1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Copyright (C) 1995 Linus Torvalds |
4 | * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs. |
5 | * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar |
6 | */ |
7 | #include <linux/sched.h> /* test_thread_flag(), ... */ |
8 | #include <linux/sched/task_stack.h> /* task_stack_*(), ... */ |
9 | #include <linux/kdebug.h> /* oops_begin/end, ... */ |
10 | #include <linux/extable.h> /* search_exception_tables */ |
11 | #include <linux/memblock.h> /* max_low_pfn */ |
12 | #include <linux/kfence.h> /* kfence_handle_page_fault */ |
13 | #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */ |
14 | #include <linux/mmiotrace.h> /* kmmio_handler, ... */ |
15 | #include <linux/perf_event.h> /* perf_sw_event */ |
16 | #include <linux/hugetlb.h> /* hstate_index_to_shift */ |
17 | #include <linux/prefetch.h> /* prefetchw */ |
18 | #include <linux/context_tracking.h> /* exception_enter(), ... */ |
19 | #include <linux/uaccess.h> /* faulthandler_disabled() */ |
20 | #include <linux/efi.h> /* efi_crash_gracefully_on_page_fault()*/ |
21 | #include <linux/mm_types.h> |
22 | #include <linux/mm.h> /* find_and_lock_vma() */ |
23 | |
24 | #include <asm/cpufeature.h> /* boot_cpu_has, ... */ |
25 | #include <asm/traps.h> /* dotraplinkage, ... */ |
26 | #include <asm/fixmap.h> /* VSYSCALL_ADDR */ |
27 | #include <asm/vsyscall.h> /* emulate_vsyscall */ |
28 | #include <asm/vm86.h> /* struct vm86 */ |
29 | #include <asm/mmu_context.h> /* vma_pkey() */ |
30 | #include <asm/efi.h> /* efi_crash_gracefully_on_page_fault()*/ |
31 | #include <asm/desc.h> /* store_idt(), ... */ |
32 | #include <asm/cpu_entry_area.h> /* exception stack */ |
33 | #include <asm/pgtable_areas.h> /* VMALLOC_START, ... */ |
34 | #include <asm/kvm_para.h> /* kvm_handle_async_pf */ |
35 | #include <asm/vdso.h> /* fixup_vdso_exception() */ |
36 | #include <asm/irq_stack.h> |
37 | |
38 | #define CREATE_TRACE_POINTS |
39 | #include <asm/trace/exceptions.h> |
40 | |
41 | /* |
42 | * Returns 0 if mmiotrace is disabled, or if the fault is not |
43 | * handled by mmiotrace: |
44 | */ |
45 | static nokprobe_inline int |
46 | kmmio_fault(struct pt_regs *regs, unsigned long addr) |
47 | { |
48 | if (unlikely(is_kmmio_active())) |
49 | if (kmmio_handler(regs, addr) == 1) |
50 | return -1; |
51 | return 0; |
52 | } |
53 | |
54 | /* |
55 | * Prefetch quirks: |
56 | * |
57 | * 32-bit mode: |
58 | * |
59 | * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch. |
60 | * Check that here and ignore it. This is AMD erratum #91. |
61 | * |
62 | * 64-bit mode: |
63 | * |
64 | * Sometimes the CPU reports invalid exceptions on prefetch. |
65 | * Check that here and ignore it. |
66 | * |
67 | * Opcode checker based on code by Richard Brunner. |
68 | */ |
69 | static inline int |
70 | check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr, |
71 | unsigned char opcode, int *prefetch) |
72 | { |
73 | unsigned char instr_hi = opcode & 0xf0; |
74 | unsigned char instr_lo = opcode & 0x0f; |
75 | |
76 | switch (instr_hi) { |
77 | case 0x20: |
78 | case 0x30: |
79 | /* |
80 | * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. |
81 | * In X86_64 long mode, the CPU will signal invalid |
82 | * opcode if some of these prefixes are present so |
83 | * X86_64 will never get here anyway |
84 | */ |
85 | return ((instr_lo & 7) == 0x6); |
86 | #ifdef CONFIG_X86_64 |
87 | case 0x40: |
88 | /* |
89 | * In 64-bit mode 0x40..0x4F are valid REX prefixes |
90 | */ |
91 | return (!user_mode(regs) || user_64bit_mode(regs)); |
92 | #endif |
93 | case 0x60: |
94 | /* 0x64 thru 0x67 are valid prefixes in all modes. */ |
95 | return (instr_lo & 0xC) == 0x4; |
96 | case 0xF0: |
97 | /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */ |
98 | return !instr_lo || (instr_lo>>1) == 1; |
99 | case 0x00: |
100 | /* Prefetch instruction is 0x0F0D or 0x0F18 */ |
101 | if (get_kernel_nofault(opcode, instr)) |
102 | return 0; |
103 | |
104 | *prefetch = (instr_lo == 0xF) && |
105 | (opcode == 0x0D || opcode == 0x18); |
106 | return 0; |
107 | default: |
108 | return 0; |
109 | } |
110 | } |
111 | |
112 | static bool is_amd_k8_pre_npt(void) |
113 | { |
114 | struct cpuinfo_x86 *c = &boot_cpu_data; |
115 | |
116 | return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) && |
117 | c->x86_vendor == X86_VENDOR_AMD && |
118 | c->x86 == 0xf && c->x86_model < 0x40); |
119 | } |
120 | |
121 | static int |
122 | is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr) |
123 | { |
124 | unsigned char *max_instr; |
125 | unsigned char *instr; |
126 | int prefetch = 0; |
127 | |
128 | /* Erratum #91 affects AMD K8, pre-NPT CPUs */ |
129 | if (!is_amd_k8_pre_npt()) |
130 | return 0; |
131 | |
132 | /* |
133 | * If it was a exec (instruction fetch) fault on NX page, then |
134 | * do not ignore the fault: |
135 | */ |
136 | if (error_code & X86_PF_INSTR) |
137 | return 0; |
138 | |
139 | instr = (void *)convert_ip_to_linear(current, regs); |
140 | max_instr = instr + 15; |
141 | |
142 | /* |
143 | * This code has historically always bailed out if IP points to a |
144 | * not-present page (e.g. due to a race). No one has ever |
145 | * complained about this. |
146 | */ |
147 | pagefault_disable(); |
148 | |
149 | while (instr < max_instr) { |
150 | unsigned char opcode; |
151 | |
152 | if (user_mode(regs)) { |
153 | if (get_user(opcode, (unsigned char __user *) instr)) |
154 | break; |
155 | } else { |
156 | if (get_kernel_nofault(opcode, instr)) |
157 | break; |
158 | } |
159 | |
160 | instr++; |
161 | |
162 | if (!check_prefetch_opcode(regs, instr, opcode, prefetch: &prefetch)) |
163 | break; |
164 | } |
165 | |
166 | pagefault_enable(); |
167 | return prefetch; |
168 | } |
169 | |
170 | DEFINE_SPINLOCK(pgd_lock); |
171 | LIST_HEAD(pgd_list); |
172 | |
173 | #ifdef CONFIG_X86_32 |
174 | static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) |
175 | { |
176 | unsigned index = pgd_index(address); |
177 | pgd_t *pgd_k; |
178 | p4d_t *p4d, *p4d_k; |
179 | pud_t *pud, *pud_k; |
180 | pmd_t *pmd, *pmd_k; |
181 | |
182 | pgd += index; |
183 | pgd_k = init_mm.pgd + index; |
184 | |
185 | if (!pgd_present(*pgd_k)) |
186 | return NULL; |
187 | |
188 | /* |
189 | * set_pgd(pgd, *pgd_k); here would be useless on PAE |
190 | * and redundant with the set_pmd() on non-PAE. As would |
191 | * set_p4d/set_pud. |
192 | */ |
193 | p4d = p4d_offset(pgd, address); |
194 | p4d_k = p4d_offset(pgd_k, address); |
195 | if (!p4d_present(*p4d_k)) |
196 | return NULL; |
197 | |
198 | pud = pud_offset(p4d, address); |
199 | pud_k = pud_offset(p4d_k, address); |
200 | if (!pud_present(*pud_k)) |
201 | return NULL; |
202 | |
203 | pmd = pmd_offset(pud, address); |
204 | pmd_k = pmd_offset(pud_k, address); |
205 | |
206 | if (pmd_present(*pmd) != pmd_present(*pmd_k)) |
207 | set_pmd(pmd, *pmd_k); |
208 | |
209 | if (!pmd_present(*pmd_k)) |
210 | return NULL; |
211 | else |
212 | BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k)); |
213 | |
214 | return pmd_k; |
215 | } |
216 | |
217 | /* |
218 | * Handle a fault on the vmalloc or module mapping area |
219 | * |
220 | * This is needed because there is a race condition between the time |
221 | * when the vmalloc mapping code updates the PMD to the point in time |
222 | * where it synchronizes this update with the other page-tables in the |
223 | * system. |
224 | * |
225 | * In this race window another thread/CPU can map an area on the same |
226 | * PMD, finds it already present and does not synchronize it with the |
227 | * rest of the system yet. As a result v[mz]alloc might return areas |
228 | * which are not mapped in every page-table in the system, causing an |
229 | * unhandled page-fault when they are accessed. |
230 | */ |
231 | static noinline int vmalloc_fault(unsigned long address) |
232 | { |
233 | unsigned long pgd_paddr; |
234 | pmd_t *pmd_k; |
235 | pte_t *pte_k; |
236 | |
237 | /* Make sure we are in vmalloc area: */ |
238 | if (!(address >= VMALLOC_START && address < VMALLOC_END)) |
239 | return -1; |
240 | |
241 | /* |
242 | * Synchronize this task's top level page-table |
243 | * with the 'reference' page table. |
244 | * |
245 | * Do _not_ use "current" here. We might be inside |
246 | * an interrupt in the middle of a task switch.. |
247 | */ |
248 | pgd_paddr = read_cr3_pa(); |
249 | pmd_k = vmalloc_sync_one(__va(pgd_paddr), address); |
250 | if (!pmd_k) |
251 | return -1; |
252 | |
253 | if (pmd_large(*pmd_k)) |
254 | return 0; |
255 | |
256 | pte_k = pte_offset_kernel(pmd_k, address); |
257 | if (!pte_present(*pte_k)) |
258 | return -1; |
259 | |
260 | return 0; |
261 | } |
262 | NOKPROBE_SYMBOL(vmalloc_fault); |
263 | |
264 | void arch_sync_kernel_mappings(unsigned long start, unsigned long end) |
265 | { |
266 | unsigned long addr; |
267 | |
268 | for (addr = start & PMD_MASK; |
269 | addr >= TASK_SIZE_MAX && addr < VMALLOC_END; |
270 | addr += PMD_SIZE) { |
271 | struct page *page; |
272 | |
273 | spin_lock(&pgd_lock); |
274 | list_for_each_entry(page, &pgd_list, lru) { |
275 | spinlock_t *pgt_lock; |
276 | |
277 | /* the pgt_lock only for Xen */ |
278 | pgt_lock = &pgd_page_get_mm(page)->page_table_lock; |
279 | |
280 | spin_lock(pgt_lock); |
281 | vmalloc_sync_one(page_address(page), addr); |
282 | spin_unlock(pgt_lock); |
283 | } |
284 | spin_unlock(&pgd_lock); |
285 | } |
286 | } |
287 | |
288 | static bool low_pfn(unsigned long pfn) |
289 | { |
290 | return pfn < max_low_pfn; |
291 | } |
292 | |
293 | static void dump_pagetable(unsigned long address) |
294 | { |
295 | pgd_t *base = __va(read_cr3_pa()); |
296 | pgd_t *pgd = &base[pgd_index(address)]; |
297 | p4d_t *p4d; |
298 | pud_t *pud; |
299 | pmd_t *pmd; |
300 | pte_t *pte; |
301 | |
302 | #ifdef CONFIG_X86_PAE |
303 | pr_info("*pdpt = %016Lx " , pgd_val(*pgd)); |
304 | if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd)) |
305 | goto out; |
306 | #define pr_pde pr_cont |
307 | #else |
308 | #define pr_pde pr_info |
309 | #endif |
310 | p4d = p4d_offset(pgd, address); |
311 | pud = pud_offset(p4d, address); |
312 | pmd = pmd_offset(pud, address); |
313 | pr_pde("*pde = %0*Lx " , sizeof(*pmd) * 2, (u64)pmd_val(*pmd)); |
314 | #undef pr_pde |
315 | |
316 | /* |
317 | * We must not directly access the pte in the highpte |
318 | * case if the page table is located in highmem. |
319 | * And let's rather not kmap-atomic the pte, just in case |
320 | * it's allocated already: |
321 | */ |
322 | if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd)) |
323 | goto out; |
324 | |
325 | pte = pte_offset_kernel(pmd, address); |
326 | pr_cont("*pte = %0*Lx " , sizeof(*pte) * 2, (u64)pte_val(*pte)); |
327 | out: |
328 | pr_cont("\n" ); |
329 | } |
330 | |
331 | #else /* CONFIG_X86_64: */ |
332 | |
333 | #ifdef CONFIG_CPU_SUP_AMD |
334 | static const char errata93_warning[] = |
335 | KERN_ERR |
336 | "******* Your BIOS seems to not contain a fix for K8 errata #93\n" |
337 | "******* Working around it, but it may cause SEGVs or burn power.\n" |
338 | "******* Please consider a BIOS update.\n" |
339 | "******* Disabling USB legacy in the BIOS may also help.\n" ; |
340 | #endif |
341 | |
342 | static int bad_address(void *p) |
343 | { |
344 | unsigned long dummy; |
345 | |
346 | return get_kernel_nofault(dummy, (unsigned long *)p); |
347 | } |
348 | |
349 | static void dump_pagetable(unsigned long address) |
350 | { |
351 | pgd_t *base = __va(read_cr3_pa()); |
352 | pgd_t *pgd = base + pgd_index(address); |
353 | p4d_t *p4d; |
354 | pud_t *pud; |
355 | pmd_t *pmd; |
356 | pte_t *pte; |
357 | |
358 | if (bad_address(p: pgd)) |
359 | goto bad; |
360 | |
361 | pr_info("PGD %lx " , pgd_val(*pgd)); |
362 | |
363 | if (!pgd_present(pgd: *pgd)) |
364 | goto out; |
365 | |
366 | p4d = p4d_offset(pgd, address); |
367 | if (bad_address(p: p4d)) |
368 | goto bad; |
369 | |
370 | pr_cont("P4D %lx " , p4d_val(*p4d)); |
371 | if (!p4d_present(p4d: *p4d) || p4d_large(p4d: *p4d)) |
372 | goto out; |
373 | |
374 | pud = pud_offset(p4d, address); |
375 | if (bad_address(p: pud)) |
376 | goto bad; |
377 | |
378 | pr_cont("PUD %lx " , pud_val(*pud)); |
379 | if (!pud_present(pud: *pud) || pud_large(pud: *pud)) |
380 | goto out; |
381 | |
382 | pmd = pmd_offset(pud, address); |
383 | if (bad_address(p: pmd)) |
384 | goto bad; |
385 | |
386 | pr_cont("PMD %lx " , pmd_val(*pmd)); |
387 | if (!pmd_present(pmd: *pmd) || pmd_large(pte: *pmd)) |
388 | goto out; |
389 | |
390 | pte = pte_offset_kernel(pmd, address); |
391 | if (bad_address(p: pte)) |
392 | goto bad; |
393 | |
394 | pr_cont("PTE %lx" , pte_val(*pte)); |
395 | out: |
396 | pr_cont("\n" ); |
397 | return; |
398 | bad: |
399 | pr_info("BAD\n" ); |
400 | } |
401 | |
402 | #endif /* CONFIG_X86_64 */ |
403 | |
404 | /* |
405 | * Workaround for K8 erratum #93 & buggy BIOS. |
406 | * |
407 | * BIOS SMM functions are required to use a specific workaround |
408 | * to avoid corruption of the 64bit RIP register on C stepping K8. |
409 | * |
410 | * A lot of BIOS that didn't get tested properly miss this. |
411 | * |
412 | * The OS sees this as a page fault with the upper 32bits of RIP cleared. |
413 | * Try to work around it here. |
414 | * |
415 | * Note we only handle faults in kernel here. |
416 | * Does nothing on 32-bit. |
417 | */ |
418 | static int is_errata93(struct pt_regs *regs, unsigned long address) |
419 | { |
420 | #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD) |
421 | if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD |
422 | || boot_cpu_data.x86 != 0xf) |
423 | return 0; |
424 | |
425 | if (user_mode(regs)) |
426 | return 0; |
427 | |
428 | if (address != regs->ip) |
429 | return 0; |
430 | |
431 | if ((address >> 32) != 0) |
432 | return 0; |
433 | |
434 | address |= 0xffffffffUL << 32; |
435 | if ((address >= (u64)_stext && address <= (u64)_etext) || |
436 | (address >= MODULES_VADDR && address <= MODULES_END)) { |
437 | printk_once(errata93_warning); |
438 | regs->ip = address; |
439 | return 1; |
440 | } |
441 | #endif |
442 | return 0; |
443 | } |
444 | |
445 | /* |
446 | * Work around K8 erratum #100 K8 in compat mode occasionally jumps |
447 | * to illegal addresses >4GB. |
448 | * |
449 | * We catch this in the page fault handler because these addresses |
450 | * are not reachable. Just detect this case and return. Any code |
451 | * segment in LDT is compatibility mode. |
452 | */ |
453 | static int is_errata100(struct pt_regs *regs, unsigned long address) |
454 | { |
455 | #ifdef CONFIG_X86_64 |
456 | if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32)) |
457 | return 1; |
458 | #endif |
459 | return 0; |
460 | } |
461 | |
462 | /* Pentium F0 0F C7 C8 bug workaround: */ |
463 | static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code, |
464 | unsigned long address) |
465 | { |
466 | #ifdef CONFIG_X86_F00F_BUG |
467 | if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) && |
468 | idt_is_f00f_address(address)) { |
469 | handle_invalid_op(regs); |
470 | return 1; |
471 | } |
472 | #endif |
473 | return 0; |
474 | } |
475 | |
476 | static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index) |
477 | { |
478 | u32 offset = (index >> 3) * sizeof(struct desc_struct); |
479 | unsigned long addr; |
480 | struct ldttss_desc desc; |
481 | |
482 | if (index == 0) { |
483 | pr_alert("%s: NULL\n" , name); |
484 | return; |
485 | } |
486 | |
487 | if (offset + sizeof(struct ldttss_desc) >= gdt->size) { |
488 | pr_alert("%s: 0x%hx -- out of bounds\n" , name, index); |
489 | return; |
490 | } |
491 | |
492 | if (copy_from_kernel_nofault(dst: &desc, src: (void *)(gdt->address + offset), |
493 | size: sizeof(struct ldttss_desc))) { |
494 | pr_alert("%s: 0x%hx -- GDT entry is not readable\n" , |
495 | name, index); |
496 | return; |
497 | } |
498 | |
499 | addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24); |
500 | #ifdef CONFIG_X86_64 |
501 | addr |= ((u64)desc.base3 << 32); |
502 | #endif |
503 | pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n" , |
504 | name, index, addr, (desc.limit0 | (desc.limit1 << 16))); |
505 | } |
506 | |
507 | static void |
508 | show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address) |
509 | { |
510 | if (!oops_may_print()) |
511 | return; |
512 | |
513 | if (error_code & X86_PF_INSTR) { |
514 | unsigned int level; |
515 | pgd_t *pgd; |
516 | pte_t *pte; |
517 | |
518 | pgd = __va(read_cr3_pa()); |
519 | pgd += pgd_index(address); |
520 | |
521 | pte = lookup_address_in_pgd(pgd, address, level: &level); |
522 | |
523 | if (pte && pte_present(a: *pte) && !pte_exec(pte: *pte)) |
524 | pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n" , |
525 | from_kuid(&init_user_ns, current_uid())); |
526 | if (pte && pte_present(a: *pte) && pte_exec(pte: *pte) && |
527 | (pgd_flags(pgd: *pgd) & _PAGE_USER) && |
528 | (__read_cr4() & X86_CR4_SMEP)) |
529 | pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n" , |
530 | from_kuid(&init_user_ns, current_uid())); |
531 | } |
532 | |
533 | if (address < PAGE_SIZE && !user_mode(regs)) |
534 | pr_alert("BUG: kernel NULL pointer dereference, address: %px\n" , |
535 | (void *)address); |
536 | else |
537 | pr_alert("BUG: unable to handle page fault for address: %px\n" , |
538 | (void *)address); |
539 | |
540 | pr_alert("#PF: %s %s in %s mode\n" , |
541 | (error_code & X86_PF_USER) ? "user" : "supervisor" , |
542 | (error_code & X86_PF_INSTR) ? "instruction fetch" : |
543 | (error_code & X86_PF_WRITE) ? "write access" : |
544 | "read access" , |
545 | user_mode(regs) ? "user" : "kernel" ); |
546 | pr_alert("#PF: error_code(0x%04lx) - %s\n" , error_code, |
547 | !(error_code & X86_PF_PROT) ? "not-present page" : |
548 | (error_code & X86_PF_RSVD) ? "reserved bit violation" : |
549 | (error_code & X86_PF_PK) ? "protection keys violation" : |
550 | "permissions violation" ); |
551 | |
552 | if (!(error_code & X86_PF_USER) && user_mode(regs)) { |
553 | struct desc_ptr idt, gdt; |
554 | u16 ldtr, tr; |
555 | |
556 | /* |
557 | * This can happen for quite a few reasons. The more obvious |
558 | * ones are faults accessing the GDT, or LDT. Perhaps |
559 | * surprisingly, if the CPU tries to deliver a benign or |
560 | * contributory exception from user code and gets a page fault |
561 | * during delivery, the page fault can be delivered as though |
562 | * it originated directly from user code. This could happen |
563 | * due to wrong permissions on the IDT, GDT, LDT, TSS, or |
564 | * kernel or IST stack. |
565 | */ |
566 | store_idt(dtr: &idt); |
567 | |
568 | /* Usable even on Xen PV -- it's just slow. */ |
569 | native_store_gdt(dtr: &gdt); |
570 | |
571 | pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n" , |
572 | idt.address, idt.size, gdt.address, gdt.size); |
573 | |
574 | store_ldt(ldtr); |
575 | show_ldttss(gdt: &gdt, name: "LDTR" , index: ldtr); |
576 | |
577 | store_tr(tr); |
578 | show_ldttss(gdt: &gdt, name: "TR" , index: tr); |
579 | } |
580 | |
581 | dump_pagetable(address); |
582 | } |
583 | |
584 | static noinline void |
585 | pgtable_bad(struct pt_regs *regs, unsigned long error_code, |
586 | unsigned long address) |
587 | { |
588 | struct task_struct *tsk; |
589 | unsigned long flags; |
590 | int sig; |
591 | |
592 | flags = oops_begin(); |
593 | tsk = current; |
594 | sig = SIGKILL; |
595 | |
596 | printk(KERN_ALERT "%s: Corrupted page table at address %lx\n" , |
597 | tsk->comm, address); |
598 | dump_pagetable(address); |
599 | |
600 | if (__die("Bad pagetable" , regs, error_code)) |
601 | sig = 0; |
602 | |
603 | oops_end(flags, regs, signr: sig); |
604 | } |
605 | |
606 | static void sanitize_error_code(unsigned long address, |
607 | unsigned long *error_code) |
608 | { |
609 | /* |
610 | * To avoid leaking information about the kernel page |
611 | * table layout, pretend that user-mode accesses to |
612 | * kernel addresses are always protection faults. |
613 | * |
614 | * NB: This means that failed vsyscalls with vsyscall=none |
615 | * will have the PROT bit. This doesn't leak any |
616 | * information and does not appear to cause any problems. |
617 | */ |
618 | if (address >= TASK_SIZE_MAX) |
619 | *error_code |= X86_PF_PROT; |
620 | } |
621 | |
622 | static void set_signal_archinfo(unsigned long address, |
623 | unsigned long error_code) |
624 | { |
625 | struct task_struct *tsk = current; |
626 | |
627 | tsk->thread.trap_nr = X86_TRAP_PF; |
628 | tsk->thread.error_code = error_code | X86_PF_USER; |
629 | tsk->thread.cr2 = address; |
630 | } |
631 | |
632 | static noinline void |
633 | page_fault_oops(struct pt_regs *regs, unsigned long error_code, |
634 | unsigned long address) |
635 | { |
636 | #ifdef CONFIG_VMAP_STACK |
637 | struct stack_info info; |
638 | #endif |
639 | unsigned long flags; |
640 | int sig; |
641 | |
642 | if (user_mode(regs)) { |
643 | /* |
644 | * Implicit kernel access from user mode? Skip the stack |
645 | * overflow and EFI special cases. |
646 | */ |
647 | goto oops; |
648 | } |
649 | |
650 | #ifdef CONFIG_VMAP_STACK |
651 | /* |
652 | * Stack overflow? During boot, we can fault near the initial |
653 | * stack in the direct map, but that's not an overflow -- check |
654 | * that we're in vmalloc space to avoid this. |
655 | */ |
656 | if (is_vmalloc_addr(x: (void *)address) && |
657 | get_stack_guard_info(stack: (void *)address, info: &info)) { |
658 | /* |
659 | * We're likely to be running with very little stack space |
660 | * left. It's plausible that we'd hit this condition but |
661 | * double-fault even before we get this far, in which case |
662 | * we're fine: the double-fault handler will deal with it. |
663 | * |
664 | * We don't want to make it all the way into the oops code |
665 | * and then double-fault, though, because we're likely to |
666 | * break the console driver and lose most of the stack dump. |
667 | */ |
668 | call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*), |
669 | handle_stack_overflow, |
670 | ASM_CALL_ARG3, |
671 | , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info)); |
672 | |
673 | unreachable(); |
674 | } |
675 | #endif |
676 | |
677 | /* |
678 | * Buggy firmware could access regions which might page fault. If |
679 | * this happens, EFI has a special OOPS path that will try to |
680 | * avoid hanging the system. |
681 | */ |
682 | if (IS_ENABLED(CONFIG_EFI)) |
683 | efi_crash_gracefully_on_page_fault(phys_addr: address); |
684 | |
685 | /* Only not-present faults should be handled by KFENCE. */ |
686 | if (!(error_code & X86_PF_PROT) && |
687 | kfence_handle_page_fault(addr: address, is_write: error_code & X86_PF_WRITE, regs)) |
688 | return; |
689 | |
690 | oops: |
691 | /* |
692 | * Oops. The kernel tried to access some bad page. We'll have to |
693 | * terminate things with extreme prejudice: |
694 | */ |
695 | flags = oops_begin(); |
696 | |
697 | show_fault_oops(regs, error_code, address); |
698 | |
699 | if (task_stack_end_corrupted(current)) |
700 | printk(KERN_EMERG "Thread overran stack, or stack corrupted\n" ); |
701 | |
702 | sig = SIGKILL; |
703 | if (__die("Oops" , regs, error_code)) |
704 | sig = 0; |
705 | |
706 | /* Executive summary in case the body of the oops scrolled away */ |
707 | printk(KERN_DEFAULT "CR2: %016lx\n" , address); |
708 | |
709 | oops_end(flags, regs, signr: sig); |
710 | } |
711 | |
712 | static noinline void |
713 | kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code, |
714 | unsigned long address, int signal, int si_code, |
715 | u32 pkey) |
716 | { |
717 | WARN_ON_ONCE(user_mode(regs)); |
718 | |
719 | /* Are we prepared to handle this kernel fault? */ |
720 | if (fixup_exception(regs, X86_TRAP_PF, error_code, fault_addr: address)) { |
721 | /* |
722 | * Any interrupt that takes a fault gets the fixup. This makes |
723 | * the below recursive fault logic only apply to a faults from |
724 | * task context. |
725 | */ |
726 | if (in_interrupt()) |
727 | return; |
728 | |
729 | /* |
730 | * Per the above we're !in_interrupt(), aka. task context. |
731 | * |
732 | * In this case we need to make sure we're not recursively |
733 | * faulting through the emulate_vsyscall() logic. |
734 | */ |
735 | if (current->thread.sig_on_uaccess_err && signal) { |
736 | sanitize_error_code(address, error_code: &error_code); |
737 | |
738 | set_signal_archinfo(address, error_code); |
739 | |
740 | if (si_code == SEGV_PKUERR) { |
741 | force_sig_pkuerr(addr: (void __user *)address, pkey); |
742 | } else { |
743 | /* XXX: hwpoison faults will set the wrong code. */ |
744 | force_sig_fault(sig: signal, code: si_code, addr: (void __user *)address); |
745 | } |
746 | } |
747 | |
748 | /* |
749 | * Barring that, we can do the fixup and be happy. |
750 | */ |
751 | return; |
752 | } |
753 | |
754 | /* |
755 | * AMD erratum #91 manifests as a spurious page fault on a PREFETCH |
756 | * instruction. |
757 | */ |
758 | if (is_prefetch(regs, error_code, addr: address)) |
759 | return; |
760 | |
761 | page_fault_oops(regs, error_code, address); |
762 | } |
763 | |
764 | /* |
765 | * Print out info about fatal segfaults, if the show_unhandled_signals |
766 | * sysctl is set: |
767 | */ |
768 | static inline void |
769 | show_signal_msg(struct pt_regs *regs, unsigned long error_code, |
770 | unsigned long address, struct task_struct *tsk) |
771 | { |
772 | const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG; |
773 | /* This is a racy snapshot, but it's better than nothing. */ |
774 | int cpu = raw_smp_processor_id(); |
775 | |
776 | if (!unhandled_signal(tsk, SIGSEGV)) |
777 | return; |
778 | |
779 | if (!printk_ratelimit()) |
780 | return; |
781 | |
782 | printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx" , |
783 | loglvl, tsk->comm, task_pid_nr(tsk), address, |
784 | (void *)regs->ip, (void *)regs->sp, error_code); |
785 | |
786 | print_vma_addr(KERN_CONT " in " , rip: regs->ip); |
787 | |
788 | /* |
789 | * Dump the likely CPU where the fatal segfault happened. |
790 | * This can help identify faulty hardware. |
791 | */ |
792 | printk(KERN_CONT " likely on CPU %d (core %d, socket %d)" , cpu, |
793 | topology_core_id(cpu), topology_physical_package_id(cpu)); |
794 | |
795 | |
796 | printk(KERN_CONT "\n" ); |
797 | |
798 | show_opcodes(regs, loglvl); |
799 | } |
800 | |
801 | /* |
802 | * The (legacy) vsyscall page is the long page in the kernel portion |
803 | * of the address space that has user-accessible permissions. |
804 | */ |
805 | static bool is_vsyscall_vaddr(unsigned long vaddr) |
806 | { |
807 | return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR); |
808 | } |
809 | |
810 | static void |
811 | __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, |
812 | unsigned long address, u32 pkey, int si_code) |
813 | { |
814 | struct task_struct *tsk = current; |
815 | |
816 | if (!user_mode(regs)) { |
817 | kernelmode_fixup_or_oops(regs, error_code, address, |
818 | SIGSEGV, si_code, pkey); |
819 | return; |
820 | } |
821 | |
822 | if (!(error_code & X86_PF_USER)) { |
823 | /* Implicit user access to kernel memory -- just oops */ |
824 | page_fault_oops(regs, error_code, address); |
825 | return; |
826 | } |
827 | |
828 | /* |
829 | * User mode accesses just cause a SIGSEGV. |
830 | * It's possible to have interrupts off here: |
831 | */ |
832 | local_irq_enable(); |
833 | |
834 | /* |
835 | * Valid to do another page fault here because this one came |
836 | * from user space: |
837 | */ |
838 | if (is_prefetch(regs, error_code, addr: address)) |
839 | return; |
840 | |
841 | if (is_errata100(regs, address)) |
842 | return; |
843 | |
844 | sanitize_error_code(address, error_code: &error_code); |
845 | |
846 | if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, fault_addr: address)) |
847 | return; |
848 | |
849 | if (likely(show_unhandled_signals)) |
850 | show_signal_msg(regs, error_code, address, tsk); |
851 | |
852 | set_signal_archinfo(address, error_code); |
853 | |
854 | if (si_code == SEGV_PKUERR) |
855 | force_sig_pkuerr(addr: (void __user *)address, pkey); |
856 | else |
857 | force_sig_fault(SIGSEGV, code: si_code, addr: (void __user *)address); |
858 | |
859 | local_irq_disable(); |
860 | } |
861 | |
862 | static noinline void |
863 | bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, |
864 | unsigned long address) |
865 | { |
866 | __bad_area_nosemaphore(regs, error_code, address, pkey: 0, SEGV_MAPERR); |
867 | } |
868 | |
869 | static void |
870 | __bad_area(struct pt_regs *regs, unsigned long error_code, |
871 | unsigned long address, u32 pkey, int si_code) |
872 | { |
873 | struct mm_struct *mm = current->mm; |
874 | /* |
875 | * Something tried to access memory that isn't in our memory map.. |
876 | * Fix it, but check if it's kernel or user first.. |
877 | */ |
878 | mmap_read_unlock(mm); |
879 | |
880 | __bad_area_nosemaphore(regs, error_code, address, pkey, si_code); |
881 | } |
882 | |
883 | static inline bool bad_area_access_from_pkeys(unsigned long error_code, |
884 | struct vm_area_struct *vma) |
885 | { |
886 | /* This code is always called on the current mm */ |
887 | bool foreign = false; |
888 | |
889 | if (!cpu_feature_enabled(X86_FEATURE_OSPKE)) |
890 | return false; |
891 | if (error_code & X86_PF_PK) |
892 | return true; |
893 | /* this checks permission keys on the VMA: */ |
894 | if (!arch_vma_access_permitted(vma, write: (error_code & X86_PF_WRITE), |
895 | execute: (error_code & X86_PF_INSTR), foreign)) |
896 | return true; |
897 | return false; |
898 | } |
899 | |
900 | static noinline void |
901 | bad_area_access_error(struct pt_regs *regs, unsigned long error_code, |
902 | unsigned long address, struct vm_area_struct *vma) |
903 | { |
904 | /* |
905 | * This OSPKE check is not strictly necessary at runtime. |
906 | * But, doing it this way allows compiler optimizations |
907 | * if pkeys are compiled out. |
908 | */ |
909 | if (bad_area_access_from_pkeys(error_code, vma)) { |
910 | /* |
911 | * A protection key fault means that the PKRU value did not allow |
912 | * access to some PTE. Userspace can figure out what PKRU was |
913 | * from the XSAVE state. This function captures the pkey from |
914 | * the vma and passes it to userspace so userspace can discover |
915 | * which protection key was set on the PTE. |
916 | * |
917 | * If we get here, we know that the hardware signaled a X86_PF_PK |
918 | * fault and that there was a VMA once we got in the fault |
919 | * handler. It does *not* guarantee that the VMA we find here |
920 | * was the one that we faulted on. |
921 | * |
922 | * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4); |
923 | * 2. T1 : set PKRU to deny access to pkey=4, touches page |
924 | * 3. T1 : faults... |
925 | * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5); |
926 | * 5. T1 : enters fault handler, takes mmap_lock, etc... |
927 | * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really |
928 | * faulted on a pte with its pkey=4. |
929 | */ |
930 | u32 pkey = vma_pkey(vma); |
931 | |
932 | __bad_area(regs, error_code, address, pkey, SEGV_PKUERR); |
933 | } else { |
934 | __bad_area(regs, error_code, address, pkey: 0, SEGV_ACCERR); |
935 | } |
936 | } |
937 | |
938 | static void |
939 | do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address, |
940 | vm_fault_t fault) |
941 | { |
942 | /* Kernel mode? Handle exceptions or die: */ |
943 | if (!user_mode(regs)) { |
944 | kernelmode_fixup_or_oops(regs, error_code, address, |
945 | SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY); |
946 | return; |
947 | } |
948 | |
949 | /* User-space => ok to do another page fault: */ |
950 | if (is_prefetch(regs, error_code, addr: address)) |
951 | return; |
952 | |
953 | sanitize_error_code(address, error_code: &error_code); |
954 | |
955 | if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, fault_addr: address)) |
956 | return; |
957 | |
958 | set_signal_archinfo(address, error_code); |
959 | |
960 | #ifdef CONFIG_MEMORY_FAILURE |
961 | if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) { |
962 | struct task_struct *tsk = current; |
963 | unsigned lsb = 0; |
964 | |
965 | pr_err( |
966 | "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n" , |
967 | tsk->comm, tsk->pid, address); |
968 | if (fault & VM_FAULT_HWPOISON_LARGE) |
969 | lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); |
970 | if (fault & VM_FAULT_HWPOISON) |
971 | lsb = PAGE_SHIFT; |
972 | force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb); |
973 | return; |
974 | } |
975 | #endif |
976 | force_sig_fault(SIGBUS, BUS_ADRERR, addr: (void __user *)address); |
977 | } |
978 | |
979 | static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte) |
980 | { |
981 | if ((error_code & X86_PF_WRITE) && !pte_write(pte: *pte)) |
982 | return 0; |
983 | |
984 | if ((error_code & X86_PF_INSTR) && !pte_exec(pte: *pte)) |
985 | return 0; |
986 | |
987 | return 1; |
988 | } |
989 | |
990 | /* |
991 | * Handle a spurious fault caused by a stale TLB entry. |
992 | * |
993 | * This allows us to lazily refresh the TLB when increasing the |
994 | * permissions of a kernel page (RO -> RW or NX -> X). Doing it |
995 | * eagerly is very expensive since that implies doing a full |
996 | * cross-processor TLB flush, even if no stale TLB entries exist |
997 | * on other processors. |
998 | * |
999 | * Spurious faults may only occur if the TLB contains an entry with |
1000 | * fewer permission than the page table entry. Non-present (P = 0) |
1001 | * and reserved bit (R = 1) faults are never spurious. |
1002 | * |
1003 | * There are no security implications to leaving a stale TLB when |
1004 | * increasing the permissions on a page. |
1005 | * |
1006 | * Returns non-zero if a spurious fault was handled, zero otherwise. |
1007 | * |
1008 | * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 |
1009 | * (Optional Invalidation). |
1010 | */ |
1011 | static noinline int |
1012 | spurious_kernel_fault(unsigned long error_code, unsigned long address) |
1013 | { |
1014 | pgd_t *pgd; |
1015 | p4d_t *p4d; |
1016 | pud_t *pud; |
1017 | pmd_t *pmd; |
1018 | pte_t *pte; |
1019 | int ret; |
1020 | |
1021 | /* |
1022 | * Only writes to RO or instruction fetches from NX may cause |
1023 | * spurious faults. |
1024 | * |
1025 | * These could be from user or supervisor accesses but the TLB |
1026 | * is only lazily flushed after a kernel mapping protection |
1027 | * change, so user accesses are not expected to cause spurious |
1028 | * faults. |
1029 | */ |
1030 | if (error_code != (X86_PF_WRITE | X86_PF_PROT) && |
1031 | error_code != (X86_PF_INSTR | X86_PF_PROT)) |
1032 | return 0; |
1033 | |
1034 | pgd = init_mm.pgd + pgd_index(address); |
1035 | if (!pgd_present(pgd: *pgd)) |
1036 | return 0; |
1037 | |
1038 | p4d = p4d_offset(pgd, address); |
1039 | if (!p4d_present(p4d: *p4d)) |
1040 | return 0; |
1041 | |
1042 | if (p4d_large(p4d: *p4d)) |
1043 | return spurious_kernel_fault_check(error_code, pte: (pte_t *) p4d); |
1044 | |
1045 | pud = pud_offset(p4d, address); |
1046 | if (!pud_present(pud: *pud)) |
1047 | return 0; |
1048 | |
1049 | if (pud_large(pud: *pud)) |
1050 | return spurious_kernel_fault_check(error_code, pte: (pte_t *) pud); |
1051 | |
1052 | pmd = pmd_offset(pud, address); |
1053 | if (!pmd_present(pmd: *pmd)) |
1054 | return 0; |
1055 | |
1056 | if (pmd_large(pte: *pmd)) |
1057 | return spurious_kernel_fault_check(error_code, pte: (pte_t *) pmd); |
1058 | |
1059 | pte = pte_offset_kernel(pmd, address); |
1060 | if (!pte_present(a: *pte)) |
1061 | return 0; |
1062 | |
1063 | ret = spurious_kernel_fault_check(error_code, pte); |
1064 | if (!ret) |
1065 | return 0; |
1066 | |
1067 | /* |
1068 | * Make sure we have permissions in PMD. |
1069 | * If not, then there's a bug in the page tables: |
1070 | */ |
1071 | ret = spurious_kernel_fault_check(error_code, pte: (pte_t *) pmd); |
1072 | WARN_ONCE(!ret, "PMD has incorrect permission bits\n" ); |
1073 | |
1074 | return ret; |
1075 | } |
1076 | NOKPROBE_SYMBOL(spurious_kernel_fault); |
1077 | |
1078 | int show_unhandled_signals = 1; |
1079 | |
1080 | static inline int |
1081 | access_error(unsigned long error_code, struct vm_area_struct *vma) |
1082 | { |
1083 | /* This is only called for the current mm, so: */ |
1084 | bool foreign = false; |
1085 | |
1086 | /* |
1087 | * Read or write was blocked by protection keys. This is |
1088 | * always an unconditional error and can never result in |
1089 | * a follow-up action to resolve the fault, like a COW. |
1090 | */ |
1091 | if (error_code & X86_PF_PK) |
1092 | return 1; |
1093 | |
1094 | /* |
1095 | * SGX hardware blocked the access. This usually happens |
1096 | * when the enclave memory contents have been destroyed, like |
1097 | * after a suspend/resume cycle. In any case, the kernel can't |
1098 | * fix the cause of the fault. Handle the fault as an access |
1099 | * error even in cases where no actual access violation |
1100 | * occurred. This allows userspace to rebuild the enclave in |
1101 | * response to the signal. |
1102 | */ |
1103 | if (unlikely(error_code & X86_PF_SGX)) |
1104 | return 1; |
1105 | |
1106 | /* |
1107 | * Make sure to check the VMA so that we do not perform |
1108 | * faults just to hit a X86_PF_PK as soon as we fill in a |
1109 | * page. |
1110 | */ |
1111 | if (!arch_vma_access_permitted(vma, write: (error_code & X86_PF_WRITE), |
1112 | execute: (error_code & X86_PF_INSTR), foreign)) |
1113 | return 1; |
1114 | |
1115 | /* |
1116 | * Shadow stack accesses (PF_SHSTK=1) are only permitted to |
1117 | * shadow stack VMAs. All other accesses result in an error. |
1118 | */ |
1119 | if (error_code & X86_PF_SHSTK) { |
1120 | if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK))) |
1121 | return 1; |
1122 | if (unlikely(!(vma->vm_flags & VM_WRITE))) |
1123 | return 1; |
1124 | return 0; |
1125 | } |
1126 | |
1127 | if (error_code & X86_PF_WRITE) { |
1128 | /* write, present and write, not present: */ |
1129 | if (unlikely(vma->vm_flags & VM_SHADOW_STACK)) |
1130 | return 1; |
1131 | if (unlikely(!(vma->vm_flags & VM_WRITE))) |
1132 | return 1; |
1133 | return 0; |
1134 | } |
1135 | |
1136 | /* read, present: */ |
1137 | if (unlikely(error_code & X86_PF_PROT)) |
1138 | return 1; |
1139 | |
1140 | /* read, not present: */ |
1141 | if (unlikely(!vma_is_accessible(vma))) |
1142 | return 1; |
1143 | |
1144 | return 0; |
1145 | } |
1146 | |
1147 | bool fault_in_kernel_space(unsigned long address) |
1148 | { |
1149 | /* |
1150 | * On 64-bit systems, the vsyscall page is at an address above |
1151 | * TASK_SIZE_MAX, but is not considered part of the kernel |
1152 | * address space. |
1153 | */ |
1154 | if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(vaddr: address)) |
1155 | return false; |
1156 | |
1157 | return address >= TASK_SIZE_MAX; |
1158 | } |
1159 | |
1160 | /* |
1161 | * Called for all faults where 'address' is part of the kernel address |
1162 | * space. Might get called for faults that originate from *code* that |
1163 | * ran in userspace or the kernel. |
1164 | */ |
1165 | static void |
1166 | do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code, |
1167 | unsigned long address) |
1168 | { |
1169 | /* |
1170 | * Protection keys exceptions only happen on user pages. We |
1171 | * have no user pages in the kernel portion of the address |
1172 | * space, so do not expect them here. |
1173 | */ |
1174 | WARN_ON_ONCE(hw_error_code & X86_PF_PK); |
1175 | |
1176 | #ifdef CONFIG_X86_32 |
1177 | /* |
1178 | * We can fault-in kernel-space virtual memory on-demand. The |
1179 | * 'reference' page table is init_mm.pgd. |
1180 | * |
1181 | * NOTE! We MUST NOT take any locks for this case. We may |
1182 | * be in an interrupt or a critical region, and should |
1183 | * only copy the information from the master page table, |
1184 | * nothing more. |
1185 | * |
1186 | * Before doing this on-demand faulting, ensure that the |
1187 | * fault is not any of the following: |
1188 | * 1. A fault on a PTE with a reserved bit set. |
1189 | * 2. A fault caused by a user-mode access. (Do not demand- |
1190 | * fault kernel memory due to user-mode accesses). |
1191 | * 3. A fault caused by a page-level protection violation. |
1192 | * (A demand fault would be on a non-present page which |
1193 | * would have X86_PF_PROT==0). |
1194 | * |
1195 | * This is only needed to close a race condition on x86-32 in |
1196 | * the vmalloc mapping/unmapping code. See the comment above |
1197 | * vmalloc_fault() for details. On x86-64 the race does not |
1198 | * exist as the vmalloc mappings don't need to be synchronized |
1199 | * there. |
1200 | */ |
1201 | if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) { |
1202 | if (vmalloc_fault(address) >= 0) |
1203 | return; |
1204 | } |
1205 | #endif |
1206 | |
1207 | if (is_f00f_bug(regs, error_code: hw_error_code, address)) |
1208 | return; |
1209 | |
1210 | /* Was the fault spurious, caused by lazy TLB invalidation? */ |
1211 | if (spurious_kernel_fault(error_code: hw_error_code, address)) |
1212 | return; |
1213 | |
1214 | /* kprobes don't want to hook the spurious faults: */ |
1215 | if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF))) |
1216 | return; |
1217 | |
1218 | /* |
1219 | * Note, despite being a "bad area", there are quite a few |
1220 | * acceptable reasons to get here, such as erratum fixups |
1221 | * and handling kernel code that can fault, like get_user(). |
1222 | * |
1223 | * Don't take the mm semaphore here. If we fixup a prefetch |
1224 | * fault we could otherwise deadlock: |
1225 | */ |
1226 | bad_area_nosemaphore(regs, error_code: hw_error_code, address); |
1227 | } |
1228 | NOKPROBE_SYMBOL(do_kern_addr_fault); |
1229 | |
1230 | /* |
1231 | * Handle faults in the user portion of the address space. Nothing in here |
1232 | * should check X86_PF_USER without a specific justification: for almost |
1233 | * all purposes, we should treat a normal kernel access to user memory |
1234 | * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction. |
1235 | * The one exception is AC flag handling, which is, per the x86 |
1236 | * architecture, special for WRUSS. |
1237 | */ |
1238 | static inline |
1239 | void do_user_addr_fault(struct pt_regs *regs, |
1240 | unsigned long error_code, |
1241 | unsigned long address) |
1242 | { |
1243 | struct vm_area_struct *vma; |
1244 | struct task_struct *tsk; |
1245 | struct mm_struct *mm; |
1246 | vm_fault_t fault; |
1247 | unsigned int flags = FAULT_FLAG_DEFAULT; |
1248 | |
1249 | tsk = current; |
1250 | mm = tsk->mm; |
1251 | |
1252 | if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) { |
1253 | /* |
1254 | * Whoops, this is kernel mode code trying to execute from |
1255 | * user memory. Unless this is AMD erratum #93, which |
1256 | * corrupts RIP such that it looks like a user address, |
1257 | * this is unrecoverable. Don't even try to look up the |
1258 | * VMA or look for extable entries. |
1259 | */ |
1260 | if (is_errata93(regs, address)) |
1261 | return; |
1262 | |
1263 | page_fault_oops(regs, error_code, address); |
1264 | return; |
1265 | } |
1266 | |
1267 | /* kprobes don't want to hook the spurious faults: */ |
1268 | if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF))) |
1269 | return; |
1270 | |
1271 | /* |
1272 | * Reserved bits are never expected to be set on |
1273 | * entries in the user portion of the page tables. |
1274 | */ |
1275 | if (unlikely(error_code & X86_PF_RSVD)) |
1276 | pgtable_bad(regs, error_code, address); |
1277 | |
1278 | /* |
1279 | * If SMAP is on, check for invalid kernel (supervisor) access to user |
1280 | * pages in the user address space. The odd case here is WRUSS, |
1281 | * which, according to the preliminary documentation, does not respect |
1282 | * SMAP and will have the USER bit set so, in all cases, SMAP |
1283 | * enforcement appears to be consistent with the USER bit. |
1284 | */ |
1285 | if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) && |
1286 | !(error_code & X86_PF_USER) && |
1287 | !(regs->flags & X86_EFLAGS_AC))) { |
1288 | /* |
1289 | * No extable entry here. This was a kernel access to an |
1290 | * invalid pointer. get_kernel_nofault() will not get here. |
1291 | */ |
1292 | page_fault_oops(regs, error_code, address); |
1293 | return; |
1294 | } |
1295 | |
1296 | /* |
1297 | * If we're in an interrupt, have no user context or are running |
1298 | * in a region with pagefaults disabled then we must not take the fault |
1299 | */ |
1300 | if (unlikely(faulthandler_disabled() || !mm)) { |
1301 | bad_area_nosemaphore(regs, error_code, address); |
1302 | return; |
1303 | } |
1304 | |
1305 | /* |
1306 | * It's safe to allow irq's after cr2 has been saved and the |
1307 | * vmalloc fault has been handled. |
1308 | * |
1309 | * User-mode registers count as a user access even for any |
1310 | * potential system fault or CPU buglet: |
1311 | */ |
1312 | if (user_mode(regs)) { |
1313 | local_irq_enable(); |
1314 | flags |= FAULT_FLAG_USER; |
1315 | } else { |
1316 | if (regs->flags & X86_EFLAGS_IF) |
1317 | local_irq_enable(); |
1318 | } |
1319 | |
1320 | perf_sw_event(event_id: PERF_COUNT_SW_PAGE_FAULTS, nr: 1, regs, addr: address); |
1321 | |
1322 | /* |
1323 | * Read-only permissions can not be expressed in shadow stack PTEs. |
1324 | * Treat all shadow stack accesses as WRITE faults. This ensures |
1325 | * that the MM will prepare everything (e.g., break COW) such that |
1326 | * maybe_mkwrite() can create a proper shadow stack PTE. |
1327 | */ |
1328 | if (error_code & X86_PF_SHSTK) |
1329 | flags |= FAULT_FLAG_WRITE; |
1330 | if (error_code & X86_PF_WRITE) |
1331 | flags |= FAULT_FLAG_WRITE; |
1332 | if (error_code & X86_PF_INSTR) |
1333 | flags |= FAULT_FLAG_INSTRUCTION; |
1334 | |
1335 | #ifdef CONFIG_X86_64 |
1336 | /* |
1337 | * Faults in the vsyscall page might need emulation. The |
1338 | * vsyscall page is at a high address (>PAGE_OFFSET), but is |
1339 | * considered to be part of the user address space. |
1340 | * |
1341 | * The vsyscall page does not have a "real" VMA, so do this |
1342 | * emulation before we go searching for VMAs. |
1343 | * |
1344 | * PKRU never rejects instruction fetches, so we don't need |
1345 | * to consider the PF_PK bit. |
1346 | */ |
1347 | if (is_vsyscall_vaddr(vaddr: address)) { |
1348 | if (emulate_vsyscall(error_code, regs, address)) |
1349 | return; |
1350 | } |
1351 | #endif |
1352 | |
1353 | if (!(flags & FAULT_FLAG_USER)) |
1354 | goto lock_mmap; |
1355 | |
1356 | vma = lock_vma_under_rcu(mm, address); |
1357 | if (!vma) |
1358 | goto lock_mmap; |
1359 | |
1360 | if (unlikely(access_error(error_code, vma))) { |
1361 | vma_end_read(vma); |
1362 | goto lock_mmap; |
1363 | } |
1364 | fault = handle_mm_fault(vma, address, flags: flags | FAULT_FLAG_VMA_LOCK, regs); |
1365 | if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED))) |
1366 | vma_end_read(vma); |
1367 | |
1368 | if (!(fault & VM_FAULT_RETRY)) { |
1369 | count_vm_vma_lock_event(VMA_LOCK_SUCCESS); |
1370 | goto done; |
1371 | } |
1372 | count_vm_vma_lock_event(VMA_LOCK_RETRY); |
1373 | |
1374 | /* Quick path to respond to signals */ |
1375 | if (fault_signal_pending(fault_flags: fault, regs)) { |
1376 | if (!user_mode(regs)) |
1377 | kernelmode_fixup_or_oops(regs, error_code, address, |
1378 | SIGBUS, BUS_ADRERR, |
1379 | ARCH_DEFAULT_PKEY); |
1380 | return; |
1381 | } |
1382 | lock_mmap: |
1383 | |
1384 | retry: |
1385 | vma = lock_mm_and_find_vma(mm, address, regs); |
1386 | if (unlikely(!vma)) { |
1387 | bad_area_nosemaphore(regs, error_code, address); |
1388 | return; |
1389 | } |
1390 | |
1391 | /* |
1392 | * Ok, we have a good vm_area for this memory access, so |
1393 | * we can handle it.. |
1394 | */ |
1395 | if (unlikely(access_error(error_code, vma))) { |
1396 | bad_area_access_error(regs, error_code, address, vma); |
1397 | return; |
1398 | } |
1399 | |
1400 | /* |
1401 | * If for any reason at all we couldn't handle the fault, |
1402 | * make sure we exit gracefully rather than endlessly redo |
1403 | * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if |
1404 | * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked. |
1405 | * |
1406 | * Note that handle_userfault() may also release and reacquire mmap_lock |
1407 | * (and not return with VM_FAULT_RETRY), when returning to userland to |
1408 | * repeat the page fault later with a VM_FAULT_NOPAGE retval |
1409 | * (potentially after handling any pending signal during the return to |
1410 | * userland). The return to userland is identified whenever |
1411 | * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags. |
1412 | */ |
1413 | fault = handle_mm_fault(vma, address, flags, regs); |
1414 | |
1415 | if (fault_signal_pending(fault_flags: fault, regs)) { |
1416 | /* |
1417 | * Quick path to respond to signals. The core mm code |
1418 | * has unlocked the mm for us if we get here. |
1419 | */ |
1420 | if (!user_mode(regs)) |
1421 | kernelmode_fixup_or_oops(regs, error_code, address, |
1422 | SIGBUS, BUS_ADRERR, |
1423 | ARCH_DEFAULT_PKEY); |
1424 | return; |
1425 | } |
1426 | |
1427 | /* The fault is fully completed (including releasing mmap lock) */ |
1428 | if (fault & VM_FAULT_COMPLETED) |
1429 | return; |
1430 | |
1431 | /* |
1432 | * If we need to retry the mmap_lock has already been released, |
1433 | * and if there is a fatal signal pending there is no guarantee |
1434 | * that we made any progress. Handle this case first. |
1435 | */ |
1436 | if (unlikely(fault & VM_FAULT_RETRY)) { |
1437 | flags |= FAULT_FLAG_TRIED; |
1438 | goto retry; |
1439 | } |
1440 | |
1441 | mmap_read_unlock(mm); |
1442 | done: |
1443 | if (likely(!(fault & VM_FAULT_ERROR))) |
1444 | return; |
1445 | |
1446 | if (fatal_signal_pending(current) && !user_mode(regs)) { |
1447 | kernelmode_fixup_or_oops(regs, error_code, address, |
1448 | signal: 0, si_code: 0, ARCH_DEFAULT_PKEY); |
1449 | return; |
1450 | } |
1451 | |
1452 | if (fault & VM_FAULT_OOM) { |
1453 | /* Kernel mode? Handle exceptions or die: */ |
1454 | if (!user_mode(regs)) { |
1455 | kernelmode_fixup_or_oops(regs, error_code, address, |
1456 | SIGSEGV, SEGV_MAPERR, |
1457 | ARCH_DEFAULT_PKEY); |
1458 | return; |
1459 | } |
1460 | |
1461 | /* |
1462 | * We ran out of memory, call the OOM killer, and return the |
1463 | * userspace (which will retry the fault, or kill us if we got |
1464 | * oom-killed): |
1465 | */ |
1466 | pagefault_out_of_memory(); |
1467 | } else { |
1468 | if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON| |
1469 | VM_FAULT_HWPOISON_LARGE)) |
1470 | do_sigbus(regs, error_code, address, fault); |
1471 | else if (fault & VM_FAULT_SIGSEGV) |
1472 | bad_area_nosemaphore(regs, error_code, address); |
1473 | else |
1474 | BUG(); |
1475 | } |
1476 | } |
1477 | NOKPROBE_SYMBOL(do_user_addr_fault); |
1478 | |
1479 | static __always_inline void |
1480 | trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code, |
1481 | unsigned long address) |
1482 | { |
1483 | if (!trace_pagefault_enabled()) |
1484 | return; |
1485 | |
1486 | if (user_mode(regs)) |
1487 | trace_page_fault_user(address, regs, error_code); |
1488 | else |
1489 | trace_page_fault_kernel(address, regs, error_code); |
1490 | } |
1491 | |
1492 | static __always_inline void |
1493 | handle_page_fault(struct pt_regs *regs, unsigned long error_code, |
1494 | unsigned long address) |
1495 | { |
1496 | trace_page_fault_entries(regs, error_code, address); |
1497 | |
1498 | if (unlikely(kmmio_fault(regs, address))) |
1499 | return; |
1500 | |
1501 | /* Was the fault on kernel-controlled part of the address space? */ |
1502 | if (unlikely(fault_in_kernel_space(address))) { |
1503 | do_kern_addr_fault(regs, hw_error_code: error_code, address); |
1504 | } else { |
1505 | do_user_addr_fault(regs, error_code, address); |
1506 | /* |
1507 | * User address page fault handling might have reenabled |
1508 | * interrupts. Fixing up all potential exit points of |
1509 | * do_user_addr_fault() and its leaf functions is just not |
1510 | * doable w/o creating an unholy mess or turning the code |
1511 | * upside down. |
1512 | */ |
1513 | local_irq_disable(); |
1514 | } |
1515 | } |
1516 | |
1517 | DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault) |
1518 | { |
1519 | unsigned long address = read_cr2(); |
1520 | irqentry_state_t state; |
1521 | |
1522 | prefetchw(x: ¤t->mm->mmap_lock); |
1523 | |
1524 | /* |
1525 | * KVM uses #PF vector to deliver 'page not present' events to guests |
1526 | * (asynchronous page fault mechanism). The event happens when a |
1527 | * userspace task is trying to access some valid (from guest's point of |
1528 | * view) memory which is not currently mapped by the host (e.g. the |
1529 | * memory is swapped out). Note, the corresponding "page ready" event |
1530 | * which is injected when the memory becomes available, is delivered via |
1531 | * an interrupt mechanism and not a #PF exception |
1532 | * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()). |
1533 | * |
1534 | * We are relying on the interrupted context being sane (valid RSP, |
1535 | * relevant locks not held, etc.), which is fine as long as the |
1536 | * interrupted context had IF=1. We are also relying on the KVM |
1537 | * async pf type field and CR2 being read consistently instead of |
1538 | * getting values from real and async page faults mixed up. |
1539 | * |
1540 | * Fingers crossed. |
1541 | * |
1542 | * The async #PF handling code takes care of idtentry handling |
1543 | * itself. |
1544 | */ |
1545 | if (kvm_handle_async_pf(regs, token: (u32)address)) |
1546 | return; |
1547 | |
1548 | /* |
1549 | * Entry handling for valid #PF from kernel mode is slightly |
1550 | * different: RCU is already watching and ct_irq_enter() must not |
1551 | * be invoked because a kernel fault on a user space address might |
1552 | * sleep. |
1553 | * |
1554 | * In case the fault hit a RCU idle region the conditional entry |
1555 | * code reenabled RCU to avoid subsequent wreckage which helps |
1556 | * debuggability. |
1557 | */ |
1558 | state = irqentry_enter(regs); |
1559 | |
1560 | instrumentation_begin(); |
1561 | handle_page_fault(regs, error_code, address); |
1562 | instrumentation_end(); |
1563 | |
1564 | irqentry_exit(regs, state); |
1565 | } |
1566 | |