1 /* 2 * Copyright (C) 1995 Linus Torvalds 3 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs. 4 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar 5 */ 6 #include <linux/sched.h> /* test_thread_flag(), ... */ 7 #include <linux/kdebug.h> /* oops_begin/end, ... */ 8 #include <linux/module.h> /* search_exception_table */ 9 #include <linux/bootmem.h> /* max_low_pfn */ 10 #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */ 11 #include <linux/mmiotrace.h> /* kmmio_handler, ... */ 12 #include <linux/perf_event.h> /* perf_sw_event */ 13 #include <linux/hugetlb.h> /* hstate_index_to_shift */ 14 #include <linux/prefetch.h> /* prefetchw */ 15 #include <linux/context_tracking.h> /* exception_enter(), ... */ 16 #include <linux/uaccess.h> /* faulthandler_disabled() */ 17 18 #include <asm/cpufeature.h> /* boot_cpu_has, ... */ 19 #include <asm/traps.h> /* dotraplinkage, ... */ 20 #include <asm/pgalloc.h> /* pgd_*(), ... */ 21 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */ 22 #include <asm/fixmap.h> /* VSYSCALL_ADDR */ 23 #include <asm/vsyscall.h> /* emulate_vsyscall */ 24 #include <asm/vm86.h> /* struct vm86 */ 25 #include <asm/mmu_context.h> /* vma_pkey() */ 26 27 #define CREATE_TRACE_POINTS 28 #include <asm/trace/exceptions.h> 29 30 /* 31 * Page fault error code bits: 32 * 33 * bit 0 == 0: no page found 1: protection fault 34 * bit 1 == 0: read access 1: write access 35 * bit 2 == 0: kernel-mode access 1: user-mode access 36 * bit 3 == 1: use of reserved bit detected 37 * bit 4 == 1: fault was an instruction fetch 38 * bit 5 == 1: protection keys block access 39 */ 40 enum x86_pf_error_code { 41 42 PF_PROT = 1 << 0, 43 PF_WRITE = 1 << 1, 44 PF_USER = 1 << 2, 45 PF_RSVD = 1 << 3, 46 PF_INSTR = 1 << 4, 47 PF_PK = 1 << 5, 48 }; 49 50 /* 51 * Returns 0 if mmiotrace is disabled, or if the fault is not 52 * handled by mmiotrace: 53 */ 54 static nokprobe_inline int 55 kmmio_fault(struct pt_regs *regs, unsigned long addr) 56 { 57 if (unlikely(is_kmmio_active())) 58 if (kmmio_handler(regs, addr) == 1) 59 return -1; 60 return 0; 61 } 62 63 static nokprobe_inline int kprobes_fault(struct pt_regs *regs) 64 { 65 int ret = 0; 66 67 /* kprobe_running() needs smp_processor_id() */ 68 if (kprobes_built_in() && !user_mode(regs)) { 69 preempt_disable(); 70 if (kprobe_running() && kprobe_fault_handler(regs, 14)) 71 ret = 1; 72 preempt_enable(); 73 } 74 75 return ret; 76 } 77 78 /* 79 * Prefetch quirks: 80 * 81 * 32-bit mode: 82 * 83 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch. 84 * Check that here and ignore it. 85 * 86 * 64-bit mode: 87 * 88 * Sometimes the CPU reports invalid exceptions on prefetch. 89 * Check that here and ignore it. 90 * 91 * Opcode checker based on code by Richard Brunner. 92 */ 93 static inline int 94 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr, 95 unsigned char opcode, int *prefetch) 96 { 97 unsigned char instr_hi = opcode & 0xf0; 98 unsigned char instr_lo = opcode & 0x0f; 99 100 switch (instr_hi) { 101 case 0x20: 102 case 0x30: 103 /* 104 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. 105 * In X86_64 long mode, the CPU will signal invalid 106 * opcode if some of these prefixes are present so 107 * X86_64 will never get here anyway 108 */ 109 return ((instr_lo & 7) == 0x6); 110 #ifdef CONFIG_X86_64 111 case 0x40: 112 /* 113 * In AMD64 long mode 0x40..0x4F are valid REX prefixes 114 * Need to figure out under what instruction mode the 115 * instruction was issued. Could check the LDT for lm, 116 * but for now it's good enough to assume that long 117 * mode only uses well known segments or kernel. 118 */ 119 return (!user_mode(regs) || user_64bit_mode(regs)); 120 #endif 121 case 0x60: 122 /* 0x64 thru 0x67 are valid prefixes in all modes. */ 123 return (instr_lo & 0xC) == 0x4; 124 case 0xF0: 125 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */ 126 return !instr_lo || (instr_lo>>1) == 1; 127 case 0x00: 128 /* Prefetch instruction is 0x0F0D or 0x0F18 */ 129 if (probe_kernel_address(instr, opcode)) 130 return 0; 131 132 *prefetch = (instr_lo == 0xF) && 133 (opcode == 0x0D || opcode == 0x18); 134 return 0; 135 default: 136 return 0; 137 } 138 } 139 140 static int 141 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr) 142 { 143 unsigned char *max_instr; 144 unsigned char *instr; 145 int prefetch = 0; 146 147 /* 148 * If it was a exec (instruction fetch) fault on NX page, then 149 * do not ignore the fault: 150 */ 151 if (error_code & PF_INSTR) 152 return 0; 153 154 instr = (void *)convert_ip_to_linear(current, regs); 155 max_instr = instr + 15; 156 157 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX) 158 return 0; 159 160 while (instr < max_instr) { 161 unsigned char opcode; 162 163 if (probe_kernel_address(instr, opcode)) 164 break; 165 166 instr++; 167 168 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch)) 169 break; 170 } 171 return prefetch; 172 } 173 174 /* 175 * A protection key fault means that the PKRU value did not allow 176 * access to some PTE. Userspace can figure out what PKRU was 177 * from the XSAVE state, and this function fills out a field in 178 * siginfo so userspace can discover which protection key was set 179 * on the PTE. 180 * 181 * If we get here, we know that the hardware signaled a PF_PK 182 * fault and that there was a VMA once we got in the fault 183 * handler. It does *not* guarantee that the VMA we find here 184 * was the one that we faulted on. 185 * 186 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4); 187 * 2. T1 : set PKRU to deny access to pkey=4, touches page 188 * 3. T1 : faults... 189 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5); 190 * 5. T1 : enters fault handler, takes mmap_sem, etc... 191 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really 192 * faulted on a pte with its pkey=4. 193 */ 194 static void fill_sig_info_pkey(int si_code, siginfo_t *info, 195 struct vm_area_struct *vma) 196 { 197 /* This is effectively an #ifdef */ 198 if (!boot_cpu_has(X86_FEATURE_OSPKE)) 199 return; 200 201 /* Fault not from Protection Keys: nothing to do */ 202 if (si_code != SEGV_PKUERR) 203 return; 204 /* 205 * force_sig_info_fault() is called from a number of 206 * contexts, some of which have a VMA and some of which 207 * do not. The PF_PK handing happens after we have a 208 * valid VMA, so we should never reach this without a 209 * valid VMA. 210 */ 211 if (!vma) { 212 WARN_ONCE(1, "PKU fault with no VMA passed in"); 213 info->si_pkey = 0; 214 return; 215 } 216 /* 217 * si_pkey should be thought of as a strong hint, but not 218 * absolutely guranteed to be 100% accurate because of 219 * the race explained above. 220 */ 221 info->si_pkey = vma_pkey(vma); 222 } 223 224 static void 225 force_sig_info_fault(int si_signo, int si_code, unsigned long address, 226 struct task_struct *tsk, struct vm_area_struct *vma, 227 int fault) 228 { 229 unsigned lsb = 0; 230 siginfo_t info; 231 232 info.si_signo = si_signo; 233 info.si_errno = 0; 234 info.si_code = si_code; 235 info.si_addr = (void __user *)address; 236 if (fault & VM_FAULT_HWPOISON_LARGE) 237 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 238 if (fault & VM_FAULT_HWPOISON) 239 lsb = PAGE_SHIFT; 240 info.si_addr_lsb = lsb; 241 242 fill_sig_info_pkey(si_code, &info, vma); 243 244 force_sig_info(si_signo, &info, tsk); 245 } 246 247 DEFINE_SPINLOCK(pgd_lock); 248 LIST_HEAD(pgd_list); 249 250 #ifdef CONFIG_X86_32 251 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) 252 { 253 unsigned index = pgd_index(address); 254 pgd_t *pgd_k; 255 pud_t *pud, *pud_k; 256 pmd_t *pmd, *pmd_k; 257 258 pgd += index; 259 pgd_k = init_mm.pgd + index; 260 261 if (!pgd_present(*pgd_k)) 262 return NULL; 263 264 /* 265 * set_pgd(pgd, *pgd_k); here would be useless on PAE 266 * and redundant with the set_pmd() on non-PAE. As would 267 * set_pud. 268 */ 269 pud = pud_offset(pgd, address); 270 pud_k = pud_offset(pgd_k, address); 271 if (!pud_present(*pud_k)) 272 return NULL; 273 274 pmd = pmd_offset(pud, address); 275 pmd_k = pmd_offset(pud_k, address); 276 if (!pmd_present(*pmd_k)) 277 return NULL; 278 279 if (!pmd_present(*pmd)) 280 set_pmd(pmd, *pmd_k); 281 else 282 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k)); 283 284 return pmd_k; 285 } 286 287 void vmalloc_sync_all(void) 288 { 289 unsigned long address; 290 291 if (SHARED_KERNEL_PMD) 292 return; 293 294 for (address = VMALLOC_START & PMD_MASK; 295 address >= TASK_SIZE && address < FIXADDR_TOP; 296 address += PMD_SIZE) { 297 struct page *page; 298 299 spin_lock(&pgd_lock); 300 list_for_each_entry(page, &pgd_list, lru) { 301 spinlock_t *pgt_lock; 302 pmd_t *ret; 303 304 /* the pgt_lock only for Xen */ 305 pgt_lock = &pgd_page_get_mm(page)->page_table_lock; 306 307 spin_lock(pgt_lock); 308 ret = vmalloc_sync_one(page_address(page), address); 309 spin_unlock(pgt_lock); 310 311 if (!ret) 312 break; 313 } 314 spin_unlock(&pgd_lock); 315 } 316 } 317 318 /* 319 * 32-bit: 320 * 321 * Handle a fault on the vmalloc or module mapping area 322 */ 323 static noinline int vmalloc_fault(unsigned long address) 324 { 325 unsigned long pgd_paddr; 326 pmd_t *pmd_k; 327 pte_t *pte_k; 328 329 /* Make sure we are in vmalloc area: */ 330 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 331 return -1; 332 333 WARN_ON_ONCE(in_nmi()); 334 335 /* 336 * Synchronize this task's top level page-table 337 * with the 'reference' page table. 338 * 339 * Do _not_ use "current" here. We might be inside 340 * an interrupt in the middle of a task switch.. 341 */ 342 pgd_paddr = read_cr3(); 343 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address); 344 if (!pmd_k) 345 return -1; 346 347 pte_k = pte_offset_kernel(pmd_k, address); 348 if (!pte_present(*pte_k)) 349 return -1; 350 351 return 0; 352 } 353 NOKPROBE_SYMBOL(vmalloc_fault); 354 355 /* 356 * Did it hit the DOS screen memory VA from vm86 mode? 357 */ 358 static inline void 359 check_v8086_mode(struct pt_regs *regs, unsigned long address, 360 struct task_struct *tsk) 361 { 362 #ifdef CONFIG_VM86 363 unsigned long bit; 364 365 if (!v8086_mode(regs) || !tsk->thread.vm86) 366 return; 367 368 bit = (address - 0xA0000) >> PAGE_SHIFT; 369 if (bit < 32) 370 tsk->thread.vm86->screen_bitmap |= 1 << bit; 371 #endif 372 } 373 374 static bool low_pfn(unsigned long pfn) 375 { 376 return pfn < max_low_pfn; 377 } 378 379 static void dump_pagetable(unsigned long address) 380 { 381 pgd_t *base = __va(read_cr3()); 382 pgd_t *pgd = &base[pgd_index(address)]; 383 pmd_t *pmd; 384 pte_t *pte; 385 386 #ifdef CONFIG_X86_PAE 387 printk("*pdpt = %016Lx ", pgd_val(*pgd)); 388 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd)) 389 goto out; 390 #endif 391 pmd = pmd_offset(pud_offset(pgd, address), address); 392 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd)); 393 394 /* 395 * We must not directly access the pte in the highpte 396 * case if the page table is located in highmem. 397 * And let's rather not kmap-atomic the pte, just in case 398 * it's allocated already: 399 */ 400 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd)) 401 goto out; 402 403 pte = pte_offset_kernel(pmd, address); 404 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte)); 405 out: 406 printk("\n"); 407 } 408 409 #else /* CONFIG_X86_64: */ 410 411 void vmalloc_sync_all(void) 412 { 413 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0); 414 } 415 416 /* 417 * 64-bit: 418 * 419 * Handle a fault on the vmalloc area 420 * 421 * This assumes no large pages in there. 422 */ 423 static noinline int vmalloc_fault(unsigned long address) 424 { 425 pgd_t *pgd, *pgd_ref; 426 pud_t *pud, *pud_ref; 427 pmd_t *pmd, *pmd_ref; 428 pte_t *pte, *pte_ref; 429 430 /* Make sure we are in vmalloc area: */ 431 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 432 return -1; 433 434 WARN_ON_ONCE(in_nmi()); 435 436 /* 437 * Copy kernel mappings over when needed. This can also 438 * happen within a race in page table update. In the later 439 * case just flush: 440 */ 441 pgd = pgd_offset(current->active_mm, address); 442 pgd_ref = pgd_offset_k(address); 443 if (pgd_none(*pgd_ref)) 444 return -1; 445 446 if (pgd_none(*pgd)) { 447 set_pgd(pgd, *pgd_ref); 448 arch_flush_lazy_mmu_mode(); 449 } else { 450 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref)); 451 } 452 453 /* 454 * Below here mismatches are bugs because these lower tables 455 * are shared: 456 */ 457 458 pud = pud_offset(pgd, address); 459 pud_ref = pud_offset(pgd_ref, address); 460 if (pud_none(*pud_ref)) 461 return -1; 462 463 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref)) 464 BUG(); 465 466 pmd = pmd_offset(pud, address); 467 pmd_ref = pmd_offset(pud_ref, address); 468 if (pmd_none(*pmd_ref)) 469 return -1; 470 471 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref)) 472 BUG(); 473 474 pte_ref = pte_offset_kernel(pmd_ref, address); 475 if (!pte_present(*pte_ref)) 476 return -1; 477 478 pte = pte_offset_kernel(pmd, address); 479 480 /* 481 * Don't use pte_page here, because the mappings can point 482 * outside mem_map, and the NUMA hash lookup cannot handle 483 * that: 484 */ 485 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref)) 486 BUG(); 487 488 return 0; 489 } 490 NOKPROBE_SYMBOL(vmalloc_fault); 491 492 #ifdef CONFIG_CPU_SUP_AMD 493 static const char errata93_warning[] = 494 KERN_ERR 495 "******* Your BIOS seems to not contain a fix for K8 errata #93\n" 496 "******* Working around it, but it may cause SEGVs or burn power.\n" 497 "******* Please consider a BIOS update.\n" 498 "******* Disabling USB legacy in the BIOS may also help.\n"; 499 #endif 500 501 /* 502 * No vm86 mode in 64-bit mode: 503 */ 504 static inline void 505 check_v8086_mode(struct pt_regs *regs, unsigned long address, 506 struct task_struct *tsk) 507 { 508 } 509 510 static int bad_address(void *p) 511 { 512 unsigned long dummy; 513 514 return probe_kernel_address((unsigned long *)p, dummy); 515 } 516 517 static void dump_pagetable(unsigned long address) 518 { 519 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK); 520 pgd_t *pgd = base + pgd_index(address); 521 pud_t *pud; 522 pmd_t *pmd; 523 pte_t *pte; 524 525 if (bad_address(pgd)) 526 goto bad; 527 528 printk("PGD %lx ", pgd_val(*pgd)); 529 530 if (!pgd_present(*pgd)) 531 goto out; 532 533 pud = pud_offset(pgd, address); 534 if (bad_address(pud)) 535 goto bad; 536 537 printk("PUD %lx ", pud_val(*pud)); 538 if (!pud_present(*pud) || pud_large(*pud)) 539 goto out; 540 541 pmd = pmd_offset(pud, address); 542 if (bad_address(pmd)) 543 goto bad; 544 545 printk("PMD %lx ", pmd_val(*pmd)); 546 if (!pmd_present(*pmd) || pmd_large(*pmd)) 547 goto out; 548 549 pte = pte_offset_kernel(pmd, address); 550 if (bad_address(pte)) 551 goto bad; 552 553 printk("PTE %lx", pte_val(*pte)); 554 out: 555 printk("\n"); 556 return; 557 bad: 558 printk("BAD\n"); 559 } 560 561 #endif /* CONFIG_X86_64 */ 562 563 /* 564 * Workaround for K8 erratum #93 & buggy BIOS. 565 * 566 * BIOS SMM functions are required to use a specific workaround 567 * to avoid corruption of the 64bit RIP register on C stepping K8. 568 * 569 * A lot of BIOS that didn't get tested properly miss this. 570 * 571 * The OS sees this as a page fault with the upper 32bits of RIP cleared. 572 * Try to work around it here. 573 * 574 * Note we only handle faults in kernel here. 575 * Does nothing on 32-bit. 576 */ 577 static int is_errata93(struct pt_regs *regs, unsigned long address) 578 { 579 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD) 580 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD 581 || boot_cpu_data.x86 != 0xf) 582 return 0; 583 584 if (address != regs->ip) 585 return 0; 586 587 if ((address >> 32) != 0) 588 return 0; 589 590 address |= 0xffffffffUL << 32; 591 if ((address >= (u64)_stext && address <= (u64)_etext) || 592 (address >= MODULES_VADDR && address <= MODULES_END)) { 593 printk_once(errata93_warning); 594 regs->ip = address; 595 return 1; 596 } 597 #endif 598 return 0; 599 } 600 601 /* 602 * Work around K8 erratum #100 K8 in compat mode occasionally jumps 603 * to illegal addresses >4GB. 604 * 605 * We catch this in the page fault handler because these addresses 606 * are not reachable. Just detect this case and return. Any code 607 * segment in LDT is compatibility mode. 608 */ 609 static int is_errata100(struct pt_regs *regs, unsigned long address) 610 { 611 #ifdef CONFIG_X86_64 612 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32)) 613 return 1; 614 #endif 615 return 0; 616 } 617 618 static int is_f00f_bug(struct pt_regs *regs, unsigned long address) 619 { 620 #ifdef CONFIG_X86_F00F_BUG 621 unsigned long nr; 622 623 /* 624 * Pentium F0 0F C7 C8 bug workaround: 625 */ 626 if (boot_cpu_has_bug(X86_BUG_F00F)) { 627 nr = (address - idt_descr.address) >> 3; 628 629 if (nr == 6) { 630 do_invalid_op(regs, 0); 631 return 1; 632 } 633 } 634 #endif 635 return 0; 636 } 637 638 static const char nx_warning[] = KERN_CRIT 639 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n"; 640 static const char smep_warning[] = KERN_CRIT 641 "unable to execute userspace code (SMEP?) (uid: %d)\n"; 642 643 static void 644 show_fault_oops(struct pt_regs *regs, unsigned long error_code, 645 unsigned long address) 646 { 647 if (!oops_may_print()) 648 return; 649 650 if (error_code & PF_INSTR) { 651 unsigned int level; 652 pgd_t *pgd; 653 pte_t *pte; 654 655 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK); 656 pgd += pgd_index(address); 657 658 pte = lookup_address_in_pgd(pgd, address, &level); 659 660 if (pte && pte_present(*pte) && !pte_exec(*pte)) 661 printk(nx_warning, 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 printk(smep_warning, from_kuid(&init_user_ns, current_uid())); 666 } 667 668 printk(KERN_ALERT "BUG: unable to handle kernel "); 669 if (address < PAGE_SIZE) 670 printk(KERN_CONT "NULL pointer dereference"); 671 else 672 printk(KERN_CONT "paging request"); 673 674 printk(KERN_CONT " at %p\n", (void *) address); 675 printk(KERN_ALERT "IP:"); 676 printk_address(regs->ip); 677 678 dump_pagetable(address); 679 } 680 681 static noinline void 682 pgtable_bad(struct pt_regs *regs, unsigned long error_code, 683 unsigned long address) 684 { 685 struct task_struct *tsk; 686 unsigned long flags; 687 int sig; 688 689 flags = oops_begin(); 690 tsk = current; 691 sig = SIGKILL; 692 693 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n", 694 tsk->comm, address); 695 dump_pagetable(address); 696 697 tsk->thread.cr2 = address; 698 tsk->thread.trap_nr = X86_TRAP_PF; 699 tsk->thread.error_code = error_code; 700 701 if (__die("Bad pagetable", regs, error_code)) 702 sig = 0; 703 704 oops_end(flags, regs, sig); 705 } 706 707 static noinline void 708 no_context(struct pt_regs *regs, unsigned long error_code, 709 unsigned long address, int signal, int si_code) 710 { 711 struct task_struct *tsk = current; 712 unsigned long flags; 713 int sig; 714 /* No context means no VMA to pass down */ 715 struct vm_area_struct *vma = NULL; 716 717 /* Are we prepared to handle this kernel fault? */ 718 if (fixup_exception(regs)) { 719 /* 720 * Any interrupt that takes a fault gets the fixup. This makes 721 * the below recursive fault logic only apply to a faults from 722 * task context. 723 */ 724 if (in_interrupt()) 725 return; 726 727 /* 728 * Per the above we're !in_interrupt(), aka. task context. 729 * 730 * In this case we need to make sure we're not recursively 731 * faulting through the emulate_vsyscall() logic. 732 */ 733 if (current_thread_info()->sig_on_uaccess_error && signal) { 734 tsk->thread.trap_nr = X86_TRAP_PF; 735 tsk->thread.error_code = error_code | PF_USER; 736 tsk->thread.cr2 = address; 737 738 /* XXX: hwpoison faults will set the wrong code. */ 739 force_sig_info_fault(signal, si_code, address, 740 tsk, vma, 0); 741 } 742 743 /* 744 * Barring that, we can do the fixup and be happy. 745 */ 746 return; 747 } 748 749 /* 750 * 32-bit: 751 * 752 * Valid to do another page fault here, because if this fault 753 * had been triggered by is_prefetch fixup_exception would have 754 * handled it. 755 * 756 * 64-bit: 757 * 758 * Hall of shame of CPU/BIOS bugs. 759 */ 760 if (is_prefetch(regs, error_code, address)) 761 return; 762 763 if (is_errata93(regs, address)) 764 return; 765 766 /* 767 * Oops. The kernel tried to access some bad page. We'll have to 768 * terminate things with extreme prejudice: 769 */ 770 flags = oops_begin(); 771 772 show_fault_oops(regs, error_code, address); 773 774 if (task_stack_end_corrupted(tsk)) 775 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 776 777 tsk->thread.cr2 = address; 778 tsk->thread.trap_nr = X86_TRAP_PF; 779 tsk->thread.error_code = error_code; 780 781 sig = SIGKILL; 782 if (__die("Oops", regs, error_code)) 783 sig = 0; 784 785 /* Executive summary in case the body of the oops scrolled away */ 786 printk(KERN_DEFAULT "CR2: %016lx\n", address); 787 788 oops_end(flags, regs, sig); 789 } 790 791 /* 792 * Print out info about fatal segfaults, if the show_unhandled_signals 793 * sysctl is set: 794 */ 795 static inline void 796 show_signal_msg(struct pt_regs *regs, unsigned long error_code, 797 unsigned long address, struct task_struct *tsk) 798 { 799 if (!unhandled_signal(tsk, SIGSEGV)) 800 return; 801 802 if (!printk_ratelimit()) 803 return; 804 805 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx", 806 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG, 807 tsk->comm, task_pid_nr(tsk), address, 808 (void *)regs->ip, (void *)regs->sp, error_code); 809 810 print_vma_addr(KERN_CONT " in ", regs->ip); 811 812 printk(KERN_CONT "\n"); 813 } 814 815 static void 816 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 817 unsigned long address, struct vm_area_struct *vma, 818 int si_code) 819 { 820 struct task_struct *tsk = current; 821 822 /* User mode accesses just cause a SIGSEGV */ 823 if (error_code & PF_USER) { 824 /* 825 * It's possible to have interrupts off here: 826 */ 827 local_irq_enable(); 828 829 /* 830 * Valid to do another page fault here because this one came 831 * from user space: 832 */ 833 if (is_prefetch(regs, error_code, address)) 834 return; 835 836 if (is_errata100(regs, address)) 837 return; 838 839 #ifdef CONFIG_X86_64 840 /* 841 * Instruction fetch faults in the vsyscall page might need 842 * emulation. 843 */ 844 if (unlikely((error_code & PF_INSTR) && 845 ((address & ~0xfff) == VSYSCALL_ADDR))) { 846 if (emulate_vsyscall(regs, address)) 847 return; 848 } 849 #endif 850 /* Kernel addresses are always protection faults: */ 851 if (address >= TASK_SIZE) 852 error_code |= PF_PROT; 853 854 if (likely(show_unhandled_signals)) 855 show_signal_msg(regs, error_code, address, tsk); 856 857 tsk->thread.cr2 = address; 858 tsk->thread.error_code = error_code; 859 tsk->thread.trap_nr = X86_TRAP_PF; 860 861 force_sig_info_fault(SIGSEGV, si_code, address, tsk, vma, 0); 862 863 return; 864 } 865 866 if (is_f00f_bug(regs, address)) 867 return; 868 869 no_context(regs, error_code, address, SIGSEGV, si_code); 870 } 871 872 static noinline void 873 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 874 unsigned long address, struct vm_area_struct *vma) 875 { 876 __bad_area_nosemaphore(regs, error_code, address, vma, SEGV_MAPERR); 877 } 878 879 static void 880 __bad_area(struct pt_regs *regs, unsigned long error_code, 881 unsigned long address, struct vm_area_struct *vma, int si_code) 882 { 883 struct mm_struct *mm = current->mm; 884 885 /* 886 * Something tried to access memory that isn't in our memory map.. 887 * Fix it, but check if it's kernel or user first.. 888 */ 889 up_read(&mm->mmap_sem); 890 891 __bad_area_nosemaphore(regs, error_code, address, vma, si_code); 892 } 893 894 static noinline void 895 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address) 896 { 897 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR); 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 (boot_cpu_has(X86_FEATURE_OSPKE) && (error_code & PF_PK)) 910 __bad_area(regs, error_code, address, vma, SEGV_PKUERR); 911 else 912 __bad_area(regs, error_code, address, vma, SEGV_ACCERR); 913 } 914 915 static void 916 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address, 917 struct vm_area_struct *vma, unsigned int fault) 918 { 919 struct task_struct *tsk = current; 920 int code = BUS_ADRERR; 921 922 /* Kernel mode? Handle exceptions or die: */ 923 if (!(error_code & PF_USER)) { 924 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR); 925 return; 926 } 927 928 /* User-space => ok to do another page fault: */ 929 if (is_prefetch(regs, error_code, address)) 930 return; 931 932 tsk->thread.cr2 = address; 933 tsk->thread.error_code = error_code; 934 tsk->thread.trap_nr = X86_TRAP_PF; 935 936 #ifdef CONFIG_MEMORY_FAILURE 937 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) { 938 printk(KERN_ERR 939 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n", 940 tsk->comm, tsk->pid, address); 941 code = BUS_MCEERR_AR; 942 } 943 #endif 944 force_sig_info_fault(SIGBUS, code, address, tsk, vma, fault); 945 } 946 947 static noinline void 948 mm_fault_error(struct pt_regs *regs, unsigned long error_code, 949 unsigned long address, struct vm_area_struct *vma, 950 unsigned int fault) 951 { 952 if (fatal_signal_pending(current) && !(error_code & PF_USER)) { 953 no_context(regs, error_code, address, 0, 0); 954 return; 955 } 956 957 if (fault & VM_FAULT_OOM) { 958 /* Kernel mode? Handle exceptions or die: */ 959 if (!(error_code & PF_USER)) { 960 no_context(regs, error_code, address, 961 SIGSEGV, SEGV_MAPERR); 962 return; 963 } 964 965 /* 966 * We ran out of memory, call the OOM killer, and return the 967 * userspace (which will retry the fault, or kill us if we got 968 * oom-killed): 969 */ 970 pagefault_out_of_memory(); 971 } else { 972 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON| 973 VM_FAULT_HWPOISON_LARGE)) 974 do_sigbus(regs, error_code, address, vma, fault); 975 else if (fault & VM_FAULT_SIGSEGV) 976 bad_area_nosemaphore(regs, error_code, address, vma); 977 else 978 BUG(); 979 } 980 } 981 982 static int spurious_fault_check(unsigned long error_code, pte_t *pte) 983 { 984 if ((error_code & PF_WRITE) && !pte_write(*pte)) 985 return 0; 986 987 if ((error_code & PF_INSTR) && !pte_exec(*pte)) 988 return 0; 989 /* 990 * Note: We do not do lazy flushing on protection key 991 * changes, so no spurious fault will ever set PF_PK. 992 */ 993 if ((error_code & PF_PK)) 994 return 1; 995 996 return 1; 997 } 998 999 /* 1000 * Handle a spurious fault caused by a stale TLB entry. 1001 * 1002 * This allows us to lazily refresh the TLB when increasing the 1003 * permissions of a kernel page (RO -> RW or NX -> X). Doing it 1004 * eagerly is very expensive since that implies doing a full 1005 * cross-processor TLB flush, even if no stale TLB entries exist 1006 * on other processors. 1007 * 1008 * Spurious faults may only occur if the TLB contains an entry with 1009 * fewer permission than the page table entry. Non-present (P = 0) 1010 * and reserved bit (R = 1) faults are never spurious. 1011 * 1012 * There are no security implications to leaving a stale TLB when 1013 * increasing the permissions on a page. 1014 * 1015 * Returns non-zero if a spurious fault was handled, zero otherwise. 1016 * 1017 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 1018 * (Optional Invalidation). 1019 */ 1020 static noinline int 1021 spurious_fault(unsigned long error_code, unsigned long address) 1022 { 1023 pgd_t *pgd; 1024 pud_t *pud; 1025 pmd_t *pmd; 1026 pte_t *pte; 1027 int ret; 1028 1029 /* 1030 * Only writes to RO or instruction fetches from NX may cause 1031 * spurious faults. 1032 * 1033 * These could be from user or supervisor accesses but the TLB 1034 * is only lazily flushed after a kernel mapping protection 1035 * change, so user accesses are not expected to cause spurious 1036 * faults. 1037 */ 1038 if (error_code != (PF_WRITE | PF_PROT) 1039 && error_code != (PF_INSTR | PF_PROT)) 1040 return 0; 1041 1042 pgd = init_mm.pgd + pgd_index(address); 1043 if (!pgd_present(*pgd)) 1044 return 0; 1045 1046 pud = pud_offset(pgd, address); 1047 if (!pud_present(*pud)) 1048 return 0; 1049 1050 if (pud_large(*pud)) 1051 return spurious_fault_check(error_code, (pte_t *) pud); 1052 1053 pmd = pmd_offset(pud, address); 1054 if (!pmd_present(*pmd)) 1055 return 0; 1056 1057 if (pmd_large(*pmd)) 1058 return spurious_fault_check(error_code, (pte_t *) pmd); 1059 1060 pte = pte_offset_kernel(pmd, address); 1061 if (!pte_present(*pte)) 1062 return 0; 1063 1064 ret = spurious_fault_check(error_code, pte); 1065 if (!ret) 1066 return 0; 1067 1068 /* 1069 * Make sure we have permissions in PMD. 1070 * If not, then there's a bug in the page tables: 1071 */ 1072 ret = spurious_fault_check(error_code, (pte_t *) pmd); 1073 WARN_ONCE(!ret, "PMD has incorrect permission bits\n"); 1074 1075 return ret; 1076 } 1077 NOKPROBE_SYMBOL(spurious_fault); 1078 1079 int show_unhandled_signals = 1; 1080 1081 static inline int 1082 access_error(unsigned long error_code, struct vm_area_struct *vma) 1083 { 1084 if (error_code & PF_WRITE) { 1085 /* write, present and write, not present: */ 1086 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1087 return 1; 1088 return 0; 1089 } 1090 1091 /* read, present: */ 1092 if (unlikely(error_code & PF_PROT)) 1093 return 1; 1094 1095 /* read, not present: */ 1096 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))) 1097 return 1; 1098 1099 return 0; 1100 } 1101 1102 static int fault_in_kernel_space(unsigned long address) 1103 { 1104 return address >= TASK_SIZE_MAX; 1105 } 1106 1107 static inline bool smap_violation(int error_code, struct pt_regs *regs) 1108 { 1109 if (!IS_ENABLED(CONFIG_X86_SMAP)) 1110 return false; 1111 1112 if (!static_cpu_has(X86_FEATURE_SMAP)) 1113 return false; 1114 1115 if (error_code & PF_USER) 1116 return false; 1117 1118 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC)) 1119 return false; 1120 1121 return true; 1122 } 1123 1124 /* 1125 * This routine handles page faults. It determines the address, 1126 * and the problem, and then passes it off to one of the appropriate 1127 * routines. 1128 * 1129 * This function must have noinline because both callers 1130 * {,trace_}do_page_fault() have notrace on. Having this an actual function 1131 * guarantees there's a function trace entry. 1132 */ 1133 static noinline void 1134 __do_page_fault(struct pt_regs *regs, unsigned long error_code, 1135 unsigned long address) 1136 { 1137 struct vm_area_struct *vma; 1138 struct task_struct *tsk; 1139 struct mm_struct *mm; 1140 int fault, major = 0; 1141 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; 1142 1143 tsk = current; 1144 mm = tsk->mm; 1145 1146 /* 1147 * Detect and handle instructions that would cause a page fault for 1148 * both a tracked kernel page and a userspace page. 1149 */ 1150 if (kmemcheck_active(regs)) 1151 kmemcheck_hide(regs); 1152 prefetchw(&mm->mmap_sem); 1153 1154 if (unlikely(kmmio_fault(regs, address))) 1155 return; 1156 1157 /* 1158 * We fault-in kernel-space virtual memory on-demand. The 1159 * 'reference' page table is init_mm.pgd. 1160 * 1161 * NOTE! We MUST NOT take any locks for this case. We may 1162 * be in an interrupt or a critical region, and should 1163 * only copy the information from the master page table, 1164 * nothing more. 1165 * 1166 * This verifies that the fault happens in kernel space 1167 * (error_code & 4) == 0, and that the fault was not a 1168 * protection error (error_code & 9) == 0. 1169 */ 1170 if (unlikely(fault_in_kernel_space(address))) { 1171 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) { 1172 if (vmalloc_fault(address) >= 0) 1173 return; 1174 1175 if (kmemcheck_fault(regs, address, error_code)) 1176 return; 1177 } 1178 1179 /* Can handle a stale RO->RW TLB: */ 1180 if (spurious_fault(error_code, address)) 1181 return; 1182 1183 /* kprobes don't want to hook the spurious faults: */ 1184 if (kprobes_fault(regs)) 1185 return; 1186 /* 1187 * Don't take the mm semaphore here. If we fixup a prefetch 1188 * fault we could otherwise deadlock: 1189 */ 1190 bad_area_nosemaphore(regs, error_code, address, NULL); 1191 1192 return; 1193 } 1194 1195 /* kprobes don't want to hook the spurious faults: */ 1196 if (unlikely(kprobes_fault(regs))) 1197 return; 1198 1199 if (unlikely(error_code & PF_RSVD)) 1200 pgtable_bad(regs, error_code, address); 1201 1202 if (unlikely(smap_violation(error_code, regs))) { 1203 bad_area_nosemaphore(regs, error_code, address, NULL); 1204 return; 1205 } 1206 1207 /* 1208 * If we're in an interrupt, have no user context or are running 1209 * in a region with pagefaults disabled then we must not take the fault 1210 */ 1211 if (unlikely(faulthandler_disabled() || !mm)) { 1212 bad_area_nosemaphore(regs, error_code, address, NULL); 1213 return; 1214 } 1215 1216 /* 1217 * It's safe to allow irq's after cr2 has been saved and the 1218 * vmalloc fault has been handled. 1219 * 1220 * User-mode registers count as a user access even for any 1221 * potential system fault or CPU buglet: 1222 */ 1223 if (user_mode(regs)) { 1224 local_irq_enable(); 1225 error_code |= PF_USER; 1226 flags |= FAULT_FLAG_USER; 1227 } else { 1228 if (regs->flags & X86_EFLAGS_IF) 1229 local_irq_enable(); 1230 } 1231 1232 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); 1233 1234 if (error_code & PF_WRITE) 1235 flags |= FAULT_FLAG_WRITE; 1236 1237 /* 1238 * When running in the kernel we expect faults to occur only to 1239 * addresses in user space. All other faults represent errors in 1240 * the kernel and should generate an OOPS. Unfortunately, in the 1241 * case of an erroneous fault occurring in a code path which already 1242 * holds mmap_sem we will deadlock attempting to validate the fault 1243 * against the address space. Luckily the kernel only validly 1244 * references user space from well defined areas of code, which are 1245 * listed in the exceptions table. 1246 * 1247 * As the vast majority of faults will be valid we will only perform 1248 * the source reference check when there is a possibility of a 1249 * deadlock. Attempt to lock the address space, if we cannot we then 1250 * validate the source. If this is invalid we can skip the address 1251 * space check, thus avoiding the deadlock: 1252 */ 1253 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 1254 if ((error_code & PF_USER) == 0 && 1255 !search_exception_tables(regs->ip)) { 1256 bad_area_nosemaphore(regs, error_code, address, NULL); 1257 return; 1258 } 1259 retry: 1260 down_read(&mm->mmap_sem); 1261 } else { 1262 /* 1263 * The above down_read_trylock() might have succeeded in 1264 * which case we'll have missed the might_sleep() from 1265 * down_read(): 1266 */ 1267 might_sleep(); 1268 } 1269 1270 vma = find_vma(mm, address); 1271 if (unlikely(!vma)) { 1272 bad_area(regs, error_code, address); 1273 return; 1274 } 1275 if (likely(vma->vm_start <= address)) 1276 goto good_area; 1277 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) { 1278 bad_area(regs, error_code, address); 1279 return; 1280 } 1281 if (error_code & PF_USER) { 1282 /* 1283 * Accessing the stack below %sp is always a bug. 1284 * The large cushion allows instructions like enter 1285 * and pusha to work. ("enter $65535, $31" pushes 1286 * 32 pointers and then decrements %sp by 65535.) 1287 */ 1288 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) { 1289 bad_area(regs, error_code, address); 1290 return; 1291 } 1292 } 1293 if (unlikely(expand_stack(vma, address))) { 1294 bad_area(regs, error_code, address); 1295 return; 1296 } 1297 1298 /* 1299 * Ok, we have a good vm_area for this memory access, so 1300 * we can handle it.. 1301 */ 1302 good_area: 1303 if (unlikely(access_error(error_code, vma))) { 1304 bad_area_access_error(regs, error_code, address, vma); 1305 return; 1306 } 1307 1308 /* 1309 * If for any reason at all we couldn't handle the fault, 1310 * make sure we exit gracefully rather than endlessly redo 1311 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if 1312 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked. 1313 */ 1314 fault = handle_mm_fault(mm, vma, address, flags); 1315 major |= fault & VM_FAULT_MAJOR; 1316 1317 /* 1318 * If we need to retry the mmap_sem has already been released, 1319 * and if there is a fatal signal pending there is no guarantee 1320 * that we made any progress. Handle this case first. 1321 */ 1322 if (unlikely(fault & VM_FAULT_RETRY)) { 1323 /* Retry at most once */ 1324 if (flags & FAULT_FLAG_ALLOW_RETRY) { 1325 flags &= ~FAULT_FLAG_ALLOW_RETRY; 1326 flags |= FAULT_FLAG_TRIED; 1327 if (!fatal_signal_pending(tsk)) 1328 goto retry; 1329 } 1330 1331 /* User mode? Just return to handle the fatal exception */ 1332 if (flags & FAULT_FLAG_USER) 1333 return; 1334 1335 /* Not returning to user mode? Handle exceptions or die: */ 1336 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR); 1337 return; 1338 } 1339 1340 up_read(&mm->mmap_sem); 1341 if (unlikely(fault & VM_FAULT_ERROR)) { 1342 mm_fault_error(regs, error_code, address, vma, fault); 1343 return; 1344 } 1345 1346 /* 1347 * Major/minor page fault accounting. If any of the events 1348 * returned VM_FAULT_MAJOR, we account it as a major fault. 1349 */ 1350 if (major) { 1351 tsk->maj_flt++; 1352 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address); 1353 } else { 1354 tsk->min_flt++; 1355 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address); 1356 } 1357 1358 check_v8086_mode(regs, address, tsk); 1359 } 1360 NOKPROBE_SYMBOL(__do_page_fault); 1361 1362 dotraplinkage void notrace 1363 do_page_fault(struct pt_regs *regs, unsigned long error_code) 1364 { 1365 unsigned long address = read_cr2(); /* Get the faulting address */ 1366 enum ctx_state prev_state; 1367 1368 /* 1369 * We must have this function tagged with __kprobes, notrace and call 1370 * read_cr2() before calling anything else. To avoid calling any kind 1371 * of tracing machinery before we've observed the CR2 value. 1372 * 1373 * exception_{enter,exit}() contain all sorts of tracepoints. 1374 */ 1375 1376 prev_state = exception_enter(); 1377 __do_page_fault(regs, error_code, address); 1378 exception_exit(prev_state); 1379 } 1380 NOKPROBE_SYMBOL(do_page_fault); 1381 1382 #ifdef CONFIG_TRACING 1383 static nokprobe_inline void 1384 trace_page_fault_entries(unsigned long address, struct pt_regs *regs, 1385 unsigned long error_code) 1386 { 1387 if (user_mode(regs)) 1388 trace_page_fault_user(address, regs, error_code); 1389 else 1390 trace_page_fault_kernel(address, regs, error_code); 1391 } 1392 1393 dotraplinkage void notrace 1394 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code) 1395 { 1396 /* 1397 * The exception_enter and tracepoint processing could 1398 * trigger another page faults (user space callchain 1399 * reading) and destroy the original cr2 value, so read 1400 * the faulting address now. 1401 */ 1402 unsigned long address = read_cr2(); 1403 enum ctx_state prev_state; 1404 1405 prev_state = exception_enter(); 1406 trace_page_fault_entries(address, regs, error_code); 1407 __do_page_fault(regs, error_code, address); 1408 exception_exit(prev_state); 1409 } 1410 NOKPROBE_SYMBOL(trace_do_page_fault); 1411 #endif /* CONFIG_TRACING */ 1412