & 4) == 0, and that the fault was not a
1027 * protection error (error_code & 9) == 0.
1028 */
1029 if (unlikely(fault_in_kernel_space(address))) {
1030 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1031 if (vmalloc_fault(address) >= 0)
1032 return;
1033
1034 if (kmemcheck_fault(regs, address, error_code))
1035 return;
1036 }
1037
1038 /* Can handle a stale RO->RW TLB: */
1039 if (spurious_fault(error_code, address))
1040 return;
1041
1042 /* kprobes don't want to hook the spurious faults: */
1043 if (notify_page_fault(regs))
1044 return;
1045 /*
1046 * Don't take the mm semaphore here. If we fixup a prefetch
1047 * fault we could otherwise deadlock:
1048 */
1049 bad_area_nosemaphore(regs, error_code, address);
1050
1051 return;
1052 }
1053
1054 /* kprobes don't want to hook the spurious faults: */
1055 if (unlikely(notify_page_fault(regs)))
1056 return;
1057 /*
1058 * It's safe to allow irq's after cr2 has been saved and the
1059 * vmalloc fault has been handled.
1060 *
1061 * User-mode registers count as a user access even for any
1062 * potential system fault or CPU buglet:
1063 */
1064 if (user_mode_vm(regs)) {
1065 local_irq_enable();
1066 error_code |= PF_USER;
1067 } else {
1068 if (regs->flags & X86_EFLAGS_IF)
1069 local_irq_enable();
1070 }
1071
1072 if (unlikely(error_code & PF_RSVD))
1073 pgtable_bad(regs, error_code, address);
1074
1075 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1076
1077 /*
1078 * If we're in an interrupt, have no user context or are running
1079 * in an atomic region then we must not take the fault:
1080 */
1081 if (unlikely(in_atomic() || !mm)) {
1082 bad_area_nosemaphore(regs, error_code, address);
1083 return;
1084 }
1085
1086 /*
1087 * When running in the kernel we expect faults to occur only to
1088 * addresses in user space. All other faults represent errors in
1089 * the kernel and should generate an OOPS. Unfortunately, in the
1090 * case of an erroneous fault occurring in a code path which already
1091 * holds mmap_sem we will deadlock attempting to validate the fault
1092 * against the address space. Luckily the kernel only validly
1093 * references user space from well defined areas of code, which are
1094 * listed in the exceptions table.
1095 *
1096 * As the vast majority of faults will be valid we will only perform
1097 * the source reference check when there is a possibility of a
1098 * deadlock. Attempt to lock the address space, if we cannot we then
1099 * validate the source. If this is invalid we can skip the address
1100 * space check, thus avoiding the deadlock:
1101 */
1102 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1103 if ((error_code & PF_USER) == 0 &&
1104 !search_exception_tables(regs->ip)) {
1105 bad_area_nosemaphore(regs, error_code, address);
1106 return;
1107 }
1108retry:
1109 down_read(&mm->mmap_sem);
1110 } else {
1111 /*
1112 * The above down_read_trylock() might have succeeded in
1113 * which case we'll have missed the might_sleep() from
1114 * down_read():
1115 */
1116 might_sleep();
1117 }
1118
1119 vma = find_vma(mm, address);
1120 if (unlikely(!vma)) {
1121 bad_area(regs, error_code, address);
1122 return;
1123 }
1124 if (likely(vma->vm_start <= address))
1125 goto good_area;
1126 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1127 bad_area(regs, error_code, address);
1128 return;
1129 }
1130 if (error_code & PF_USER) {
1131 /*
1132 * Accessing the stack below %sp is always a bug.
1133 * The large cushion allows instructions like enter
1134 * and pusha to work. ("enter $65535, $31" pushes
1135 * 32 pointers and then decrements %sp by 65535.)
1136 */
1137 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1138 bad_area(regs, error_code, address);
1139 return;
1140 }
1141 }
1142 if (unlikely(expand_stack(vma, address))) {
1143 bad_area(regs, error_code, address);
1144 return;
1145 }
1146
1147 /*
1148 * Ok, we have a good vm_area for this memory access, so
1149 * we can handle it..
1150 */
1151good_area:
1152 if (unlikely(access_error(error_code, vma))) {
1153 bad_area_access_error(regs, error_code, address);
1154 return;
1155 }
1156
1157 /*
1158 * If for any reason at all we couldn't handle the fault,
1159 * make sure we exit gracefully rather than endlessly redo
1160 * the fault:
1161 */
1162 fault = handle_mm_fault(mm, vma, address, flags);
1163
1164 if (unlikely(fault & (VM_FAULT_RETRY|VM_FAULT_ERROR))) {
1165 if (mm_fault_error(regs, error_code, address, fault))
1166 return;
1167 }
1168
1169 /*
1170 * Major/minor page fault accounting is only done on the
1171 * initial attempt. If we go through a retry, it is extremely
1172 * likely that the page will be found in page cache at that point.
1173 */
1174 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1175 if (fault & VM_FAULT_MAJOR) {
1176 tsk->maj_flt++;
1177 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1178 regs, address);
1179 } else {
1180 tsk->min_flt++;
1181 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1182 regs, address);
1183 }
1184 if (fault & VM_FAULT_RETRY) {
1185 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1186 * of starvation. */
1187 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1188 goto retry;
1189 }
1190 }
1191
1192 check_v8086_mode(regs, address, tsk);
1193
1194 up_read(&mm->mmap_sem);
1195}
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/magic.h> /* STACK_END_MAGIC */
7#include <linux/sched.h> /* test_thread_flag(), ... */
8#include <linux/kdebug.h> /* oops_begin/end, ... */
9#include <linux/module.h> /* search_exception_table */
10#include <linux/bootmem.h> /* max_low_pfn */
11#include <linux/kprobes.h> /* __kprobes, ... */
12#include <linux/mmiotrace.h> /* kmmio_handler, ... */
13#include <linux/perf_event.h> /* perf_sw_event */
14#include <linux/hugetlb.h> /* hstate_index_to_shift */
15#include <linux/prefetch.h> /* prefetchw */
16#include <linux/context_tracking.h> /* exception_enter(), ... */
17
18#include <asm/traps.h> /* dotraplinkage, ... */
19#include <asm/pgalloc.h> /* pgd_*(), ... */
20#include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
21#include <asm/fixmap.h> /* VSYSCALL_START */
22
23#define CREATE_TRACE_POINTS
24#include <asm/trace/exceptions.h>
25
26/*
27 * Page fault error code bits:
28 *
29 * bit 0 == 0: no page found 1: protection fault
30 * bit 1 == 0: read access 1: write access
31 * bit 2 == 0: kernel-mode access 1: user-mode access
32 * bit 3 == 1: use of reserved bit detected
33 * bit 4 == 1: fault was an instruction fetch
34 */
35enum x86_pf_error_code {
36
37 PF_PROT = 1 << 0,
38 PF_WRITE = 1 << 1,
39 PF_USER = 1 << 2,
40 PF_RSVD = 1 << 3,
41 PF_INSTR = 1 << 4,
42};
43
44/*
45 * Returns 0 if mmiotrace is disabled, or if the fault is not
46 * handled by mmiotrace:
47 */
48static inline int __kprobes
49kmmio_fault(struct pt_regs *regs, unsigned long addr)
50{
51 if (unlikely(is_kmmio_active()))
52 if (kmmio_handler(regs, addr) == 1)
53 return -1;
54 return 0;
55}
56
57static inline int __kprobes kprobes_fault(struct pt_regs *regs)
58{
59 int ret = 0;
60
61 /* kprobe_running() needs smp_processor_id() */
62 if (kprobes_built_in() && !user_mode_vm(regs)) {
63 preempt_disable();
64 if (kprobe_running() && kprobe_fault_handler(regs, 14))
65 ret = 1;
66 preempt_enable();
67 }
68
69 return ret;
70}
71
72/*
73 * Prefetch quirks:
74 *
75 * 32-bit mode:
76 *
77 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
78 * Check that here and ignore it.
79 *
80 * 64-bit mode:
81 *
82 * Sometimes the CPU reports invalid exceptions on prefetch.
83 * Check that here and ignore it.
84 *
85 * Opcode checker based on code by Richard Brunner.
86 */
87static inline int
88check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
89 unsigned char opcode, int *prefetch)
90{
91 unsigned char instr_hi = opcode & 0xf0;
92 unsigned char instr_lo = opcode & 0x0f;
93
94 switch (instr_hi) {
95 case 0x20:
96 case 0x30:
97 /*
98 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
99 * In X86_64 long mode, the CPU will signal invalid
100 * opcode if some of these prefixes are present so
101 * X86_64 will never get here anyway
102 */
103 return ((instr_lo & 7) == 0x6);
104#ifdef CONFIG_X86_64
105 case 0x40:
106 /*
107 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
108 * Need to figure out under what instruction mode the
109 * instruction was issued. Could check the LDT for lm,
110 * but for now it's good enough to assume that long
111 * mode only uses well known segments or kernel.
112 */
113 return (!user_mode(regs) || user_64bit_mode(regs));
114#endif
115 case 0x60:
116 /* 0x64 thru 0x67 are valid prefixes in all modes. */
117 return (instr_lo & 0xC) == 0x4;
118 case 0xF0:
119 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
120 return !instr_lo || (instr_lo>>1) == 1;
121 case 0x00:
122 /* Prefetch instruction is 0x0F0D or 0x0F18 */
123 if (probe_kernel_address(instr, opcode))
124 return 0;
125
126 *prefetch = (instr_lo == 0xF) &&
127 (opcode == 0x0D || opcode == 0x18);
128 return 0;
129 default:
130 return 0;
131 }
132}
133
134static int
135is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
136{
137 unsigned char *max_instr;
138 unsigned char *instr;
139 int prefetch = 0;
140
141 /*
142 * If it was a exec (instruction fetch) fault on NX page, then
143 * do not ignore the fault:
144 */
145 if (error_code & PF_INSTR)
146 return 0;
147
148 instr = (void *)convert_ip_to_linear(current, regs);
149 max_instr = instr + 15;
150
151 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
152 return 0;
153
154 while (instr < max_instr) {
155 unsigned char opcode;
156
157 if (probe_kernel_address(instr, opcode))
158 break;
159
160 instr++;
161
162 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
163 break;
164 }
165 return prefetch;
166}
167
168static void
169force_sig_info_fault(int si_signo, int si_code, unsigned long address,
170 struct task_struct *tsk, int fault)
171{
172 unsigned lsb = 0;
173 siginfo_t info;
174
175 info.si_signo = si_signo;
176 info.si_errno = 0;
177 info.si_code = si_code;
178 info.si_addr = (void __user *)address;
179 if (fault & VM_FAULT_HWPOISON_LARGE)
180 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
181 if (fault & VM_FAULT_HWPOISON)
182 lsb = PAGE_SHIFT;
183 info.si_addr_lsb = lsb;
184
185 force_sig_info(si_signo, &info, tsk);
186}
187
188DEFINE_SPINLOCK(pgd_lock);
189LIST_HEAD(pgd_list);
190
191#ifdef CONFIG_X86_32
192static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
193{
194 unsigned index = pgd_index(address);
195 pgd_t *pgd_k;
196 pud_t *pud, *pud_k;
197 pmd_t *pmd, *pmd_k;
198
199 pgd += index;
200 pgd_k = init_mm.pgd + index;
201
202 if (!pgd_present(*pgd_k))
203 return NULL;
204
205 /*
206 * set_pgd(pgd, *pgd_k); here would be useless on PAE
207 * and redundant with the set_pmd() on non-PAE. As would
208 * set_pud.
209 */
210 pud = pud_offset(pgd, address);
211 pud_k = pud_offset(pgd_k, address);
212 if (!pud_present(*pud_k))
213 return NULL;
214
215 pmd = pmd_offset(pud, address);
216 pmd_k = pmd_offset(pud_k, address);
217 if (!pmd_present(*pmd_k))
218 return NULL;
219
220 if (!pmd_present(*pmd))
221 set_pmd(pmd, *pmd_k);
222 else
223 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
224
225 return pmd_k;
226}
227
228void vmalloc_sync_all(void)
229{
230 unsigned long address;
231
232 if (SHARED_KERNEL_PMD)
233 return;
234
235 for (address = VMALLOC_START & PMD_MASK;
236 address >= TASK_SIZE && address < FIXADDR_TOP;
237 address += PMD_SIZE) {
238 struct page *page;
239
240 spin_lock(&pgd_lock);
241 list_for_each_entry(page, &pgd_list, lru) {
242 spinlock_t *pgt_lock;
243 pmd_t *ret;
244
245 /* the pgt_lock only for Xen */
246 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
247
248 spin_lock(pgt_lock);
249 ret = vmalloc_sync_one(page_address(page), address);
250 spin_unlock(pgt_lock);
251
252 if (!ret)
253 break;
254 }
255 spin_unlock(&pgd_lock);
256 }
257}
258
259/*
260 * 32-bit:
261 *
262 * Handle a fault on the vmalloc or module mapping area
263 */
264static noinline __kprobes int vmalloc_fault(unsigned long address)
265{
266 unsigned long pgd_paddr;
267 pmd_t *pmd_k;
268 pte_t *pte_k;
269
270 /* Make sure we are in vmalloc area: */
271 if (!(address >= VMALLOC_START && address < VMALLOC_END))
272 return -1;
273
274 WARN_ON_ONCE(in_nmi());
275
276 /*
277 * Synchronize this task's top level page-table
278 * with the 'reference' page table.
279 *
280 * Do _not_ use "current" here. We might be inside
281 * an interrupt in the middle of a task switch..
282 */
283 pgd_paddr = read_cr3();
284 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
285 if (!pmd_k)
286 return -1;
287
288 pte_k = pte_offset_kernel(pmd_k, address);
289 if (!pte_present(*pte_k))
290 return -1;
291
292 return 0;
293}
294
295/*
296 * Did it hit the DOS screen memory VA from vm86 mode?
297 */
298static inline void
299check_v8086_mode(struct pt_regs *regs, unsigned long address,
300 struct task_struct *tsk)
301{
302 unsigned long bit;
303
304 if (!v8086_mode(regs))
305 return;
306
307 bit = (address - 0xA0000) >> PAGE_SHIFT;
308 if (bit < 32)
309 tsk->thread.screen_bitmap |= 1 << bit;
310}
311
312static bool low_pfn(unsigned long pfn)
313{
314 return pfn < max_low_pfn;
315}
316
317static void dump_pagetable(unsigned long address)
318{
319 pgd_t *base = __va(read_cr3());
320 pgd_t *pgd = &base[pgd_index(address)];
321 pmd_t *pmd;
322 pte_t *pte;
323
324#ifdef CONFIG_X86_PAE
325 printk("*pdpt = %016Lx ", pgd_val(*pgd));
326 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
327 goto out;
328#endif
329 pmd = pmd_offset(pud_offset(pgd, address), address);
330 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
331
332 /*
333 * We must not directly access the pte in the highpte
334 * case if the page table is located in highmem.
335 * And let's rather not kmap-atomic the pte, just in case
336 * it's allocated already:
337 */
338 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
339 goto out;
340
341 pte = pte_offset_kernel(pmd, address);
342 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
343out:
344 printk("\n");
345}
346
347#else /* CONFIG_X86_64: */
348
349void vmalloc_sync_all(void)
350{
351 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
352}
353
354/*
355 * 64-bit:
356 *
357 * Handle a fault on the vmalloc area
358 *
359 * This assumes no large pages in there.
360 */
361static noinline __kprobes int vmalloc_fault(unsigned long address)
362{
363 pgd_t *pgd, *pgd_ref;
364 pud_t *pud, *pud_ref;
365 pmd_t *pmd, *pmd_ref;
366 pte_t *pte, *pte_ref;
367
368 /* Make sure we are in vmalloc area: */
369 if (!(address >= VMALLOC_START && address < VMALLOC_END))
370 return -1;
371
372 WARN_ON_ONCE(in_nmi());
373
374 /*
375 * Copy kernel mappings over when needed. This can also
376 * happen within a race in page table update. In the later
377 * case just flush:
378 */
379 pgd = pgd_offset(current->active_mm, address);
380 pgd_ref = pgd_offset_k(address);
381 if (pgd_none(*pgd_ref))
382 return -1;
383
384 if (pgd_none(*pgd)) {
385 set_pgd(pgd, *pgd_ref);
386 arch_flush_lazy_mmu_mode();
387 } else {
388 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
389 }
390
391 /*
392 * Below here mismatches are bugs because these lower tables
393 * are shared:
394 */
395
396 pud = pud_offset(pgd, address);
397 pud_ref = pud_offset(pgd_ref, address);
398 if (pud_none(*pud_ref))
399 return -1;
400
401 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
402 BUG();
403
404 pmd = pmd_offset(pud, address);
405 pmd_ref = pmd_offset(pud_ref, address);
406 if (pmd_none(*pmd_ref))
407 return -1;
408
409 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
410 BUG();
411
412 pte_ref = pte_offset_kernel(pmd_ref, address);
413 if (!pte_present(*pte_ref))
414 return -1;
415
416 pte = pte_offset_kernel(pmd, address);
417
418 /*
419 * Don't use pte_page here, because the mappings can point
420 * outside mem_map, and the NUMA hash lookup cannot handle
421 * that:
422 */
423 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
424 BUG();
425
426 return 0;
427}
428
429#ifdef CONFIG_CPU_SUP_AMD
430static const char errata93_warning[] =
431KERN_ERR
432"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
433"******* Working around it, but it may cause SEGVs or burn power.\n"
434"******* Please consider a BIOS update.\n"
435"******* Disabling USB legacy in the BIOS may also help.\n";
436#endif
437
438/*
439 * No vm86 mode in 64-bit mode:
440 */
441static inline void
442check_v8086_mode(struct pt_regs *regs, unsigned long address,
443 struct task_struct *tsk)
444{
445}
446
447static int bad_address(void *p)
448{
449 unsigned long dummy;
450
451 return probe_kernel_address((unsigned long *)p, dummy);
452}
453
454static void dump_pagetable(unsigned long address)
455{
456 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
457 pgd_t *pgd = base + pgd_index(address);
458 pud_t *pud;
459 pmd_t *pmd;
460 pte_t *pte;
461
462 if (bad_address(pgd))
463 goto bad;
464
465 printk("PGD %lx ", pgd_val(*pgd));
466
467 if (!pgd_present(*pgd))
468 goto out;
469
470 pud = pud_offset(pgd, address);
471 if (bad_address(pud))
472 goto bad;
473
474 printk("PUD %lx ", pud_val(*pud));
475 if (!pud_present(*pud) || pud_large(*pud))
476 goto out;
477
478 pmd = pmd_offset(pud, address);
479 if (bad_address(pmd))
480 goto bad;
481
482 printk("PMD %lx ", pmd_val(*pmd));
483 if (!pmd_present(*pmd) || pmd_large(*pmd))
484 goto out;
485
486 pte = pte_offset_kernel(pmd, address);
487 if (bad_address(pte))
488 goto bad;
489
490 printk("PTE %lx", pte_val(*pte));
491out:
492 printk("\n");
493 return;
494bad:
495 printk("BAD\n");
496}
497
498#endif /* CONFIG_X86_64 */
499
500/*
501 * Workaround for K8 erratum #93 & buggy BIOS.
502 *
503 * BIOS SMM functions are required to use a specific workaround
504 * to avoid corruption of the 64bit RIP register on C stepping K8.
505 *
506 * A lot of BIOS that didn't get tested properly miss this.
507 *
508 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
509 * Try to work around it here.
510 *
511 * Note we only handle faults in kernel here.
512 * Does nothing on 32-bit.
513 */
514static int is_errata93(struct pt_regs *regs, unsigned long address)
515{
516#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
517 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
518 || boot_cpu_data.x86 != 0xf)
519 return 0;
520
521 if (address != regs->ip)
522 return 0;
523
524 if ((address >> 32) != 0)
525 return 0;
526
527 address |= 0xffffffffUL << 32;
528 if ((address >= (u64)_stext && address <= (u64)_etext) ||
529 (address >= MODULES_VADDR && address <= MODULES_END)) {
530 printk_once(errata93_warning);
531 regs->ip = address;
532 return 1;
533 }
534#endif
535 return 0;
536}
537
538/*
539 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
540 * to illegal addresses >4GB.
541 *
542 * We catch this in the page fault handler because these addresses
543 * are not reachable. Just detect this case and return. Any code
544 * segment in LDT is compatibility mode.
545 */
546static int is_errata100(struct pt_regs *regs, unsigned long address)
547{
548#ifdef CONFIG_X86_64
549 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
550 return 1;
551#endif
552 return 0;
553}
554
555static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
556{
557#ifdef CONFIG_X86_F00F_BUG
558 unsigned long nr;
559
560 /*
561 * Pentium F0 0F C7 C8 bug workaround:
562 */
563 if (boot_cpu_has_bug(X86_BUG_F00F)) {
564 nr = (address - idt_descr.address) >> 3;
565
566 if (nr == 6) {
567 do_invalid_op(regs, 0);
568 return 1;
569 }
570 }
571#endif
572 return 0;
573}
574
575static const char nx_warning[] = KERN_CRIT
576"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
577
578static void
579show_fault_oops(struct pt_regs *regs, unsigned long error_code,
580 unsigned long address)
581{
582 if (!oops_may_print())
583 return;
584
585 if (error_code & PF_INSTR) {
586 unsigned int level;
587 pgd_t *pgd;
588 pte_t *pte;
589
590 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
591 pgd += pgd_index(address);
592
593 pte = lookup_address_in_pgd(pgd, address, &level);
594
595 if (pte && pte_present(*pte) && !pte_exec(*pte))
596 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
597 }
598
599 printk(KERN_ALERT "BUG: unable to handle kernel ");
600 if (address < PAGE_SIZE)
601 printk(KERN_CONT "NULL pointer dereference");
602 else
603 printk(KERN_CONT "paging request");
604
605 printk(KERN_CONT " at %p\n", (void *) address);
606 printk(KERN_ALERT "IP:");
607 printk_address(regs->ip);
608
609 dump_pagetable(address);
610}
611
612static noinline void
613pgtable_bad(struct pt_regs *regs, unsigned long error_code,
614 unsigned long address)
615{
616 struct task_struct *tsk;
617 unsigned long flags;
618 int sig;
619
620 flags = oops_begin();
621 tsk = current;
622 sig = SIGKILL;
623
624 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
625 tsk->comm, address);
626 dump_pagetable(address);
627
628 tsk->thread.cr2 = address;
629 tsk->thread.trap_nr = X86_TRAP_PF;
630 tsk->thread.error_code = error_code;
631
632 if (__die("Bad pagetable", regs, error_code))
633 sig = 0;
634
635 oops_end(flags, regs, sig);
636}
637
638static noinline void
639no_context(struct pt_regs *regs, unsigned long error_code,
640 unsigned long address, int signal, int si_code)
641{
642 struct task_struct *tsk = current;
643 unsigned long *stackend;
644 unsigned long flags;
645 int sig;
646
647 /* Are we prepared to handle this kernel fault? */
648 if (fixup_exception(regs)) {
649 /*
650 * Any interrupt that takes a fault gets the fixup. This makes
651 * the below recursive fault logic only apply to a faults from
652 * task context.
653 */
654 if (in_interrupt())
655 return;
656
657 /*
658 * Per the above we're !in_interrupt(), aka. task context.
659 *
660 * In this case we need to make sure we're not recursively
661 * faulting through the emulate_vsyscall() logic.
662 */
663 if (current_thread_info()->sig_on_uaccess_error && signal) {
664 tsk->thread.trap_nr = X86_TRAP_PF;
665 tsk->thread.error_code = error_code | PF_USER;
666 tsk->thread.cr2 = address;
667
668 /* XXX: hwpoison faults will set the wrong code. */
669 force_sig_info_fault(signal, si_code, address, tsk, 0);
670 }
671
672 /*
673 * Barring that, we can do the fixup and be happy.
674 */
675 return;
676 }
677
678 /*
679 * 32-bit:
680 *
681 * Valid to do another page fault here, because if this fault
682 * had been triggered by is_prefetch fixup_exception would have
683 * handled it.
684 *
685 * 64-bit:
686 *
687 * Hall of shame of CPU/BIOS bugs.
688 */
689 if (is_prefetch(regs, error_code, address))
690 return;
691
692 if (is_errata93(regs, address))
693 return;
694
695 /*
696 * Oops. The kernel tried to access some bad page. We'll have to
697 * terminate things with extreme prejudice:
698 */
699 flags = oops_begin();
700
701 show_fault_oops(regs, error_code, address);
702
703 stackend = end_of_stack(tsk);
704 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
705 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
706
707 tsk->thread.cr2 = address;
708 tsk->thread.trap_nr = X86_TRAP_PF;
709 tsk->thread.error_code = error_code;
710
711 sig = SIGKILL;
712 if (__die("Oops", regs, error_code))
713 sig = 0;
714
715 /* Executive summary in case the body of the oops scrolled away */
716 printk(KERN_DEFAULT "CR2: %016lx\n", address);
717
718 oops_end(flags, regs, sig);
719}
720
721/*
722 * Print out info about fatal segfaults, if the show_unhandled_signals
723 * sysctl is set:
724 */
725static inline void
726show_signal_msg(struct pt_regs *regs, unsigned long error_code,
727 unsigned long address, struct task_struct *tsk)
728{
729 if (!unhandled_signal(tsk, SIGSEGV))
730 return;
731
732 if (!printk_ratelimit())
733 return;
734
735 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
736 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
737 tsk->comm, task_pid_nr(tsk), address,
738 (void *)regs->ip, (void *)regs->sp, error_code);
739
740 print_vma_addr(KERN_CONT " in ", regs->ip);
741
742 printk(KERN_CONT "\n");
743}
744
745static void
746__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
747 unsigned long address, int si_code)
748{
749 struct task_struct *tsk = current;
750
751 /* User mode accesses just cause a SIGSEGV */
752 if (error_code & PF_USER) {
753 /*
754 * It's possible to have interrupts off here:
755 */
756 local_irq_enable();
757
758 /*
759 * Valid to do another page fault here because this one came
760 * from user space:
761 */
762 if (is_prefetch(regs, error_code, address))
763 return;
764
765 if (is_errata100(regs, address))
766 return;
767
768#ifdef CONFIG_X86_64
769 /*
770 * Instruction fetch faults in the vsyscall page might need
771 * emulation.
772 */
773 if (unlikely((error_code & PF_INSTR) &&
774 ((address & ~0xfff) == VSYSCALL_START))) {
775 if (emulate_vsyscall(regs, address))
776 return;
777 }
778#endif
779 /* Kernel addresses are always protection faults: */
780 if (address >= TASK_SIZE)
781 error_code |= PF_PROT;
782
783 if (likely(show_unhandled_signals))
784 show_signal_msg(regs, error_code, address, tsk);
785
786 tsk->thread.cr2 = address;
787 tsk->thread.error_code = error_code;
788 tsk->thread.trap_nr = X86_TRAP_PF;
789
790 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
791
792 return;
793 }
794
795 if (is_f00f_bug(regs, address))
796 return;
797
798 no_context(regs, error_code, address, SIGSEGV, si_code);
799}
800
801static noinline void
802bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
803 unsigned long address)
804{
805 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
806}
807
808static void
809__bad_area(struct pt_regs *regs, unsigned long error_code,
810 unsigned long address, int si_code)
811{
812 struct mm_struct *mm = current->mm;
813
814 /*
815 * Something tried to access memory that isn't in our memory map..
816 * Fix it, but check if it's kernel or user first..
817 */
818 up_read(&mm->mmap_sem);
819
820 __bad_area_nosemaphore(regs, error_code, address, si_code);
821}
822
823static noinline void
824bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
825{
826 __bad_area(regs, error_code, address, SEGV_MAPERR);
827}
828
829static noinline void
830bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
831 unsigned long address)
832{
833 __bad_area(regs, error_code, address, SEGV_ACCERR);
834}
835
836static void
837do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
838 unsigned int fault)
839{
840 struct task_struct *tsk = current;
841 struct mm_struct *mm = tsk->mm;
842 int code = BUS_ADRERR;
843
844 up_read(&mm->mmap_sem);
845
846 /* Kernel mode? Handle exceptions or die: */
847 if (!(error_code & PF_USER)) {
848 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
849 return;
850 }
851
852 /* User-space => ok to do another page fault: */
853 if (is_prefetch(regs, error_code, address))
854 return;
855
856 tsk->thread.cr2 = address;
857 tsk->thread.error_code = error_code;
858 tsk->thread.trap_nr = X86_TRAP_PF;
859
860#ifdef CONFIG_MEMORY_FAILURE
861 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
862 printk(KERN_ERR
863 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
864 tsk->comm, tsk->pid, address);
865 code = BUS_MCEERR_AR;
866 }
867#endif
868 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
869}
870
871static noinline void
872mm_fault_error(struct pt_regs *regs, unsigned long error_code,
873 unsigned long address, unsigned int fault)
874{
875 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
876 up_read(¤t->mm->mmap_sem);
877 no_context(regs, error_code, address, 0, 0);
878 return;
879 }
880
881 if (fault & VM_FAULT_OOM) {
882 /* Kernel mode? Handle exceptions or die: */
883 if (!(error_code & PF_USER)) {
884 up_read(¤t->mm->mmap_sem);
885 no_context(regs, error_code, address,
886 SIGSEGV, SEGV_MAPERR);
887 return;
888 }
889
890 up_read(¤t->mm->mmap_sem);
891
892 /*
893 * We ran out of memory, call the OOM killer, and return the
894 * userspace (which will retry the fault, or kill us if we got
895 * oom-killed):
896 */
897 pagefault_out_of_memory();
898 } else {
899 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
900 VM_FAULT_HWPOISON_LARGE))
901 do_sigbus(regs, error_code, address, fault);
902 else
903 BUG();
904 }
905}
906
907static int spurious_fault_check(unsigned long error_code, pte_t *pte)
908{
909 if ((error_code & PF_WRITE) && !pte_write(*pte))
910 return 0;
911
912 if ((error_code & PF_INSTR) && !pte_exec(*pte))
913 return 0;
914
915 return 1;
916}
917
918/*
919 * Handle a spurious fault caused by a stale TLB entry.
920 *
921 * This allows us to lazily refresh the TLB when increasing the
922 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
923 * eagerly is very expensive since that implies doing a full
924 * cross-processor TLB flush, even if no stale TLB entries exist
925 * on other processors.
926 *
927 * There are no security implications to leaving a stale TLB when
928 * increasing the permissions on a page.
929 */
930static noinline __kprobes int
931spurious_fault(unsigned long error_code, unsigned long address)
932{
933 pgd_t *pgd;
934 pud_t *pud;
935 pmd_t *pmd;
936 pte_t *pte;
937 int ret;
938
939 /* Reserved-bit violation or user access to kernel space? */
940 if (error_code & (PF_USER | PF_RSVD))
941 return 0;
942
943 pgd = init_mm.pgd + pgd_index(address);
944 if (!pgd_present(*pgd))
945 return 0;
946
947 pud = pud_offset(pgd, address);
948 if (!pud_present(*pud))
949 return 0;
950
951 if (pud_large(*pud))
952 return spurious_fault_check(error_code, (pte_t *) pud);
953
954 pmd = pmd_offset(pud, address);
955 if (!pmd_present(*pmd))
956 return 0;
957
958 if (pmd_large(*pmd))
959 return spurious_fault_check(error_code, (pte_t *) pmd);
960
961 pte = pte_offset_kernel(pmd, address);
962 if (!pte_present(*pte))
963 return 0;
964
965 ret = spurious_fault_check(error_code, pte);
966 if (!ret)
967 return 0;
968
969 /*
970 * Make sure we have permissions in PMD.
971 * If not, then there's a bug in the page tables:
972 */
973 ret = spurious_fault_check(error_code, (pte_t *) pmd);
974 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
975
976 return ret;
977}
978
979int show_unhandled_signals = 1;
980
981static inline int
982access_error(unsigned long error_code, struct vm_area_struct *vma)
983{
984 if (error_code & PF_WRITE) {
985 /* write, present and write, not present: */
986 if (unlikely(!(vma->vm_flags & VM_WRITE)))
987 return 1;
988 return 0;
989 }
990
991 /* read, present: */
992 if (unlikely(error_code & PF_PROT))
993 return 1;
994
995 /* read, not present: */
996 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
997 return 1;
998
999 return 0;
1000}
1001
1002static int fault_in_kernel_space(unsigned long address)
1003{
1004 return address >= TASK_SIZE_MAX;
1005}
1006
1007static inline bool smap_violation(int error_code, struct pt_regs *regs)
1008{
1009 if (!IS_ENABLED(CONFIG_X86_SMAP))
1010 return false;
1011
1012 if (!static_cpu_has(X86_FEATURE_SMAP))
1013 return false;
1014
1015 if (error_code & PF_USER)
1016 return false;
1017
1018 if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
1019 return false;
1020
1021 return true;
1022}
1023
1024/*
1025 * This routine handles page faults. It determines the address,
1026 * and the problem, and then passes it off to one of the appropriate
1027 * routines.
1028 *
1029 * This function must have noinline because both callers
1030 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1031 * guarantees there's a function trace entry.
1032 */
1033static void __kprobes noinline
1034__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1035 unsigned long address)
1036{
1037 struct vm_area_struct *vma;
1038 struct task_struct *tsk;
1039 struct mm_struct *mm;
1040 int fault;
1041 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1042
1043 tsk = current;
1044 mm = tsk->mm;
1045
1046 /*
1047 * Detect and handle instructions that would cause a page fault for
1048 * both a tracked kernel page and a userspace page.
1049 */
1050 if (kmemcheck_active(regs))
1051 kmemcheck_hide(regs);
1052 prefetchw(&mm->mmap_sem);
1053
1054 if (unlikely(kmmio_fault(regs, address)))
1055 return;
1056
1057 /*
1058 * We fault-in kernel-space virtual memory on-demand. The
1059 * 'reference' page table is init_mm.pgd.
1060 *
1061 * NOTE! We MUST NOT take any locks for this case. We may
1062 * be in an interrupt or a critical region, and should
1063 * only copy the information from the master page table,
1064 * nothing more.
1065 *
1066 * This verifies that the fault happens in kernel space
1067 * (error_code & 4) == 0, and that the fault was not a
1068 * protection error (error_code & 9) == 0.
1069 */
1070 if (unlikely(fault_in_kernel_space(address))) {
1071 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1072 if (vmalloc_fault(address) >= 0)
1073 return;
1074
1075 if (kmemcheck_fault(regs, address, error_code))
1076 return;
1077 }
1078
1079 /* Can handle a stale RO->RW TLB: */
1080 if (spurious_fault(error_code, address))
1081 return;
1082
1083 /* kprobes don't want to hook the spurious faults: */
1084 if (kprobes_fault(regs))
1085 return;
1086 /*
1087 * Don't take the mm semaphore here. If we fixup a prefetch
1088 * fault we could otherwise deadlock:
1089 */
1090 bad_area_nosemaphore(regs, error_code, address);
1091
1092 return;
1093 }
1094
1095 /* kprobes don't want to hook the spurious faults: */
1096 if (unlikely(kprobes_fault(regs)))
1097 return;
1098
1099 if (unlikely(error_code & PF_RSVD))
1100 pgtable_bad(regs, error_code, address);
1101
1102 if (unlikely(smap_violation(error_code, regs))) {
1103 bad_area_nosemaphore(regs, error_code, address);
1104 return;
1105 }
1106
1107 /*
1108 * If we're in an interrupt, have no user context or are running
1109 * in an atomic region then we must not take the fault:
1110 */
1111 if (unlikely(in_atomic() || !mm)) {
1112 bad_area_nosemaphore(regs, error_code, address);
1113 return;
1114 }
1115
1116 /*
1117 * It's safe to allow irq's after cr2 has been saved and the
1118 * vmalloc fault has been handled.
1119 *
1120 * User-mode registers count as a user access even for any
1121 * potential system fault or CPU buglet:
1122 */
1123 if (user_mode_vm(regs)) {
1124 local_irq_enable();
1125 error_code |= PF_USER;
1126 flags |= FAULT_FLAG_USER;
1127 } else {
1128 if (regs->flags & X86_EFLAGS_IF)
1129 local_irq_enable();
1130 }
1131
1132 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1133
1134 if (error_code & PF_WRITE)
1135 flags |= FAULT_FLAG_WRITE;
1136
1137 /*
1138 * When running in the kernel we expect faults to occur only to
1139 * addresses in user space. All other faults represent errors in
1140 * the kernel and should generate an OOPS. Unfortunately, in the
1141 * case of an erroneous fault occurring in a code path which already
1142 * holds mmap_sem we will deadlock attempting to validate the fault
1143 * against the address space. Luckily the kernel only validly
1144 * references user space from well defined areas of code, which are
1145 * listed in the exceptions table.
1146 *
1147 * As the vast majority of faults will be valid we will only perform
1148 * the source reference check when there is a possibility of a
1149 * deadlock. Attempt to lock the address space, if we cannot we then
1150 * validate the source. If this is invalid we can skip the address
1151 * space check, thus avoiding the deadlock:
1152 */
1153 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1154 if ((error_code & PF_USER) == 0 &&
1155 !search_exception_tables(regs->ip)) {
1156 bad_area_nosemaphore(regs, error_code, address);
1157 return;
1158 }
1159retry:
1160 down_read(&mm->mmap_sem);
1161 } else {
1162 /*
1163 * The above down_read_trylock() might have succeeded in
1164 * which case we'll have missed the might_sleep() from
1165 * down_read():
1166 */
1167 might_sleep();
1168 }
1169
1170 vma = find_vma(mm, address);
1171 if (unlikely(!vma)) {
1172 bad_area(regs, error_code, address);
1173 return;
1174 }
1175 if (likely(vma->vm_start <= address))
1176 goto good_area;
1177 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1178 bad_area(regs, error_code, address);
1179 return;
1180 }
1181 if (error_code & PF_USER) {
1182 /*
1183 * Accessing the stack below %sp is always a bug.
1184 * The large cushion allows instructions like enter
1185 * and pusha to work. ("enter $65535, $31" pushes
1186 * 32 pointers and then decrements %sp by 65535.)
1187 */
1188 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1189 bad_area(regs, error_code, address);
1190 return;
1191 }
1192 }
1193 if (unlikely(expand_stack(vma, address))) {
1194 bad_area(regs, error_code, address);
1195 return;
1196 }
1197
1198 /*
1199 * Ok, we have a good vm_area for this memory access, so
1200 * we can handle it..
1201 */
1202good_area:
1203 if (unlikely(access_error(error_code, vma))) {
1204 bad_area_access_error(regs, error_code, address);
1205 return;
1206 }
1207
1208 /*
1209 * If for any reason at all we couldn't handle the fault,
1210 * make sure we exit gracefully rather than endlessly redo
1211 * the fault:
1212 */
1213 fault = handle_mm_fault(mm, vma, address, flags);
1214
1215 /*
1216 * If we need to retry but a fatal signal is pending, handle the
1217 * signal first. We do not need to release the mmap_sem because it
1218 * would already be released in __lock_page_or_retry in mm/filemap.c.
1219 */
1220 if (unlikely((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)))
1221 return;
1222
1223 if (unlikely(fault & VM_FAULT_ERROR)) {
1224 mm_fault_error(regs, error_code, address, fault);
1225 return;
1226 }
1227
1228 /*
1229 * Major/minor page fault accounting is only done on the
1230 * initial attempt. If we go through a retry, it is extremely
1231 * likely that the page will be found in page cache at that point.
1232 */
1233 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1234 if (fault & VM_FAULT_MAJOR) {
1235 tsk->maj_flt++;
1236 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1237 regs, address);
1238 } else {
1239 tsk->min_flt++;
1240 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1241 regs, address);
1242 }
1243 if (fault & VM_FAULT_RETRY) {
1244 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1245 * of starvation. */
1246 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1247 flags |= FAULT_FLAG_TRIED;
1248 goto retry;
1249 }
1250 }
1251
1252 check_v8086_mode(regs, address, tsk);
1253
1254 up_read(&mm->mmap_sem);
1255}
1256
1257dotraplinkage void __kprobes notrace
1258do_page_fault(struct pt_regs *regs, unsigned long error_code)
1259{
1260 unsigned long address = read_cr2(); /* Get the faulting address */
1261 enum ctx_state prev_state;
1262
1263 /*
1264 * We must have this function tagged with __kprobes, notrace and call
1265 * read_cr2() before calling anything else. To avoid calling any kind
1266 * of tracing machinery before we've observed the CR2 value.
1267 *
1268 * exception_{enter,exit}() contain all sorts of tracepoints.
1269 */
1270
1271 prev_state = exception_enter();
1272 __do_page_fault(regs, error_code, address);
1273 exception_exit(prev_state);
1274}
1275
1276#ifdef CONFIG_TRACING
1277static void trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1278 unsigned long error_code)
1279{
1280 if (user_mode(regs))
1281 trace_page_fault_user(address, regs, error_code);
1282 else
1283 trace_page_fault_kernel(address, regs, error_code);
1284}
1285
1286dotraplinkage void __kprobes notrace
1287trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1288{
1289 /*
1290 * The exception_enter and tracepoint processing could
1291 * trigger another page faults (user space callchain
1292 * reading) and destroy the original cr2 value, so read
1293 * the faulting address now.
1294 */
1295 unsigned long address = read_cr2();
1296 enum ctx_state prev_state;
1297
1298 prev_state = exception_enter();
1299 trace_page_fault_entries(address, regs, error_code);
1300 __do_page_fault(regs, error_code, address);
1301 exception_exit(prev_state);
1302}
1303#endif /* CONFIG_TRACING */