Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 1995 Linus Torvalds
4 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
5 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
6 */
7#include <linux/sched.h> /* test_thread_flag(), ... */
8#include <linux/sched/task_stack.h> /* task_stack_*(), ... */
9#include <linux/kdebug.h> /* oops_begin/end, ... */
10#include <linux/extable.h> /* search_exception_tables */
11#include <linux/memblock.h> /* max_low_pfn */
12#include <linux/kfence.h> /* kfence_handle_page_fault */
13#include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
14#include <linux/mmiotrace.h> /* kmmio_handler, ... */
15#include <linux/perf_event.h> /* perf_sw_event */
16#include <linux/hugetlb.h> /* hstate_index_to_shift */
17#include <linux/prefetch.h> /* prefetchw */
18#include <linux/context_tracking.h> /* exception_enter(), ... */
19#include <linux/uaccess.h> /* faulthandler_disabled() */
20#include <linux/efi.h> /* efi_crash_gracefully_on_page_fault()*/
21#include <linux/mm_types.h>
22#include <linux/mm.h> /* find_and_lock_vma() */
23#include <linux/vmalloc.h>
24
25#include <asm/cpufeature.h> /* boot_cpu_has, ... */
26#include <asm/traps.h> /* dotraplinkage, ... */
27#include <asm/fixmap.h> /* VSYSCALL_ADDR */
28#include <asm/vsyscall.h> /* emulate_vsyscall */
29#include <asm/vm86.h> /* struct vm86 */
30#include <asm/mmu_context.h> /* vma_pkey() */
31#include <asm/efi.h> /* efi_crash_gracefully_on_page_fault()*/
32#include <asm/desc.h> /* store_idt(), ... */
33#include <asm/cpu_entry_area.h> /* exception stack */
34#include <asm/pgtable_areas.h> /* VMALLOC_START, ... */
35#include <asm/kvm_para.h> /* kvm_handle_async_pf */
36#include <asm/vdso.h> /* fixup_vdso_exception() */
37#include <asm/irq_stack.h>
38#include <asm/fred.h>
39#include <asm/sev.h> /* snp_dump_hva_rmpentry() */
40
41#define CREATE_TRACE_POINTS
42#include <asm/trace/exceptions.h>
43
44/*
45 * Returns 0 if mmiotrace is disabled, or if the fault is not
46 * handled by mmiotrace:
47 */
48static nokprobe_inline int
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
57/*
58 * Prefetch quirks:
59 *
60 * 32-bit mode:
61 *
62 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
63 * Check that here and ignore it. This is AMD erratum #91.
64 *
65 * 64-bit mode:
66 *
67 * Sometimes the CPU reports invalid exceptions on prefetch.
68 * Check that here and ignore it.
69 *
70 * Opcode checker based on code by Richard Brunner.
71 */
72static inline int
73check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
74 unsigned char opcode, int *prefetch)
75{
76 unsigned char instr_hi = opcode & 0xf0;
77 unsigned char instr_lo = opcode & 0x0f;
78
79 switch (instr_hi) {
80 case 0x20:
81 case 0x30:
82 /*
83 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
84 * In X86_64 long mode, the CPU will signal invalid
85 * opcode if some of these prefixes are present so
86 * X86_64 will never get here anyway
87 */
88 return ((instr_lo & 7) == 0x6);
89#ifdef CONFIG_X86_64
90 case 0x40:
91 /*
92 * In 64-bit mode 0x40..0x4F are valid REX prefixes
93 */
94 return (!user_mode(regs) || user_64bit_mode(regs));
95#endif
96 case 0x60:
97 /* 0x64 thru 0x67 are valid prefixes in all modes. */
98 return (instr_lo & 0xC) == 0x4;
99 case 0xF0:
100 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
101 return !instr_lo || (instr_lo>>1) == 1;
102 case 0x00:
103 /* Prefetch instruction is 0x0F0D or 0x0F18 */
104 if (get_kernel_nofault(opcode, instr))
105 return 0;
106
107 *prefetch = (instr_lo == 0xF) &&
108 (opcode == 0x0D || opcode == 0x18);
109 return 0;
110 default:
111 return 0;
112 }
113}
114
115static bool is_amd_k8_pre_npt(void)
116{
117 struct cpuinfo_x86 *c = &boot_cpu_data;
118
119 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
120 c->x86_vendor == X86_VENDOR_AMD &&
121 c->x86 == 0xf && c->x86_model < 0x40);
122}
123
124static int
125is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
126{
127 unsigned char *max_instr;
128 unsigned char *instr;
129 int prefetch = 0;
130
131 /* Erratum #91 affects AMD K8, pre-NPT CPUs */
132 if (!is_amd_k8_pre_npt())
133 return 0;
134
135 /*
136 * If it was a exec (instruction fetch) fault on NX page, then
137 * do not ignore the fault:
138 */
139 if (error_code & X86_PF_INSTR)
140 return 0;
141
142 instr = (void *)convert_ip_to_linear(current, regs);
143 max_instr = instr + 15;
144
145 /*
146 * This code has historically always bailed out if IP points to a
147 * not-present page (e.g. due to a race). No one has ever
148 * complained about this.
149 */
150 pagefault_disable();
151
152 while (instr < max_instr) {
153 unsigned char opcode;
154
155 if (user_mode(regs)) {
156 if (get_user(opcode, (unsigned char __user *) instr))
157 break;
158 } else {
159 if (get_kernel_nofault(opcode, instr))
160 break;
161 }
162
163 instr++;
164
165 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
166 break;
167 }
168
169 pagefault_enable();
170 return prefetch;
171}
172
173DEFINE_SPINLOCK(pgd_lock);
174LIST_HEAD(pgd_list);
175
176#ifdef CONFIG_X86_32
177static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
178{
179 unsigned index = pgd_index(address);
180 pgd_t *pgd_k;
181 p4d_t *p4d, *p4d_k;
182 pud_t *pud, *pud_k;
183 pmd_t *pmd, *pmd_k;
184
185 pgd += index;
186 pgd_k = init_mm.pgd + index;
187
188 if (!pgd_present(*pgd_k))
189 return NULL;
190
191 /*
192 * set_pgd(pgd, *pgd_k); here would be useless on PAE
193 * and redundant with the set_pmd() on non-PAE. As would
194 * set_p4d/set_pud.
195 */
196 p4d = p4d_offset(pgd, address);
197 p4d_k = p4d_offset(pgd_k, address);
198 if (!p4d_present(*p4d_k))
199 return NULL;
200
201 pud = pud_offset(p4d, address);
202 pud_k = pud_offset(p4d_k, address);
203 if (!pud_present(*pud_k))
204 return NULL;
205
206 pmd = pmd_offset(pud, address);
207 pmd_k = pmd_offset(pud_k, address);
208
209 if (pmd_present(*pmd) != pmd_present(*pmd_k))
210 set_pmd(pmd, *pmd_k);
211
212 if (!pmd_present(*pmd_k))
213 return NULL;
214 else
215 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
216
217 return pmd_k;
218}
219
220/*
221 * Handle a fault on the vmalloc or module mapping area
222 *
223 * This is needed because there is a race condition between the time
224 * when the vmalloc mapping code updates the PMD to the point in time
225 * where it synchronizes this update with the other page-tables in the
226 * system.
227 *
228 * In this race window another thread/CPU can map an area on the same
229 * PMD, finds it already present and does not synchronize it with the
230 * rest of the system yet. As a result v[mz]alloc might return areas
231 * which are not mapped in every page-table in the system, causing an
232 * unhandled page-fault when they are accessed.
233 */
234static noinline int vmalloc_fault(unsigned long address)
235{
236 unsigned long pgd_paddr;
237 pmd_t *pmd_k;
238 pte_t *pte_k;
239
240 /* Make sure we are in vmalloc area: */
241 if (!(address >= VMALLOC_START && address < VMALLOC_END))
242 return -1;
243
244 /*
245 * Synchronize this task's top level page-table
246 * with the 'reference' page table.
247 *
248 * Do _not_ use "current" here. We might be inside
249 * an interrupt in the middle of a task switch..
250 */
251 pgd_paddr = read_cr3_pa();
252 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
253 if (!pmd_k)
254 return -1;
255
256 if (pmd_leaf(*pmd_k))
257 return 0;
258
259 pte_k = pte_offset_kernel(pmd_k, address);
260 if (!pte_present(*pte_k))
261 return -1;
262
263 return 0;
264}
265NOKPROBE_SYMBOL(vmalloc_fault);
266
267void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
268{
269 unsigned long addr;
270
271 for (addr = start & PMD_MASK;
272 addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
273 addr += PMD_SIZE) {
274 struct page *page;
275
276 spin_lock(&pgd_lock);
277 list_for_each_entry(page, &pgd_list, lru) {
278 spinlock_t *pgt_lock;
279
280 /* the pgt_lock only for Xen */
281 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
282
283 spin_lock(pgt_lock);
284 vmalloc_sync_one(page_address(page), addr);
285 spin_unlock(pgt_lock);
286 }
287 spin_unlock(&pgd_lock);
288 }
289}
290
291static bool low_pfn(unsigned long pfn)
292{
293 return pfn < max_low_pfn;
294}
295
296static void dump_pagetable(unsigned long address)
297{
298 pgd_t *base = __va(read_cr3_pa());
299 pgd_t *pgd = &base[pgd_index(address)];
300 p4d_t *p4d;
301 pud_t *pud;
302 pmd_t *pmd;
303 pte_t *pte;
304
305#ifdef CONFIG_X86_PAE
306 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
307 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
308 goto out;
309#define pr_pde pr_cont
310#else
311#define pr_pde pr_info
312#endif
313 p4d = p4d_offset(pgd, address);
314 pud = pud_offset(p4d, address);
315 pmd = pmd_offset(pud, address);
316 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
317#undef pr_pde
318
319 /*
320 * We must not directly access the pte in the highpte
321 * case if the page table is located in highmem.
322 * And let's rather not kmap-atomic the pte, just in case
323 * it's allocated already:
324 */
325 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_leaf(*pmd))
326 goto out;
327
328 pte = pte_offset_kernel(pmd, address);
329 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
330out:
331 pr_cont("\n");
332}
333
334#else /* CONFIG_X86_64: */
335
336#ifdef CONFIG_CPU_SUP_AMD
337static const char errata93_warning[] =
338KERN_ERR
339"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
340"******* Working around it, but it may cause SEGVs or burn power.\n"
341"******* Please consider a BIOS update.\n"
342"******* Disabling USB legacy in the BIOS may also help.\n";
343#endif
344
345static int bad_address(void *p)
346{
347 unsigned long dummy;
348
349 return get_kernel_nofault(dummy, (unsigned long *)p);
350}
351
352static void dump_pagetable(unsigned long address)
353{
354 pgd_t *base = __va(read_cr3_pa());
355 pgd_t *pgd = base + pgd_index(address);
356 p4d_t *p4d;
357 pud_t *pud;
358 pmd_t *pmd;
359 pte_t *pte;
360
361 if (bad_address(pgd))
362 goto bad;
363
364 pr_info("PGD %lx ", pgd_val(*pgd));
365
366 if (!pgd_present(*pgd))
367 goto out;
368
369 p4d = p4d_offset(pgd, address);
370 if (bad_address(p4d))
371 goto bad;
372
373 pr_cont("P4D %lx ", p4d_val(*p4d));
374 if (!p4d_present(*p4d) || p4d_leaf(*p4d))
375 goto out;
376
377 pud = pud_offset(p4d, address);
378 if (bad_address(pud))
379 goto bad;
380
381 pr_cont("PUD %lx ", pud_val(*pud));
382 if (!pud_present(*pud) || pud_leaf(*pud))
383 goto out;
384
385 pmd = pmd_offset(pud, address);
386 if (bad_address(pmd))
387 goto bad;
388
389 pr_cont("PMD %lx ", pmd_val(*pmd));
390 if (!pmd_present(*pmd) || pmd_leaf(*pmd))
391 goto out;
392
393 pte = pte_offset_kernel(pmd, address);
394 if (bad_address(pte))
395 goto bad;
396
397 pr_cont("PTE %lx", pte_val(*pte));
398out:
399 pr_cont("\n");
400 return;
401bad:
402 pr_info("BAD\n");
403}
404
405#endif /* CONFIG_X86_64 */
406
407/*
408 * Workaround for K8 erratum #93 & buggy BIOS.
409 *
410 * BIOS SMM functions are required to use a specific workaround
411 * to avoid corruption of the 64bit RIP register on C stepping K8.
412 *
413 * A lot of BIOS that didn't get tested properly miss this.
414 *
415 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
416 * Try to work around it here.
417 *
418 * Note we only handle faults in kernel here.
419 * Does nothing on 32-bit.
420 */
421static int is_errata93(struct pt_regs *regs, unsigned long address)
422{
423#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
424 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
425 || boot_cpu_data.x86 != 0xf)
426 return 0;
427
428 if (user_mode(regs))
429 return 0;
430
431 if (address != regs->ip)
432 return 0;
433
434 if ((address >> 32) != 0)
435 return 0;
436
437 address |= 0xffffffffUL << 32;
438 if ((address >= (u64)_stext && address <= (u64)_etext) ||
439 (address >= MODULES_VADDR && address <= MODULES_END)) {
440 printk_once(errata93_warning);
441 regs->ip = address;
442 return 1;
443 }
444#endif
445 return 0;
446}
447
448/*
449 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
450 * to illegal addresses >4GB.
451 *
452 * We catch this in the page fault handler because these addresses
453 * are not reachable. Just detect this case and return. Any code
454 * segment in LDT is compatibility mode.
455 */
456static int is_errata100(struct pt_regs *regs, unsigned long address)
457{
458#ifdef CONFIG_X86_64
459 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
460 return 1;
461#endif
462 return 0;
463}
464
465/* Pentium F0 0F C7 C8 bug workaround: */
466static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
467 unsigned long address)
468{
469#ifdef CONFIG_X86_F00F_BUG
470 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
471 idt_is_f00f_address(address)) {
472 handle_invalid_op(regs);
473 return 1;
474 }
475#endif
476 return 0;
477}
478
479static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
480{
481 u32 offset = (index >> 3) * sizeof(struct desc_struct);
482 unsigned long addr;
483 struct ldttss_desc desc;
484
485 if (index == 0) {
486 pr_alert("%s: NULL\n", name);
487 return;
488 }
489
490 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
491 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
492 return;
493 }
494
495 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
496 sizeof(struct ldttss_desc))) {
497 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
498 name, index);
499 return;
500 }
501
502 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
503#ifdef CONFIG_X86_64
504 addr |= ((u64)desc.base3 << 32);
505#endif
506 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
507 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
508}
509
510static void
511show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
512{
513 if (!oops_may_print())
514 return;
515
516 if (error_code & X86_PF_INSTR) {
517 unsigned int level;
518 bool nx, rw;
519 pgd_t *pgd;
520 pte_t *pte;
521
522 pgd = __va(read_cr3_pa());
523 pgd += pgd_index(address);
524
525 pte = lookup_address_in_pgd_attr(pgd, address, &level, &nx, &rw);
526
527 if (pte && pte_present(*pte) && (!pte_exec(*pte) || nx))
528 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
529 from_kuid(&init_user_ns, current_uid()));
530 if (pte && pte_present(*pte) && pte_exec(*pte) && !nx &&
531 (pgd_flags(*pgd) & _PAGE_USER) &&
532 (__read_cr4() & X86_CR4_SMEP))
533 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
534 from_kuid(&init_user_ns, current_uid()));
535 }
536
537 if (address < PAGE_SIZE && !user_mode(regs))
538 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
539 (void *)address);
540 else
541 pr_alert("BUG: unable to handle page fault for address: %px\n",
542 (void *)address);
543
544 pr_alert("#PF: %s %s in %s mode\n",
545 (error_code & X86_PF_USER) ? "user" : "supervisor",
546 (error_code & X86_PF_INSTR) ? "instruction fetch" :
547 (error_code & X86_PF_WRITE) ? "write access" :
548 "read access",
549 user_mode(regs) ? "user" : "kernel");
550 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
551 !(error_code & X86_PF_PROT) ? "not-present page" :
552 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
553 (error_code & X86_PF_PK) ? "protection keys violation" :
554 (error_code & X86_PF_RMP) ? "RMP violation" :
555 "permissions violation");
556
557 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
558 struct desc_ptr idt, gdt;
559 u16 ldtr, tr;
560
561 /*
562 * This can happen for quite a few reasons. The more obvious
563 * ones are faults accessing the GDT, or LDT. Perhaps
564 * surprisingly, if the CPU tries to deliver a benign or
565 * contributory exception from user code and gets a page fault
566 * during delivery, the page fault can be delivered as though
567 * it originated directly from user code. This could happen
568 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
569 * kernel or IST stack.
570 */
571 store_idt(&idt);
572
573 /* Usable even on Xen PV -- it's just slow. */
574 native_store_gdt(&gdt);
575
576 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
577 idt.address, idt.size, gdt.address, gdt.size);
578
579 store_ldt(ldtr);
580 show_ldttss(&gdt, "LDTR", ldtr);
581
582 store_tr(tr);
583 show_ldttss(&gdt, "TR", tr);
584 }
585
586 dump_pagetable(address);
587
588 if (error_code & X86_PF_RMP)
589 snp_dump_hva_rmpentry(address);
590}
591
592static noinline void
593pgtable_bad(struct pt_regs *regs, unsigned long error_code,
594 unsigned long address)
595{
596 struct task_struct *tsk;
597 unsigned long flags;
598 int sig;
599
600 flags = oops_begin();
601 tsk = current;
602 sig = SIGKILL;
603
604 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
605 tsk->comm, address);
606 dump_pagetable(address);
607
608 if (__die("Bad pagetable", regs, error_code))
609 sig = 0;
610
611 oops_end(flags, regs, sig);
612}
613
614static void sanitize_error_code(unsigned long address,
615 unsigned long *error_code)
616{
617 /*
618 * To avoid leaking information about the kernel page
619 * table layout, pretend that user-mode accesses to
620 * kernel addresses are always protection faults.
621 *
622 * NB: This means that failed vsyscalls with vsyscall=none
623 * will have the PROT bit. This doesn't leak any
624 * information and does not appear to cause any problems.
625 */
626 if (address >= TASK_SIZE_MAX)
627 *error_code |= X86_PF_PROT;
628}
629
630static void set_signal_archinfo(unsigned long address,
631 unsigned long error_code)
632{
633 struct task_struct *tsk = current;
634
635 tsk->thread.trap_nr = X86_TRAP_PF;
636 tsk->thread.error_code = error_code | X86_PF_USER;
637 tsk->thread.cr2 = address;
638}
639
640static noinline void
641page_fault_oops(struct pt_regs *regs, unsigned long error_code,
642 unsigned long address)
643{
644#ifdef CONFIG_VMAP_STACK
645 struct stack_info info;
646#endif
647 unsigned long flags;
648 int sig;
649
650 if (user_mode(regs)) {
651 /*
652 * Implicit kernel access from user mode? Skip the stack
653 * overflow and EFI special cases.
654 */
655 goto oops;
656 }
657
658#ifdef CONFIG_VMAP_STACK
659 /*
660 * Stack overflow? During boot, we can fault near the initial
661 * stack in the direct map, but that's not an overflow -- check
662 * that we're in vmalloc space to avoid this.
663 */
664 if (is_vmalloc_addr((void *)address) &&
665 get_stack_guard_info((void *)address, &info)) {
666 /*
667 * We're likely to be running with very little stack space
668 * left. It's plausible that we'd hit this condition but
669 * double-fault even before we get this far, in which case
670 * we're fine: the double-fault handler will deal with it.
671 *
672 * We don't want to make it all the way into the oops code
673 * and then double-fault, though, because we're likely to
674 * break the console driver and lose most of the stack dump.
675 */
676 call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
677 handle_stack_overflow,
678 ASM_CALL_ARG3,
679 , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
680
681 BUG();
682 }
683#endif
684
685 /*
686 * Buggy firmware could access regions which might page fault. If
687 * this happens, EFI has a special OOPS path that will try to
688 * avoid hanging the system.
689 */
690 if (IS_ENABLED(CONFIG_EFI))
691 efi_crash_gracefully_on_page_fault(address);
692
693 /* Only not-present faults should be handled by KFENCE. */
694 if (!(error_code & X86_PF_PROT) &&
695 kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
696 return;
697
698oops:
699 /*
700 * Oops. The kernel tried to access some bad page. We'll have to
701 * terminate things with extreme prejudice:
702 */
703 flags = oops_begin();
704
705 show_fault_oops(regs, error_code, address);
706
707 if (task_stack_end_corrupted(current))
708 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
709
710 sig = SIGKILL;
711 if (__die("Oops", regs, error_code))
712 sig = 0;
713
714 /* Executive summary in case the body of the oops scrolled away */
715 printk(KERN_DEFAULT "CR2: %016lx\n", address);
716
717 oops_end(flags, regs, sig);
718}
719
720static noinline void
721kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
722 unsigned long address, int signal, int si_code,
723 u32 pkey)
724{
725 WARN_ON_ONCE(user_mode(regs));
726
727 /* Are we prepared to handle this kernel fault? */
728 if (fixup_exception(regs, X86_TRAP_PF, error_code, address))
729 return;
730
731 /*
732 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
733 * instruction.
734 */
735 if (is_prefetch(regs, error_code, address))
736 return;
737
738 page_fault_oops(regs, error_code, address);
739}
740
741/*
742 * Print out info about fatal segfaults, if the show_unhandled_signals
743 * sysctl is set:
744 */
745static inline void
746show_signal_msg(struct pt_regs *regs, unsigned long error_code,
747 unsigned long address, struct task_struct *tsk)
748{
749 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
750 /* This is a racy snapshot, but it's better than nothing. */
751 int cpu = raw_smp_processor_id();
752
753 if (!unhandled_signal(tsk, SIGSEGV))
754 return;
755
756 if (!printk_ratelimit())
757 return;
758
759 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
760 loglvl, tsk->comm, task_pid_nr(tsk), address,
761 (void *)regs->ip, (void *)regs->sp, error_code);
762
763 print_vma_addr(KERN_CONT " in ", regs->ip);
764
765 /*
766 * Dump the likely CPU where the fatal segfault happened.
767 * This can help identify faulty hardware.
768 */
769 printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu,
770 topology_core_id(cpu), topology_physical_package_id(cpu));
771
772
773 printk(KERN_CONT "\n");
774
775 show_opcodes(regs, loglvl);
776}
777
778static void
779__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
780 unsigned long address, u32 pkey, int si_code)
781{
782 struct task_struct *tsk = current;
783
784 if (!user_mode(regs)) {
785 kernelmode_fixup_or_oops(regs, error_code, address,
786 SIGSEGV, si_code, pkey);
787 return;
788 }
789
790 if (!(error_code & X86_PF_USER)) {
791 /* Implicit user access to kernel memory -- just oops */
792 page_fault_oops(regs, error_code, address);
793 return;
794 }
795
796 /*
797 * User mode accesses just cause a SIGSEGV.
798 * It's possible to have interrupts off here:
799 */
800 local_irq_enable();
801
802 /*
803 * Valid to do another page fault here because this one came
804 * from user space:
805 */
806 if (is_prefetch(regs, error_code, address))
807 return;
808
809 if (is_errata100(regs, address))
810 return;
811
812 sanitize_error_code(address, &error_code);
813
814 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
815 return;
816
817 if (likely(show_unhandled_signals))
818 show_signal_msg(regs, error_code, address, tsk);
819
820 set_signal_archinfo(address, error_code);
821
822 if (si_code == SEGV_PKUERR)
823 force_sig_pkuerr((void __user *)address, pkey);
824 else
825 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
826
827 local_irq_disable();
828}
829
830static noinline void
831bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
832 unsigned long address)
833{
834 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
835}
836
837static void
838__bad_area(struct pt_regs *regs, unsigned long error_code,
839 unsigned long address, struct mm_struct *mm,
840 struct vm_area_struct *vma, u32 pkey, int si_code)
841{
842 /*
843 * Something tried to access memory that isn't in our memory map..
844 * Fix it, but check if it's kernel or user first..
845 */
846 if (mm)
847 mmap_read_unlock(mm);
848 else
849 vma_end_read(vma);
850
851 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
852}
853
854static inline bool bad_area_access_from_pkeys(unsigned long error_code,
855 struct vm_area_struct *vma)
856{
857 /* This code is always called on the current mm */
858 bool foreign = false;
859
860 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
861 return false;
862 if (error_code & X86_PF_PK)
863 return true;
864 /* this checks permission keys on the VMA: */
865 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
866 (error_code & X86_PF_INSTR), foreign))
867 return true;
868 return false;
869}
870
871static noinline void
872bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
873 unsigned long address, struct mm_struct *mm,
874 struct vm_area_struct *vma)
875{
876 /*
877 * This OSPKE check is not strictly necessary at runtime.
878 * But, doing it this way allows compiler optimizations
879 * if pkeys are compiled out.
880 */
881 if (bad_area_access_from_pkeys(error_code, vma)) {
882 /*
883 * A protection key fault means that the PKRU value did not allow
884 * access to some PTE. Userspace can figure out what PKRU was
885 * from the XSAVE state. This function captures the pkey from
886 * the vma and passes it to userspace so userspace can discover
887 * which protection key was set on the PTE.
888 *
889 * If we get here, we know that the hardware signaled a X86_PF_PK
890 * fault and that there was a VMA once we got in the fault
891 * handler. It does *not* guarantee that the VMA we find here
892 * was the one that we faulted on.
893 *
894 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
895 * 2. T1 : set PKRU to deny access to pkey=4, touches page
896 * 3. T1 : faults...
897 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
898 * 5. T1 : enters fault handler, takes mmap_lock, etc...
899 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
900 * faulted on a pte with its pkey=4.
901 */
902 u32 pkey = vma_pkey(vma);
903
904 __bad_area(regs, error_code, address, mm, vma, pkey, SEGV_PKUERR);
905 } else {
906 __bad_area(regs, error_code, address, mm, vma, 0, SEGV_ACCERR);
907 }
908}
909
910static void
911do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
912 vm_fault_t fault)
913{
914 /* Kernel mode? Handle exceptions or die: */
915 if (!user_mode(regs)) {
916 kernelmode_fixup_or_oops(regs, error_code, address,
917 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
918 return;
919 }
920
921 /* User-space => ok to do another page fault: */
922 if (is_prefetch(regs, error_code, address))
923 return;
924
925 sanitize_error_code(address, &error_code);
926
927 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
928 return;
929
930 set_signal_archinfo(address, error_code);
931
932#ifdef CONFIG_MEMORY_FAILURE
933 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
934 struct task_struct *tsk = current;
935 unsigned lsb = 0;
936
937 pr_err(
938 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
939 tsk->comm, tsk->pid, address);
940 if (fault & VM_FAULT_HWPOISON_LARGE)
941 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
942 if (fault & VM_FAULT_HWPOISON)
943 lsb = PAGE_SHIFT;
944 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
945 return;
946 }
947#endif
948 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
949}
950
951static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
952{
953 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
954 return 0;
955
956 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
957 return 0;
958
959 return 1;
960}
961
962/*
963 * Handle a spurious fault caused by a stale TLB entry.
964 *
965 * This allows us to lazily refresh the TLB when increasing the
966 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
967 * eagerly is very expensive since that implies doing a full
968 * cross-processor TLB flush, even if no stale TLB entries exist
969 * on other processors.
970 *
971 * Spurious faults may only occur if the TLB contains an entry with
972 * fewer permission than the page table entry. Non-present (P = 0)
973 * and reserved bit (R = 1) faults are never spurious.
974 *
975 * There are no security implications to leaving a stale TLB when
976 * increasing the permissions on a page.
977 *
978 * Returns non-zero if a spurious fault was handled, zero otherwise.
979 *
980 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
981 * (Optional Invalidation).
982 */
983static noinline int
984spurious_kernel_fault(unsigned long error_code, unsigned long address)
985{
986 pgd_t *pgd;
987 p4d_t *p4d;
988 pud_t *pud;
989 pmd_t *pmd;
990 pte_t *pte;
991 int ret;
992
993 /*
994 * Only writes to RO or instruction fetches from NX may cause
995 * spurious faults.
996 *
997 * These could be from user or supervisor accesses but the TLB
998 * is only lazily flushed after a kernel mapping protection
999 * change, so user accesses are not expected to cause spurious
1000 * faults.
1001 */
1002 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1003 error_code != (X86_PF_INSTR | X86_PF_PROT))
1004 return 0;
1005
1006 pgd = init_mm.pgd + pgd_index(address);
1007 if (!pgd_present(*pgd))
1008 return 0;
1009
1010 p4d = p4d_offset(pgd, address);
1011 if (!p4d_present(*p4d))
1012 return 0;
1013
1014 if (p4d_leaf(*p4d))
1015 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1016
1017 pud = pud_offset(p4d, address);
1018 if (!pud_present(*pud))
1019 return 0;
1020
1021 if (pud_leaf(*pud))
1022 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1023
1024 pmd = pmd_offset(pud, address);
1025 if (!pmd_present(*pmd))
1026 return 0;
1027
1028 if (pmd_leaf(*pmd))
1029 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1030
1031 pte = pte_offset_kernel(pmd, address);
1032 if (!pte_present(*pte))
1033 return 0;
1034
1035 ret = spurious_kernel_fault_check(error_code, pte);
1036 if (!ret)
1037 return 0;
1038
1039 /*
1040 * Make sure we have permissions in PMD.
1041 * If not, then there's a bug in the page tables:
1042 */
1043 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1044 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1045
1046 return ret;
1047}
1048NOKPROBE_SYMBOL(spurious_kernel_fault);
1049
1050int show_unhandled_signals = 1;
1051
1052static inline int
1053access_error(unsigned long error_code, struct vm_area_struct *vma)
1054{
1055 /* This is only called for the current mm, so: */
1056 bool foreign = false;
1057
1058 /*
1059 * Read or write was blocked by protection keys. This is
1060 * always an unconditional error and can never result in
1061 * a follow-up action to resolve the fault, like a COW.
1062 */
1063 if (error_code & X86_PF_PK)
1064 return 1;
1065
1066 /*
1067 * SGX hardware blocked the access. This usually happens
1068 * when the enclave memory contents have been destroyed, like
1069 * after a suspend/resume cycle. In any case, the kernel can't
1070 * fix the cause of the fault. Handle the fault as an access
1071 * error even in cases where no actual access violation
1072 * occurred. This allows userspace to rebuild the enclave in
1073 * response to the signal.
1074 */
1075 if (unlikely(error_code & X86_PF_SGX))
1076 return 1;
1077
1078 /*
1079 * Make sure to check the VMA so that we do not perform
1080 * faults just to hit a X86_PF_PK as soon as we fill in a
1081 * page.
1082 */
1083 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1084 (error_code & X86_PF_INSTR), foreign))
1085 return 1;
1086
1087 /*
1088 * Shadow stack accesses (PF_SHSTK=1) are only permitted to
1089 * shadow stack VMAs. All other accesses result in an error.
1090 */
1091 if (error_code & X86_PF_SHSTK) {
1092 if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK)))
1093 return 1;
1094 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1095 return 1;
1096 return 0;
1097 }
1098
1099 if (error_code & X86_PF_WRITE) {
1100 /* write, present and write, not present: */
1101 if (unlikely(vma->vm_flags & VM_SHADOW_STACK))
1102 return 1;
1103 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1104 return 1;
1105 return 0;
1106 }
1107
1108 /* read, present: */
1109 if (unlikely(error_code & X86_PF_PROT))
1110 return 1;
1111
1112 /* read, not present: */
1113 if (unlikely(!vma_is_accessible(vma)))
1114 return 1;
1115
1116 return 0;
1117}
1118
1119bool fault_in_kernel_space(unsigned long address)
1120{
1121 /*
1122 * On 64-bit systems, the vsyscall page is at an address above
1123 * TASK_SIZE_MAX, but is not considered part of the kernel
1124 * address space.
1125 */
1126 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1127 return false;
1128
1129 return address >= TASK_SIZE_MAX;
1130}
1131
1132/*
1133 * Called for all faults where 'address' is part of the kernel address
1134 * space. Might get called for faults that originate from *code* that
1135 * ran in userspace or the kernel.
1136 */
1137static void
1138do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1139 unsigned long address)
1140{
1141 /*
1142 * Protection keys exceptions only happen on user pages. We
1143 * have no user pages in the kernel portion of the address
1144 * space, so do not expect them here.
1145 */
1146 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1147
1148#ifdef CONFIG_X86_32
1149 /*
1150 * We can fault-in kernel-space virtual memory on-demand. The
1151 * 'reference' page table is init_mm.pgd.
1152 *
1153 * NOTE! We MUST NOT take any locks for this case. We may
1154 * be in an interrupt or a critical region, and should
1155 * only copy the information from the master page table,
1156 * nothing more.
1157 *
1158 * Before doing this on-demand faulting, ensure that the
1159 * fault is not any of the following:
1160 * 1. A fault on a PTE with a reserved bit set.
1161 * 2. A fault caused by a user-mode access. (Do not demand-
1162 * fault kernel memory due to user-mode accesses).
1163 * 3. A fault caused by a page-level protection violation.
1164 * (A demand fault would be on a non-present page which
1165 * would have X86_PF_PROT==0).
1166 *
1167 * This is only needed to close a race condition on x86-32 in
1168 * the vmalloc mapping/unmapping code. See the comment above
1169 * vmalloc_fault() for details. On x86-64 the race does not
1170 * exist as the vmalloc mappings don't need to be synchronized
1171 * there.
1172 */
1173 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1174 if (vmalloc_fault(address) >= 0)
1175 return;
1176 }
1177#endif
1178
1179 if (is_f00f_bug(regs, hw_error_code, address))
1180 return;
1181
1182 /* Was the fault spurious, caused by lazy TLB invalidation? */
1183 if (spurious_kernel_fault(hw_error_code, address))
1184 return;
1185
1186 /* kprobes don't want to hook the spurious faults: */
1187 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1188 return;
1189
1190 /*
1191 * Note, despite being a "bad area", there are quite a few
1192 * acceptable reasons to get here, such as erratum fixups
1193 * and handling kernel code that can fault, like get_user().
1194 *
1195 * Don't take the mm semaphore here. If we fixup a prefetch
1196 * fault we could otherwise deadlock:
1197 */
1198 bad_area_nosemaphore(regs, hw_error_code, address);
1199}
1200NOKPROBE_SYMBOL(do_kern_addr_fault);
1201
1202/*
1203 * Handle faults in the user portion of the address space. Nothing in here
1204 * should check X86_PF_USER without a specific justification: for almost
1205 * all purposes, we should treat a normal kernel access to user memory
1206 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1207 * The one exception is AC flag handling, which is, per the x86
1208 * architecture, special for WRUSS.
1209 */
1210static inline
1211void do_user_addr_fault(struct pt_regs *regs,
1212 unsigned long error_code,
1213 unsigned long address)
1214{
1215 struct vm_area_struct *vma;
1216 struct task_struct *tsk;
1217 struct mm_struct *mm;
1218 vm_fault_t fault;
1219 unsigned int flags = FAULT_FLAG_DEFAULT;
1220
1221 tsk = current;
1222 mm = tsk->mm;
1223
1224 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1225 /*
1226 * Whoops, this is kernel mode code trying to execute from
1227 * user memory. Unless this is AMD erratum #93, which
1228 * corrupts RIP such that it looks like a user address,
1229 * this is unrecoverable. Don't even try to look up the
1230 * VMA or look for extable entries.
1231 */
1232 if (is_errata93(regs, address))
1233 return;
1234
1235 page_fault_oops(regs, error_code, address);
1236 return;
1237 }
1238
1239 /* kprobes don't want to hook the spurious faults: */
1240 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1241 return;
1242
1243 /*
1244 * Reserved bits are never expected to be set on
1245 * entries in the user portion of the page tables.
1246 */
1247 if (unlikely(error_code & X86_PF_RSVD))
1248 pgtable_bad(regs, error_code, address);
1249
1250 /*
1251 * If SMAP is on, check for invalid kernel (supervisor) access to user
1252 * pages in the user address space. The odd case here is WRUSS,
1253 * which, according to the preliminary documentation, does not respect
1254 * SMAP and will have the USER bit set so, in all cases, SMAP
1255 * enforcement appears to be consistent with the USER bit.
1256 */
1257 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1258 !(error_code & X86_PF_USER) &&
1259 !(regs->flags & X86_EFLAGS_AC))) {
1260 /*
1261 * No extable entry here. This was a kernel access to an
1262 * invalid pointer. get_kernel_nofault() will not get here.
1263 */
1264 page_fault_oops(regs, error_code, address);
1265 return;
1266 }
1267
1268 /*
1269 * If we're in an interrupt, have no user context or are running
1270 * in a region with pagefaults disabled then we must not take the fault
1271 */
1272 if (unlikely(faulthandler_disabled() || !mm)) {
1273 bad_area_nosemaphore(regs, error_code, address);
1274 return;
1275 }
1276
1277 /* Legacy check - remove this after verifying that it doesn't trigger */
1278 if (WARN_ON_ONCE(!(regs->flags & X86_EFLAGS_IF))) {
1279 bad_area_nosemaphore(regs, error_code, address);
1280 return;
1281 }
1282
1283 local_irq_enable();
1284
1285 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1286
1287 /*
1288 * Read-only permissions can not be expressed in shadow stack PTEs.
1289 * Treat all shadow stack accesses as WRITE faults. This ensures
1290 * that the MM will prepare everything (e.g., break COW) such that
1291 * maybe_mkwrite() can create a proper shadow stack PTE.
1292 */
1293 if (error_code & X86_PF_SHSTK)
1294 flags |= FAULT_FLAG_WRITE;
1295 if (error_code & X86_PF_WRITE)
1296 flags |= FAULT_FLAG_WRITE;
1297 if (error_code & X86_PF_INSTR)
1298 flags |= FAULT_FLAG_INSTRUCTION;
1299
1300 /*
1301 * We set FAULT_FLAG_USER based on the register state, not
1302 * based on X86_PF_USER. User space accesses that cause
1303 * system page faults are still user accesses.
1304 */
1305 if (user_mode(regs))
1306 flags |= FAULT_FLAG_USER;
1307
1308#ifdef CONFIG_X86_64
1309 /*
1310 * Faults in the vsyscall page might need emulation. The
1311 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1312 * considered to be part of the user address space.
1313 *
1314 * The vsyscall page does not have a "real" VMA, so do this
1315 * emulation before we go searching for VMAs.
1316 *
1317 * PKRU never rejects instruction fetches, so we don't need
1318 * to consider the PF_PK bit.
1319 */
1320 if (is_vsyscall_vaddr(address)) {
1321 if (emulate_vsyscall(error_code, regs, address))
1322 return;
1323 }
1324#endif
1325
1326 if (!(flags & FAULT_FLAG_USER))
1327 goto lock_mmap;
1328
1329 vma = lock_vma_under_rcu(mm, address);
1330 if (!vma)
1331 goto lock_mmap;
1332
1333 if (unlikely(access_error(error_code, vma))) {
1334 bad_area_access_error(regs, error_code, address, NULL, vma);
1335 count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1336 return;
1337 }
1338 fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
1339 if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
1340 vma_end_read(vma);
1341
1342 if (!(fault & VM_FAULT_RETRY)) {
1343 count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1344 goto done;
1345 }
1346 count_vm_vma_lock_event(VMA_LOCK_RETRY);
1347 if (fault & VM_FAULT_MAJOR)
1348 flags |= FAULT_FLAG_TRIED;
1349
1350 /* Quick path to respond to signals */
1351 if (fault_signal_pending(fault, regs)) {
1352 if (!user_mode(regs))
1353 kernelmode_fixup_or_oops(regs, error_code, address,
1354 SIGBUS, BUS_ADRERR,
1355 ARCH_DEFAULT_PKEY);
1356 return;
1357 }
1358lock_mmap:
1359
1360retry:
1361 vma = lock_mm_and_find_vma(mm, address, regs);
1362 if (unlikely(!vma)) {
1363 bad_area_nosemaphore(regs, error_code, address);
1364 return;
1365 }
1366
1367 /*
1368 * Ok, we have a good vm_area for this memory access, so
1369 * we can handle it..
1370 */
1371 if (unlikely(access_error(error_code, vma))) {
1372 bad_area_access_error(regs, error_code, address, mm, vma);
1373 return;
1374 }
1375
1376 /*
1377 * If for any reason at all we couldn't handle the fault,
1378 * make sure we exit gracefully rather than endlessly redo
1379 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1380 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1381 *
1382 * Note that handle_userfault() may also release and reacquire mmap_lock
1383 * (and not return with VM_FAULT_RETRY), when returning to userland to
1384 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1385 * (potentially after handling any pending signal during the return to
1386 * userland). The return to userland is identified whenever
1387 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1388 */
1389 fault = handle_mm_fault(vma, address, flags, regs);
1390
1391 if (fault_signal_pending(fault, regs)) {
1392 /*
1393 * Quick path to respond to signals. The core mm code
1394 * has unlocked the mm for us if we get here.
1395 */
1396 if (!user_mode(regs))
1397 kernelmode_fixup_or_oops(regs, error_code, address,
1398 SIGBUS, BUS_ADRERR,
1399 ARCH_DEFAULT_PKEY);
1400 return;
1401 }
1402
1403 /* The fault is fully completed (including releasing mmap lock) */
1404 if (fault & VM_FAULT_COMPLETED)
1405 return;
1406
1407 /*
1408 * If we need to retry the mmap_lock has already been released,
1409 * and if there is a fatal signal pending there is no guarantee
1410 * that we made any progress. Handle this case first.
1411 */
1412 if (unlikely(fault & VM_FAULT_RETRY)) {
1413 flags |= FAULT_FLAG_TRIED;
1414 goto retry;
1415 }
1416
1417 mmap_read_unlock(mm);
1418done:
1419 if (likely(!(fault & VM_FAULT_ERROR)))
1420 return;
1421
1422 if (fatal_signal_pending(current) && !user_mode(regs)) {
1423 kernelmode_fixup_or_oops(regs, error_code, address,
1424 0, 0, ARCH_DEFAULT_PKEY);
1425 return;
1426 }
1427
1428 if (fault & VM_FAULT_OOM) {
1429 /* Kernel mode? Handle exceptions or die: */
1430 if (!user_mode(regs)) {
1431 kernelmode_fixup_or_oops(regs, error_code, address,
1432 SIGSEGV, SEGV_MAPERR,
1433 ARCH_DEFAULT_PKEY);
1434 return;
1435 }
1436
1437 /*
1438 * We ran out of memory, call the OOM killer, and return the
1439 * userspace (which will retry the fault, or kill us if we got
1440 * oom-killed):
1441 */
1442 pagefault_out_of_memory();
1443 } else {
1444 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1445 VM_FAULT_HWPOISON_LARGE))
1446 do_sigbus(regs, error_code, address, fault);
1447 else if (fault & VM_FAULT_SIGSEGV)
1448 bad_area_nosemaphore(regs, error_code, address);
1449 else
1450 BUG();
1451 }
1452}
1453NOKPROBE_SYMBOL(do_user_addr_fault);
1454
1455static __always_inline void
1456trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1457 unsigned long address)
1458{
1459 if (!trace_pagefault_enabled())
1460 return;
1461
1462 if (user_mode(regs))
1463 trace_page_fault_user(address, regs, error_code);
1464 else
1465 trace_page_fault_kernel(address, regs, error_code);
1466}
1467
1468static __always_inline void
1469handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1470 unsigned long address)
1471{
1472 trace_page_fault_entries(regs, error_code, address);
1473
1474 if (unlikely(kmmio_fault(regs, address)))
1475 return;
1476
1477 /* Was the fault on kernel-controlled part of the address space? */
1478 if (unlikely(fault_in_kernel_space(address))) {
1479 do_kern_addr_fault(regs, error_code, address);
1480 } else {
1481 do_user_addr_fault(regs, error_code, address);
1482 /*
1483 * User address page fault handling might have reenabled
1484 * interrupts. Fixing up all potential exit points of
1485 * do_user_addr_fault() and its leaf functions is just not
1486 * doable w/o creating an unholy mess or turning the code
1487 * upside down.
1488 */
1489 local_irq_disable();
1490 }
1491}
1492
1493DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1494{
1495 irqentry_state_t state;
1496 unsigned long address;
1497
1498 address = cpu_feature_enabled(X86_FEATURE_FRED) ? fred_event_data(regs) : read_cr2();
1499
1500 prefetchw(¤t->mm->mmap_lock);
1501
1502 /*
1503 * KVM uses #PF vector to deliver 'page not present' events to guests
1504 * (asynchronous page fault mechanism). The event happens when a
1505 * userspace task is trying to access some valid (from guest's point of
1506 * view) memory which is not currently mapped by the host (e.g. the
1507 * memory is swapped out). Note, the corresponding "page ready" event
1508 * which is injected when the memory becomes available, is delivered via
1509 * an interrupt mechanism and not a #PF exception
1510 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1511 *
1512 * We are relying on the interrupted context being sane (valid RSP,
1513 * relevant locks not held, etc.), which is fine as long as the
1514 * interrupted context had IF=1. We are also relying on the KVM
1515 * async pf type field and CR2 being read consistently instead of
1516 * getting values from real and async page faults mixed up.
1517 *
1518 * Fingers crossed.
1519 *
1520 * The async #PF handling code takes care of idtentry handling
1521 * itself.
1522 */
1523 if (kvm_handle_async_pf(regs, (u32)address))
1524 return;
1525
1526 /*
1527 * Entry handling for valid #PF from kernel mode is slightly
1528 * different: RCU is already watching and ct_irq_enter() must not
1529 * be invoked because a kernel fault on a user space address might
1530 * sleep.
1531 *
1532 * In case the fault hit a RCU idle region the conditional entry
1533 * code reenabled RCU to avoid subsequent wreckage which helps
1534 * debuggability.
1535 */
1536 state = irqentry_enter(regs);
1537
1538 instrumentation_begin();
1539 handle_page_fault(regs, error_code, address);
1540 instrumentation_end();
1541
1542 irqentry_exit(regs, state);
1543}
1/*
2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
4 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
5 */
6#include <linux/sched.h> /* test_thread_flag(), ... */
7#include <linux/kdebug.h> /* oops_begin/end, ... */
8#include <linux/module.h> /* search_exception_table */
9#include <linux/bootmem.h> /* max_low_pfn */
10#include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
11#include <linux/mmiotrace.h> /* kmmio_handler, ... */
12#include <linux/perf_event.h> /* perf_sw_event */
13#include <linux/hugetlb.h> /* hstate_index_to_shift */
14#include <linux/prefetch.h> /* prefetchw */
15#include <linux/context_tracking.h> /* exception_enter(), ... */
16#include <linux/uaccess.h> /* faulthandler_disabled() */
17
18#include <asm/cpufeature.h> /* boot_cpu_has, ... */
19#include <asm/traps.h> /* dotraplinkage, ... */
20#include <asm/pgalloc.h> /* pgd_*(), ... */
21#include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
22#include <asm/fixmap.h> /* VSYSCALL_ADDR */
23#include <asm/vsyscall.h> /* emulate_vsyscall */
24#include <asm/vm86.h> /* struct vm86 */
25#include <asm/mmu_context.h> /* vma_pkey() */
26
27#define CREATE_TRACE_POINTS
28#include <asm/trace/exceptions.h>
29
30/*
31 * Page fault error code bits:
32 *
33 * bit 0 == 0: no page found 1: protection fault
34 * bit 1 == 0: read access 1: write access
35 * bit 2 == 0: kernel-mode access 1: user-mode access
36 * bit 3 == 1: use of reserved bit detected
37 * bit 4 == 1: fault was an instruction fetch
38 * bit 5 == 1: protection keys block access
39 */
40enum x86_pf_error_code {
41
42 PF_PROT = 1 << 0,
43 PF_WRITE = 1 << 1,
44 PF_USER = 1 << 2,
45 PF_RSVD = 1 << 3,
46 PF_INSTR = 1 << 4,
47 PF_PK = 1 << 5,
48};
49
50/*
51 * Returns 0 if mmiotrace is disabled, or if the fault is not
52 * handled by mmiotrace:
53 */
54static nokprobe_inline int
55kmmio_fault(struct pt_regs *regs, unsigned long addr)
56{
57 if (unlikely(is_kmmio_active()))
58 if (kmmio_handler(regs, addr) == 1)
59 return -1;
60 return 0;
61}
62
63static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
64{
65 int ret = 0;
66
67 /* kprobe_running() needs smp_processor_id() */
68 if (kprobes_built_in() && !user_mode(regs)) {
69 preempt_disable();
70 if (kprobe_running() && kprobe_fault_handler(regs, 14))
71 ret = 1;
72 preempt_enable();
73 }
74
75 return ret;
76}
77
78/*
79 * Prefetch quirks:
80 *
81 * 32-bit mode:
82 *
83 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
84 * Check that here and ignore it.
85 *
86 * 64-bit mode:
87 *
88 * Sometimes the CPU reports invalid exceptions on prefetch.
89 * Check that here and ignore it.
90 *
91 * Opcode checker based on code by Richard Brunner.
92 */
93static inline int
94check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
95 unsigned char opcode, int *prefetch)
96{
97 unsigned char instr_hi = opcode & 0xf0;
98 unsigned char instr_lo = opcode & 0x0f;
99
100 switch (instr_hi) {
101 case 0x20:
102 case 0x30:
103 /*
104 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
105 * In X86_64 long mode, the CPU will signal invalid
106 * opcode if some of these prefixes are present so
107 * X86_64 will never get here anyway
108 */
109 return ((instr_lo & 7) == 0x6);
110#ifdef CONFIG_X86_64
111 case 0x40:
112 /*
113 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
114 * Need to figure out under what instruction mode the
115 * instruction was issued. Could check the LDT for lm,
116 * but for now it's good enough to assume that long
117 * mode only uses well known segments or kernel.
118 */
119 return (!user_mode(regs) || user_64bit_mode(regs));
120#endif
121 case 0x60:
122 /* 0x64 thru 0x67 are valid prefixes in all modes. */
123 return (instr_lo & 0xC) == 0x4;
124 case 0xF0:
125 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
126 return !instr_lo || (instr_lo>>1) == 1;
127 case 0x00:
128 /* Prefetch instruction is 0x0F0D or 0x0F18 */
129 if (probe_kernel_address(instr, opcode))
130 return 0;
131
132 *prefetch = (instr_lo == 0xF) &&
133 (opcode == 0x0D || opcode == 0x18);
134 return 0;
135 default:
136 return 0;
137 }
138}
139
140static int
141is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
142{
143 unsigned char *max_instr;
144 unsigned char *instr;
145 int prefetch = 0;
146
147 /*
148 * If it was a exec (instruction fetch) fault on NX page, then
149 * do not ignore the fault:
150 */
151 if (error_code & PF_INSTR)
152 return 0;
153
154 instr = (void *)convert_ip_to_linear(current, regs);
155 max_instr = instr + 15;
156
157 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
158 return 0;
159
160 while (instr < max_instr) {
161 unsigned char opcode;
162
163 if (probe_kernel_address(instr, opcode))
164 break;
165
166 instr++;
167
168 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
169 break;
170 }
171 return prefetch;
172}
173
174/*
175 * A protection key fault means that the PKRU value did not allow
176 * access to some PTE. Userspace can figure out what PKRU was
177 * from the XSAVE state, and this function fills out a field in
178 * siginfo so userspace can discover which protection key was set
179 * on the PTE.
180 *
181 * If we get here, we know that the hardware signaled a PF_PK
182 * fault and that there was a VMA once we got in the fault
183 * handler. It does *not* guarantee that the VMA we find here
184 * was the one that we faulted on.
185 *
186 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
187 * 2. T1 : set PKRU to deny access to pkey=4, touches page
188 * 3. T1 : faults...
189 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
190 * 5. T1 : enters fault handler, takes mmap_sem, etc...
191 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
192 * faulted on a pte with its pkey=4.
193 */
194static void fill_sig_info_pkey(int si_code, siginfo_t *info,
195 struct vm_area_struct *vma)
196{
197 /* This is effectively an #ifdef */
198 if (!boot_cpu_has(X86_FEATURE_OSPKE))
199 return;
200
201 /* Fault not from Protection Keys: nothing to do */
202 if (si_code != SEGV_PKUERR)
203 return;
204 /*
205 * force_sig_info_fault() is called from a number of
206 * contexts, some of which have a VMA and some of which
207 * do not. The PF_PK handing happens after we have a
208 * valid VMA, so we should never reach this without a
209 * valid VMA.
210 */
211 if (!vma) {
212 WARN_ONCE(1, "PKU fault with no VMA passed in");
213 info->si_pkey = 0;
214 return;
215 }
216 /*
217 * si_pkey should be thought of as a strong hint, but not
218 * absolutely guranteed to be 100% accurate because of
219 * the race explained above.
220 */
221 info->si_pkey = vma_pkey(vma);
222}
223
224static void
225force_sig_info_fault(int si_signo, int si_code, unsigned long address,
226 struct task_struct *tsk, struct vm_area_struct *vma,
227 int fault)
228{
229 unsigned lsb = 0;
230 siginfo_t info;
231
232 info.si_signo = si_signo;
233 info.si_errno = 0;
234 info.si_code = si_code;
235 info.si_addr = (void __user *)address;
236 if (fault & VM_FAULT_HWPOISON_LARGE)
237 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
238 if (fault & VM_FAULT_HWPOISON)
239 lsb = PAGE_SHIFT;
240 info.si_addr_lsb = lsb;
241
242 fill_sig_info_pkey(si_code, &info, vma);
243
244 force_sig_info(si_signo, &info, tsk);
245}
246
247DEFINE_SPINLOCK(pgd_lock);
248LIST_HEAD(pgd_list);
249
250#ifdef CONFIG_X86_32
251static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
252{
253 unsigned index = pgd_index(address);
254 pgd_t *pgd_k;
255 pud_t *pud, *pud_k;
256 pmd_t *pmd, *pmd_k;
257
258 pgd += index;
259 pgd_k = init_mm.pgd + index;
260
261 if (!pgd_present(*pgd_k))
262 return NULL;
263
264 /*
265 * set_pgd(pgd, *pgd_k); here would be useless on PAE
266 * and redundant with the set_pmd() on non-PAE. As would
267 * set_pud.
268 */
269 pud = pud_offset(pgd, address);
270 pud_k = pud_offset(pgd_k, address);
271 if (!pud_present(*pud_k))
272 return NULL;
273
274 pmd = pmd_offset(pud, address);
275 pmd_k = pmd_offset(pud_k, address);
276 if (!pmd_present(*pmd_k))
277 return NULL;
278
279 if (!pmd_present(*pmd))
280 set_pmd(pmd, *pmd_k);
281 else
282 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
283
284 return pmd_k;
285}
286
287void vmalloc_sync_all(void)
288{
289 unsigned long address;
290
291 if (SHARED_KERNEL_PMD)
292 return;
293
294 for (address = VMALLOC_START & PMD_MASK;
295 address >= TASK_SIZE && address < FIXADDR_TOP;
296 address += PMD_SIZE) {
297 struct page *page;
298
299 spin_lock(&pgd_lock);
300 list_for_each_entry(page, &pgd_list, lru) {
301 spinlock_t *pgt_lock;
302 pmd_t *ret;
303
304 /* the pgt_lock only for Xen */
305 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
306
307 spin_lock(pgt_lock);
308 ret = vmalloc_sync_one(page_address(page), address);
309 spin_unlock(pgt_lock);
310
311 if (!ret)
312 break;
313 }
314 spin_unlock(&pgd_lock);
315 }
316}
317
318/*
319 * 32-bit:
320 *
321 * Handle a fault on the vmalloc or module mapping area
322 */
323static noinline int vmalloc_fault(unsigned long address)
324{
325 unsigned long pgd_paddr;
326 pmd_t *pmd_k;
327 pte_t *pte_k;
328
329 /* Make sure we are in vmalloc area: */
330 if (!(address >= VMALLOC_START && address < VMALLOC_END))
331 return -1;
332
333 WARN_ON_ONCE(in_nmi());
334
335 /*
336 * Synchronize this task's top level page-table
337 * with the 'reference' page table.
338 *
339 * Do _not_ use "current" here. We might be inside
340 * an interrupt in the middle of a task switch..
341 */
342 pgd_paddr = read_cr3();
343 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
344 if (!pmd_k)
345 return -1;
346
347 if (pmd_huge(*pmd_k))
348 return 0;
349
350 pte_k = pte_offset_kernel(pmd_k, address);
351 if (!pte_present(*pte_k))
352 return -1;
353
354 return 0;
355}
356NOKPROBE_SYMBOL(vmalloc_fault);
357
358/*
359 * Did it hit the DOS screen memory VA from vm86 mode?
360 */
361static inline void
362check_v8086_mode(struct pt_regs *regs, unsigned long address,
363 struct task_struct *tsk)
364{
365#ifdef CONFIG_VM86
366 unsigned long bit;
367
368 if (!v8086_mode(regs) || !tsk->thread.vm86)
369 return;
370
371 bit = (address - 0xA0000) >> PAGE_SHIFT;
372 if (bit < 32)
373 tsk->thread.vm86->screen_bitmap |= 1 << bit;
374#endif
375}
376
377static bool low_pfn(unsigned long pfn)
378{
379 return pfn < max_low_pfn;
380}
381
382static void dump_pagetable(unsigned long address)
383{
384 pgd_t *base = __va(read_cr3());
385 pgd_t *pgd = &base[pgd_index(address)];
386 pmd_t *pmd;
387 pte_t *pte;
388
389#ifdef CONFIG_X86_PAE
390 printk("*pdpt = %016Lx ", pgd_val(*pgd));
391 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
392 goto out;
393#endif
394 pmd = pmd_offset(pud_offset(pgd, address), address);
395 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
396
397 /*
398 * We must not directly access the pte in the highpte
399 * case if the page table is located in highmem.
400 * And let's rather not kmap-atomic the pte, just in case
401 * it's allocated already:
402 */
403 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
404 goto out;
405
406 pte = pte_offset_kernel(pmd, address);
407 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
408out:
409 printk("\n");
410}
411
412#else /* CONFIG_X86_64: */
413
414void vmalloc_sync_all(void)
415{
416 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0);
417}
418
419/*
420 * 64-bit:
421 *
422 * Handle a fault on the vmalloc area
423 */
424static noinline int vmalloc_fault(unsigned long address)
425{
426 pgd_t *pgd, *pgd_ref;
427 pud_t *pud, *pud_ref;
428 pmd_t *pmd, *pmd_ref;
429 pte_t *pte, *pte_ref;
430
431 /* Make sure we are in vmalloc area: */
432 if (!(address >= VMALLOC_START && address < VMALLOC_END))
433 return -1;
434
435 WARN_ON_ONCE(in_nmi());
436
437 /*
438 * Copy kernel mappings over when needed. This can also
439 * happen within a race in page table update. In the later
440 * case just flush:
441 */
442 pgd = pgd_offset(current->active_mm, address);
443 pgd_ref = pgd_offset_k(address);
444 if (pgd_none(*pgd_ref))
445 return -1;
446
447 if (pgd_none(*pgd)) {
448 set_pgd(pgd, *pgd_ref);
449 arch_flush_lazy_mmu_mode();
450 } else {
451 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
452 }
453
454 /*
455 * Below here mismatches are bugs because these lower tables
456 * are shared:
457 */
458
459 pud = pud_offset(pgd, address);
460 pud_ref = pud_offset(pgd_ref, address);
461 if (pud_none(*pud_ref))
462 return -1;
463
464 if (pud_none(*pud) || pud_pfn(*pud) != pud_pfn(*pud_ref))
465 BUG();
466
467 if (pud_huge(*pud))
468 return 0;
469
470 pmd = pmd_offset(pud, address);
471 pmd_ref = pmd_offset(pud_ref, address);
472 if (pmd_none(*pmd_ref))
473 return -1;
474
475 if (pmd_none(*pmd) || pmd_pfn(*pmd) != pmd_pfn(*pmd_ref))
476 BUG();
477
478 if (pmd_huge(*pmd))
479 return 0;
480
481 pte_ref = pte_offset_kernel(pmd_ref, address);
482 if (!pte_present(*pte_ref))
483 return -1;
484
485 pte = pte_offset_kernel(pmd, address);
486
487 /*
488 * Don't use pte_page here, because the mappings can point
489 * outside mem_map, and the NUMA hash lookup cannot handle
490 * that:
491 */
492 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
493 BUG();
494
495 return 0;
496}
497NOKPROBE_SYMBOL(vmalloc_fault);
498
499#ifdef CONFIG_CPU_SUP_AMD
500static const char errata93_warning[] =
501KERN_ERR
502"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
503"******* Working around it, but it may cause SEGVs or burn power.\n"
504"******* Please consider a BIOS update.\n"
505"******* Disabling USB legacy in the BIOS may also help.\n";
506#endif
507
508/*
509 * No vm86 mode in 64-bit mode:
510 */
511static inline void
512check_v8086_mode(struct pt_regs *regs, unsigned long address,
513 struct task_struct *tsk)
514{
515}
516
517static int bad_address(void *p)
518{
519 unsigned long dummy;
520
521 return probe_kernel_address((unsigned long *)p, dummy);
522}
523
524static void dump_pagetable(unsigned long address)
525{
526 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
527 pgd_t *pgd = base + pgd_index(address);
528 pud_t *pud;
529 pmd_t *pmd;
530 pte_t *pte;
531
532 if (bad_address(pgd))
533 goto bad;
534
535 printk("PGD %lx ", pgd_val(*pgd));
536
537 if (!pgd_present(*pgd))
538 goto out;
539
540 pud = pud_offset(pgd, address);
541 if (bad_address(pud))
542 goto bad;
543
544 printk("PUD %lx ", pud_val(*pud));
545 if (!pud_present(*pud) || pud_large(*pud))
546 goto out;
547
548 pmd = pmd_offset(pud, address);
549 if (bad_address(pmd))
550 goto bad;
551
552 printk("PMD %lx ", pmd_val(*pmd));
553 if (!pmd_present(*pmd) || pmd_large(*pmd))
554 goto out;
555
556 pte = pte_offset_kernel(pmd, address);
557 if (bad_address(pte))
558 goto bad;
559
560 printk("PTE %lx", pte_val(*pte));
561out:
562 printk("\n");
563 return;
564bad:
565 printk("BAD\n");
566}
567
568#endif /* CONFIG_X86_64 */
569
570/*
571 * Workaround for K8 erratum #93 & buggy BIOS.
572 *
573 * BIOS SMM functions are required to use a specific workaround
574 * to avoid corruption of the 64bit RIP register on C stepping K8.
575 *
576 * A lot of BIOS that didn't get tested properly miss this.
577 *
578 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
579 * Try to work around it here.
580 *
581 * Note we only handle faults in kernel here.
582 * Does nothing on 32-bit.
583 */
584static int is_errata93(struct pt_regs *regs, unsigned long address)
585{
586#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
587 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
588 || boot_cpu_data.x86 != 0xf)
589 return 0;
590
591 if (address != regs->ip)
592 return 0;
593
594 if ((address >> 32) != 0)
595 return 0;
596
597 address |= 0xffffffffUL << 32;
598 if ((address >= (u64)_stext && address <= (u64)_etext) ||
599 (address >= MODULES_VADDR && address <= MODULES_END)) {
600 printk_once(errata93_warning);
601 regs->ip = address;
602 return 1;
603 }
604#endif
605 return 0;
606}
607
608/*
609 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
610 * to illegal addresses >4GB.
611 *
612 * We catch this in the page fault handler because these addresses
613 * are not reachable. Just detect this case and return. Any code
614 * segment in LDT is compatibility mode.
615 */
616static int is_errata100(struct pt_regs *regs, unsigned long address)
617{
618#ifdef CONFIG_X86_64
619 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
620 return 1;
621#endif
622 return 0;
623}
624
625static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
626{
627#ifdef CONFIG_X86_F00F_BUG
628 unsigned long nr;
629
630 /*
631 * Pentium F0 0F C7 C8 bug workaround:
632 */
633 if (boot_cpu_has_bug(X86_BUG_F00F)) {
634 nr = (address - idt_descr.address) >> 3;
635
636 if (nr == 6) {
637 do_invalid_op(regs, 0);
638 return 1;
639 }
640 }
641#endif
642 return 0;
643}
644
645static const char nx_warning[] = KERN_CRIT
646"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
647static const char smep_warning[] = KERN_CRIT
648"unable to execute userspace code (SMEP?) (uid: %d)\n";
649
650static void
651show_fault_oops(struct pt_regs *regs, unsigned long error_code,
652 unsigned long address)
653{
654 if (!oops_may_print())
655 return;
656
657 if (error_code & PF_INSTR) {
658 unsigned int level;
659 pgd_t *pgd;
660 pte_t *pte;
661
662 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
663 pgd += pgd_index(address);
664
665 pte = lookup_address_in_pgd(pgd, address, &level);
666
667 if (pte && pte_present(*pte) && !pte_exec(*pte))
668 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
669 if (pte && pte_present(*pte) && pte_exec(*pte) &&
670 (pgd_flags(*pgd) & _PAGE_USER) &&
671 (__read_cr4() & X86_CR4_SMEP))
672 printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
673 }
674
675 printk(KERN_ALERT "BUG: unable to handle kernel ");
676 if (address < PAGE_SIZE)
677 printk(KERN_CONT "NULL pointer dereference");
678 else
679 printk(KERN_CONT "paging request");
680
681 printk(KERN_CONT " at %p\n", (void *) address);
682 printk(KERN_ALERT "IP:");
683 printk_address(regs->ip);
684
685 dump_pagetable(address);
686}
687
688static noinline void
689pgtable_bad(struct pt_regs *regs, unsigned long error_code,
690 unsigned long address)
691{
692 struct task_struct *tsk;
693 unsigned long flags;
694 int sig;
695
696 flags = oops_begin();
697 tsk = current;
698 sig = SIGKILL;
699
700 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
701 tsk->comm, address);
702 dump_pagetable(address);
703
704 tsk->thread.cr2 = address;
705 tsk->thread.trap_nr = X86_TRAP_PF;
706 tsk->thread.error_code = error_code;
707
708 if (__die("Bad pagetable", regs, error_code))
709 sig = 0;
710
711 oops_end(flags, regs, sig);
712}
713
714static noinline void
715no_context(struct pt_regs *regs, unsigned long error_code,
716 unsigned long address, int signal, int si_code)
717{
718 struct task_struct *tsk = current;
719 unsigned long flags;
720 int sig;
721 /* No context means no VMA to pass down */
722 struct vm_area_struct *vma = NULL;
723
724 /* Are we prepared to handle this kernel fault? */
725 if (fixup_exception(regs, X86_TRAP_PF)) {
726 /*
727 * Any interrupt that takes a fault gets the fixup. This makes
728 * the below recursive fault logic only apply to a faults from
729 * task context.
730 */
731 if (in_interrupt())
732 return;
733
734 /*
735 * Per the above we're !in_interrupt(), aka. task context.
736 *
737 * In this case we need to make sure we're not recursively
738 * faulting through the emulate_vsyscall() logic.
739 */
740 if (current_thread_info()->sig_on_uaccess_error && signal) {
741 tsk->thread.trap_nr = X86_TRAP_PF;
742 tsk->thread.error_code = error_code | PF_USER;
743 tsk->thread.cr2 = address;
744
745 /* XXX: hwpoison faults will set the wrong code. */
746 force_sig_info_fault(signal, si_code, address,
747 tsk, vma, 0);
748 }
749
750 /*
751 * Barring that, we can do the fixup and be happy.
752 */
753 return;
754 }
755
756 /*
757 * 32-bit:
758 *
759 * Valid to do another page fault here, because if this fault
760 * had been triggered by is_prefetch fixup_exception would have
761 * handled it.
762 *
763 * 64-bit:
764 *
765 * Hall of shame of CPU/BIOS bugs.
766 */
767 if (is_prefetch(regs, error_code, address))
768 return;
769
770 if (is_errata93(regs, address))
771 return;
772
773 /*
774 * Oops. The kernel tried to access some bad page. We'll have to
775 * terminate things with extreme prejudice:
776 */
777 flags = oops_begin();
778
779 show_fault_oops(regs, error_code, address);
780
781 if (task_stack_end_corrupted(tsk))
782 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
783
784 tsk->thread.cr2 = address;
785 tsk->thread.trap_nr = X86_TRAP_PF;
786 tsk->thread.error_code = error_code;
787
788 sig = SIGKILL;
789 if (__die("Oops", regs, error_code))
790 sig = 0;
791
792 /* Executive summary in case the body of the oops scrolled away */
793 printk(KERN_DEFAULT "CR2: %016lx\n", address);
794
795 oops_end(flags, regs, sig);
796}
797
798/*
799 * Print out info about fatal segfaults, if the show_unhandled_signals
800 * sysctl is set:
801 */
802static inline void
803show_signal_msg(struct pt_regs *regs, unsigned long error_code,
804 unsigned long address, struct task_struct *tsk)
805{
806 if (!unhandled_signal(tsk, SIGSEGV))
807 return;
808
809 if (!printk_ratelimit())
810 return;
811
812 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
813 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
814 tsk->comm, task_pid_nr(tsk), address,
815 (void *)regs->ip, (void *)regs->sp, error_code);
816
817 print_vma_addr(KERN_CONT " in ", regs->ip);
818
819 printk(KERN_CONT "\n");
820}
821
822static void
823__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
824 unsigned long address, struct vm_area_struct *vma,
825 int si_code)
826{
827 struct task_struct *tsk = current;
828
829 /* User mode accesses just cause a SIGSEGV */
830 if (error_code & PF_USER) {
831 /*
832 * It's possible to have interrupts off here:
833 */
834 local_irq_enable();
835
836 /*
837 * Valid to do another page fault here because this one came
838 * from user space:
839 */
840 if (is_prefetch(regs, error_code, address))
841 return;
842
843 if (is_errata100(regs, address))
844 return;
845
846#ifdef CONFIG_X86_64
847 /*
848 * Instruction fetch faults in the vsyscall page might need
849 * emulation.
850 */
851 if (unlikely((error_code & PF_INSTR) &&
852 ((address & ~0xfff) == VSYSCALL_ADDR))) {
853 if (emulate_vsyscall(regs, address))
854 return;
855 }
856#endif
857 /* Kernel addresses are always protection faults: */
858 if (address >= TASK_SIZE)
859 error_code |= PF_PROT;
860
861 if (likely(show_unhandled_signals))
862 show_signal_msg(regs, error_code, address, tsk);
863
864 tsk->thread.cr2 = address;
865 tsk->thread.error_code = error_code;
866 tsk->thread.trap_nr = X86_TRAP_PF;
867
868 force_sig_info_fault(SIGSEGV, si_code, address, tsk, vma, 0);
869
870 return;
871 }
872
873 if (is_f00f_bug(regs, address))
874 return;
875
876 no_context(regs, error_code, address, SIGSEGV, si_code);
877}
878
879static noinline void
880bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
881 unsigned long address, struct vm_area_struct *vma)
882{
883 __bad_area_nosemaphore(regs, error_code, address, vma, SEGV_MAPERR);
884}
885
886static void
887__bad_area(struct pt_regs *regs, unsigned long error_code,
888 unsigned long address, struct vm_area_struct *vma, int si_code)
889{
890 struct mm_struct *mm = current->mm;
891
892 /*
893 * Something tried to access memory that isn't in our memory map..
894 * Fix it, but check if it's kernel or user first..
895 */
896 up_read(&mm->mmap_sem);
897
898 __bad_area_nosemaphore(regs, error_code, address, vma, si_code);
899}
900
901static noinline void
902bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
903{
904 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
905}
906
907static inline bool bad_area_access_from_pkeys(unsigned long error_code,
908 struct vm_area_struct *vma)
909{
910 /* This code is always called on the current mm */
911 bool foreign = false;
912
913 if (!boot_cpu_has(X86_FEATURE_OSPKE))
914 return false;
915 if (error_code & PF_PK)
916 return true;
917 /* this checks permission keys on the VMA: */
918 if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
919 (error_code & PF_INSTR), foreign))
920 return true;
921 return false;
922}
923
924static noinline void
925bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
926 unsigned long address, struct vm_area_struct *vma)
927{
928 /*
929 * This OSPKE check is not strictly necessary at runtime.
930 * But, doing it this way allows compiler optimizations
931 * if pkeys are compiled out.
932 */
933 if (bad_area_access_from_pkeys(error_code, vma))
934 __bad_area(regs, error_code, address, vma, SEGV_PKUERR);
935 else
936 __bad_area(regs, error_code, address, vma, SEGV_ACCERR);
937}
938
939static void
940do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
941 struct vm_area_struct *vma, unsigned int fault)
942{
943 struct task_struct *tsk = current;
944 int code = BUS_ADRERR;
945
946 /* Kernel mode? Handle exceptions or die: */
947 if (!(error_code & PF_USER)) {
948 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
949 return;
950 }
951
952 /* User-space => ok to do another page fault: */
953 if (is_prefetch(regs, error_code, address))
954 return;
955
956 tsk->thread.cr2 = address;
957 tsk->thread.error_code = error_code;
958 tsk->thread.trap_nr = X86_TRAP_PF;
959
960#ifdef CONFIG_MEMORY_FAILURE
961 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
962 printk(KERN_ERR
963 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
964 tsk->comm, tsk->pid, address);
965 code = BUS_MCEERR_AR;
966 }
967#endif
968 force_sig_info_fault(SIGBUS, code, address, tsk, vma, fault);
969}
970
971static noinline void
972mm_fault_error(struct pt_regs *regs, unsigned long error_code,
973 unsigned long address, struct vm_area_struct *vma,
974 unsigned int fault)
975{
976 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
977 no_context(regs, error_code, address, 0, 0);
978 return;
979 }
980
981 if (fault & VM_FAULT_OOM) {
982 /* Kernel mode? Handle exceptions or die: */
983 if (!(error_code & PF_USER)) {
984 no_context(regs, error_code, address,
985 SIGSEGV, SEGV_MAPERR);
986 return;
987 }
988
989 /*
990 * We ran out of memory, call the OOM killer, and return the
991 * userspace (which will retry the fault, or kill us if we got
992 * oom-killed):
993 */
994 pagefault_out_of_memory();
995 } else {
996 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
997 VM_FAULT_HWPOISON_LARGE))
998 do_sigbus(regs, error_code, address, vma, fault);
999 else if (fault & VM_FAULT_SIGSEGV)
1000 bad_area_nosemaphore(regs, error_code, address, vma);
1001 else
1002 BUG();
1003 }
1004}
1005
1006static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1007{
1008 if ((error_code & PF_WRITE) && !pte_write(*pte))
1009 return 0;
1010
1011 if ((error_code & PF_INSTR) && !pte_exec(*pte))
1012 return 0;
1013 /*
1014 * Note: We do not do lazy flushing on protection key
1015 * changes, so no spurious fault will ever set PF_PK.
1016 */
1017 if ((error_code & PF_PK))
1018 return 1;
1019
1020 return 1;
1021}
1022
1023/*
1024 * Handle a spurious fault caused by a stale TLB entry.
1025 *
1026 * This allows us to lazily refresh the TLB when increasing the
1027 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1028 * eagerly is very expensive since that implies doing a full
1029 * cross-processor TLB flush, even if no stale TLB entries exist
1030 * on other processors.
1031 *
1032 * Spurious faults may only occur if the TLB contains an entry with
1033 * fewer permission than the page table entry. Non-present (P = 0)
1034 * and reserved bit (R = 1) faults are never spurious.
1035 *
1036 * There are no security implications to leaving a stale TLB when
1037 * increasing the permissions on a page.
1038 *
1039 * Returns non-zero if a spurious fault was handled, zero otherwise.
1040 *
1041 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1042 * (Optional Invalidation).
1043 */
1044static noinline int
1045spurious_fault(unsigned long error_code, unsigned long address)
1046{
1047 pgd_t *pgd;
1048 pud_t *pud;
1049 pmd_t *pmd;
1050 pte_t *pte;
1051 int ret;
1052
1053 /*
1054 * Only writes to RO or instruction fetches from NX may cause
1055 * spurious faults.
1056 *
1057 * These could be from user or supervisor accesses but the TLB
1058 * is only lazily flushed after a kernel mapping protection
1059 * change, so user accesses are not expected to cause spurious
1060 * faults.
1061 */
1062 if (error_code != (PF_WRITE | PF_PROT)
1063 && error_code != (PF_INSTR | PF_PROT))
1064 return 0;
1065
1066 pgd = init_mm.pgd + pgd_index(address);
1067 if (!pgd_present(*pgd))
1068 return 0;
1069
1070 pud = pud_offset(pgd, address);
1071 if (!pud_present(*pud))
1072 return 0;
1073
1074 if (pud_large(*pud))
1075 return spurious_fault_check(error_code, (pte_t *) pud);
1076
1077 pmd = pmd_offset(pud, address);
1078 if (!pmd_present(*pmd))
1079 return 0;
1080
1081 if (pmd_large(*pmd))
1082 return spurious_fault_check(error_code, (pte_t *) pmd);
1083
1084 pte = pte_offset_kernel(pmd, address);
1085 if (!pte_present(*pte))
1086 return 0;
1087
1088 ret = spurious_fault_check(error_code, pte);
1089 if (!ret)
1090 return 0;
1091
1092 /*
1093 * Make sure we have permissions in PMD.
1094 * If not, then there's a bug in the page tables:
1095 */
1096 ret = spurious_fault_check(error_code, (pte_t *) pmd);
1097 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1098
1099 return ret;
1100}
1101NOKPROBE_SYMBOL(spurious_fault);
1102
1103int show_unhandled_signals = 1;
1104
1105static inline int
1106access_error(unsigned long error_code, struct vm_area_struct *vma)
1107{
1108 /* This is only called for the current mm, so: */
1109 bool foreign = false;
1110 /*
1111 * Make sure to check the VMA so that we do not perform
1112 * faults just to hit a PF_PK as soon as we fill in a
1113 * page.
1114 */
1115 if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
1116 (error_code & PF_INSTR), foreign))
1117 return 1;
1118
1119 if (error_code & PF_WRITE) {
1120 /* write, present and write, not present: */
1121 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1122 return 1;
1123 return 0;
1124 }
1125
1126 /* read, present: */
1127 if (unlikely(error_code & PF_PROT))
1128 return 1;
1129
1130 /* read, not present: */
1131 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1132 return 1;
1133
1134 return 0;
1135}
1136
1137static int fault_in_kernel_space(unsigned long address)
1138{
1139 return address >= TASK_SIZE_MAX;
1140}
1141
1142static inline bool smap_violation(int error_code, struct pt_regs *regs)
1143{
1144 if (!IS_ENABLED(CONFIG_X86_SMAP))
1145 return false;
1146
1147 if (!static_cpu_has(X86_FEATURE_SMAP))
1148 return false;
1149
1150 if (error_code & PF_USER)
1151 return false;
1152
1153 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1154 return false;
1155
1156 return true;
1157}
1158
1159/*
1160 * This routine handles page faults. It determines the address,
1161 * and the problem, and then passes it off to one of the appropriate
1162 * routines.
1163 *
1164 * This function must have noinline because both callers
1165 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1166 * guarantees there's a function trace entry.
1167 */
1168static noinline void
1169__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1170 unsigned long address)
1171{
1172 struct vm_area_struct *vma;
1173 struct task_struct *tsk;
1174 struct mm_struct *mm;
1175 int fault, major = 0;
1176 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1177
1178 tsk = current;
1179 mm = tsk->mm;
1180
1181 /*
1182 * Detect and handle instructions that would cause a page fault for
1183 * both a tracked kernel page and a userspace page.
1184 */
1185 if (kmemcheck_active(regs))
1186 kmemcheck_hide(regs);
1187 prefetchw(&mm->mmap_sem);
1188
1189 if (unlikely(kmmio_fault(regs, address)))
1190 return;
1191
1192 /*
1193 * We fault-in kernel-space virtual memory on-demand. The
1194 * 'reference' page table is init_mm.pgd.
1195 *
1196 * NOTE! We MUST NOT take any locks for this case. We may
1197 * be in an interrupt or a critical region, and should
1198 * only copy the information from the master page table,
1199 * nothing more.
1200 *
1201 * This verifies that the fault happens in kernel space
1202 * (error_code & 4) == 0, and that the fault was not a
1203 * protection error (error_code & 9) == 0.
1204 */
1205 if (unlikely(fault_in_kernel_space(address))) {
1206 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1207 if (vmalloc_fault(address) >= 0)
1208 return;
1209
1210 if (kmemcheck_fault(regs, address, error_code))
1211 return;
1212 }
1213
1214 /* Can handle a stale RO->RW TLB: */
1215 if (spurious_fault(error_code, address))
1216 return;
1217
1218 /* kprobes don't want to hook the spurious faults: */
1219 if (kprobes_fault(regs))
1220 return;
1221 /*
1222 * Don't take the mm semaphore here. If we fixup a prefetch
1223 * fault we could otherwise deadlock:
1224 */
1225 bad_area_nosemaphore(regs, error_code, address, NULL);
1226
1227 return;
1228 }
1229
1230 /* kprobes don't want to hook the spurious faults: */
1231 if (unlikely(kprobes_fault(regs)))
1232 return;
1233
1234 if (unlikely(error_code & PF_RSVD))
1235 pgtable_bad(regs, error_code, address);
1236
1237 if (unlikely(smap_violation(error_code, regs))) {
1238 bad_area_nosemaphore(regs, error_code, address, NULL);
1239 return;
1240 }
1241
1242 /*
1243 * If we're in an interrupt, have no user context or are running
1244 * in a region with pagefaults disabled then we must not take the fault
1245 */
1246 if (unlikely(faulthandler_disabled() || !mm)) {
1247 bad_area_nosemaphore(regs, error_code, address, NULL);
1248 return;
1249 }
1250
1251 /*
1252 * It's safe to allow irq's after cr2 has been saved and the
1253 * vmalloc fault has been handled.
1254 *
1255 * User-mode registers count as a user access even for any
1256 * potential system fault or CPU buglet:
1257 */
1258 if (user_mode(regs)) {
1259 local_irq_enable();
1260 error_code |= PF_USER;
1261 flags |= FAULT_FLAG_USER;
1262 } else {
1263 if (regs->flags & X86_EFLAGS_IF)
1264 local_irq_enable();
1265 }
1266
1267 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1268
1269 if (error_code & PF_WRITE)
1270 flags |= FAULT_FLAG_WRITE;
1271 if (error_code & PF_INSTR)
1272 flags |= FAULT_FLAG_INSTRUCTION;
1273
1274 /*
1275 * When running in the kernel we expect faults to occur only to
1276 * addresses in user space. All other faults represent errors in
1277 * the kernel and should generate an OOPS. Unfortunately, in the
1278 * case of an erroneous fault occurring in a code path which already
1279 * holds mmap_sem we will deadlock attempting to validate the fault
1280 * against the address space. Luckily the kernel only validly
1281 * references user space from well defined areas of code, which are
1282 * listed in the exceptions table.
1283 *
1284 * As the vast majority of faults will be valid we will only perform
1285 * the source reference check when there is a possibility of a
1286 * deadlock. Attempt to lock the address space, if we cannot we then
1287 * validate the source. If this is invalid we can skip the address
1288 * space check, thus avoiding the deadlock:
1289 */
1290 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1291 if ((error_code & PF_USER) == 0 &&
1292 !search_exception_tables(regs->ip)) {
1293 bad_area_nosemaphore(regs, error_code, address, NULL);
1294 return;
1295 }
1296retry:
1297 down_read(&mm->mmap_sem);
1298 } else {
1299 /*
1300 * The above down_read_trylock() might have succeeded in
1301 * which case we'll have missed the might_sleep() from
1302 * down_read():
1303 */
1304 might_sleep();
1305 }
1306
1307 vma = find_vma(mm, address);
1308 if (unlikely(!vma)) {
1309 bad_area(regs, error_code, address);
1310 return;
1311 }
1312 if (likely(vma->vm_start <= address))
1313 goto good_area;
1314 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1315 bad_area(regs, error_code, address);
1316 return;
1317 }
1318 if (error_code & PF_USER) {
1319 /*
1320 * Accessing the stack below %sp is always a bug.
1321 * The large cushion allows instructions like enter
1322 * and pusha to work. ("enter $65535, $31" pushes
1323 * 32 pointers and then decrements %sp by 65535.)
1324 */
1325 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1326 bad_area(regs, error_code, address);
1327 return;
1328 }
1329 }
1330 if (unlikely(expand_stack(vma, address))) {
1331 bad_area(regs, error_code, address);
1332 return;
1333 }
1334
1335 /*
1336 * Ok, we have a good vm_area for this memory access, so
1337 * we can handle it..
1338 */
1339good_area:
1340 if (unlikely(access_error(error_code, vma))) {
1341 bad_area_access_error(regs, error_code, address, vma);
1342 return;
1343 }
1344
1345 /*
1346 * If for any reason at all we couldn't handle the fault,
1347 * make sure we exit gracefully rather than endlessly redo
1348 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1349 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1350 */
1351 fault = handle_mm_fault(mm, vma, address, flags);
1352 major |= fault & VM_FAULT_MAJOR;
1353
1354 /*
1355 * If we need to retry the mmap_sem has already been released,
1356 * and if there is a fatal signal pending there is no guarantee
1357 * that we made any progress. Handle this case first.
1358 */
1359 if (unlikely(fault & VM_FAULT_RETRY)) {
1360 /* Retry at most once */
1361 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1362 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1363 flags |= FAULT_FLAG_TRIED;
1364 if (!fatal_signal_pending(tsk))
1365 goto retry;
1366 }
1367
1368 /* User mode? Just return to handle the fatal exception */
1369 if (flags & FAULT_FLAG_USER)
1370 return;
1371
1372 /* Not returning to user mode? Handle exceptions or die: */
1373 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1374 return;
1375 }
1376
1377 up_read(&mm->mmap_sem);
1378 if (unlikely(fault & VM_FAULT_ERROR)) {
1379 mm_fault_error(regs, error_code, address, vma, fault);
1380 return;
1381 }
1382
1383 /*
1384 * Major/minor page fault accounting. If any of the events
1385 * returned VM_FAULT_MAJOR, we account it as a major fault.
1386 */
1387 if (major) {
1388 tsk->maj_flt++;
1389 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1390 } else {
1391 tsk->min_flt++;
1392 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1393 }
1394
1395 check_v8086_mode(regs, address, tsk);
1396}
1397NOKPROBE_SYMBOL(__do_page_fault);
1398
1399dotraplinkage void notrace
1400do_page_fault(struct pt_regs *regs, unsigned long error_code)
1401{
1402 unsigned long address = read_cr2(); /* Get the faulting address */
1403 enum ctx_state prev_state;
1404
1405 /*
1406 * We must have this function tagged with __kprobes, notrace and call
1407 * read_cr2() before calling anything else. To avoid calling any kind
1408 * of tracing machinery before we've observed the CR2 value.
1409 *
1410 * exception_{enter,exit}() contain all sorts of tracepoints.
1411 */
1412
1413 prev_state = exception_enter();
1414 __do_page_fault(regs, error_code, address);
1415 exception_exit(prev_state);
1416}
1417NOKPROBE_SYMBOL(do_page_fault);
1418
1419#ifdef CONFIG_TRACING
1420static nokprobe_inline void
1421trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1422 unsigned long error_code)
1423{
1424 if (user_mode(regs))
1425 trace_page_fault_user(address, regs, error_code);
1426 else
1427 trace_page_fault_kernel(address, regs, error_code);
1428}
1429
1430dotraplinkage void notrace
1431trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1432{
1433 /*
1434 * The exception_enter and tracepoint processing could
1435 * trigger another page faults (user space callchain
1436 * reading) and destroy the original cr2 value, so read
1437 * the faulting address now.
1438 */
1439 unsigned long address = read_cr2();
1440 enum ctx_state prev_state;
1441
1442 prev_state = exception_enter();
1443 trace_page_fault_entries(address, regs, error_code);
1444 __do_page_fault(regs, error_code, address);
1445 exception_exit(prev_state);
1446}
1447NOKPROBE_SYMBOL(trace_do_page_fault);
1448#endif /* CONFIG_TRACING */