1 /* 2 * Copyright (C) 1995 Linus Torvalds 3 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs. 4 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar 5 */ 6 #include <linux/sched.h> /* test_thread_flag(), ... */ 7 #include <linux/kdebug.h> /* oops_begin/end, ... */ 8 #include <linux/module.h> /* search_exception_table */ 9 #include <linux/bootmem.h> /* max_low_pfn */ 10 #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */ 11 #include <linux/mmiotrace.h> /* kmmio_handler, ... */ 12 #include <linux/perf_event.h> /* perf_sw_event */ 13 #include <linux/hugetlb.h> /* hstate_index_to_shift */ 14 #include <linux/prefetch.h> /* prefetchw */ 15 #include <linux/context_tracking.h> /* exception_enter(), ... */ 16 #include <linux/uaccess.h> /* faulthandler_disabled() */ 17 18 #include <asm/traps.h> /* dotraplinkage, ... */ 19 #include <asm/pgalloc.h> /* pgd_*(), ... */ 20 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */ 21 #include <asm/fixmap.h> /* VSYSCALL_ADDR */ 22 #include <asm/vsyscall.h> /* emulate_vsyscall */ 23 #include <asm/vm86.h> /* struct vm86 */ 24 25 #define CREATE_TRACE_POINTS 26 #include <asm/trace/exceptions.h> 27 28 /* 29 * Page fault error code bits: 30 * 31 * bit 0 == 0: no page found 1: protection fault 32 * bit 1 == 0: read access 1: write access 33 * bit 2 == 0: kernel-mode access 1: user-mode access 34 * bit 3 == 1: use of reserved bit detected 35 * bit 4 == 1: fault was an instruction fetch 36 * bit 5 == 1: protection keys block access 37 */ 38 enum x86_pf_error_code { 39 40 PF_PROT = 1 << 0, 41 PF_WRITE = 1 << 1, 42 PF_USER = 1 << 2, 43 PF_RSVD = 1 << 3, 44 PF_INSTR = 1 << 4, 45 PF_PK = 1 << 5, 46 }; 47 48 /* 49 * Returns 0 if mmiotrace is disabled, or if the fault is not 50 * handled by mmiotrace: 51 */ 52 static nokprobe_inline int 53 kmmio_fault(struct pt_regs *regs, unsigned long addr) 54 { 55 if (unlikely(is_kmmio_active())) 56 if (kmmio_handler(regs, addr) == 1) 57 return -1; 58 return 0; 59 } 60 61 static nokprobe_inline int kprobes_fault(struct pt_regs *regs) 62 { 63 int ret = 0; 64 65 /* kprobe_running() needs smp_processor_id() */ 66 if (kprobes_built_in() && !user_mode(regs)) { 67 preempt_disable(); 68 if (kprobe_running() && kprobe_fault_handler(regs, 14)) 69 ret = 1; 70 preempt_enable(); 71 } 72 73 return ret; 74 } 75 76 /* 77 * Prefetch quirks: 78 * 79 * 32-bit mode: 80 * 81 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch. 82 * Check that here and ignore it. 83 * 84 * 64-bit mode: 85 * 86 * Sometimes the CPU reports invalid exceptions on prefetch. 87 * Check that here and ignore it. 88 * 89 * Opcode checker based on code by Richard Brunner. 90 */ 91 static inline int 92 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr, 93 unsigned char opcode, int *prefetch) 94 { 95 unsigned char instr_hi = opcode & 0xf0; 96 unsigned char instr_lo = opcode & 0x0f; 97 98 switch (instr_hi) { 99 case 0x20: 100 case 0x30: 101 /* 102 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. 103 * In X86_64 long mode, the CPU will signal invalid 104 * opcode if some of these prefixes are present so 105 * X86_64 will never get here anyway 106 */ 107 return ((instr_lo & 7) == 0x6); 108 #ifdef CONFIG_X86_64 109 case 0x40: 110 /* 111 * In AMD64 long mode 0x40..0x4F are valid REX prefixes 112 * Need to figure out under what instruction mode the 113 * instruction was issued. Could check the LDT for lm, 114 * but for now it's good enough to assume that long 115 * mode only uses well known segments or kernel. 116 */ 117 return (!user_mode(regs) || user_64bit_mode(regs)); 118 #endif 119 case 0x60: 120 /* 0x64 thru 0x67 are valid prefixes in all modes. */ 121 return (instr_lo & 0xC) == 0x4; 122 case 0xF0: 123 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */ 124 return !instr_lo || (instr_lo>>1) == 1; 125 case 0x00: 126 /* Prefetch instruction is 0x0F0D or 0x0F18 */ 127 if (probe_kernel_address(instr, opcode)) 128 return 0; 129 130 *prefetch = (instr_lo == 0xF) && 131 (opcode == 0x0D || opcode == 0x18); 132 return 0; 133 default: 134 return 0; 135 } 136 } 137 138 static int 139 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr) 140 { 141 unsigned char *max_instr; 142 unsigned char *instr; 143 int prefetch = 0; 144 145 /* 146 * If it was a exec (instruction fetch) fault on NX page, then 147 * do not ignore the fault: 148 */ 149 if (error_code & PF_INSTR) 150 return 0; 151 152 instr = (void *)convert_ip_to_linear(current, regs); 153 max_instr = instr + 15; 154 155 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX) 156 return 0; 157 158 while (instr < max_instr) { 159 unsigned char opcode; 160 161 if (probe_kernel_address(instr, opcode)) 162 break; 163 164 instr++; 165 166 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch)) 167 break; 168 } 169 return prefetch; 170 } 171 172 static void 173 force_sig_info_fault(int si_signo, int si_code, unsigned long address, 174 struct task_struct *tsk, struct vm_area_struct *vma, 175 int fault) 176 { 177 unsigned lsb = 0; 178 siginfo_t info; 179 180 info.si_signo = si_signo; 181 info.si_errno = 0; 182 info.si_code = si_code; 183 info.si_addr = (void __user *)address; 184 if (fault & VM_FAULT_HWPOISON_LARGE) 185 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 186 if (fault & VM_FAULT_HWPOISON) 187 lsb = PAGE_SHIFT; 188 info.si_addr_lsb = lsb; 189 190 force_sig_info(si_signo, &info, tsk); 191 } 192 193 DEFINE_SPINLOCK(pgd_lock); 194 LIST_HEAD(pgd_list); 195 196 #ifdef CONFIG_X86_32 197 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) 198 { 199 unsigned index = pgd_index(address); 200 pgd_t *pgd_k; 201 pud_t *pud, *pud_k; 202 pmd_t *pmd, *pmd_k; 203 204 pgd += index; 205 pgd_k = init_mm.pgd + index; 206 207 if (!pgd_present(*pgd_k)) 208 return NULL; 209 210 /* 211 * set_pgd(pgd, *pgd_k); here would be useless on PAE 212 * and redundant with the set_pmd() on non-PAE. As would 213 * set_pud. 214 */ 215 pud = pud_offset(pgd, address); 216 pud_k = pud_offset(pgd_k, address); 217 if (!pud_present(*pud_k)) 218 return NULL; 219 220 pmd = pmd_offset(pud, address); 221 pmd_k = pmd_offset(pud_k, address); 222 if (!pmd_present(*pmd_k)) 223 return NULL; 224 225 if (!pmd_present(*pmd)) 226 set_pmd(pmd, *pmd_k); 227 else 228 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k)); 229 230 return pmd_k; 231 } 232 233 void vmalloc_sync_all(void) 234 { 235 unsigned long address; 236 237 if (SHARED_KERNEL_PMD) 238 return; 239 240 for (address = VMALLOC_START & PMD_MASK; 241 address >= TASK_SIZE && address < FIXADDR_TOP; 242 address += PMD_SIZE) { 243 struct page *page; 244 245 spin_lock(&pgd_lock); 246 list_for_each_entry(page, &pgd_list, lru) { 247 spinlock_t *pgt_lock; 248 pmd_t *ret; 249 250 /* the pgt_lock only for Xen */ 251 pgt_lock = &pgd_page_get_mm(page)->page_table_lock; 252 253 spin_lock(pgt_lock); 254 ret = vmalloc_sync_one(page_address(page), address); 255 spin_unlock(pgt_lock); 256 257 if (!ret) 258 break; 259 } 260 spin_unlock(&pgd_lock); 261 } 262 } 263 264 /* 265 * 32-bit: 266 * 267 * Handle a fault on the vmalloc or module mapping area 268 */ 269 static noinline int vmalloc_fault(unsigned long address) 270 { 271 unsigned long pgd_paddr; 272 pmd_t *pmd_k; 273 pte_t *pte_k; 274 275 /* Make sure we are in vmalloc area: */ 276 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 277 return -1; 278 279 WARN_ON_ONCE(in_nmi()); 280 281 /* 282 * Synchronize this task's top level page-table 283 * with the 'reference' page table. 284 * 285 * Do _not_ use "current" here. We might be inside 286 * an interrupt in the middle of a task switch.. 287 */ 288 pgd_paddr = read_cr3(); 289 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address); 290 if (!pmd_k) 291 return -1; 292 293 pte_k = pte_offset_kernel(pmd_k, address); 294 if (!pte_present(*pte_k)) 295 return -1; 296 297 return 0; 298 } 299 NOKPROBE_SYMBOL(vmalloc_fault); 300 301 /* 302 * Did it hit the DOS screen memory VA from vm86 mode? 303 */ 304 static inline void 305 check_v8086_mode(struct pt_regs *regs, unsigned long address, 306 struct task_struct *tsk) 307 { 308 #ifdef CONFIG_VM86 309 unsigned long bit; 310 311 if (!v8086_mode(regs) || !tsk->thread.vm86) 312 return; 313 314 bit = (address - 0xA0000) >> PAGE_SHIFT; 315 if (bit < 32) 316 tsk->thread.vm86->screen_bitmap |= 1 << bit; 317 #endif 318 } 319 320 static bool low_pfn(unsigned long pfn) 321 { 322 return pfn < max_low_pfn; 323 } 324 325 static void dump_pagetable(unsigned long address) 326 { 327 pgd_t *base = __va(read_cr3()); 328 pgd_t *pgd = &base[pgd_index(address)]; 329 pmd_t *pmd; 330 pte_t *pte; 331 332 #ifdef CONFIG_X86_PAE 333 printk("*pdpt = %016Lx ", pgd_val(*pgd)); 334 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd)) 335 goto out; 336 #endif 337 pmd = pmd_offset(pud_offset(pgd, address), address); 338 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd)); 339 340 /* 341 * We must not directly access the pte in the highpte 342 * case if the page table is located in highmem. 343 * And let's rather not kmap-atomic the pte, just in case 344 * it's allocated already: 345 */ 346 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd)) 347 goto out; 348 349 pte = pte_offset_kernel(pmd, address); 350 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte)); 351 out: 352 printk("\n"); 353 } 354 355 #else /* CONFIG_X86_64: */ 356 357 void vmalloc_sync_all(void) 358 { 359 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0); 360 } 361 362 /* 363 * 64-bit: 364 * 365 * Handle a fault on the vmalloc area 366 * 367 * This assumes no large pages in there. 368 */ 369 static noinline int vmalloc_fault(unsigned long address) 370 { 371 pgd_t *pgd, *pgd_ref; 372 pud_t *pud, *pud_ref; 373 pmd_t *pmd, *pmd_ref; 374 pte_t *pte, *pte_ref; 375 376 /* Make sure we are in vmalloc area: */ 377 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 378 return -1; 379 380 WARN_ON_ONCE(in_nmi()); 381 382 /* 383 * Copy kernel mappings over when needed. This can also 384 * happen within a race in page table update. In the later 385 * case just flush: 386 */ 387 pgd = pgd_offset(current->active_mm, address); 388 pgd_ref = pgd_offset_k(address); 389 if (pgd_none(*pgd_ref)) 390 return -1; 391 392 if (pgd_none(*pgd)) { 393 set_pgd(pgd, *pgd_ref); 394 arch_flush_lazy_mmu_mode(); 395 } else { 396 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref)); 397 } 398 399 /* 400 * Below here mismatches are bugs because these lower tables 401 * are shared: 402 */ 403 404 pud = pud_offset(pgd, address); 405 pud_ref = pud_offset(pgd_ref, address); 406 if (pud_none(*pud_ref)) 407 return -1; 408 409 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref)) 410 BUG(); 411 412 pmd = pmd_offset(pud, address); 413 pmd_ref = pmd_offset(pud_ref, address); 414 if (pmd_none(*pmd_ref)) 415 return -1; 416 417 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref)) 418 BUG(); 419 420 pte_ref = pte_offset_kernel(pmd_ref, address); 421 if (!pte_present(*pte_ref)) 422 return -1; 423 424 pte = pte_offset_kernel(pmd, address); 425 426 /* 427 * Don't use pte_page here, because the mappings can point 428 * outside mem_map, and the NUMA hash lookup cannot handle 429 * that: 430 */ 431 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref)) 432 BUG(); 433 434 return 0; 435 } 436 NOKPROBE_SYMBOL(vmalloc_fault); 437 438 #ifdef CONFIG_CPU_SUP_AMD 439 static const char errata93_warning[] = 440 KERN_ERR 441 "******* Your BIOS seems to not contain a fix for K8 errata #93\n" 442 "******* Working around it, but it may cause SEGVs or burn power.\n" 443 "******* Please consider a BIOS update.\n" 444 "******* Disabling USB legacy in the BIOS may also help.\n"; 445 #endif 446 447 /* 448 * No vm86 mode in 64-bit mode: 449 */ 450 static inline void 451 check_v8086_mode(struct pt_regs *regs, unsigned long address, 452 struct task_struct *tsk) 453 { 454 } 455 456 static int bad_address(void *p) 457 { 458 unsigned long dummy; 459 460 return probe_kernel_address((unsigned long *)p, dummy); 461 } 462 463 static void dump_pagetable(unsigned long address) 464 { 465 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK); 466 pgd_t *pgd = base + pgd_index(address); 467 pud_t *pud; 468 pmd_t *pmd; 469 pte_t *pte; 470 471 if (bad_address(pgd)) 472 goto bad; 473 474 printk("PGD %lx ", pgd_val(*pgd)); 475 476 if (!pgd_present(*pgd)) 477 goto out; 478 479 pud = pud_offset(pgd, address); 480 if (bad_address(pud)) 481 goto bad; 482 483 printk("PUD %lx ", pud_val(*pud)); 484 if (!pud_present(*pud) || pud_large(*pud)) 485 goto out; 486 487 pmd = pmd_offset(pud, address); 488 if (bad_address(pmd)) 489 goto bad; 490 491 printk("PMD %lx ", pmd_val(*pmd)); 492 if (!pmd_present(*pmd) || pmd_large(*pmd)) 493 goto out; 494 495 pte = pte_offset_kernel(pmd, address); 496 if (bad_address(pte)) 497 goto bad; 498 499 printk("PTE %lx", pte_val(*pte)); 500 out: 501 printk("\n"); 502 return; 503 bad: 504 printk("BAD\n"); 505 } 506 507 #endif /* CONFIG_X86_64 */ 508 509 /* 510 * Workaround for K8 erratum #93 & buggy BIOS. 511 * 512 * BIOS SMM functions are required to use a specific workaround 513 * to avoid corruption of the 64bit RIP register on C stepping K8. 514 * 515 * A lot of BIOS that didn't get tested properly miss this. 516 * 517 * The OS sees this as a page fault with the upper 32bits of RIP cleared. 518 * Try to work around it here. 519 * 520 * Note we only handle faults in kernel here. 521 * Does nothing on 32-bit. 522 */ 523 static int is_errata93(struct pt_regs *regs, unsigned long address) 524 { 525 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD) 526 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD 527 || boot_cpu_data.x86 != 0xf) 528 return 0; 529 530 if (address != regs->ip) 531 return 0; 532 533 if ((address >> 32) != 0) 534 return 0; 535 536 address |= 0xffffffffUL << 32; 537 if ((address >= (u64)_stext && address <= (u64)_etext) || 538 (address >= MODULES_VADDR && address <= MODULES_END)) { 539 printk_once(errata93_warning); 540 regs->ip = address; 541 return 1; 542 } 543 #endif 544 return 0; 545 } 546 547 /* 548 * Work around K8 erratum #100 K8 in compat mode occasionally jumps 549 * to illegal addresses >4GB. 550 * 551 * We catch this in the page fault handler because these addresses 552 * are not reachable. Just detect this case and return. Any code 553 * segment in LDT is compatibility mode. 554 */ 555 static int is_errata100(struct pt_regs *regs, unsigned long address) 556 { 557 #ifdef CONFIG_X86_64 558 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32)) 559 return 1; 560 #endif 561 return 0; 562 } 563 564 static int is_f00f_bug(struct pt_regs *regs, unsigned long address) 565 { 566 #ifdef CONFIG_X86_F00F_BUG 567 unsigned long nr; 568 569 /* 570 * Pentium F0 0F C7 C8 bug workaround: 571 */ 572 if (boot_cpu_has_bug(X86_BUG_F00F)) { 573 nr = (address - idt_descr.address) >> 3; 574 575 if (nr == 6) { 576 do_invalid_op(regs, 0); 577 return 1; 578 } 579 } 580 #endif 581 return 0; 582 } 583 584 static const char nx_warning[] = KERN_CRIT 585 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n"; 586 static const char smep_warning[] = KERN_CRIT 587 "unable to execute userspace code (SMEP?) (uid: %d)\n"; 588 589 static void 590 show_fault_oops(struct pt_regs *regs, unsigned long error_code, 591 unsigned long address) 592 { 593 if (!oops_may_print()) 594 return; 595 596 if (error_code & PF_INSTR) { 597 unsigned int level; 598 pgd_t *pgd; 599 pte_t *pte; 600 601 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK); 602 pgd += pgd_index(address); 603 604 pte = lookup_address_in_pgd(pgd, address, &level); 605 606 if (pte && pte_present(*pte) && !pte_exec(*pte)) 607 printk(nx_warning, from_kuid(&init_user_ns, current_uid())); 608 if (pte && pte_present(*pte) && pte_exec(*pte) && 609 (pgd_flags(*pgd) & _PAGE_USER) && 610 (__read_cr4() & X86_CR4_SMEP)) 611 printk(smep_warning, from_kuid(&init_user_ns, current_uid())); 612 } 613 614 printk(KERN_ALERT "BUG: unable to handle kernel "); 615 if (address < PAGE_SIZE) 616 printk(KERN_CONT "NULL pointer dereference"); 617 else 618 printk(KERN_CONT "paging request"); 619 620 printk(KERN_CONT " at %p\n", (void *) address); 621 printk(KERN_ALERT "IP:"); 622 printk_address(regs->ip); 623 624 dump_pagetable(address); 625 } 626 627 static noinline void 628 pgtable_bad(struct pt_regs *regs, unsigned long error_code, 629 unsigned long address) 630 { 631 struct task_struct *tsk; 632 unsigned long flags; 633 int sig; 634 635 flags = oops_begin(); 636 tsk = current; 637 sig = SIGKILL; 638 639 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n", 640 tsk->comm, address); 641 dump_pagetable(address); 642 643 tsk->thread.cr2 = address; 644 tsk->thread.trap_nr = X86_TRAP_PF; 645 tsk->thread.error_code = error_code; 646 647 if (__die("Bad pagetable", regs, error_code)) 648 sig = 0; 649 650 oops_end(flags, regs, sig); 651 } 652 653 static noinline void 654 no_context(struct pt_regs *regs, unsigned long error_code, 655 unsigned long address, int signal, int si_code) 656 { 657 struct task_struct *tsk = current; 658 unsigned long flags; 659 int sig; 660 /* No context means no VMA to pass down */ 661 struct vm_area_struct *vma = NULL; 662 663 /* Are we prepared to handle this kernel fault? */ 664 if (fixup_exception(regs)) { 665 /* 666 * Any interrupt that takes a fault gets the fixup. This makes 667 * the below recursive fault logic only apply to a faults from 668 * task context. 669 */ 670 if (in_interrupt()) 671 return; 672 673 /* 674 * Per the above we're !in_interrupt(), aka. task context. 675 * 676 * In this case we need to make sure we're not recursively 677 * faulting through the emulate_vsyscall() logic. 678 */ 679 if (current_thread_info()->sig_on_uaccess_error && signal) { 680 tsk->thread.trap_nr = X86_TRAP_PF; 681 tsk->thread.error_code = error_code | PF_USER; 682 tsk->thread.cr2 = address; 683 684 /* XXX: hwpoison faults will set the wrong code. */ 685 force_sig_info_fault(signal, si_code, address, 686 tsk, vma, 0); 687 } 688 689 /* 690 * Barring that, we can do the fixup and be happy. 691 */ 692 return; 693 } 694 695 /* 696 * 32-bit: 697 * 698 * Valid to do another page fault here, because if this fault 699 * had been triggered by is_prefetch fixup_exception would have 700 * handled it. 701 * 702 * 64-bit: 703 * 704 * Hall of shame of CPU/BIOS bugs. 705 */ 706 if (is_prefetch(regs, error_code, address)) 707 return; 708 709 if (is_errata93(regs, address)) 710 return; 711 712 /* 713 * Oops. The kernel tried to access some bad page. We'll have to 714 * terminate things with extreme prejudice: 715 */ 716 flags = oops_begin(); 717 718 show_fault_oops(regs, error_code, address); 719 720 if (task_stack_end_corrupted(tsk)) 721 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 722 723 tsk->thread.cr2 = address; 724 tsk->thread.trap_nr = X86_TRAP_PF; 725 tsk->thread.error_code = error_code; 726 727 sig = SIGKILL; 728 if (__die("Oops", regs, error_code)) 729 sig = 0; 730 731 /* Executive summary in case the body of the oops scrolled away */ 732 printk(KERN_DEFAULT "CR2: %016lx\n", address); 733 734 oops_end(flags, regs, sig); 735 } 736 737 /* 738 * Print out info about fatal segfaults, if the show_unhandled_signals 739 * sysctl is set: 740 */ 741 static inline void 742 show_signal_msg(struct pt_regs *regs, unsigned long error_code, 743 unsigned long address, struct task_struct *tsk) 744 { 745 if (!unhandled_signal(tsk, SIGSEGV)) 746 return; 747 748 if (!printk_ratelimit()) 749 return; 750 751 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx", 752 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG, 753 tsk->comm, task_pid_nr(tsk), address, 754 (void *)regs->ip, (void *)regs->sp, error_code); 755 756 print_vma_addr(KERN_CONT " in ", regs->ip); 757 758 printk(KERN_CONT "\n"); 759 } 760 761 static void 762 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 763 unsigned long address, struct vm_area_struct *vma, 764 int si_code) 765 { 766 struct task_struct *tsk = current; 767 768 /* User mode accesses just cause a SIGSEGV */ 769 if (error_code & PF_USER) { 770 /* 771 * It's possible to have interrupts off here: 772 */ 773 local_irq_enable(); 774 775 /* 776 * Valid to do another page fault here because this one came 777 * from user space: 778 */ 779 if (is_prefetch(regs, error_code, address)) 780 return; 781 782 if (is_errata100(regs, address)) 783 return; 784 785 #ifdef CONFIG_X86_64 786 /* 787 * Instruction fetch faults in the vsyscall page might need 788 * emulation. 789 */ 790 if (unlikely((error_code & PF_INSTR) && 791 ((address & ~0xfff) == VSYSCALL_ADDR))) { 792 if (emulate_vsyscall(regs, address)) 793 return; 794 } 795 #endif 796 /* Kernel addresses are always protection faults: */ 797 if (address >= TASK_SIZE) 798 error_code |= PF_PROT; 799 800 if (likely(show_unhandled_signals)) 801 show_signal_msg(regs, error_code, address, tsk); 802 803 tsk->thread.cr2 = address; 804 tsk->thread.error_code = error_code; 805 tsk->thread.trap_nr = X86_TRAP_PF; 806 807 force_sig_info_fault(SIGSEGV, si_code, address, tsk, vma, 0); 808 809 return; 810 } 811 812 if (is_f00f_bug(regs, address)) 813 return; 814 815 no_context(regs, error_code, address, SIGSEGV, si_code); 816 } 817 818 static noinline void 819 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 820 unsigned long address, struct vm_area_struct *vma) 821 { 822 __bad_area_nosemaphore(regs, error_code, address, vma, SEGV_MAPERR); 823 } 824 825 static void 826 __bad_area(struct pt_regs *regs, unsigned long error_code, 827 unsigned long address, struct vm_area_struct *vma, int si_code) 828 { 829 struct mm_struct *mm = current->mm; 830 831 /* 832 * Something tried to access memory that isn't in our memory map.. 833 * Fix it, but check if it's kernel or user first.. 834 */ 835 up_read(&mm->mmap_sem); 836 837 __bad_area_nosemaphore(regs, error_code, address, vma, si_code); 838 } 839 840 static noinline void 841 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address) 842 { 843 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR); 844 } 845 846 static noinline void 847 bad_area_access_error(struct pt_regs *regs, unsigned long error_code, 848 unsigned long address, struct vm_area_struct *vma) 849 { 850 __bad_area(regs, error_code, address, vma, SEGV_ACCERR); 851 } 852 853 static void 854 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address, 855 struct vm_area_struct *vma, unsigned int fault) 856 { 857 struct task_struct *tsk = current; 858 int code = BUS_ADRERR; 859 860 /* Kernel mode? Handle exceptions or die: */ 861 if (!(error_code & PF_USER)) { 862 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR); 863 return; 864 } 865 866 /* User-space => ok to do another page fault: */ 867 if (is_prefetch(regs, error_code, address)) 868 return; 869 870 tsk->thread.cr2 = address; 871 tsk->thread.error_code = error_code; 872 tsk->thread.trap_nr = X86_TRAP_PF; 873 874 #ifdef CONFIG_MEMORY_FAILURE 875 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) { 876 printk(KERN_ERR 877 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n", 878 tsk->comm, tsk->pid, address); 879 code = BUS_MCEERR_AR; 880 } 881 #endif 882 force_sig_info_fault(SIGBUS, code, address, tsk, vma, fault); 883 } 884 885 static noinline void 886 mm_fault_error(struct pt_regs *regs, unsigned long error_code, 887 unsigned long address, struct vm_area_struct *vma, 888 unsigned int fault) 889 { 890 if (fatal_signal_pending(current) && !(error_code & PF_USER)) { 891 no_context(regs, error_code, address, 0, 0); 892 return; 893 } 894 895 if (fault & VM_FAULT_OOM) { 896 /* Kernel mode? Handle exceptions or die: */ 897 if (!(error_code & PF_USER)) { 898 no_context(regs, error_code, address, 899 SIGSEGV, SEGV_MAPERR); 900 return; 901 } 902 903 /* 904 * We ran out of memory, call the OOM killer, and return the 905 * userspace (which will retry the fault, or kill us if we got 906 * oom-killed): 907 */ 908 pagefault_out_of_memory(); 909 } else { 910 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON| 911 VM_FAULT_HWPOISON_LARGE)) 912 do_sigbus(regs, error_code, address, vma, fault); 913 else if (fault & VM_FAULT_SIGSEGV) 914 bad_area_nosemaphore(regs, error_code, address, vma); 915 else 916 BUG(); 917 } 918 } 919 920 static int spurious_fault_check(unsigned long error_code, pte_t *pte) 921 { 922 if ((error_code & PF_WRITE) && !pte_write(*pte)) 923 return 0; 924 925 if ((error_code & PF_INSTR) && !pte_exec(*pte)) 926 return 0; 927 /* 928 * Note: We do not do lazy flushing on protection key 929 * changes, so no spurious fault will ever set PF_PK. 930 */ 931 if ((error_code & PF_PK)) 932 return 1; 933 934 return 1; 935 } 936 937 /* 938 * Handle a spurious fault caused by a stale TLB entry. 939 * 940 * This allows us to lazily refresh the TLB when increasing the 941 * permissions of a kernel page (RO -> RW or NX -> X). Doing it 942 * eagerly is very expensive since that implies doing a full 943 * cross-processor TLB flush, even if no stale TLB entries exist 944 * on other processors. 945 * 946 * Spurious faults may only occur if the TLB contains an entry with 947 * fewer permission than the page table entry. Non-present (P = 0) 948 * and reserved bit (R = 1) faults are never spurious. 949 * 950 * There are no security implications to leaving a stale TLB when 951 * increasing the permissions on a page. 952 * 953 * Returns non-zero if a spurious fault was handled, zero otherwise. 954 * 955 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 956 * (Optional Invalidation). 957 */ 958 static noinline int 959 spurious_fault(unsigned long error_code, unsigned long address) 960 { 961 pgd_t *pgd; 962 pud_t *pud; 963 pmd_t *pmd; 964 pte_t *pte; 965 int ret; 966 967 /* 968 * Only writes to RO or instruction fetches from NX may cause 969 * spurious faults. 970 * 971 * These could be from user or supervisor accesses but the TLB 972 * is only lazily flushed after a kernel mapping protection 973 * change, so user accesses are not expected to cause spurious 974 * faults. 975 */ 976 if (error_code != (PF_WRITE | PF_PROT) 977 && error_code != (PF_INSTR | PF_PROT)) 978 return 0; 979 980 pgd = init_mm.pgd + pgd_index(address); 981 if (!pgd_present(*pgd)) 982 return 0; 983 984 pud = pud_offset(pgd, address); 985 if (!pud_present(*pud)) 986 return 0; 987 988 if (pud_large(*pud)) 989 return spurious_fault_check(error_code, (pte_t *) pud); 990 991 pmd = pmd_offset(pud, address); 992 if (!pmd_present(*pmd)) 993 return 0; 994 995 if (pmd_large(*pmd)) 996 return spurious_fault_check(error_code, (pte_t *) pmd); 997 998 pte = pte_offset_kernel(pmd, address); 999 if (!pte_present(*pte)) 1000 return 0; 1001 1002 ret = spurious_fault_check(error_code, pte); 1003 if (!ret) 1004 return 0; 1005 1006 /* 1007 * Make sure we have permissions in PMD. 1008 * If not, then there's a bug in the page tables: 1009 */ 1010 ret = spurious_fault_check(error_code, (pte_t *) pmd); 1011 WARN_ONCE(!ret, "PMD has incorrect permission bits\n"); 1012 1013 return ret; 1014 } 1015 NOKPROBE_SYMBOL(spurious_fault); 1016 1017 int show_unhandled_signals = 1; 1018 1019 static inline int 1020 access_error(unsigned long error_code, struct vm_area_struct *vma) 1021 { 1022 if (error_code & PF_WRITE) { 1023 /* write, present and write, not present: */ 1024 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1025 return 1; 1026 return 0; 1027 } 1028 1029 /* read, present: */ 1030 if (unlikely(error_code & PF_PROT)) 1031 return 1; 1032 1033 /* read, not present: */ 1034 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))) 1035 return 1; 1036 1037 return 0; 1038 } 1039 1040 static int fault_in_kernel_space(unsigned long address) 1041 { 1042 return address >= TASK_SIZE_MAX; 1043 } 1044 1045 static inline bool smap_violation(int error_code, struct pt_regs *regs) 1046 { 1047 if (!IS_ENABLED(CONFIG_X86_SMAP)) 1048 return false; 1049 1050 if (!static_cpu_has(X86_FEATURE_SMAP)) 1051 return false; 1052 1053 if (error_code & PF_USER) 1054 return false; 1055 1056 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC)) 1057 return false; 1058 1059 return true; 1060 } 1061 1062 /* 1063 * This routine handles page faults. It determines the address, 1064 * and the problem, and then passes it off to one of the appropriate 1065 * routines. 1066 * 1067 * This function must have noinline because both callers 1068 * {,trace_}do_page_fault() have notrace on. Having this an actual function 1069 * guarantees there's a function trace entry. 1070 */ 1071 static noinline void 1072 __do_page_fault(struct pt_regs *regs, unsigned long error_code, 1073 unsigned long address) 1074 { 1075 struct vm_area_struct *vma; 1076 struct task_struct *tsk; 1077 struct mm_struct *mm; 1078 int fault, major = 0; 1079 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; 1080 1081 tsk = current; 1082 mm = tsk->mm; 1083 1084 /* 1085 * Detect and handle instructions that would cause a page fault for 1086 * both a tracked kernel page and a userspace page. 1087 */ 1088 if (kmemcheck_active(regs)) 1089 kmemcheck_hide(regs); 1090 prefetchw(&mm->mmap_sem); 1091 1092 if (unlikely(kmmio_fault(regs, address))) 1093 return; 1094 1095 /* 1096 * We fault-in kernel-space virtual memory on-demand. The 1097 * 'reference' page table is init_mm.pgd. 1098 * 1099 * NOTE! We MUST NOT take any locks for this case. We may 1100 * be in an interrupt or a critical region, and should 1101 * only copy the information from the master page table, 1102 * nothing more. 1103 * 1104 * This verifies that the fault happens in kernel space 1105 * (error_code & 4) == 0, and that the fault was not a 1106 * protection error (error_code & 9) == 0. 1107 */ 1108 if (unlikely(fault_in_kernel_space(address))) { 1109 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) { 1110 if (vmalloc_fault(address) >= 0) 1111 return; 1112 1113 if (kmemcheck_fault(regs, address, error_code)) 1114 return; 1115 } 1116 1117 /* Can handle a stale RO->RW TLB: */ 1118 if (spurious_fault(error_code, address)) 1119 return; 1120 1121 /* kprobes don't want to hook the spurious faults: */ 1122 if (kprobes_fault(regs)) 1123 return; 1124 /* 1125 * Don't take the mm semaphore here. If we fixup a prefetch 1126 * fault we could otherwise deadlock: 1127 */ 1128 bad_area_nosemaphore(regs, error_code, address, NULL); 1129 1130 return; 1131 } 1132 1133 /* kprobes don't want to hook the spurious faults: */ 1134 if (unlikely(kprobes_fault(regs))) 1135 return; 1136 1137 if (unlikely(error_code & PF_RSVD)) 1138 pgtable_bad(regs, error_code, address); 1139 1140 if (unlikely(smap_violation(error_code, regs))) { 1141 bad_area_nosemaphore(regs, error_code, address, NULL); 1142 return; 1143 } 1144 1145 /* 1146 * If we're in an interrupt, have no user context or are running 1147 * in a region with pagefaults disabled then we must not take the fault 1148 */ 1149 if (unlikely(faulthandler_disabled() || !mm)) { 1150 bad_area_nosemaphore(regs, error_code, address, NULL); 1151 return; 1152 } 1153 1154 /* 1155 * It's safe to allow irq's after cr2 has been saved and the 1156 * vmalloc fault has been handled. 1157 * 1158 * User-mode registers count as a user access even for any 1159 * potential system fault or CPU buglet: 1160 */ 1161 if (user_mode(regs)) { 1162 local_irq_enable(); 1163 error_code |= PF_USER; 1164 flags |= FAULT_FLAG_USER; 1165 } else { 1166 if (regs->flags & X86_EFLAGS_IF) 1167 local_irq_enable(); 1168 } 1169 1170 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); 1171 1172 if (error_code & PF_WRITE) 1173 flags |= FAULT_FLAG_WRITE; 1174 1175 /* 1176 * When running in the kernel we expect faults to occur only to 1177 * addresses in user space. All other faults represent errors in 1178 * the kernel and should generate an OOPS. Unfortunately, in the 1179 * case of an erroneous fault occurring in a code path which already 1180 * holds mmap_sem we will deadlock attempting to validate the fault 1181 * against the address space. Luckily the kernel only validly 1182 * references user space from well defined areas of code, which are 1183 * listed in the exceptions table. 1184 * 1185 * As the vast majority of faults will be valid we will only perform 1186 * the source reference check when there is a possibility of a 1187 * deadlock. Attempt to lock the address space, if we cannot we then 1188 * validate the source. If this is invalid we can skip the address 1189 * space check, thus avoiding the deadlock: 1190 */ 1191 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 1192 if ((error_code & PF_USER) == 0 && 1193 !search_exception_tables(regs->ip)) { 1194 bad_area_nosemaphore(regs, error_code, address, NULL); 1195 return; 1196 } 1197 retry: 1198 down_read(&mm->mmap_sem); 1199 } else { 1200 /* 1201 * The above down_read_trylock() might have succeeded in 1202 * which case we'll have missed the might_sleep() from 1203 * down_read(): 1204 */ 1205 might_sleep(); 1206 } 1207 1208 vma = find_vma(mm, address); 1209 if (unlikely(!vma)) { 1210 bad_area(regs, error_code, address); 1211 return; 1212 } 1213 if (likely(vma->vm_start <= address)) 1214 goto good_area; 1215 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) { 1216 bad_area(regs, error_code, address); 1217 return; 1218 } 1219 if (error_code & PF_USER) { 1220 /* 1221 * Accessing the stack below %sp is always a bug. 1222 * The large cushion allows instructions like enter 1223 * and pusha to work. ("enter $65535, $31" pushes 1224 * 32 pointers and then decrements %sp by 65535.) 1225 */ 1226 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) { 1227 bad_area(regs, error_code, address); 1228 return; 1229 } 1230 } 1231 if (unlikely(expand_stack(vma, address))) { 1232 bad_area(regs, error_code, address); 1233 return; 1234 } 1235 1236 /* 1237 * Ok, we have a good vm_area for this memory access, so 1238 * we can handle it.. 1239 */ 1240 good_area: 1241 if (unlikely(access_error(error_code, vma))) { 1242 bad_area_access_error(regs, error_code, address, vma); 1243 return; 1244 } 1245 1246 /* 1247 * If for any reason at all we couldn't handle the fault, 1248 * make sure we exit gracefully rather than endlessly redo 1249 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if 1250 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked. 1251 */ 1252 fault = handle_mm_fault(mm, vma, address, flags); 1253 major |= fault & VM_FAULT_MAJOR; 1254 1255 /* 1256 * If we need to retry the mmap_sem has already been released, 1257 * and if there is a fatal signal pending there is no guarantee 1258 * that we made any progress. Handle this case first. 1259 */ 1260 if (unlikely(fault & VM_FAULT_RETRY)) { 1261 /* Retry at most once */ 1262 if (flags & FAULT_FLAG_ALLOW_RETRY) { 1263 flags &= ~FAULT_FLAG_ALLOW_RETRY; 1264 flags |= FAULT_FLAG_TRIED; 1265 if (!fatal_signal_pending(tsk)) 1266 goto retry; 1267 } 1268 1269 /* User mode? Just return to handle the fatal exception */ 1270 if (flags & FAULT_FLAG_USER) 1271 return; 1272 1273 /* Not returning to user mode? Handle exceptions or die: */ 1274 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR); 1275 return; 1276 } 1277 1278 up_read(&mm->mmap_sem); 1279 if (unlikely(fault & VM_FAULT_ERROR)) { 1280 mm_fault_error(regs, error_code, address, vma, fault); 1281 return; 1282 } 1283 1284 /* 1285 * Major/minor page fault accounting. If any of the events 1286 * returned VM_FAULT_MAJOR, we account it as a major fault. 1287 */ 1288 if (major) { 1289 tsk->maj_flt++; 1290 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 1291 } else { 1292 tsk->min_flt++; 1293 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 1294 } 1295 1296 check_v8086_mode(regs, address, tsk); 1297 } 1298 NOKPROBE_SYMBOL(__do_page_fault); 1299 1300 dotraplinkage void notrace 1301 do_page_fault(struct pt_regs *regs, unsigned long error_code) 1302 { 1303 unsigned long address = read_cr2(); /* Get the faulting address */ 1304 enum ctx_state prev_state; 1305 1306 /* 1307 * We must have this function tagged with __kprobes, notrace and call 1308 * read_cr2() before calling anything else. To avoid calling any kind 1309 * of tracing machinery before we've observed the CR2 value. 1310 * 1311 * exception_{enter,exit}() contain all sorts of tracepoints. 1312 */ 1313 1314 prev_state = exception_enter(); 1315 __do_page_fault(regs, error_code, address); 1316 exception_exit(prev_state); 1317 } 1318 NOKPROBE_SYMBOL(do_page_fault); 1319 1320 #ifdef CONFIG_TRACING 1321 static nokprobe_inline void 1322 trace_page_fault_entries(unsigned long address, struct pt_regs *regs, 1323 unsigned long error_code) 1324 { 1325 if (user_mode(regs)) 1326 trace_page_fault_user(address, regs, error_code); 1327 else 1328 trace_page_fault_kernel(address, regs, error_code); 1329 } 1330 1331 dotraplinkage void notrace 1332 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code) 1333 { 1334 /* 1335 * The exception_enter and tracepoint processing could 1336 * trigger another page faults (user space callchain 1337 * reading) and destroy the original cr2 value, so read 1338 * the faulting address now. 1339 */ 1340 unsigned long address = read_cr2(); 1341 enum ctx_state prev_state; 1342 1343 prev_state = exception_enter(); 1344 trace_page_fault_entries(address, regs, error_code); 1345 __do_page_fault(regs, error_code, address); 1346 exception_exit(prev_state); 1347 } 1348 NOKPROBE_SYMBOL(trace_do_page_fault); 1349 #endif /* CONFIG_TRACING */ 1350