projects / deliverable / linux.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
mm: page_alloc: reduce cost of the fair zone allocation policy
[deliverable/linux.git] / arch / x86 / mm / fault.c
... / ...
CommitLineData
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> /* NOKPROBE_SYMBOL, ... */
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_ADDR */
22#include <asm/vsyscall.h> /* emulate_vsyscall */
23
24#define CREATE_TRACE_POINTS
25#include <asm/trace/exceptions.h>
26
27/*
28 * Page fault error code bits:
29 *
30 * bit 0 == 0: no page found 1: protection fault
31 * bit 1 == 0: read access 1: write access
32 * bit 2 == 0: kernel-mode access 1: user-mode access
33 * bit 3 == 1: use of reserved bit detected
34 * bit 4 == 1: fault was an instruction fetch
35 */
36enum x86_pf_error_code {
37
38 PF_PROT = 1 << 0,
39 PF_WRITE = 1 << 1,
40 PF_USER = 1 << 2,
41 PF_RSVD = 1 << 3,
42 PF_INSTR = 1 << 4,
43};
44
45/*
46 * Returns 0 if mmiotrace is disabled, or if the fault is not
47 * handled by mmiotrace:
48 */
49static nokprobe_inline int
50kmmio_fault(struct pt_regs *regs, unsigned long addr)
51{
52 if (unlikely(is_kmmio_active()))
53 if (kmmio_handler(regs, addr) == 1)
54 return -1;
55 return 0;
56}
57
58static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
59{
60 int ret = 0;
61
62 /* kprobe_running() needs smp_processor_id() */
63 if (kprobes_built_in() && !user_mode_vm(regs)) {
64 preempt_disable();
65 if (kprobe_running() && kprobe_fault_handler(regs, 14))
66 ret = 1;
67 preempt_enable();
68 }
69
70 return ret;
71}
72
73/*
74 * Prefetch quirks:
75 *
76 * 32-bit mode:
77 *
78 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
79 * Check that here and ignore it.
80 *
81 * 64-bit mode:
82 *
83 * Sometimes the CPU reports invalid exceptions on prefetch.
84 * Check that here and ignore it.
85 *
86 * Opcode checker based on code by Richard Brunner.
87 */
88static inline int
89check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
90 unsigned char opcode, int *prefetch)
91{
92 unsigned char instr_hi = opcode & 0xf0;
93 unsigned char instr_lo = opcode & 0x0f;
94
95 switch (instr_hi) {
96 case 0x20:
97 case 0x30:
98 /*
99 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
100 * In X86_64 long mode, the CPU will signal invalid
101 * opcode if some of these prefixes are present so
102 * X86_64 will never get here anyway
103 */
104 return ((instr_lo & 7) == 0x6);
105#ifdef CONFIG_X86_64
106 case 0x40:
107 /*
108 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
109 * Need to figure out under what instruction mode the
110 * instruction was issued. Could check the LDT for lm,
111 * but for now it's good enough to assume that long
112 * mode only uses well known segments or kernel.
113 */
114 return (!user_mode(regs) || user_64bit_mode(regs));
115#endif
116 case 0x60:
117 /* 0x64 thru 0x67 are valid prefixes in all modes. */
118 return (instr_lo & 0xC) == 0x4;
119 case 0xF0:
120 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
121 return !instr_lo || (instr_lo>>1) == 1;
122 case 0x00:
123 /* Prefetch instruction is 0x0F0D or 0x0F18 */
124 if (probe_kernel_address(instr, opcode))
125 return 0;
126
127 *prefetch = (instr_lo == 0xF) &&
128 (opcode == 0x0D || opcode == 0x18);
129 return 0;
130 default:
131 return 0;
132 }
133}
134
135static int
136is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
137{
138 unsigned char *max_instr;
139 unsigned char *instr;
140 int prefetch = 0;
141
142 /*
143 * If it was a exec (instruction fetch) fault on NX page, then
144 * do not ignore the fault:
145 */
146 if (error_code & PF_INSTR)
147 return 0;
148
149 instr = (void *)convert_ip_to_linear(current, regs);
150 max_instr = instr + 15;
151
152 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
153 return 0;
154
155 while (instr < max_instr) {
156 unsigned char opcode;
157
158 if (probe_kernel_address(instr, opcode))
159 break;
160
161 instr++;
162
163 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
164 break;
165 }
166 return prefetch;
167}
168
169static void
170force_sig_info_fault(int si_signo, int si_code, unsigned long address,
171 struct task_struct *tsk, int fault)
172{
173 unsigned lsb = 0;
174 siginfo_t info;
175
176 info.si_signo = si_signo;
177 info.si_errno = 0;
178 info.si_code = si_code;
179 info.si_addr = (void __user *)address;
180 if (fault & VM_FAULT_HWPOISON_LARGE)
181 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
182 if (fault & VM_FAULT_HWPOISON)
183 lsb = PAGE_SHIFT;
184 info.si_addr_lsb = lsb;
185
186 force_sig_info(si_signo, &info, tsk);
187}
188
189DEFINE_SPINLOCK(pgd_lock);
190LIST_HEAD(pgd_list);
191
192#ifdef CONFIG_X86_32
193static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
194{
195 unsigned index = pgd_index(address);
196 pgd_t *pgd_k;
197 pud_t *pud, *pud_k;
198 pmd_t *pmd, *pmd_k;
199
200 pgd += index;
201 pgd_k = init_mm.pgd + index;
202
203 if (!pgd_present(*pgd_k))
204 return NULL;
205
206 /*
207 * set_pgd(pgd, *pgd_k); here would be useless on PAE
208 * and redundant with the set_pmd() on non-PAE. As would
209 * set_pud.
210 */
211 pud = pud_offset(pgd, address);
212 pud_k = pud_offset(pgd_k, address);
213 if (!pud_present(*pud_k))
214 return NULL;
215
216 pmd = pmd_offset(pud, address);
217 pmd_k = pmd_offset(pud_k, address);
218 if (!pmd_present(*pmd_k))
219 return NULL;
220
221 if (!pmd_present(*pmd))
222 set_pmd(pmd, *pmd_k);
223 else
224 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
225
226 return pmd_k;
227}
228
229void vmalloc_sync_all(void)
230{
231 unsigned long address;
232
233 if (SHARED_KERNEL_PMD)
234 return;
235
236 for (address = VMALLOC_START & PMD_MASK;
237 address >= TASK_SIZE && address < FIXADDR_TOP;
238 address += PMD_SIZE) {
239 struct page *page;
240
241 spin_lock(&pgd_lock);
242 list_for_each_entry(page, &pgd_list, lru) {
243 spinlock_t *pgt_lock;
244 pmd_t *ret;
245
246 /* the pgt_lock only for Xen */
247 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
248
249 spin_lock(pgt_lock);
250 ret = vmalloc_sync_one(page_address(page), address);
251 spin_unlock(pgt_lock);
252
253 if (!ret)
254 break;
255 }
256 spin_unlock(&pgd_lock);
257 }
258}
259
260/*
261 * 32-bit:
262 *
263 * Handle a fault on the vmalloc or module mapping area
264 */
265static noinline int vmalloc_fault(unsigned long address)
266{
267 unsigned long pgd_paddr;
268 pmd_t *pmd_k;
269 pte_t *pte_k;
270
271 /* Make sure we are in vmalloc area: */
272 if (!(address >= VMALLOC_START && address < VMALLOC_END))
273 return -1;
274
275 WARN_ON_ONCE(in_nmi());
276
277 /*
278 * Synchronize this task's top level page-table
279 * with the 'reference' page table.
280 *
281 * Do _not_ use "current" here. We might be inside
282 * an interrupt in the middle of a task switch..
283 */
284 pgd_paddr = read_cr3();
285 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
286 if (!pmd_k)
287 return -1;
288
289 pte_k = pte_offset_kernel(pmd_k, address);
290 if (!pte_present(*pte_k))
291 return -1;
292
293 return 0;
294}
295NOKPROBE_SYMBOL(vmalloc_fault);
296
297/*
298 * Did it hit the DOS screen memory VA from vm86 mode?
299 */
300static inline void
301check_v8086_mode(struct pt_regs *regs, unsigned long address,
302 struct task_struct *tsk)
303{
304 unsigned long bit;
305
306 if (!v8086_mode(regs))
307 return;
308
309 bit = (address - 0xA0000) >> PAGE_SHIFT;
310 if (bit < 32)
311 tsk->thread.screen_bitmap |= 1 << bit;
312}
313
314static bool low_pfn(unsigned long pfn)
315{
316 return pfn < max_low_pfn;
317}
318
319static void dump_pagetable(unsigned long address)
320{
321 pgd_t *base = __va(read_cr3());
322 pgd_t *pgd = &base[pgd_index(address)];
323 pmd_t *pmd;
324 pte_t *pte;
325
326#ifdef CONFIG_X86_PAE
327 printk("*pdpt = %016Lx ", pgd_val(*pgd));
328 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
329 goto out;
330#endif
331 pmd = pmd_offset(pud_offset(pgd, address), address);
332 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
333
334 /*
335 * We must not directly access the pte in the highpte
336 * case if the page table is located in highmem.
337 * And let's rather not kmap-atomic the pte, just in case
338 * it's allocated already:
339 */
340 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
341 goto out;
342
343 pte = pte_offset_kernel(pmd, address);
344 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
345out:
346 printk("\n");
347}
348
349#else /* CONFIG_X86_64: */
350
351void vmalloc_sync_all(void)
352{
353 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
354}
355
356/*
357 * 64-bit:
358 *
359 * Handle a fault on the vmalloc area
360 *
361 * This assumes no large pages in there.
362 */
363static noinline int vmalloc_fault(unsigned long address)
364{
365 pgd_t *pgd, *pgd_ref;
366 pud_t *pud, *pud_ref;
367 pmd_t *pmd, *pmd_ref;
368 pte_t *pte, *pte_ref;
369
370 /* Make sure we are in vmalloc area: */
371 if (!(address >= VMALLOC_START && address < VMALLOC_END))
372 return -1;
373
374 WARN_ON_ONCE(in_nmi());
375
376 /*
377 * Copy kernel mappings over when needed. This can also
378 * happen within a race in page table update. In the later
379 * case just flush:
380 */
381 pgd = pgd_offset(current->active_mm, address);
382 pgd_ref = pgd_offset_k(address);
383 if (pgd_none(*pgd_ref))
384 return -1;
385
386 if (pgd_none(*pgd)) {
387 set_pgd(pgd, *pgd_ref);
388 arch_flush_lazy_mmu_mode();
389 } else {
390 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
391 }
392
393 /*
394 * Below here mismatches are bugs because these lower tables
395 * are shared:
396 */
397
398 pud = pud_offset(pgd, address);
399 pud_ref = pud_offset(pgd_ref, address);
400 if (pud_none(*pud_ref))
401 return -1;
402
403 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
404 BUG();
405
406 pmd = pmd_offset(pud, address);
407 pmd_ref = pmd_offset(pud_ref, address);
408 if (pmd_none(*pmd_ref))
409 return -1;
410
411 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
412 BUG();
413
414 pte_ref = pte_offset_kernel(pmd_ref, address);
415 if (!pte_present(*pte_ref))
416 return -1;
417
418 pte = pte_offset_kernel(pmd, address);
419
420 /*
421 * Don't use pte_page here, because the mappings can point
422 * outside mem_map, and the NUMA hash lookup cannot handle
423 * that:
424 */
425 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
426 BUG();
427
428 return 0;
429}
430NOKPROBE_SYMBOL(vmalloc_fault);
431
432#ifdef CONFIG_CPU_SUP_AMD
433static const char errata93_warning[] =
434KERN_ERR
435"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
436"******* Working around it, but it may cause SEGVs or burn power.\n"
437"******* Please consider a BIOS update.\n"
438"******* Disabling USB legacy in the BIOS may also help.\n";
439#endif
440
441/*
442 * No vm86 mode in 64-bit mode:
443 */
444static inline void
445check_v8086_mode(struct pt_regs *regs, unsigned long address,
446 struct task_struct *tsk)
447{
448}
449
450static int bad_address(void *p)
451{
452 unsigned long dummy;
453
454 return probe_kernel_address((unsigned long *)p, dummy);
455}
456
457static void dump_pagetable(unsigned long address)
458{
459 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
460 pgd_t *pgd = base + pgd_index(address);
461 pud_t *pud;
462 pmd_t *pmd;
463 pte_t *pte;
464
465 if (bad_address(pgd))
466 goto bad;
467
468 printk("PGD %lx ", pgd_val(*pgd));
469
470 if (!pgd_present(*pgd))
471 goto out;
472
473 pud = pud_offset(pgd, address);
474 if (bad_address(pud))
475 goto bad;
476
477 printk("PUD %lx ", pud_val(*pud));
478 if (!pud_present(*pud) || pud_large(*pud))
479 goto out;
480
481 pmd = pmd_offset(pud, address);
482 if (bad_address(pmd))
483 goto bad;
484
485 printk("PMD %lx ", pmd_val(*pmd));
486 if (!pmd_present(*pmd) || pmd_large(*pmd))
487 goto out;
488
489 pte = pte_offset_kernel(pmd, address);
490 if (bad_address(pte))
491 goto bad;
492
493 printk("PTE %lx", pte_val(*pte));
494out:
495 printk("\n");
496 return;
497bad:
498 printk("BAD\n");
499}
500
501#endif /* CONFIG_X86_64 */
502
503/*
504 * Workaround for K8 erratum #93 & buggy BIOS.
505 *
506 * BIOS SMM functions are required to use a specific workaround
507 * to avoid corruption of the 64bit RIP register on C stepping K8.
508 *
509 * A lot of BIOS that didn't get tested properly miss this.
510 *
511 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
512 * Try to work around it here.
513 *
514 * Note we only handle faults in kernel here.
515 * Does nothing on 32-bit.
516 */
517static int is_errata93(struct pt_regs *regs, unsigned long address)
518{
519#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
520 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
521 || boot_cpu_data.x86 != 0xf)
522 return 0;
523
524 if (address != regs->ip)
525 return 0;
526
527 if ((address >> 32) != 0)
528 return 0;
529
530 address |= 0xffffffffUL << 32;
531 if ((address >= (u64)_stext && address <= (u64)_etext) ||
532 (address >= MODULES_VADDR && address <= MODULES_END)) {
533 printk_once(errata93_warning);
534 regs->ip = address;
535 return 1;
536 }
537#endif
538 return 0;
539}
540
541/*
542 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
543 * to illegal addresses >4GB.
544 *
545 * We catch this in the page fault handler because these addresses
546 * are not reachable. Just detect this case and return. Any code
547 * segment in LDT is compatibility mode.
548 */
549static int is_errata100(struct pt_regs *regs, unsigned long address)
550{
551#ifdef CONFIG_X86_64
552 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
553 return 1;
554#endif
555 return 0;
556}
557
558static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
559{
560#ifdef CONFIG_X86_F00F_BUG
561 unsigned long nr;
562
563 /*
564 * Pentium F0 0F C7 C8 bug workaround:
565 */
566 if (boot_cpu_has_bug(X86_BUG_F00F)) {
567 nr = (address - idt_descr.address) >> 3;
568
569 if (nr == 6) {
570 do_invalid_op(regs, 0);
571 return 1;
572 }
573 }
574#endif
575 return 0;
576}
577
578static const char nx_warning[] = KERN_CRIT
579"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
580static const char smep_warning[] = KERN_CRIT
581"unable to execute userspace code (SMEP?) (uid: %d)\n";
582
583static void
584show_fault_oops(struct pt_regs *regs, unsigned long error_code,
585 unsigned long address)
586{
587 if (!oops_may_print())
588 return;
589
590 if (error_code & PF_INSTR) {
591 unsigned int level;
592 pgd_t *pgd;
593 pte_t *pte;
594
595 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
596 pgd += pgd_index(address);
597
598 pte = lookup_address_in_pgd(pgd, address, &level);
599
600 if (pte && pte_present(*pte) && !pte_exec(*pte))
601 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
602 if (pte && pte_present(*pte) && pte_exec(*pte) &&
603 (pgd_flags(*pgd) & _PAGE_USER) &&
604 (read_cr4() & X86_CR4_SMEP))
605 printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
606 }
607
608 printk(KERN_ALERT "BUG: unable to handle kernel ");
609 if (address < PAGE_SIZE)
610 printk(KERN_CONT "NULL pointer dereference");
611 else
612 printk(KERN_CONT "paging request");
613
614 printk(KERN_CONT " at %p\n", (void *) address);
615 printk(KERN_ALERT "IP:");
616 printk_address(regs->ip);
617
618 dump_pagetable(address);
619}
620
621static noinline void
622pgtable_bad(struct pt_regs *regs, unsigned long error_code,
623 unsigned long address)
624{
625 struct task_struct *tsk;
626 unsigned long flags;
627 int sig;
628
629 flags = oops_begin();
630 tsk = current;
631 sig = SIGKILL;
632
633 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
634 tsk->comm, address);
635 dump_pagetable(address);
636
637 tsk->thread.cr2 = address;
638 tsk->thread.trap_nr = X86_TRAP_PF;
639 tsk->thread.error_code = error_code;
640
641 if (__die("Bad pagetable", regs, error_code))
642 sig = 0;
643
644 oops_end(flags, regs, sig);
645}
646
647static noinline void
648no_context(struct pt_regs *regs, unsigned long error_code,
649 unsigned long address, int signal, int si_code)
650{
651 struct task_struct *tsk = current;
652 unsigned long *stackend;
653 unsigned long flags;
654 int sig;
655
656 /* Are we prepared to handle this kernel fault? */
657 if (fixup_exception(regs)) {
658 /*
659 * Any interrupt that takes a fault gets the fixup. This makes
660 * the below recursive fault logic only apply to a faults from
661 * task context.
662 */
663 if (in_interrupt())
664 return;
665
666 /*
667 * Per the above we're !in_interrupt(), aka. task context.
668 *
669 * In this case we need to make sure we're not recursively
670 * faulting through the emulate_vsyscall() logic.
671 */
672 if (current_thread_info()->sig_on_uaccess_error && signal) {
673 tsk->thread.trap_nr = X86_TRAP_PF;
674 tsk->thread.error_code = error_code | PF_USER;
675 tsk->thread.cr2 = address;
676
677 /* XXX: hwpoison faults will set the wrong code. */
678 force_sig_info_fault(signal, si_code, address, tsk, 0);
679 }
680
681 /*
682 * Barring that, we can do the fixup and be happy.
683 */
684 return;
685 }
686
687 /*
688 * 32-bit:
689 *
690 * Valid to do another page fault here, because if this fault
691 * had been triggered by is_prefetch fixup_exception would have
692 * handled it.
693 *
694 * 64-bit:
695 *
696 * Hall of shame of CPU/BIOS bugs.
697 */
698 if (is_prefetch(regs, error_code, address))
699 return;
700
701 if (is_errata93(regs, address))
702 return;
703
704 /*
705 * Oops. The kernel tried to access some bad page. We'll have to
706 * terminate things with extreme prejudice:
707 */
708 flags = oops_begin();
709
710 show_fault_oops(regs, error_code, address);
711
712 stackend = end_of_stack(tsk);
713 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
714 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
715
716 tsk->thread.cr2 = address;
717 tsk->thread.trap_nr = X86_TRAP_PF;
718 tsk->thread.error_code = error_code;
719
720 sig = SIGKILL;
721 if (__die("Oops", regs, error_code))
722 sig = 0;
723
724 /* Executive summary in case the body of the oops scrolled away */
725 printk(KERN_DEFAULT "CR2: %016lx\n", address);
726
727 oops_end(flags, regs, sig);
728}
729
730/*
731 * Print out info about fatal segfaults, if the show_unhandled_signals
732 * sysctl is set:
733 */
734static inline void
735show_signal_msg(struct pt_regs *regs, unsigned long error_code,
736 unsigned long address, struct task_struct *tsk)
737{
738 if (!unhandled_signal(tsk, SIGSEGV))
739 return;
740
741 if (!printk_ratelimit())
742 return;
743
744 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
745 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
746 tsk->comm, task_pid_nr(tsk), address,
747 (void *)regs->ip, (void *)regs->sp, error_code);
748
749 print_vma_addr(KERN_CONT " in ", regs->ip);
750
751 printk(KERN_CONT "\n");
752}
753
754static void
755__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
756 unsigned long address, int si_code)
757{
758 struct task_struct *tsk = current;
759
760 /* User mode accesses just cause a SIGSEGV */
761 if (error_code & PF_USER) {
762 /*
763 * It's possible to have interrupts off here:
764 */
765 local_irq_enable();
766
767 /*
768 * Valid to do another page fault here because this one came
769 * from user space:
770 */
771 if (is_prefetch(regs, error_code, address))
772 return;
773
774 if (is_errata100(regs, address))
775 return;
776
777#ifdef CONFIG_X86_64
778 /*
779 * Instruction fetch faults in the vsyscall page might need
780 * emulation.
781 */
782 if (unlikely((error_code & PF_INSTR) &&
783 ((address & ~0xfff) == VSYSCALL_ADDR))) {
784 if (emulate_vsyscall(regs, address))
785 return;
786 }
787#endif
788 /* Kernel addresses are always protection faults: */
789 if (address >= TASK_SIZE)
790 error_code |= PF_PROT;
791
792 if (likely(show_unhandled_signals))
793 show_signal_msg(regs, error_code, address, tsk);
794
795 tsk->thread.cr2 = address;
796 tsk->thread.error_code = error_code;
797 tsk->thread.trap_nr = X86_TRAP_PF;
798
799 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
800
801 return;
802 }
803
804 if (is_f00f_bug(regs, address))
805 return;
806
807 no_context(regs, error_code, address, SIGSEGV, si_code);
808}
809
810static noinline void
811bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
812 unsigned long address)
813{
814 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
815}
816
817static void
818__bad_area(struct pt_regs *regs, unsigned long error_code,
819 unsigned long address, int si_code)
820{
821 struct mm_struct *mm = current->mm;
822
823 /*
824 * Something tried to access memory that isn't in our memory map..
825 * Fix it, but check if it's kernel or user first..
826 */
827 up_read(&mm->mmap_sem);
828
829 __bad_area_nosemaphore(regs, error_code, address, si_code);
830}
831
832static noinline void
833bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
834{
835 __bad_area(regs, error_code, address, SEGV_MAPERR);
836}
837
838static noinline void
839bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
840 unsigned long address)
841{
842 __bad_area(regs, error_code, address, SEGV_ACCERR);
843}
844
845static void
846do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
847 unsigned int fault)
848{
849 struct task_struct *tsk = current;
850 struct mm_struct *mm = tsk->mm;
851 int code = BUS_ADRERR;
852
853 up_read(&mm->mmap_sem);
854
855 /* Kernel mode? Handle exceptions or die: */
856 if (!(error_code & PF_USER)) {
857 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
858 return;
859 }
860
861 /* User-space => ok to do another page fault: */
862 if (is_prefetch(regs, error_code, address))
863 return;
864
865 tsk->thread.cr2 = address;
866 tsk->thread.error_code = error_code;
867 tsk->thread.trap_nr = X86_TRAP_PF;
868
869#ifdef CONFIG_MEMORY_FAILURE
870 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
871 printk(KERN_ERR
872 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
873 tsk->comm, tsk->pid, address);
874 code = BUS_MCEERR_AR;
875 }
876#endif
877 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
878}
879
880static noinline void
881mm_fault_error(struct pt_regs *regs, unsigned long error_code,
882 unsigned long address, unsigned int fault)
883{
884 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
885 up_read(&current->mm->mmap_sem);
886 no_context(regs, error_code, address, 0, 0);
887 return;
888 }
889
890 if (fault & VM_FAULT_OOM) {
891 /* Kernel mode? Handle exceptions or die: */
892 if (!(error_code & PF_USER)) {
893 up_read(&current->mm->mmap_sem);
894 no_context(regs, error_code, address,
895 SIGSEGV, SEGV_MAPERR);
896 return;
897 }
898
899 up_read(&current->mm->mmap_sem);
900
901 /*
902 * We ran out of memory, call the OOM killer, and return the
903 * userspace (which will retry the fault, or kill us if we got
904 * oom-killed):
905 */
906 pagefault_out_of_memory();
907 } else {
908 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
909 VM_FAULT_HWPOISON_LARGE))
910 do_sigbus(regs, error_code, address, fault);
911 else
912 BUG();
913 }
914}
915
916static int spurious_fault_check(unsigned long error_code, pte_t *pte)
917{
918 if ((error_code & PF_WRITE) && !pte_write(*pte))
919 return 0;
920
921 if ((error_code & PF_INSTR) && !pte_exec(*pte))
922 return 0;
923
924 return 1;
925}
926
927/*
928 * Handle a spurious fault caused by a stale TLB entry.
929 *
930 * This allows us to lazily refresh the TLB when increasing the
931 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
932 * eagerly is very expensive since that implies doing a full
933 * cross-processor TLB flush, even if no stale TLB entries exist
934 * on other processors.
935 *
936 * There are no security implications to leaving a stale TLB when
937 * increasing the permissions on a page.
938 */
939static noinline int
940spurious_fault(unsigned long error_code, unsigned long address)
941{
942 pgd_t *pgd;
943 pud_t *pud;
944 pmd_t *pmd;
945 pte_t *pte;
946 int ret;
947
948 /* Reserved-bit violation or user access to kernel space? */
949 if (error_code & (PF_USER | PF_RSVD))
950 return 0;
951
952 pgd = init_mm.pgd + pgd_index(address);
953 if (!pgd_present(*pgd))
954 return 0;
955
956 pud = pud_offset(pgd, address);
957 if (!pud_present(*pud))
958 return 0;
959
960 if (pud_large(*pud))
961 return spurious_fault_check(error_code, (pte_t *) pud);
962
963 pmd = pmd_offset(pud, address);
964 if (!pmd_present(*pmd))
965 return 0;
966
967 if (pmd_large(*pmd))
968 return spurious_fault_check(error_code, (pte_t *) pmd);
969
970 pte = pte_offset_kernel(pmd, address);
971 if (!pte_present(*pte))
972 return 0;
973
974 ret = spurious_fault_check(error_code, pte);
975 if (!ret)
976 return 0;
977
978 /*
979 * Make sure we have permissions in PMD.
980 * If not, then there's a bug in the page tables:
981 */
982 ret = spurious_fault_check(error_code, (pte_t *) pmd);
983 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
984
985 return ret;
986}
987NOKPROBE_SYMBOL(spurious_fault);
988
989int show_unhandled_signals = 1;
990
991static inline int
992access_error(unsigned long error_code, struct vm_area_struct *vma)
993{
994 if (error_code & PF_WRITE) {
995 /* write, present and write, not present: */
996 if (unlikely(!(vma->vm_flags & VM_WRITE)))
997 return 1;
998 return 0;
999 }
1000
1001 /* read, present: */
1002 if (unlikely(error_code & PF_PROT))
1003 return 1;
1004
1005 /* read, not present: */
1006 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1007 return 1;
1008
1009 return 0;
1010}
1011
1012static int fault_in_kernel_space(unsigned long address)
1013{
1014 return address >= TASK_SIZE_MAX;
1015}
1016
1017static inline bool smap_violation(int error_code, struct pt_regs *regs)
1018{
1019 if (!IS_ENABLED(CONFIG_X86_SMAP))
1020 return false;
1021
1022 if (!static_cpu_has(X86_FEATURE_SMAP))
1023 return false;
1024
1025 if (error_code & PF_USER)
1026 return false;
1027
1028 if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
1029 return false;
1030
1031 return true;
1032}
1033
1034/*
1035 * This routine handles page faults. It determines the address,
1036 * and the problem, and then passes it off to one of the appropriate
1037 * routines.
1038 *
1039 * This function must have noinline because both callers
1040 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1041 * guarantees there's a function trace entry.
1042 */
1043static noinline void
1044__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1045 unsigned long address)
1046{
1047 struct vm_area_struct *vma;
1048 struct task_struct *tsk;
1049 struct mm_struct *mm;
1050 int fault;
1051 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1052
1053 tsk = current;
1054 mm = tsk->mm;
1055
1056 /*
1057 * Detect and handle instructions that would cause a page fault for
1058 * both a tracked kernel page and a userspace page.
1059 */
1060 if (kmemcheck_active(regs))
1061 kmemcheck_hide(regs);
1062 prefetchw(&mm->mmap_sem);
1063
1064 if (unlikely(kmmio_fault(regs, address)))
1065 return;
1066
1067 /*
1068 * We fault-in kernel-space virtual memory on-demand. The
1069 * 'reference' page table is init_mm.pgd.
1070 *
1071 * NOTE! We MUST NOT take any locks for this case. We may
1072 * be in an interrupt or a critical region, and should
1073 * only copy the information from the master page table,
1074 * nothing more.
1075 *
1076 * This verifies that the fault happens in kernel space
1077 * (error_code & 4) == 0, and that the fault was not a
1078 * protection error (error_code & 9) == 0.
1079 */
1080 if (unlikely(fault_in_kernel_space(address))) {
1081 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1082 if (vmalloc_fault(address) >= 0)
1083 return;
1084
1085 if (kmemcheck_fault(regs, address, error_code))
1086 return;
1087 }
1088
1089 /* Can handle a stale RO->RW TLB: */
1090 if (spurious_fault(error_code, address))
1091 return;
1092
1093 /* kprobes don't want to hook the spurious faults: */
1094 if (kprobes_fault(regs))
1095 return;
1096 /*
1097 * Don't take the mm semaphore here. If we fixup a prefetch
1098 * fault we could otherwise deadlock:
1099 */
1100 bad_area_nosemaphore(regs, error_code, address);
1101
1102 return;
1103 }
1104
1105 /* kprobes don't want to hook the spurious faults: */
1106 if (unlikely(kprobes_fault(regs)))
1107 return;
1108
1109 if (unlikely(error_code & PF_RSVD))
1110 pgtable_bad(regs, error_code, address);
1111
1112 if (unlikely(smap_violation(error_code, regs))) {
1113 bad_area_nosemaphore(regs, error_code, address);
1114 return;
1115 }
1116
1117 /*
1118 * If we're in an interrupt, have no user context or are running
1119 * in an atomic region then we must not take the fault:
1120 */
1121 if (unlikely(in_atomic() || !mm)) {
1122 bad_area_nosemaphore(regs, error_code, address);
1123 return;
1124 }
1125
1126 /*
1127 * It's safe to allow irq's after cr2 has been saved and the
1128 * vmalloc fault has been handled.
1129 *
1130 * User-mode registers count as a user access even for any
1131 * potential system fault or CPU buglet:
1132 */
1133 if (user_mode_vm(regs)) {
1134 local_irq_enable();
1135 error_code |= PF_USER;
1136 flags |= FAULT_FLAG_USER;
1137 } else {
1138 if (regs->flags & X86_EFLAGS_IF)
1139 local_irq_enable();
1140 }
1141
1142 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1143
1144 if (error_code & PF_WRITE)
1145 flags |= FAULT_FLAG_WRITE;
1146
1147 /*
1148 * When running in the kernel we expect faults to occur only to
1149 * addresses in user space. All other faults represent errors in
1150 * the kernel and should generate an OOPS. Unfortunately, in the
1151 * case of an erroneous fault occurring in a code path which already
1152 * holds mmap_sem we will deadlock attempting to validate the fault
1153 * against the address space. Luckily the kernel only validly
1154 * references user space from well defined areas of code, which are
1155 * listed in the exceptions table.
1156 *
1157 * As the vast majority of faults will be valid we will only perform
1158 * the source reference check when there is a possibility of a
1159 * deadlock. Attempt to lock the address space, if we cannot we then
1160 * validate the source. If this is invalid we can skip the address
1161 * space check, thus avoiding the deadlock:
1162 */
1163 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1164 if ((error_code & PF_USER) == 0 &&
1165 !search_exception_tables(regs->ip)) {
1166 bad_area_nosemaphore(regs, error_code, address);
1167 return;
1168 }
1169retry:
1170 down_read(&mm->mmap_sem);
1171 } else {
1172 /*
1173 * The above down_read_trylock() might have succeeded in
1174 * which case we'll have missed the might_sleep() from
1175 * down_read():
1176 */
1177 might_sleep();
1178 }
1179
1180 vma = find_vma(mm, address);
1181 if (unlikely(!vma)) {
1182 bad_area(regs, error_code, address);
1183 return;
1184 }
1185 if (likely(vma->vm_start <= address))
1186 goto good_area;
1187 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1188 bad_area(regs, error_code, address);
1189 return;
1190 }
1191 if (error_code & PF_USER) {
1192 /*
1193 * Accessing the stack below %sp is always a bug.
1194 * The large cushion allows instructions like enter
1195 * and pusha to work. ("enter $65535, $31" pushes
1196 * 32 pointers and then decrements %sp by 65535.)
1197 */
1198 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1199 bad_area(regs, error_code, address);
1200 return;
1201 }
1202 }
1203 if (unlikely(expand_stack(vma, address))) {
1204 bad_area(regs, error_code, address);
1205 return;
1206 }
1207
1208 /*
1209 * Ok, we have a good vm_area for this memory access, so
1210 * we can handle it..
1211 */
1212good_area:
1213 if (unlikely(access_error(error_code, vma))) {
1214 bad_area_access_error(regs, error_code, address);
1215 return;
1216 }
1217
1218 /*
1219 * If for any reason at all we couldn't handle the fault,
1220 * make sure we exit gracefully rather than endlessly redo
1221 * the fault:
1222 */
1223 fault = handle_mm_fault(mm, vma, address, flags);
1224
1225 /*
1226 * If we need to retry but a fatal signal is pending, handle the
1227 * signal first. We do not need to release the mmap_sem because it
1228 * would already be released in __lock_page_or_retry in mm/filemap.c.
1229 */
1230 if (unlikely((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)))
1231 return;
1232
1233 if (unlikely(fault & VM_FAULT_ERROR)) {
1234 mm_fault_error(regs, error_code, address, fault);
1235 return;
1236 }
1237
1238 /*
1239 * Major/minor page fault accounting is only done on the
1240 * initial attempt. If we go through a retry, it is extremely
1241 * likely that the page will be found in page cache at that point.
1242 */
1243 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1244 if (fault & VM_FAULT_MAJOR) {
1245 tsk->maj_flt++;
1246 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1247 regs, address);
1248 } else {
1249 tsk->min_flt++;
1250 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1251 regs, address);
1252 }
1253 if (fault & VM_FAULT_RETRY) {
1254 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1255 * of starvation. */
1256 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1257 flags |= FAULT_FLAG_TRIED;
1258 goto retry;
1259 }
1260 }
1261
1262 check_v8086_mode(regs, address, tsk);
1263
1264 up_read(&mm->mmap_sem);
1265}
1266NOKPROBE_SYMBOL(__do_page_fault);
1267
1268dotraplinkage void notrace
1269do_page_fault(struct pt_regs *regs, unsigned long error_code)
1270{
1271 unsigned long address = read_cr2(); /* Get the faulting address */
1272 enum ctx_state prev_state;
1273
1274 /*
1275 * We must have this function tagged with __kprobes, notrace and call
1276 * read_cr2() before calling anything else. To avoid calling any kind
1277 * of tracing machinery before we've observed the CR2 value.
1278 *
1279 * exception_{enter,exit}() contain all sorts of tracepoints.
1280 */
1281
1282 prev_state = exception_enter();
1283 __do_page_fault(regs, error_code, address);
1284 exception_exit(prev_state);
1285}
1286NOKPROBE_SYMBOL(do_page_fault);
1287
1288#ifdef CONFIG_TRACING
1289static nokprobe_inline void
1290trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1291 unsigned long error_code)
1292{
1293 if (user_mode(regs))
1294 trace_page_fault_user(address, regs, error_code);
1295 else
1296 trace_page_fault_kernel(address, regs, error_code);
1297}
1298
1299dotraplinkage void notrace
1300trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1301{
1302 /*
1303 * The exception_enter and tracepoint processing could
1304 * trigger another page faults (user space callchain
1305 * reading) and destroy the original cr2 value, so read
1306 * the faulting address now.
1307 */
1308 unsigned long address = read_cr2();
1309 enum ctx_state prev_state;
1310
1311 prev_state = exception_enter();
1312 trace_page_fault_entries(address, regs, error_code);
1313 __do_page_fault(regs, error_code, address);
1314 exception_exit(prev_state);
1315}
1316NOKPROBE_SYMBOL(trace_do_page_fault);
1317#endif /* CONFIG_TRACING */
Linux kernel - Deliverables
This page took 0.030548 seconds and 5 git commands to generate.