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/bootmem.h> /* max_low_pfn */ 12 #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */ 13 #include <linux/mmiotrace.h> /* kmmio_handler, ... */ 14 #include <linux/perf_event.h> /* perf_sw_event */ 15 #include <linux/hugetlb.h> /* hstate_index_to_shift */ 16 #include <linux/prefetch.h> /* prefetchw */ 17 #include <linux/context_tracking.h> /* exception_enter(), ... */ 18 #include <linux/uaccess.h> /* faulthandler_disabled() */ 19 20 #include <asm/cpufeature.h> /* boot_cpu_has, ... */ 21 #include <asm/traps.h> /* dotraplinkage, ... */ 22 #include <asm/pgalloc.h> /* pgd_*(), ... */ 23 #include <asm/fixmap.h> /* VSYSCALL_ADDR */ 24 #include <asm/vsyscall.h> /* emulate_vsyscall */ 25 #include <asm/vm86.h> /* struct vm86 */ 26 #include <asm/mmu_context.h> /* vma_pkey() */ 27 28 #define CREATE_TRACE_POINTS 29 #include <asm/trace/exceptions.h> 30 31 /* 32 * Returns 0 if mmiotrace is disabled, or if the fault is not 33 * handled by mmiotrace: 34 */ 35 static nokprobe_inline int 36 kmmio_fault(struct pt_regs *regs, unsigned long addr) 37 { 38 if (unlikely(is_kmmio_active())) 39 if (kmmio_handler(regs, addr) == 1) 40 return -1; 41 return 0; 42 } 43 44 static nokprobe_inline int kprobes_fault(struct pt_regs *regs) 45 { 46 int ret = 0; 47 48 /* kprobe_running() needs smp_processor_id() */ 49 if (kprobes_built_in() && !user_mode(regs)) { 50 preempt_disable(); 51 if (kprobe_running() && kprobe_fault_handler(regs, 14)) 52 ret = 1; 53 preempt_enable(); 54 } 55 56 return ret; 57 } 58 59 /* 60 * Prefetch quirks: 61 * 62 * 32-bit mode: 63 * 64 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch. 65 * Check that here and ignore it. 66 * 67 * 64-bit mode: 68 * 69 * Sometimes the CPU reports invalid exceptions on prefetch. 70 * Check that here and ignore it. 71 * 72 * Opcode checker based on code by Richard Brunner. 73 */ 74 static inline int 75 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr, 76 unsigned char opcode, int *prefetch) 77 { 78 unsigned char instr_hi = opcode & 0xf0; 79 unsigned char instr_lo = opcode & 0x0f; 80 81 switch (instr_hi) { 82 case 0x20: 83 case 0x30: 84 /* 85 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. 86 * In X86_64 long mode, the CPU will signal invalid 87 * opcode if some of these prefixes are present so 88 * X86_64 will never get here anyway 89 */ 90 return ((instr_lo & 7) == 0x6); 91 #ifdef CONFIG_X86_64 92 case 0x40: 93 /* 94 * In AMD64 long mode 0x40..0x4F are valid REX prefixes 95 * Need to figure out under what instruction mode the 96 * instruction was issued. Could check the LDT for lm, 97 * but for now it's good enough to assume that long 98 * mode only uses well known segments or kernel. 99 */ 100 return (!user_mode(regs) || user_64bit_mode(regs)); 101 #endif 102 case 0x60: 103 /* 0x64 thru 0x67 are valid prefixes in all modes. */ 104 return (instr_lo & 0xC) == 0x4; 105 case 0xF0: 106 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */ 107 return !instr_lo || (instr_lo>>1) == 1; 108 case 0x00: 109 /* Prefetch instruction is 0x0F0D or 0x0F18 */ 110 if (probe_kernel_address(instr, opcode)) 111 return 0; 112 113 *prefetch = (instr_lo == 0xF) && 114 (opcode == 0x0D || opcode == 0x18); 115 return 0; 116 default: 117 return 0; 118 } 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 /* 129 * If it was a exec (instruction fetch) fault on NX page, then 130 * do not ignore the fault: 131 */ 132 if (error_code & X86_PF_INSTR) 133 return 0; 134 135 instr = (void *)convert_ip_to_linear(current, regs); 136 max_instr = instr + 15; 137 138 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX) 139 return 0; 140 141 while (instr < max_instr) { 142 unsigned char opcode; 143 144 if (probe_kernel_address(instr, opcode)) 145 break; 146 147 instr++; 148 149 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch)) 150 break; 151 } 152 return prefetch; 153 } 154 155 /* 156 * A protection key fault means that the PKRU value did not allow 157 * access to some PTE. Userspace can figure out what PKRU was 158 * from the XSAVE state, and this function fills out a field in 159 * siginfo so userspace can discover which protection key was set 160 * on the PTE. 161 * 162 * If we get here, we know that the hardware signaled a X86_PF_PK 163 * fault and that there was a VMA once we got in the fault 164 * handler. It does *not* guarantee that the VMA we find here 165 * was the one that we faulted on. 166 * 167 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4); 168 * 2. T1 : set PKRU to deny access to pkey=4, touches page 169 * 3. T1 : faults... 170 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5); 171 * 5. T1 : enters fault handler, takes mmap_sem, etc... 172 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really 173 * faulted on a pte with its pkey=4. 174 */ 175 static void fill_sig_info_pkey(int si_signo, int si_code, siginfo_t *info, 176 u32 *pkey) 177 { 178 /* This is effectively an #ifdef */ 179 if (!boot_cpu_has(X86_FEATURE_OSPKE)) 180 return; 181 182 /* Fault not from Protection Keys: nothing to do */ 183 if ((si_code != SEGV_PKUERR) || (si_signo != SIGSEGV)) 184 return; 185 /* 186 * force_sig_info_fault() is called from a number of 187 * contexts, some of which have a VMA and some of which 188 * do not. The X86_PF_PK handing happens after we have a 189 * valid VMA, so we should never reach this without a 190 * valid VMA. 191 */ 192 if (!pkey) { 193 WARN_ONCE(1, "PKU fault with no VMA passed in"); 194 info->si_pkey = 0; 195 return; 196 } 197 /* 198 * si_pkey should be thought of as a strong hint, but not 199 * absolutely guranteed to be 100% accurate because of 200 * the race explained above. 201 */ 202 info->si_pkey = *pkey; 203 } 204 205 static void 206 force_sig_info_fault(int si_signo, int si_code, unsigned long address, 207 struct task_struct *tsk, u32 *pkey, int fault) 208 { 209 unsigned lsb = 0; 210 siginfo_t info; 211 212 clear_siginfo(&info); 213 info.si_signo = si_signo; 214 info.si_errno = 0; 215 info.si_code = si_code; 216 info.si_addr = (void __user *)address; 217 if (fault & VM_FAULT_HWPOISON_LARGE) 218 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 219 if (fault & VM_FAULT_HWPOISON) 220 lsb = PAGE_SHIFT; 221 info.si_addr_lsb = lsb; 222 223 fill_sig_info_pkey(si_signo, si_code, &info, pkey); 224 225 force_sig_info(si_signo, &info, tsk); 226 } 227 228 DEFINE_SPINLOCK(pgd_lock); 229 LIST_HEAD(pgd_list); 230 231 #ifdef CONFIG_X86_32 232 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) 233 { 234 unsigned index = pgd_index(address); 235 pgd_t *pgd_k; 236 p4d_t *p4d, *p4d_k; 237 pud_t *pud, *pud_k; 238 pmd_t *pmd, *pmd_k; 239 240 pgd += index; 241 pgd_k = init_mm.pgd + index; 242 243 if (!pgd_present(*pgd_k)) 244 return NULL; 245 246 /* 247 * set_pgd(pgd, *pgd_k); here would be useless on PAE 248 * and redundant with the set_pmd() on non-PAE. As would 249 * set_p4d/set_pud. 250 */ 251 p4d = p4d_offset(pgd, address); 252 p4d_k = p4d_offset(pgd_k, address); 253 if (!p4d_present(*p4d_k)) 254 return NULL; 255 256 pud = pud_offset(p4d, address); 257 pud_k = pud_offset(p4d_k, address); 258 if (!pud_present(*pud_k)) 259 return NULL; 260 261 pmd = pmd_offset(pud, address); 262 pmd_k = pmd_offset(pud_k, address); 263 if (!pmd_present(*pmd_k)) 264 return NULL; 265 266 if (!pmd_present(*pmd)) 267 set_pmd(pmd, *pmd_k); 268 else 269 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k)); 270 271 return pmd_k; 272 } 273 274 void vmalloc_sync_all(void) 275 { 276 unsigned long address; 277 278 if (SHARED_KERNEL_PMD) 279 return; 280 281 for (address = VMALLOC_START & PMD_MASK; 282 address >= TASK_SIZE_MAX && address < FIXADDR_TOP; 283 address += PMD_SIZE) { 284 struct page *page; 285 286 spin_lock(&pgd_lock); 287 list_for_each_entry(page, &pgd_list, lru) { 288 spinlock_t *pgt_lock; 289 pmd_t *ret; 290 291 /* the pgt_lock only for Xen */ 292 pgt_lock = &pgd_page_get_mm(page)->page_table_lock; 293 294 spin_lock(pgt_lock); 295 ret = vmalloc_sync_one(page_address(page), address); 296 spin_unlock(pgt_lock); 297 298 if (!ret) 299 break; 300 } 301 spin_unlock(&pgd_lock); 302 } 303 } 304 305 /* 306 * 32-bit: 307 * 308 * Handle a fault on the vmalloc or module mapping area 309 */ 310 static noinline int vmalloc_fault(unsigned long address) 311 { 312 unsigned long pgd_paddr; 313 pmd_t *pmd_k; 314 pte_t *pte_k; 315 316 /* Make sure we are in vmalloc area: */ 317 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 318 return -1; 319 320 /* 321 * Synchronize this task's top level page-table 322 * with the 'reference' page table. 323 * 324 * Do _not_ use "current" here. We might be inside 325 * an interrupt in the middle of a task switch.. 326 */ 327 pgd_paddr = read_cr3_pa(); 328 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address); 329 if (!pmd_k) 330 return -1; 331 332 if (pmd_large(*pmd_k)) 333 return 0; 334 335 pte_k = pte_offset_kernel(pmd_k, address); 336 if (!pte_present(*pte_k)) 337 return -1; 338 339 return 0; 340 } 341 NOKPROBE_SYMBOL(vmalloc_fault); 342 343 /* 344 * Did it hit the DOS screen memory VA from vm86 mode? 345 */ 346 static inline void 347 check_v8086_mode(struct pt_regs *regs, unsigned long address, 348 struct task_struct *tsk) 349 { 350 #ifdef CONFIG_VM86 351 unsigned long bit; 352 353 if (!v8086_mode(regs) || !tsk->thread.vm86) 354 return; 355 356 bit = (address - 0xA0000) >> PAGE_SHIFT; 357 if (bit < 32) 358 tsk->thread.vm86->screen_bitmap |= 1 << bit; 359 #endif 360 } 361 362 static bool low_pfn(unsigned long pfn) 363 { 364 return pfn < max_low_pfn; 365 } 366 367 static void dump_pagetable(unsigned long address) 368 { 369 pgd_t *base = __va(read_cr3_pa()); 370 pgd_t *pgd = &base[pgd_index(address)]; 371 p4d_t *p4d; 372 pud_t *pud; 373 pmd_t *pmd; 374 pte_t *pte; 375 376 #ifdef CONFIG_X86_PAE 377 pr_info("*pdpt = %016Lx ", pgd_val(*pgd)); 378 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd)) 379 goto out; 380 #define pr_pde pr_cont 381 #else 382 #define pr_pde pr_info 383 #endif 384 p4d = p4d_offset(pgd, address); 385 pud = pud_offset(p4d, address); 386 pmd = pmd_offset(pud, address); 387 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd)); 388 #undef pr_pde 389 390 /* 391 * We must not directly access the pte in the highpte 392 * case if the page table is located in highmem. 393 * And let's rather not kmap-atomic the pte, just in case 394 * it's allocated already: 395 */ 396 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd)) 397 goto out; 398 399 pte = pte_offset_kernel(pmd, address); 400 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte)); 401 out: 402 pr_cont("\n"); 403 } 404 405 #else /* CONFIG_X86_64: */ 406 407 void vmalloc_sync_all(void) 408 { 409 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END); 410 } 411 412 /* 413 * 64-bit: 414 * 415 * Handle a fault on the vmalloc area 416 */ 417 static noinline int vmalloc_fault(unsigned long address) 418 { 419 pgd_t *pgd, *pgd_k; 420 p4d_t *p4d, *p4d_k; 421 pud_t *pud; 422 pmd_t *pmd; 423 pte_t *pte; 424 425 /* Make sure we are in vmalloc area: */ 426 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 427 return -1; 428 429 WARN_ON_ONCE(in_nmi()); 430 431 /* 432 * Copy kernel mappings over when needed. This can also 433 * happen within a race in page table update. In the later 434 * case just flush: 435 */ 436 pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address); 437 pgd_k = pgd_offset_k(address); 438 if (pgd_none(*pgd_k)) 439 return -1; 440 441 if (pgtable_l5_enabled()) { 442 if (pgd_none(*pgd)) { 443 set_pgd(pgd, *pgd_k); 444 arch_flush_lazy_mmu_mode(); 445 } else { 446 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_k)); 447 } 448 } 449 450 /* With 4-level paging, copying happens on the p4d level. */ 451 p4d = p4d_offset(pgd, address); 452 p4d_k = p4d_offset(pgd_k, address); 453 if (p4d_none(*p4d_k)) 454 return -1; 455 456 if (p4d_none(*p4d) && !pgtable_l5_enabled()) { 457 set_p4d(p4d, *p4d_k); 458 arch_flush_lazy_mmu_mode(); 459 } else { 460 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_k)); 461 } 462 463 BUILD_BUG_ON(CONFIG_PGTABLE_LEVELS < 4); 464 465 pud = pud_offset(p4d, address); 466 if (pud_none(*pud)) 467 return -1; 468 469 if (pud_large(*pud)) 470 return 0; 471 472 pmd = pmd_offset(pud, address); 473 if (pmd_none(*pmd)) 474 return -1; 475 476 if (pmd_large(*pmd)) 477 return 0; 478 479 pte = pte_offset_kernel(pmd, address); 480 if (!pte_present(*pte)) 481 return -1; 482 483 return 0; 484 } 485 NOKPROBE_SYMBOL(vmalloc_fault); 486 487 #ifdef CONFIG_CPU_SUP_AMD 488 static const char errata93_warning[] = 489 KERN_ERR 490 "******* Your BIOS seems to not contain a fix for K8 errata #93\n" 491 "******* Working around it, but it may cause SEGVs or burn power.\n" 492 "******* Please consider a BIOS update.\n" 493 "******* Disabling USB legacy in the BIOS may also help.\n"; 494 #endif 495 496 /* 497 * No vm86 mode in 64-bit mode: 498 */ 499 static inline void 500 check_v8086_mode(struct pt_regs *regs, unsigned long address, 501 struct task_struct *tsk) 502 { 503 } 504 505 static int bad_address(void *p) 506 { 507 unsigned long dummy; 508 509 return probe_kernel_address((unsigned long *)p, dummy); 510 } 511 512 static void dump_pagetable(unsigned long address) 513 { 514 pgd_t *base = __va(read_cr3_pa()); 515 pgd_t *pgd = base + pgd_index(address); 516 p4d_t *p4d; 517 pud_t *pud; 518 pmd_t *pmd; 519 pte_t *pte; 520 521 if (bad_address(pgd)) 522 goto bad; 523 524 pr_info("PGD %lx ", pgd_val(*pgd)); 525 526 if (!pgd_present(*pgd)) 527 goto out; 528 529 p4d = p4d_offset(pgd, address); 530 if (bad_address(p4d)) 531 goto bad; 532 533 pr_cont("P4D %lx ", p4d_val(*p4d)); 534 if (!p4d_present(*p4d) || p4d_large(*p4d)) 535 goto out; 536 537 pud = pud_offset(p4d, address); 538 if (bad_address(pud)) 539 goto bad; 540 541 pr_cont("PUD %lx ", pud_val(*pud)); 542 if (!pud_present(*pud) || pud_large(*pud)) 543 goto out; 544 545 pmd = pmd_offset(pud, address); 546 if (bad_address(pmd)) 547 goto bad; 548 549 pr_cont("PMD %lx ", pmd_val(*pmd)); 550 if (!pmd_present(*pmd) || pmd_large(*pmd)) 551 goto out; 552 553 pte = pte_offset_kernel(pmd, address); 554 if (bad_address(pte)) 555 goto bad; 556 557 pr_cont("PTE %lx", pte_val(*pte)); 558 out: 559 pr_cont("\n"); 560 return; 561 bad: 562 pr_info("BAD\n"); 563 } 564 565 #endif /* CONFIG_X86_64 */ 566 567 /* 568 * Workaround for K8 erratum #93 & buggy BIOS. 569 * 570 * BIOS SMM functions are required to use a specific workaround 571 * to avoid corruption of the 64bit RIP register on C stepping K8. 572 * 573 * A lot of BIOS that didn't get tested properly miss this. 574 * 575 * The OS sees this as a page fault with the upper 32bits of RIP cleared. 576 * Try to work around it here. 577 * 578 * Note we only handle faults in kernel here. 579 * Does nothing on 32-bit. 580 */ 581 static int is_errata93(struct pt_regs *regs, unsigned long address) 582 { 583 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD) 584 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD 585 || boot_cpu_data.x86 != 0xf) 586 return 0; 587 588 if (address != regs->ip) 589 return 0; 590 591 if ((address >> 32) != 0) 592 return 0; 593 594 address |= 0xffffffffUL << 32; 595 if ((address >= (u64)_stext && address <= (u64)_etext) || 596 (address >= MODULES_VADDR && address <= MODULES_END)) { 597 printk_once(errata93_warning); 598 regs->ip = address; 599 return 1; 600 } 601 #endif 602 return 0; 603 } 604 605 /* 606 * Work around K8 erratum #100 K8 in compat mode occasionally jumps 607 * to illegal addresses >4GB. 608 * 609 * We catch this in the page fault handler because these addresses 610 * are not reachable. Just detect this case and return. Any code 611 * segment in LDT is compatibility mode. 612 */ 613 static int is_errata100(struct pt_regs *regs, unsigned long address) 614 { 615 #ifdef CONFIG_X86_64 616 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32)) 617 return 1; 618 #endif 619 return 0; 620 } 621 622 static int is_f00f_bug(struct pt_regs *regs, unsigned long address) 623 { 624 #ifdef CONFIG_X86_F00F_BUG 625 unsigned long nr; 626 627 /* 628 * Pentium F0 0F C7 C8 bug workaround: 629 */ 630 if (boot_cpu_has_bug(X86_BUG_F00F)) { 631 nr = (address - idt_descr.address) >> 3; 632 633 if (nr == 6) { 634 do_invalid_op(regs, 0); 635 return 1; 636 } 637 } 638 #endif 639 return 0; 640 } 641 642 static void 643 show_fault_oops(struct pt_regs *regs, unsigned long error_code, 644 unsigned long address) 645 { 646 if (!oops_may_print()) 647 return; 648 649 if (error_code & X86_PF_INSTR) { 650 unsigned int level; 651 pgd_t *pgd; 652 pte_t *pte; 653 654 pgd = __va(read_cr3_pa()); 655 pgd += pgd_index(address); 656 657 pte = lookup_address_in_pgd(pgd, address, &level); 658 659 if (pte && pte_present(*pte) && !pte_exec(*pte)) 660 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", 661 from_kuid(&init_user_ns, current_uid())); 662 if (pte && pte_present(*pte) && pte_exec(*pte) && 663 (pgd_flags(*pgd) & _PAGE_USER) && 664 (__read_cr4() & X86_CR4_SMEP)) 665 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n", 666 from_kuid(&init_user_ns, current_uid())); 667 } 668 669 pr_alert("BUG: unable to handle kernel %s at %px\n", 670 address < PAGE_SIZE ? "NULL pointer dereference" : "paging request", 671 (void *)address); 672 673 dump_pagetable(address); 674 } 675 676 static noinline void 677 pgtable_bad(struct pt_regs *regs, unsigned long error_code, 678 unsigned long address) 679 { 680 struct task_struct *tsk; 681 unsigned long flags; 682 int sig; 683 684 flags = oops_begin(); 685 tsk = current; 686 sig = SIGKILL; 687 688 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n", 689 tsk->comm, address); 690 dump_pagetable(address); 691 692 tsk->thread.cr2 = address; 693 tsk->thread.trap_nr = X86_TRAP_PF; 694 tsk->thread.error_code = error_code; 695 696 if (__die("Bad pagetable", regs, error_code)) 697 sig = 0; 698 699 oops_end(flags, regs, sig); 700 } 701 702 static noinline void 703 no_context(struct pt_regs *regs, unsigned long error_code, 704 unsigned long address, int signal, int si_code) 705 { 706 struct task_struct *tsk = current; 707 unsigned long flags; 708 int sig; 709 710 /* Are we prepared to handle this kernel fault? */ 711 if (fixup_exception(regs, X86_TRAP_PF)) { 712 /* 713 * Any interrupt that takes a fault gets the fixup. This makes 714 * the below recursive fault logic only apply to a faults from 715 * task context. 716 */ 717 if (in_interrupt()) 718 return; 719 720 /* 721 * Per the above we're !in_interrupt(), aka. task context. 722 * 723 * In this case we need to make sure we're not recursively 724 * faulting through the emulate_vsyscall() logic. 725 */ 726 if (current->thread.sig_on_uaccess_err && signal) { 727 tsk->thread.trap_nr = X86_TRAP_PF; 728 tsk->thread.error_code = error_code | X86_PF_USER; 729 tsk->thread.cr2 = address; 730 731 /* XXX: hwpoison faults will set the wrong code. */ 732 force_sig_info_fault(signal, si_code, address, 733 tsk, NULL, 0); 734 } 735 736 /* 737 * Barring that, we can do the fixup and be happy. 738 */ 739 return; 740 } 741 742 #ifdef CONFIG_VMAP_STACK 743 /* 744 * Stack overflow? During boot, we can fault near the initial 745 * stack in the direct map, but that's not an overflow -- check 746 * that we're in vmalloc space to avoid this. 747 */ 748 if (is_vmalloc_addr((void *)address) && 749 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) || 750 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) { 751 unsigned long stack = this_cpu_read(orig_ist.ist[DOUBLEFAULT_STACK]) - sizeof(void *); 752 /* 753 * We're likely to be running with very little stack space 754 * left. It's plausible that we'd hit this condition but 755 * double-fault even before we get this far, in which case 756 * we're fine: the double-fault handler will deal with it. 757 * 758 * We don't want to make it all the way into the oops code 759 * and then double-fault, though, because we're likely to 760 * break the console driver and lose most of the stack dump. 761 */ 762 asm volatile ("movq %[stack], %%rsp\n\t" 763 "call handle_stack_overflow\n\t" 764 "1: jmp 1b" 765 : ASM_CALL_CONSTRAINT 766 : "D" ("kernel stack overflow (page fault)"), 767 "S" (regs), "d" (address), 768 [stack] "rm" (stack)); 769 unreachable(); 770 } 771 #endif 772 773 /* 774 * 32-bit: 775 * 776 * Valid to do another page fault here, because if this fault 777 * had been triggered by is_prefetch fixup_exception would have 778 * handled it. 779 * 780 * 64-bit: 781 * 782 * Hall of shame of CPU/BIOS bugs. 783 */ 784 if (is_prefetch(regs, error_code, address)) 785 return; 786 787 if (is_errata93(regs, address)) 788 return; 789 790 /* 791 * Oops. The kernel tried to access some bad page. We'll have to 792 * terminate things with extreme prejudice: 793 */ 794 flags = oops_begin(); 795 796 show_fault_oops(regs, error_code, address); 797 798 if (task_stack_end_corrupted(tsk)) 799 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 800 801 tsk->thread.cr2 = address; 802 tsk->thread.trap_nr = X86_TRAP_PF; 803 tsk->thread.error_code = error_code; 804 805 sig = SIGKILL; 806 if (__die("Oops", regs, error_code)) 807 sig = 0; 808 809 /* Executive summary in case the body of the oops scrolled away */ 810 printk(KERN_DEFAULT "CR2: %016lx\n", address); 811 812 oops_end(flags, regs, sig); 813 } 814 815 /* 816 * Print out info about fatal segfaults, if the show_unhandled_signals 817 * sysctl is set: 818 */ 819 static inline void 820 show_signal_msg(struct pt_regs *regs, unsigned long error_code, 821 unsigned long address, struct task_struct *tsk) 822 { 823 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG; 824 825 if (!unhandled_signal(tsk, SIGSEGV)) 826 return; 827 828 if (!printk_ratelimit()) 829 return; 830 831 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx", 832 loglvl, tsk->comm, task_pid_nr(tsk), address, 833 (void *)regs->ip, (void *)regs->sp, error_code); 834 835 print_vma_addr(KERN_CONT " in ", regs->ip); 836 837 printk(KERN_CONT "\n"); 838 839 show_opcodes((u8 *)regs->ip, loglvl); 840 } 841 842 static void 843 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 844 unsigned long address, u32 *pkey, int si_code) 845 { 846 struct task_struct *tsk = current; 847 848 /* User mode accesses just cause a SIGSEGV */ 849 if (error_code & X86_PF_USER) { 850 /* 851 * It's possible to have interrupts off here: 852 */ 853 local_irq_enable(); 854 855 /* 856 * Valid to do another page fault here because this one came 857 * from user space: 858 */ 859 if (is_prefetch(regs, error_code, address)) 860 return; 861 862 if (is_errata100(regs, address)) 863 return; 864 865 #ifdef CONFIG_X86_64 866 /* 867 * Instruction fetch faults in the vsyscall page might need 868 * emulation. 869 */ 870 if (unlikely((error_code & X86_PF_INSTR) && 871 ((address & ~0xfff) == VSYSCALL_ADDR))) { 872 if (emulate_vsyscall(regs, address)) 873 return; 874 } 875 #endif 876 877 /* 878 * To avoid leaking information about the kernel page table 879 * layout, pretend that user-mode accesses to kernel addresses 880 * are always protection faults. 881 */ 882 if (address >= TASK_SIZE_MAX) 883 error_code |= X86_PF_PROT; 884 885 if (likely(show_unhandled_signals)) 886 show_signal_msg(regs, error_code, address, tsk); 887 888 tsk->thread.cr2 = address; 889 tsk->thread.error_code = error_code; 890 tsk->thread.trap_nr = X86_TRAP_PF; 891 892 force_sig_info_fault(SIGSEGV, si_code, address, tsk, pkey, 0); 893 894 return; 895 } 896 897 if (is_f00f_bug(regs, address)) 898 return; 899 900 no_context(regs, error_code, address, SIGSEGV, si_code); 901 } 902 903 static noinline void 904 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 905 unsigned long address, u32 *pkey) 906 { 907 __bad_area_nosemaphore(regs, error_code, address, pkey, SEGV_MAPERR); 908 } 909 910 static void 911 __bad_area(struct pt_regs *regs, unsigned long error_code, 912 unsigned long address, struct vm_area_struct *vma, int si_code) 913 { 914 struct mm_struct *mm = current->mm; 915 u32 pkey; 916 917 if (vma) 918 pkey = vma_pkey(vma); 919 920 /* 921 * Something tried to access memory that isn't in our memory map.. 922 * Fix it, but check if it's kernel or user first.. 923 */ 924 up_read(&mm->mmap_sem); 925 926 __bad_area_nosemaphore(regs, error_code, address, 927 (vma) ? &pkey : NULL, si_code); 928 } 929 930 static noinline void 931 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address) 932 { 933 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR); 934 } 935 936 static inline bool bad_area_access_from_pkeys(unsigned long error_code, 937 struct vm_area_struct *vma) 938 { 939 /* This code is always called on the current mm */ 940 bool foreign = false; 941 942 if (!boot_cpu_has(X86_FEATURE_OSPKE)) 943 return false; 944 if (error_code & X86_PF_PK) 945 return true; 946 /* this checks permission keys on the VMA: */ 947 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 948 (error_code & X86_PF_INSTR), foreign)) 949 return true; 950 return false; 951 } 952 953 static noinline void 954 bad_area_access_error(struct pt_regs *regs, unsigned long error_code, 955 unsigned long address, struct vm_area_struct *vma) 956 { 957 /* 958 * This OSPKE check is not strictly necessary at runtime. 959 * But, doing it this way allows compiler optimizations 960 * if pkeys are compiled out. 961 */ 962 if (bad_area_access_from_pkeys(error_code, vma)) 963 __bad_area(regs, error_code, address, vma, SEGV_PKUERR); 964 else 965 __bad_area(regs, error_code, address, vma, SEGV_ACCERR); 966 } 967 968 static void 969 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address, 970 u32 *pkey, unsigned int fault) 971 { 972 struct task_struct *tsk = current; 973 int code = BUS_ADRERR; 974 975 /* Kernel mode? Handle exceptions or die: */ 976 if (!(error_code & X86_PF_USER)) { 977 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR); 978 return; 979 } 980 981 /* User-space => ok to do another page fault: */ 982 if (is_prefetch(regs, error_code, address)) 983 return; 984 985 tsk->thread.cr2 = address; 986 tsk->thread.error_code = error_code; 987 tsk->thread.trap_nr = X86_TRAP_PF; 988 989 #ifdef CONFIG_MEMORY_FAILURE 990 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) { 991 printk(KERN_ERR 992 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n", 993 tsk->comm, tsk->pid, address); 994 code = BUS_MCEERR_AR; 995 } 996 #endif 997 force_sig_info_fault(SIGBUS, code, address, tsk, pkey, fault); 998 } 999 1000 static noinline void 1001 mm_fault_error(struct pt_regs *regs, unsigned long error_code, 1002 unsigned long address, u32 *pkey, unsigned int fault) 1003 { 1004 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) { 1005 no_context(regs, error_code, address, 0, 0); 1006 return; 1007 } 1008 1009 if (fault & VM_FAULT_OOM) { 1010 /* Kernel mode? Handle exceptions or die: */ 1011 if (!(error_code & X86_PF_USER)) { 1012 no_context(regs, error_code, address, 1013 SIGSEGV, SEGV_MAPERR); 1014 return; 1015 } 1016 1017 /* 1018 * We ran out of memory, call the OOM killer, and return the 1019 * userspace (which will retry the fault, or kill us if we got 1020 * oom-killed): 1021 */ 1022 pagefault_out_of_memory(); 1023 } else { 1024 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON| 1025 VM_FAULT_HWPOISON_LARGE)) 1026 do_sigbus(regs, error_code, address, pkey, fault); 1027 else if (fault & VM_FAULT_SIGSEGV) 1028 bad_area_nosemaphore(regs, error_code, address, pkey); 1029 else 1030 BUG(); 1031 } 1032 } 1033 1034 static int spurious_fault_check(unsigned long error_code, pte_t *pte) 1035 { 1036 if ((error_code & X86_PF_WRITE) && !pte_write(*pte)) 1037 return 0; 1038 1039 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte)) 1040 return 0; 1041 /* 1042 * Note: We do not do lazy flushing on protection key 1043 * changes, so no spurious fault will ever set X86_PF_PK. 1044 */ 1045 if ((error_code & X86_PF_PK)) 1046 return 1; 1047 1048 return 1; 1049 } 1050 1051 /* 1052 * Handle a spurious fault caused by a stale TLB entry. 1053 * 1054 * This allows us to lazily refresh the TLB when increasing the 1055 * permissions of a kernel page (RO -> RW or NX -> X). Doing it 1056 * eagerly is very expensive since that implies doing a full 1057 * cross-processor TLB flush, even if no stale TLB entries exist 1058 * on other processors. 1059 * 1060 * Spurious faults may only occur if the TLB contains an entry with 1061 * fewer permission than the page table entry. Non-present (P = 0) 1062 * and reserved bit (R = 1) faults are never spurious. 1063 * 1064 * There are no security implications to leaving a stale TLB when 1065 * increasing the permissions on a page. 1066 * 1067 * Returns non-zero if a spurious fault was handled, zero otherwise. 1068 * 1069 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 1070 * (Optional Invalidation). 1071 */ 1072 static noinline int 1073 spurious_fault(unsigned long error_code, unsigned long address) 1074 { 1075 pgd_t *pgd; 1076 p4d_t *p4d; 1077 pud_t *pud; 1078 pmd_t *pmd; 1079 pte_t *pte; 1080 int ret; 1081 1082 /* 1083 * Only writes to RO or instruction fetches from NX may cause 1084 * spurious faults. 1085 * 1086 * These could be from user or supervisor accesses but the TLB 1087 * is only lazily flushed after a kernel mapping protection 1088 * change, so user accesses are not expected to cause spurious 1089 * faults. 1090 */ 1091 if (error_code != (X86_PF_WRITE | X86_PF_PROT) && 1092 error_code != (X86_PF_INSTR | X86_PF_PROT)) 1093 return 0; 1094 1095 pgd = init_mm.pgd + pgd_index(address); 1096 if (!pgd_present(*pgd)) 1097 return 0; 1098 1099 p4d = p4d_offset(pgd, address); 1100 if (!p4d_present(*p4d)) 1101 return 0; 1102 1103 if (p4d_large(*p4d)) 1104 return spurious_fault_check(error_code, (pte_t *) p4d); 1105 1106 pud = pud_offset(p4d, address); 1107 if (!pud_present(*pud)) 1108 return 0; 1109 1110 if (pud_large(*pud)) 1111 return spurious_fault_check(error_code, (pte_t *) pud); 1112 1113 pmd = pmd_offset(pud, address); 1114 if (!pmd_present(*pmd)) 1115 return 0; 1116 1117 if (pmd_large(*pmd)) 1118 return spurious_fault_check(error_code, (pte_t *) pmd); 1119 1120 pte = pte_offset_kernel(pmd, address); 1121 if (!pte_present(*pte)) 1122 return 0; 1123 1124 ret = spurious_fault_check(error_code, pte); 1125 if (!ret) 1126 return 0; 1127 1128 /* 1129 * Make sure we have permissions in PMD. 1130 * If not, then there's a bug in the page tables: 1131 */ 1132 ret = spurious_fault_check(error_code, (pte_t *) pmd); 1133 WARN_ONCE(!ret, "PMD has incorrect permission bits\n"); 1134 1135 return ret; 1136 } 1137 NOKPROBE_SYMBOL(spurious_fault); 1138 1139 int show_unhandled_signals = 1; 1140 1141 static inline int 1142 access_error(unsigned long error_code, struct vm_area_struct *vma) 1143 { 1144 /* This is only called for the current mm, so: */ 1145 bool foreign = false; 1146 1147 /* 1148 * Read or write was blocked by protection keys. This is 1149 * always an unconditional error and can never result in 1150 * a follow-up action to resolve the fault, like a COW. 1151 */ 1152 if (error_code & X86_PF_PK) 1153 return 1; 1154 1155 /* 1156 * Make sure to check the VMA so that we do not perform 1157 * faults just to hit a X86_PF_PK as soon as we fill in a 1158 * page. 1159 */ 1160 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 1161 (error_code & X86_PF_INSTR), foreign)) 1162 return 1; 1163 1164 if (error_code & X86_PF_WRITE) { 1165 /* write, present and write, not present: */ 1166 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1167 return 1; 1168 return 0; 1169 } 1170 1171 /* read, present: */ 1172 if (unlikely(error_code & X86_PF_PROT)) 1173 return 1; 1174 1175 /* read, not present: */ 1176 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))) 1177 return 1; 1178 1179 return 0; 1180 } 1181 1182 static int fault_in_kernel_space(unsigned long address) 1183 { 1184 return address >= TASK_SIZE_MAX; 1185 } 1186 1187 static inline bool smap_violation(int error_code, struct pt_regs *regs) 1188 { 1189 if (!IS_ENABLED(CONFIG_X86_SMAP)) 1190 return false; 1191 1192 if (!static_cpu_has(X86_FEATURE_SMAP)) 1193 return false; 1194 1195 if (error_code & X86_PF_USER) 1196 return false; 1197 1198 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC)) 1199 return false; 1200 1201 return true; 1202 } 1203 1204 /* 1205 * This routine handles page faults. It determines the address, 1206 * and the problem, and then passes it off to one of the appropriate 1207 * routines. 1208 */ 1209 static noinline void 1210 __do_page_fault(struct pt_regs *regs, unsigned long error_code, 1211 unsigned long address) 1212 { 1213 struct vm_area_struct *vma; 1214 struct task_struct *tsk; 1215 struct mm_struct *mm; 1216 int fault, major = 0; 1217 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; 1218 u32 pkey; 1219 1220 tsk = current; 1221 mm = tsk->mm; 1222 1223 prefetchw(&mm->mmap_sem); 1224 1225 if (unlikely(kmmio_fault(regs, address))) 1226 return; 1227 1228 /* 1229 * We fault-in kernel-space virtual memory on-demand. The 1230 * 'reference' page table is init_mm.pgd. 1231 * 1232 * NOTE! We MUST NOT take any locks for this case. We may 1233 * be in an interrupt or a critical region, and should 1234 * only copy the information from the master page table, 1235 * nothing more. 1236 * 1237 * This verifies that the fault happens in kernel space 1238 * (error_code & 4) == 0, and that the fault was not a 1239 * protection error (error_code & 9) == 0. 1240 */ 1241 if (unlikely(fault_in_kernel_space(address))) { 1242 if (!(error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) { 1243 if (vmalloc_fault(address) >= 0) 1244 return; 1245 } 1246 1247 /* Can handle a stale RO->RW TLB: */ 1248 if (spurious_fault(error_code, address)) 1249 return; 1250 1251 /* kprobes don't want to hook the spurious faults: */ 1252 if (kprobes_fault(regs)) 1253 return; 1254 /* 1255 * Don't take the mm semaphore here. If we fixup a prefetch 1256 * fault we could otherwise deadlock: 1257 */ 1258 bad_area_nosemaphore(regs, error_code, address, NULL); 1259 1260 return; 1261 } 1262 1263 /* kprobes don't want to hook the spurious faults: */ 1264 if (unlikely(kprobes_fault(regs))) 1265 return; 1266 1267 if (unlikely(error_code & X86_PF_RSVD)) 1268 pgtable_bad(regs, error_code, address); 1269 1270 if (unlikely(smap_violation(error_code, regs))) { 1271 bad_area_nosemaphore(regs, error_code, address, NULL); 1272 return; 1273 } 1274 1275 /* 1276 * If we're in an interrupt, have no user context or are running 1277 * in a region with pagefaults disabled then we must not take the fault 1278 */ 1279 if (unlikely(faulthandler_disabled() || !mm)) { 1280 bad_area_nosemaphore(regs, error_code, address, NULL); 1281 return; 1282 } 1283 1284 /* 1285 * It's safe to allow irq's after cr2 has been saved and the 1286 * vmalloc fault has been handled. 1287 * 1288 * User-mode registers count as a user access even for any 1289 * potential system fault or CPU buglet: 1290 */ 1291 if (user_mode(regs)) { 1292 local_irq_enable(); 1293 error_code |= X86_PF_USER; 1294 flags |= FAULT_FLAG_USER; 1295 } else { 1296 if (regs->flags & X86_EFLAGS_IF) 1297 local_irq_enable(); 1298 } 1299 1300 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); 1301 1302 if (error_code & X86_PF_WRITE) 1303 flags |= FAULT_FLAG_WRITE; 1304 if (error_code & X86_PF_INSTR) 1305 flags |= FAULT_FLAG_INSTRUCTION; 1306 1307 /* 1308 * When running in the kernel we expect faults to occur only to 1309 * addresses in user space. All other faults represent errors in 1310 * the kernel and should generate an OOPS. Unfortunately, in the 1311 * case of an erroneous fault occurring in a code path which already 1312 * holds mmap_sem we will deadlock attempting to validate the fault 1313 * against the address space. Luckily the kernel only validly 1314 * references user space from well defined areas of code, which are 1315 * listed in the exceptions table. 1316 * 1317 * As the vast majority of faults will be valid we will only perform 1318 * the source reference check when there is a possibility of a 1319 * deadlock. Attempt to lock the address space, if we cannot we then 1320 * validate the source. If this is invalid we can skip the address 1321 * space check, thus avoiding the deadlock: 1322 */ 1323 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 1324 if (!(error_code & X86_PF_USER) && 1325 !search_exception_tables(regs->ip)) { 1326 bad_area_nosemaphore(regs, error_code, address, NULL); 1327 return; 1328 } 1329 retry: 1330 down_read(&mm->mmap_sem); 1331 } else { 1332 /* 1333 * The above down_read_trylock() might have succeeded in 1334 * which case we'll have missed the might_sleep() from 1335 * down_read(): 1336 */ 1337 might_sleep(); 1338 } 1339 1340 vma = find_vma(mm, address); 1341 if (unlikely(!vma)) { 1342 bad_area(regs, error_code, address); 1343 return; 1344 } 1345 if (likely(vma->vm_start <= address)) 1346 goto good_area; 1347 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) { 1348 bad_area(regs, error_code, address); 1349 return; 1350 } 1351 if (error_code & X86_PF_USER) { 1352 /* 1353 * Accessing the stack below %sp is always a bug. 1354 * The large cushion allows instructions like enter 1355 * and pusha to work. ("enter $65535, $31" pushes 1356 * 32 pointers and then decrements %sp by 65535.) 1357 */ 1358 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) { 1359 bad_area(regs, error_code, address); 1360 return; 1361 } 1362 } 1363 if (unlikely(expand_stack(vma, address))) { 1364 bad_area(regs, error_code, address); 1365 return; 1366 } 1367 1368 /* 1369 * Ok, we have a good vm_area for this memory access, so 1370 * we can handle it.. 1371 */ 1372 good_area: 1373 if (unlikely(access_error(error_code, vma))) { 1374 bad_area_access_error(regs, error_code, address, vma); 1375 return; 1376 } 1377 1378 /* 1379 * If for any reason at all we couldn't handle the fault, 1380 * make sure we exit gracefully rather than endlessly redo 1381 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if 1382 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked. 1383 * 1384 * Note that handle_userfault() may also release and reacquire mmap_sem 1385 * (and not return with VM_FAULT_RETRY), when returning to userland to 1386 * repeat the page fault later with a VM_FAULT_NOPAGE retval 1387 * (potentially after handling any pending signal during the return to 1388 * userland). The return to userland is identified whenever 1389 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags. 1390 * Thus we have to be careful about not touching vma after handling the 1391 * fault, so we read the pkey beforehand. 1392 */ 1393 pkey = vma_pkey(vma); 1394 fault = handle_mm_fault(vma, address, flags); 1395 major |= fault & VM_FAULT_MAJOR; 1396 1397 /* 1398 * If we need to retry the mmap_sem has already been released, 1399 * and if there is a fatal signal pending there is no guarantee 1400 * that we made any progress. Handle this case first. 1401 */ 1402 if (unlikely(fault & VM_FAULT_RETRY)) { 1403 /* Retry at most once */ 1404 if (flags & FAULT_FLAG_ALLOW_RETRY) { 1405 flags &= ~FAULT_FLAG_ALLOW_RETRY; 1406 flags |= FAULT_FLAG_TRIED; 1407 if (!fatal_signal_pending(tsk)) 1408 goto retry; 1409 } 1410 1411 /* User mode? Just return to handle the fatal exception */ 1412 if (flags & FAULT_FLAG_USER) 1413 return; 1414 1415 /* Not returning to user mode? Handle exceptions or die: */ 1416 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR); 1417 return; 1418 } 1419 1420 up_read(&mm->mmap_sem); 1421 if (unlikely(fault & VM_FAULT_ERROR)) { 1422 mm_fault_error(regs, error_code, address, &pkey, fault); 1423 return; 1424 } 1425 1426 /* 1427 * Major/minor page fault accounting. If any of the events 1428 * returned VM_FAULT_MAJOR, we account it as a major fault. 1429 */ 1430 if (major) { 1431 tsk->maj_flt++; 1432 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 1433 } else { 1434 tsk->min_flt++; 1435 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 1436 } 1437 1438 check_v8086_mode(regs, address, tsk); 1439 } 1440 NOKPROBE_SYMBOL(__do_page_fault); 1441 1442 static nokprobe_inline void 1443 trace_page_fault_entries(unsigned long address, struct pt_regs *regs, 1444 unsigned long error_code) 1445 { 1446 if (user_mode(regs)) 1447 trace_page_fault_user(address, regs, error_code); 1448 else 1449 trace_page_fault_kernel(address, regs, error_code); 1450 } 1451 1452 /* 1453 * We must have this function blacklisted from kprobes, tagged with notrace 1454 * and call read_cr2() before calling anything else. To avoid calling any 1455 * kind of tracing machinery before we've observed the CR2 value. 1456 * 1457 * exception_{enter,exit}() contains all sorts of tracepoints. 1458 */ 1459 dotraplinkage void notrace 1460 do_page_fault(struct pt_regs *regs, unsigned long error_code) 1461 { 1462 unsigned long address = read_cr2(); /* Get the faulting address */ 1463 enum ctx_state prev_state; 1464 1465 prev_state = exception_enter(); 1466 if (trace_pagefault_enabled()) 1467 trace_page_fault_entries(address, regs, error_code); 1468 1469 __do_page_fault(regs, error_code, address); 1470 exception_exit(prev_state); 1471 } 1472 NOKPROBE_SYMBOL(do_page_fault); 1473