1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 1995 Linus Torvalds 4 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs. 5 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar 6 */ 7 #include <linux/sched.h> /* test_thread_flag(), ... */ 8 #include <linux/sched/task_stack.h> /* task_stack_*(), ... */ 9 #include <linux/kdebug.h> /* oops_begin/end, ... */ 10 #include <linux/extable.h> /* search_exception_tables */ 11 #include <linux/memblock.h> /* max_low_pfn */ 12 #include <linux/kfence.h> /* kfence_handle_page_fault */ 13 #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */ 14 #include <linux/mmiotrace.h> /* kmmio_handler, ... */ 15 #include <linux/perf_event.h> /* perf_sw_event */ 16 #include <linux/hugetlb.h> /* hstate_index_to_shift */ 17 #include <linux/prefetch.h> /* prefetchw */ 18 #include <linux/context_tracking.h> /* exception_enter(), ... */ 19 #include <linux/uaccess.h> /* faulthandler_disabled() */ 20 #include <linux/efi.h> /* efi_crash_gracefully_on_page_fault()*/ 21 #include <linux/mm_types.h> 22 #include <linux/mm.h> /* find_and_lock_vma() */ 23 24 #include <asm/cpufeature.h> /* boot_cpu_has, ... */ 25 #include <asm/traps.h> /* dotraplinkage, ... */ 26 #include <asm/fixmap.h> /* VSYSCALL_ADDR */ 27 #include <asm/vsyscall.h> /* emulate_vsyscall */ 28 #include <asm/vm86.h> /* struct vm86 */ 29 #include <asm/mmu_context.h> /* vma_pkey() */ 30 #include <asm/efi.h> /* efi_crash_gracefully_on_page_fault()*/ 31 #include <asm/desc.h> /* store_idt(), ... */ 32 #include <asm/cpu_entry_area.h> /* exception stack */ 33 #include <asm/pgtable_areas.h> /* VMALLOC_START, ... */ 34 #include <asm/kvm_para.h> /* kvm_handle_async_pf */ 35 #include <asm/vdso.h> /* fixup_vdso_exception() */ 36 #include <asm/irq_stack.h> 37 #include <asm/fred.h> 38 39 #define CREATE_TRACE_POINTS 40 #include <asm/trace/exceptions.h> 41 42 /* 43 * Returns 0 if mmiotrace is disabled, or if the fault is not 44 * handled by mmiotrace: 45 */ 46 static nokprobe_inline int 47 kmmio_fault(struct pt_regs *regs, unsigned long addr) 48 { 49 if (unlikely(is_kmmio_active())) 50 if (kmmio_handler(regs, addr) == 1) 51 return -1; 52 return 0; 53 } 54 55 /* 56 * Prefetch quirks: 57 * 58 * 32-bit mode: 59 * 60 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch. 61 * Check that here and ignore it. This is AMD erratum #91. 62 * 63 * 64-bit mode: 64 * 65 * Sometimes the CPU reports invalid exceptions on prefetch. 66 * Check that here and ignore it. 67 * 68 * Opcode checker based on code by Richard Brunner. 69 */ 70 static inline int 71 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr, 72 unsigned char opcode, int *prefetch) 73 { 74 unsigned char instr_hi = opcode & 0xf0; 75 unsigned char instr_lo = opcode & 0x0f; 76 77 switch (instr_hi) { 78 case 0x20: 79 case 0x30: 80 /* 81 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. 82 * In X86_64 long mode, the CPU will signal invalid 83 * opcode if some of these prefixes are present so 84 * X86_64 will never get here anyway 85 */ 86 return ((instr_lo & 7) == 0x6); 87 #ifdef CONFIG_X86_64 88 case 0x40: 89 /* 90 * In 64-bit mode 0x40..0x4F are valid REX prefixes 91 */ 92 return (!user_mode(regs) || user_64bit_mode(regs)); 93 #endif 94 case 0x60: 95 /* 0x64 thru 0x67 are valid prefixes in all modes. */ 96 return (instr_lo & 0xC) == 0x4; 97 case 0xF0: 98 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */ 99 return !instr_lo || (instr_lo>>1) == 1; 100 case 0x00: 101 /* Prefetch instruction is 0x0F0D or 0x0F18 */ 102 if (get_kernel_nofault(opcode, instr)) 103 return 0; 104 105 *prefetch = (instr_lo == 0xF) && 106 (opcode == 0x0D || opcode == 0x18); 107 return 0; 108 default: 109 return 0; 110 } 111 } 112 113 static bool is_amd_k8_pre_npt(void) 114 { 115 struct cpuinfo_x86 *c = &boot_cpu_data; 116 117 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) && 118 c->x86_vendor == X86_VENDOR_AMD && 119 c->x86 == 0xf && c->x86_model < 0x40); 120 } 121 122 static int 123 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr) 124 { 125 unsigned char *max_instr; 126 unsigned char *instr; 127 int prefetch = 0; 128 129 /* Erratum #91 affects AMD K8, pre-NPT CPUs */ 130 if (!is_amd_k8_pre_npt()) 131 return 0; 132 133 /* 134 * If it was a exec (instruction fetch) fault on NX page, then 135 * do not ignore the fault: 136 */ 137 if (error_code & X86_PF_INSTR) 138 return 0; 139 140 instr = (void *)convert_ip_to_linear(current, regs); 141 max_instr = instr + 15; 142 143 /* 144 * This code has historically always bailed out if IP points to a 145 * not-present page (e.g. due to a race). No one has ever 146 * complained about this. 147 */ 148 pagefault_disable(); 149 150 while (instr < max_instr) { 151 unsigned char opcode; 152 153 if (user_mode(regs)) { 154 if (get_user(opcode, (unsigned char __user *) instr)) 155 break; 156 } else { 157 if (get_kernel_nofault(opcode, instr)) 158 break; 159 } 160 161 instr++; 162 163 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch)) 164 break; 165 } 166 167 pagefault_enable(); 168 return prefetch; 169 } 170 171 DEFINE_SPINLOCK(pgd_lock); 172 LIST_HEAD(pgd_list); 173 174 #ifdef CONFIG_X86_32 175 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) 176 { 177 unsigned index = pgd_index(address); 178 pgd_t *pgd_k; 179 p4d_t *p4d, *p4d_k; 180 pud_t *pud, *pud_k; 181 pmd_t *pmd, *pmd_k; 182 183 pgd += index; 184 pgd_k = init_mm.pgd + index; 185 186 if (!pgd_present(*pgd_k)) 187 return NULL; 188 189 /* 190 * set_pgd(pgd, *pgd_k); here would be useless on PAE 191 * and redundant with the set_pmd() on non-PAE. As would 192 * set_p4d/set_pud. 193 */ 194 p4d = p4d_offset(pgd, address); 195 p4d_k = p4d_offset(pgd_k, address); 196 if (!p4d_present(*p4d_k)) 197 return NULL; 198 199 pud = pud_offset(p4d, address); 200 pud_k = pud_offset(p4d_k, address); 201 if (!pud_present(*pud_k)) 202 return NULL; 203 204 pmd = pmd_offset(pud, address); 205 pmd_k = pmd_offset(pud_k, address); 206 207 if (pmd_present(*pmd) != pmd_present(*pmd_k)) 208 set_pmd(pmd, *pmd_k); 209 210 if (!pmd_present(*pmd_k)) 211 return NULL; 212 else 213 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k)); 214 215 return pmd_k; 216 } 217 218 /* 219 * Handle a fault on the vmalloc or module mapping area 220 * 221 * This is needed because there is a race condition between the time 222 * when the vmalloc mapping code updates the PMD to the point in time 223 * where it synchronizes this update with the other page-tables in the 224 * system. 225 * 226 * In this race window another thread/CPU can map an area on the same 227 * PMD, finds it already present and does not synchronize it with the 228 * rest of the system yet. As a result v[mz]alloc might return areas 229 * which are not mapped in every page-table in the system, causing an 230 * unhandled page-fault when they are accessed. 231 */ 232 static noinline int vmalloc_fault(unsigned long address) 233 { 234 unsigned long pgd_paddr; 235 pmd_t *pmd_k; 236 pte_t *pte_k; 237 238 /* Make sure we are in vmalloc area: */ 239 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 240 return -1; 241 242 /* 243 * Synchronize this task's top level page-table 244 * with the 'reference' page table. 245 * 246 * Do _not_ use "current" here. We might be inside 247 * an interrupt in the middle of a task switch.. 248 */ 249 pgd_paddr = read_cr3_pa(); 250 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address); 251 if (!pmd_k) 252 return -1; 253 254 if (pmd_large(*pmd_k)) 255 return 0; 256 257 pte_k = pte_offset_kernel(pmd_k, address); 258 if (!pte_present(*pte_k)) 259 return -1; 260 261 return 0; 262 } 263 NOKPROBE_SYMBOL(vmalloc_fault); 264 265 void arch_sync_kernel_mappings(unsigned long start, unsigned long end) 266 { 267 unsigned long addr; 268 269 for (addr = start & PMD_MASK; 270 addr >= TASK_SIZE_MAX && addr < VMALLOC_END; 271 addr += PMD_SIZE) { 272 struct page *page; 273 274 spin_lock(&pgd_lock); 275 list_for_each_entry(page, &pgd_list, lru) { 276 spinlock_t *pgt_lock; 277 278 /* the pgt_lock only for Xen */ 279 pgt_lock = &pgd_page_get_mm(page)->page_table_lock; 280 281 spin_lock(pgt_lock); 282 vmalloc_sync_one(page_address(page), addr); 283 spin_unlock(pgt_lock); 284 } 285 spin_unlock(&pgd_lock); 286 } 287 } 288 289 static bool low_pfn(unsigned long pfn) 290 { 291 return pfn < max_low_pfn; 292 } 293 294 static void dump_pagetable(unsigned long address) 295 { 296 pgd_t *base = __va(read_cr3_pa()); 297 pgd_t *pgd = &base[pgd_index(address)]; 298 p4d_t *p4d; 299 pud_t *pud; 300 pmd_t *pmd; 301 pte_t *pte; 302 303 #ifdef CONFIG_X86_PAE 304 pr_info("*pdpt = %016Lx ", pgd_val(*pgd)); 305 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd)) 306 goto out; 307 #define pr_pde pr_cont 308 #else 309 #define pr_pde pr_info 310 #endif 311 p4d = p4d_offset(pgd, address); 312 pud = pud_offset(p4d, address); 313 pmd = pmd_offset(pud, address); 314 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd)); 315 #undef pr_pde 316 317 /* 318 * We must not directly access the pte in the highpte 319 * case if the page table is located in highmem. 320 * And let's rather not kmap-atomic the pte, just in case 321 * it's allocated already: 322 */ 323 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd)) 324 goto out; 325 326 pte = pte_offset_kernel(pmd, address); 327 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte)); 328 out: 329 pr_cont("\n"); 330 } 331 332 #else /* CONFIG_X86_64: */ 333 334 #ifdef CONFIG_CPU_SUP_AMD 335 static const char errata93_warning[] = 336 KERN_ERR 337 "******* Your BIOS seems to not contain a fix for K8 errata #93\n" 338 "******* Working around it, but it may cause SEGVs or burn power.\n" 339 "******* Please consider a BIOS update.\n" 340 "******* Disabling USB legacy in the BIOS may also help.\n"; 341 #endif 342 343 static int bad_address(void *p) 344 { 345 unsigned long dummy; 346 347 return get_kernel_nofault(dummy, (unsigned long *)p); 348 } 349 350 static void dump_pagetable(unsigned long address) 351 { 352 pgd_t *base = __va(read_cr3_pa()); 353 pgd_t *pgd = base + pgd_index(address); 354 p4d_t *p4d; 355 pud_t *pud; 356 pmd_t *pmd; 357 pte_t *pte; 358 359 if (bad_address(pgd)) 360 goto bad; 361 362 pr_info("PGD %lx ", pgd_val(*pgd)); 363 364 if (!pgd_present(*pgd)) 365 goto out; 366 367 p4d = p4d_offset(pgd, address); 368 if (bad_address(p4d)) 369 goto bad; 370 371 pr_cont("P4D %lx ", p4d_val(*p4d)); 372 if (!p4d_present(*p4d) || p4d_large(*p4d)) 373 goto out; 374 375 pud = pud_offset(p4d, address); 376 if (bad_address(pud)) 377 goto bad; 378 379 pr_cont("PUD %lx ", pud_val(*pud)); 380 if (!pud_present(*pud) || pud_large(*pud)) 381 goto out; 382 383 pmd = pmd_offset(pud, address); 384 if (bad_address(pmd)) 385 goto bad; 386 387 pr_cont("PMD %lx ", pmd_val(*pmd)); 388 if (!pmd_present(*pmd) || pmd_large(*pmd)) 389 goto out; 390 391 pte = pte_offset_kernel(pmd, address); 392 if (bad_address(pte)) 393 goto bad; 394 395 pr_cont("PTE %lx", pte_val(*pte)); 396 out: 397 pr_cont("\n"); 398 return; 399 bad: 400 pr_info("BAD\n"); 401 } 402 403 #endif /* CONFIG_X86_64 */ 404 405 /* 406 * Workaround for K8 erratum #93 & buggy BIOS. 407 * 408 * BIOS SMM functions are required to use a specific workaround 409 * to avoid corruption of the 64bit RIP register on C stepping K8. 410 * 411 * A lot of BIOS that didn't get tested properly miss this. 412 * 413 * The OS sees this as a page fault with the upper 32bits of RIP cleared. 414 * Try to work around it here. 415 * 416 * Note we only handle faults in kernel here. 417 * Does nothing on 32-bit. 418 */ 419 static int is_errata93(struct pt_regs *regs, unsigned long address) 420 { 421 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD) 422 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD 423 || boot_cpu_data.x86 != 0xf) 424 return 0; 425 426 if (user_mode(regs)) 427 return 0; 428 429 if (address != regs->ip) 430 return 0; 431 432 if ((address >> 32) != 0) 433 return 0; 434 435 address |= 0xffffffffUL << 32; 436 if ((address >= (u64)_stext && address <= (u64)_etext) || 437 (address >= MODULES_VADDR && address <= MODULES_END)) { 438 printk_once(errata93_warning); 439 regs->ip = address; 440 return 1; 441 } 442 #endif 443 return 0; 444 } 445 446 /* 447 * Work around K8 erratum #100 K8 in compat mode occasionally jumps 448 * to illegal addresses >4GB. 449 * 450 * We catch this in the page fault handler because these addresses 451 * are not reachable. Just detect this case and return. Any code 452 * segment in LDT is compatibility mode. 453 */ 454 static int is_errata100(struct pt_regs *regs, unsigned long address) 455 { 456 #ifdef CONFIG_X86_64 457 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32)) 458 return 1; 459 #endif 460 return 0; 461 } 462 463 /* Pentium F0 0F C7 C8 bug workaround: */ 464 static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code, 465 unsigned long address) 466 { 467 #ifdef CONFIG_X86_F00F_BUG 468 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) && 469 idt_is_f00f_address(address)) { 470 handle_invalid_op(regs); 471 return 1; 472 } 473 #endif 474 return 0; 475 } 476 477 static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index) 478 { 479 u32 offset = (index >> 3) * sizeof(struct desc_struct); 480 unsigned long addr; 481 struct ldttss_desc desc; 482 483 if (index == 0) { 484 pr_alert("%s: NULL\n", name); 485 return; 486 } 487 488 if (offset + sizeof(struct ldttss_desc) >= gdt->size) { 489 pr_alert("%s: 0x%hx -- out of bounds\n", name, index); 490 return; 491 } 492 493 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset), 494 sizeof(struct ldttss_desc))) { 495 pr_alert("%s: 0x%hx -- GDT entry is not readable\n", 496 name, index); 497 return; 498 } 499 500 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24); 501 #ifdef CONFIG_X86_64 502 addr |= ((u64)desc.base3 << 32); 503 #endif 504 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n", 505 name, index, addr, (desc.limit0 | (desc.limit1 << 16))); 506 } 507 508 static void 509 show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address) 510 { 511 if (!oops_may_print()) 512 return; 513 514 if (error_code & X86_PF_INSTR) { 515 unsigned int level; 516 pgd_t *pgd; 517 pte_t *pte; 518 519 pgd = __va(read_cr3_pa()); 520 pgd += pgd_index(address); 521 522 pte = lookup_address_in_pgd(pgd, address, &level); 523 524 if (pte && pte_present(*pte) && !pte_exec(*pte)) 525 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", 526 from_kuid(&init_user_ns, current_uid())); 527 if (pte && pte_present(*pte) && pte_exec(*pte) && 528 (pgd_flags(*pgd) & _PAGE_USER) && 529 (__read_cr4() & X86_CR4_SMEP)) 530 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n", 531 from_kuid(&init_user_ns, current_uid())); 532 } 533 534 if (address < PAGE_SIZE && !user_mode(regs)) 535 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n", 536 (void *)address); 537 else 538 pr_alert("BUG: unable to handle page fault for address: %px\n", 539 (void *)address); 540 541 pr_alert("#PF: %s %s in %s mode\n", 542 (error_code & X86_PF_USER) ? "user" : "supervisor", 543 (error_code & X86_PF_INSTR) ? "instruction fetch" : 544 (error_code & X86_PF_WRITE) ? "write access" : 545 "read access", 546 user_mode(regs) ? "user" : "kernel"); 547 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code, 548 !(error_code & X86_PF_PROT) ? "not-present page" : 549 (error_code & X86_PF_RSVD) ? "reserved bit violation" : 550 (error_code & X86_PF_PK) ? "protection keys violation" : 551 "permissions violation"); 552 553 if (!(error_code & X86_PF_USER) && user_mode(regs)) { 554 struct desc_ptr idt, gdt; 555 u16 ldtr, tr; 556 557 /* 558 * This can happen for quite a few reasons. The more obvious 559 * ones are faults accessing the GDT, or LDT. Perhaps 560 * surprisingly, if the CPU tries to deliver a benign or 561 * contributory exception from user code and gets a page fault 562 * during delivery, the page fault can be delivered as though 563 * it originated directly from user code. This could happen 564 * due to wrong permissions on the IDT, GDT, LDT, TSS, or 565 * kernel or IST stack. 566 */ 567 store_idt(&idt); 568 569 /* Usable even on Xen PV -- it's just slow. */ 570 native_store_gdt(&gdt); 571 572 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n", 573 idt.address, idt.size, gdt.address, gdt.size); 574 575 store_ldt(ldtr); 576 show_ldttss(&gdt, "LDTR", ldtr); 577 578 store_tr(tr); 579 show_ldttss(&gdt, "TR", tr); 580 } 581 582 dump_pagetable(address); 583 } 584 585 static noinline void 586 pgtable_bad(struct pt_regs *regs, unsigned long error_code, 587 unsigned long address) 588 { 589 struct task_struct *tsk; 590 unsigned long flags; 591 int sig; 592 593 flags = oops_begin(); 594 tsk = current; 595 sig = SIGKILL; 596 597 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n", 598 tsk->comm, address); 599 dump_pagetable(address); 600 601 if (__die("Bad pagetable", regs, error_code)) 602 sig = 0; 603 604 oops_end(flags, regs, sig); 605 } 606 607 static void sanitize_error_code(unsigned long address, 608 unsigned long *error_code) 609 { 610 /* 611 * To avoid leaking information about the kernel page 612 * table layout, pretend that user-mode accesses to 613 * kernel addresses are always protection faults. 614 * 615 * NB: This means that failed vsyscalls with vsyscall=none 616 * will have the PROT bit. This doesn't leak any 617 * information and does not appear to cause any problems. 618 */ 619 if (address >= TASK_SIZE_MAX) 620 *error_code |= X86_PF_PROT; 621 } 622 623 static void set_signal_archinfo(unsigned long address, 624 unsigned long error_code) 625 { 626 struct task_struct *tsk = current; 627 628 tsk->thread.trap_nr = X86_TRAP_PF; 629 tsk->thread.error_code = error_code | X86_PF_USER; 630 tsk->thread.cr2 = address; 631 } 632 633 static noinline void 634 page_fault_oops(struct pt_regs *regs, unsigned long error_code, 635 unsigned long address) 636 { 637 #ifdef CONFIG_VMAP_STACK 638 struct stack_info info; 639 #endif 640 unsigned long flags; 641 int sig; 642 643 if (user_mode(regs)) { 644 /* 645 * Implicit kernel access from user mode? Skip the stack 646 * overflow and EFI special cases. 647 */ 648 goto oops; 649 } 650 651 #ifdef CONFIG_VMAP_STACK 652 /* 653 * Stack overflow? During boot, we can fault near the initial 654 * stack in the direct map, but that's not an overflow -- check 655 * that we're in vmalloc space to avoid this. 656 */ 657 if (is_vmalloc_addr((void *)address) && 658 get_stack_guard_info((void *)address, &info)) { 659 /* 660 * We're likely to be running with very little stack space 661 * left. It's plausible that we'd hit this condition but 662 * double-fault even before we get this far, in which case 663 * we're fine: the double-fault handler will deal with it. 664 * 665 * We don't want to make it all the way into the oops code 666 * and then double-fault, though, because we're likely to 667 * break the console driver and lose most of the stack dump. 668 */ 669 call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*), 670 handle_stack_overflow, 671 ASM_CALL_ARG3, 672 , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info)); 673 674 unreachable(); 675 } 676 #endif 677 678 /* 679 * Buggy firmware could access regions which might page fault. If 680 * this happens, EFI has a special OOPS path that will try to 681 * avoid hanging the system. 682 */ 683 if (IS_ENABLED(CONFIG_EFI)) 684 efi_crash_gracefully_on_page_fault(address); 685 686 /* Only not-present faults should be handled by KFENCE. */ 687 if (!(error_code & X86_PF_PROT) && 688 kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs)) 689 return; 690 691 oops: 692 /* 693 * Oops. The kernel tried to access some bad page. We'll have to 694 * terminate things with extreme prejudice: 695 */ 696 flags = oops_begin(); 697 698 show_fault_oops(regs, error_code, address); 699 700 if (task_stack_end_corrupted(current)) 701 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 702 703 sig = SIGKILL; 704 if (__die("Oops", regs, error_code)) 705 sig = 0; 706 707 /* Executive summary in case the body of the oops scrolled away */ 708 printk(KERN_DEFAULT "CR2: %016lx\n", address); 709 710 oops_end(flags, regs, sig); 711 } 712 713 static noinline void 714 kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code, 715 unsigned long address, int signal, int si_code, 716 u32 pkey) 717 { 718 WARN_ON_ONCE(user_mode(regs)); 719 720 /* Are we prepared to handle this kernel fault? */ 721 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) { 722 /* 723 * Any interrupt that takes a fault gets the fixup. This makes 724 * the below recursive fault logic only apply to a faults from 725 * task context. 726 */ 727 if (in_interrupt()) 728 return; 729 730 /* 731 * Per the above we're !in_interrupt(), aka. task context. 732 * 733 * In this case we need to make sure we're not recursively 734 * faulting through the emulate_vsyscall() logic. 735 */ 736 if (current->thread.sig_on_uaccess_err && signal) { 737 sanitize_error_code(address, &error_code); 738 739 set_signal_archinfo(address, error_code); 740 741 if (si_code == SEGV_PKUERR) { 742 force_sig_pkuerr((void __user *)address, pkey); 743 } else { 744 /* XXX: hwpoison faults will set the wrong code. */ 745 force_sig_fault(signal, si_code, (void __user *)address); 746 } 747 } 748 749 /* 750 * Barring that, we can do the fixup and be happy. 751 */ 752 return; 753 } 754 755 /* 756 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH 757 * instruction. 758 */ 759 if (is_prefetch(regs, error_code, address)) 760 return; 761 762 page_fault_oops(regs, error_code, address); 763 } 764 765 /* 766 * Print out info about fatal segfaults, if the show_unhandled_signals 767 * sysctl is set: 768 */ 769 static inline void 770 show_signal_msg(struct pt_regs *regs, unsigned long error_code, 771 unsigned long address, struct task_struct *tsk) 772 { 773 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG; 774 /* This is a racy snapshot, but it's better than nothing. */ 775 int cpu = raw_smp_processor_id(); 776 777 if (!unhandled_signal(tsk, SIGSEGV)) 778 return; 779 780 if (!printk_ratelimit()) 781 return; 782 783 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx", 784 loglvl, tsk->comm, task_pid_nr(tsk), address, 785 (void *)regs->ip, (void *)regs->sp, error_code); 786 787 print_vma_addr(KERN_CONT " in ", regs->ip); 788 789 /* 790 * Dump the likely CPU where the fatal segfault happened. 791 * This can help identify faulty hardware. 792 */ 793 printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu, 794 topology_core_id(cpu), topology_physical_package_id(cpu)); 795 796 797 printk(KERN_CONT "\n"); 798 799 show_opcodes(regs, loglvl); 800 } 801 802 static void 803 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 804 unsigned long address, u32 pkey, int si_code) 805 { 806 struct task_struct *tsk = current; 807 808 if (!user_mode(regs)) { 809 kernelmode_fixup_or_oops(regs, error_code, address, 810 SIGSEGV, si_code, pkey); 811 return; 812 } 813 814 if (!(error_code & X86_PF_USER)) { 815 /* Implicit user access to kernel memory -- just oops */ 816 page_fault_oops(regs, error_code, address); 817 return; 818 } 819 820 /* 821 * User mode accesses just cause a SIGSEGV. 822 * It's possible to have interrupts off here: 823 */ 824 local_irq_enable(); 825 826 /* 827 * Valid to do another page fault here because this one came 828 * from user space: 829 */ 830 if (is_prefetch(regs, error_code, address)) 831 return; 832 833 if (is_errata100(regs, address)) 834 return; 835 836 sanitize_error_code(address, &error_code); 837 838 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address)) 839 return; 840 841 if (likely(show_unhandled_signals)) 842 show_signal_msg(regs, error_code, address, tsk); 843 844 set_signal_archinfo(address, error_code); 845 846 if (si_code == SEGV_PKUERR) 847 force_sig_pkuerr((void __user *)address, pkey); 848 else 849 force_sig_fault(SIGSEGV, si_code, (void __user *)address); 850 851 local_irq_disable(); 852 } 853 854 static noinline void 855 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 856 unsigned long address) 857 { 858 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR); 859 } 860 861 static void 862 __bad_area(struct pt_regs *regs, unsigned long error_code, 863 unsigned long address, u32 pkey, int si_code) 864 { 865 struct mm_struct *mm = current->mm; 866 /* 867 * Something tried to access memory that isn't in our memory map.. 868 * Fix it, but check if it's kernel or user first.. 869 */ 870 mmap_read_unlock(mm); 871 872 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code); 873 } 874 875 static inline bool bad_area_access_from_pkeys(unsigned long error_code, 876 struct vm_area_struct *vma) 877 { 878 /* This code is always called on the current mm */ 879 bool foreign = false; 880 881 if (!cpu_feature_enabled(X86_FEATURE_OSPKE)) 882 return false; 883 if (error_code & X86_PF_PK) 884 return true; 885 /* this checks permission keys on the VMA: */ 886 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 887 (error_code & X86_PF_INSTR), foreign)) 888 return true; 889 return false; 890 } 891 892 static noinline void 893 bad_area_access_error(struct pt_regs *regs, unsigned long error_code, 894 unsigned long address, struct vm_area_struct *vma) 895 { 896 /* 897 * This OSPKE check is not strictly necessary at runtime. 898 * But, doing it this way allows compiler optimizations 899 * if pkeys are compiled out. 900 */ 901 if (bad_area_access_from_pkeys(error_code, vma)) { 902 /* 903 * A protection key fault means that the PKRU value did not allow 904 * access to some PTE. Userspace can figure out what PKRU was 905 * from the XSAVE state. This function captures the pkey from 906 * the vma and passes it to userspace so userspace can discover 907 * which protection key was set on the PTE. 908 * 909 * If we get here, we know that the hardware signaled a X86_PF_PK 910 * fault and that there was a VMA once we got in the fault 911 * handler. It does *not* guarantee that the VMA we find here 912 * was the one that we faulted on. 913 * 914 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4); 915 * 2. T1 : set PKRU to deny access to pkey=4, touches page 916 * 3. T1 : faults... 917 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5); 918 * 5. T1 : enters fault handler, takes mmap_lock, etc... 919 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really 920 * faulted on a pte with its pkey=4. 921 */ 922 u32 pkey = vma_pkey(vma); 923 924 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR); 925 } else { 926 __bad_area(regs, error_code, address, 0, SEGV_ACCERR); 927 } 928 } 929 930 static void 931 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address, 932 vm_fault_t fault) 933 { 934 /* Kernel mode? Handle exceptions or die: */ 935 if (!user_mode(regs)) { 936 kernelmode_fixup_or_oops(regs, error_code, address, 937 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY); 938 return; 939 } 940 941 /* User-space => ok to do another page fault: */ 942 if (is_prefetch(regs, error_code, address)) 943 return; 944 945 sanitize_error_code(address, &error_code); 946 947 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address)) 948 return; 949 950 set_signal_archinfo(address, error_code); 951 952 #ifdef CONFIG_MEMORY_FAILURE 953 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) { 954 struct task_struct *tsk = current; 955 unsigned lsb = 0; 956 957 pr_err( 958 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n", 959 tsk->comm, tsk->pid, address); 960 if (fault & VM_FAULT_HWPOISON_LARGE) 961 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 962 if (fault & VM_FAULT_HWPOISON) 963 lsb = PAGE_SHIFT; 964 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb); 965 return; 966 } 967 #endif 968 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address); 969 } 970 971 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte) 972 { 973 if ((error_code & X86_PF_WRITE) && !pte_write(*pte)) 974 return 0; 975 976 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte)) 977 return 0; 978 979 return 1; 980 } 981 982 /* 983 * Handle a spurious fault caused by a stale TLB entry. 984 * 985 * This allows us to lazily refresh the TLB when increasing the 986 * permissions of a kernel page (RO -> RW or NX -> X). Doing it 987 * eagerly is very expensive since that implies doing a full 988 * cross-processor TLB flush, even if no stale TLB entries exist 989 * on other processors. 990 * 991 * Spurious faults may only occur if the TLB contains an entry with 992 * fewer permission than the page table entry. Non-present (P = 0) 993 * and reserved bit (R = 1) faults are never spurious. 994 * 995 * There are no security implications to leaving a stale TLB when 996 * increasing the permissions on a page. 997 * 998 * Returns non-zero if a spurious fault was handled, zero otherwise. 999 * 1000 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 1001 * (Optional Invalidation). 1002 */ 1003 static noinline int 1004 spurious_kernel_fault(unsigned long error_code, unsigned long address) 1005 { 1006 pgd_t *pgd; 1007 p4d_t *p4d; 1008 pud_t *pud; 1009 pmd_t *pmd; 1010 pte_t *pte; 1011 int ret; 1012 1013 /* 1014 * Only writes to RO or instruction fetches from NX may cause 1015 * spurious faults. 1016 * 1017 * These could be from user or supervisor accesses but the TLB 1018 * is only lazily flushed after a kernel mapping protection 1019 * change, so user accesses are not expected to cause spurious 1020 * faults. 1021 */ 1022 if (error_code != (X86_PF_WRITE | X86_PF_PROT) && 1023 error_code != (X86_PF_INSTR | X86_PF_PROT)) 1024 return 0; 1025 1026 pgd = init_mm.pgd + pgd_index(address); 1027 if (!pgd_present(*pgd)) 1028 return 0; 1029 1030 p4d = p4d_offset(pgd, address); 1031 if (!p4d_present(*p4d)) 1032 return 0; 1033 1034 if (p4d_large(*p4d)) 1035 return spurious_kernel_fault_check(error_code, (pte_t *) p4d); 1036 1037 pud = pud_offset(p4d, address); 1038 if (!pud_present(*pud)) 1039 return 0; 1040 1041 if (pud_large(*pud)) 1042 return spurious_kernel_fault_check(error_code, (pte_t *) pud); 1043 1044 pmd = pmd_offset(pud, address); 1045 if (!pmd_present(*pmd)) 1046 return 0; 1047 1048 if (pmd_large(*pmd)) 1049 return spurious_kernel_fault_check(error_code, (pte_t *) pmd); 1050 1051 pte = pte_offset_kernel(pmd, address); 1052 if (!pte_present(*pte)) 1053 return 0; 1054 1055 ret = spurious_kernel_fault_check(error_code, pte); 1056 if (!ret) 1057 return 0; 1058 1059 /* 1060 * Make sure we have permissions in PMD. 1061 * If not, then there's a bug in the page tables: 1062 */ 1063 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd); 1064 WARN_ONCE(!ret, "PMD has incorrect permission bits\n"); 1065 1066 return ret; 1067 } 1068 NOKPROBE_SYMBOL(spurious_kernel_fault); 1069 1070 int show_unhandled_signals = 1; 1071 1072 static inline int 1073 access_error(unsigned long error_code, struct vm_area_struct *vma) 1074 { 1075 /* This is only called for the current mm, so: */ 1076 bool foreign = false; 1077 1078 /* 1079 * Read or write was blocked by protection keys. This is 1080 * always an unconditional error and can never result in 1081 * a follow-up action to resolve the fault, like a COW. 1082 */ 1083 if (error_code & X86_PF_PK) 1084 return 1; 1085 1086 /* 1087 * SGX hardware blocked the access. This usually happens 1088 * when the enclave memory contents have been destroyed, like 1089 * after a suspend/resume cycle. In any case, the kernel can't 1090 * fix the cause of the fault. Handle the fault as an access 1091 * error even in cases where no actual access violation 1092 * occurred. This allows userspace to rebuild the enclave in 1093 * response to the signal. 1094 */ 1095 if (unlikely(error_code & X86_PF_SGX)) 1096 return 1; 1097 1098 /* 1099 * Make sure to check the VMA so that we do not perform 1100 * faults just to hit a X86_PF_PK as soon as we fill in a 1101 * page. 1102 */ 1103 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 1104 (error_code & X86_PF_INSTR), foreign)) 1105 return 1; 1106 1107 /* 1108 * Shadow stack accesses (PF_SHSTK=1) are only permitted to 1109 * shadow stack VMAs. All other accesses result in an error. 1110 */ 1111 if (error_code & X86_PF_SHSTK) { 1112 if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK))) 1113 return 1; 1114 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1115 return 1; 1116 return 0; 1117 } 1118 1119 if (error_code & X86_PF_WRITE) { 1120 /* write, present and write, not present: */ 1121 if (unlikely(vma->vm_flags & VM_SHADOW_STACK)) 1122 return 1; 1123 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1124 return 1; 1125 return 0; 1126 } 1127 1128 /* read, present: */ 1129 if (unlikely(error_code & X86_PF_PROT)) 1130 return 1; 1131 1132 /* read, not present: */ 1133 if (unlikely(!vma_is_accessible(vma))) 1134 return 1; 1135 1136 return 0; 1137 } 1138 1139 bool fault_in_kernel_space(unsigned long address) 1140 { 1141 /* 1142 * On 64-bit systems, the vsyscall page is at an address above 1143 * TASK_SIZE_MAX, but is not considered part of the kernel 1144 * address space. 1145 */ 1146 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address)) 1147 return false; 1148 1149 return address >= TASK_SIZE_MAX; 1150 } 1151 1152 /* 1153 * Called for all faults where 'address' is part of the kernel address 1154 * space. Might get called for faults that originate from *code* that 1155 * ran in userspace or the kernel. 1156 */ 1157 static void 1158 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code, 1159 unsigned long address) 1160 { 1161 /* 1162 * Protection keys exceptions only happen on user pages. We 1163 * have no user pages in the kernel portion of the address 1164 * space, so do not expect them here. 1165 */ 1166 WARN_ON_ONCE(hw_error_code & X86_PF_PK); 1167 1168 #ifdef CONFIG_X86_32 1169 /* 1170 * We can fault-in kernel-space virtual memory on-demand. The 1171 * 'reference' page table is init_mm.pgd. 1172 * 1173 * NOTE! We MUST NOT take any locks for this case. We may 1174 * be in an interrupt or a critical region, and should 1175 * only copy the information from the master page table, 1176 * nothing more. 1177 * 1178 * Before doing this on-demand faulting, ensure that the 1179 * fault is not any of the following: 1180 * 1. A fault on a PTE with a reserved bit set. 1181 * 2. A fault caused by a user-mode access. (Do not demand- 1182 * fault kernel memory due to user-mode accesses). 1183 * 3. A fault caused by a page-level protection violation. 1184 * (A demand fault would be on a non-present page which 1185 * would have X86_PF_PROT==0). 1186 * 1187 * This is only needed to close a race condition on x86-32 in 1188 * the vmalloc mapping/unmapping code. See the comment above 1189 * vmalloc_fault() for details. On x86-64 the race does not 1190 * exist as the vmalloc mappings don't need to be synchronized 1191 * there. 1192 */ 1193 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) { 1194 if (vmalloc_fault(address) >= 0) 1195 return; 1196 } 1197 #endif 1198 1199 if (is_f00f_bug(regs, hw_error_code, address)) 1200 return; 1201 1202 /* Was the fault spurious, caused by lazy TLB invalidation? */ 1203 if (spurious_kernel_fault(hw_error_code, address)) 1204 return; 1205 1206 /* kprobes don't want to hook the spurious faults: */ 1207 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF))) 1208 return; 1209 1210 /* 1211 * Note, despite being a "bad area", there are quite a few 1212 * acceptable reasons to get here, such as erratum fixups 1213 * and handling kernel code that can fault, like get_user(). 1214 * 1215 * Don't take the mm semaphore here. If we fixup a prefetch 1216 * fault we could otherwise deadlock: 1217 */ 1218 bad_area_nosemaphore(regs, hw_error_code, address); 1219 } 1220 NOKPROBE_SYMBOL(do_kern_addr_fault); 1221 1222 /* 1223 * Handle faults in the user portion of the address space. Nothing in here 1224 * should check X86_PF_USER without a specific justification: for almost 1225 * all purposes, we should treat a normal kernel access to user memory 1226 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction. 1227 * The one exception is AC flag handling, which is, per the x86 1228 * architecture, special for WRUSS. 1229 */ 1230 static inline 1231 void do_user_addr_fault(struct pt_regs *regs, 1232 unsigned long error_code, 1233 unsigned long address) 1234 { 1235 struct vm_area_struct *vma; 1236 struct task_struct *tsk; 1237 struct mm_struct *mm; 1238 vm_fault_t fault; 1239 unsigned int flags = FAULT_FLAG_DEFAULT; 1240 1241 tsk = current; 1242 mm = tsk->mm; 1243 1244 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) { 1245 /* 1246 * Whoops, this is kernel mode code trying to execute from 1247 * user memory. Unless this is AMD erratum #93, which 1248 * corrupts RIP such that it looks like a user address, 1249 * this is unrecoverable. Don't even try to look up the 1250 * VMA or look for extable entries. 1251 */ 1252 if (is_errata93(regs, address)) 1253 return; 1254 1255 page_fault_oops(regs, error_code, address); 1256 return; 1257 } 1258 1259 /* kprobes don't want to hook the spurious faults: */ 1260 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF))) 1261 return; 1262 1263 /* 1264 * Reserved bits are never expected to be set on 1265 * entries in the user portion of the page tables. 1266 */ 1267 if (unlikely(error_code & X86_PF_RSVD)) 1268 pgtable_bad(regs, error_code, address); 1269 1270 /* 1271 * If SMAP is on, check for invalid kernel (supervisor) access to user 1272 * pages in the user address space. The odd case here is WRUSS, 1273 * which, according to the preliminary documentation, does not respect 1274 * SMAP and will have the USER bit set so, in all cases, SMAP 1275 * enforcement appears to be consistent with the USER bit. 1276 */ 1277 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) && 1278 !(error_code & X86_PF_USER) && 1279 !(regs->flags & X86_EFLAGS_AC))) { 1280 /* 1281 * No extable entry here. This was a kernel access to an 1282 * invalid pointer. get_kernel_nofault() will not get here. 1283 */ 1284 page_fault_oops(regs, error_code, address); 1285 return; 1286 } 1287 1288 /* 1289 * If we're in an interrupt, have no user context or are running 1290 * in a region with pagefaults disabled then we must not take the fault 1291 */ 1292 if (unlikely(faulthandler_disabled() || !mm)) { 1293 bad_area_nosemaphore(regs, error_code, address); 1294 return; 1295 } 1296 1297 /* 1298 * It's safe to allow irq's after cr2 has been saved and the 1299 * vmalloc fault has been handled. 1300 * 1301 * User-mode registers count as a user access even for any 1302 * potential system fault or CPU buglet: 1303 */ 1304 if (user_mode(regs)) { 1305 local_irq_enable(); 1306 flags |= FAULT_FLAG_USER; 1307 } else { 1308 if (regs->flags & X86_EFLAGS_IF) 1309 local_irq_enable(); 1310 } 1311 1312 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); 1313 1314 /* 1315 * Read-only permissions can not be expressed in shadow stack PTEs. 1316 * Treat all shadow stack accesses as WRITE faults. This ensures 1317 * that the MM will prepare everything (e.g., break COW) such that 1318 * maybe_mkwrite() can create a proper shadow stack PTE. 1319 */ 1320 if (error_code & X86_PF_SHSTK) 1321 flags |= FAULT_FLAG_WRITE; 1322 if (error_code & X86_PF_WRITE) 1323 flags |= FAULT_FLAG_WRITE; 1324 if (error_code & X86_PF_INSTR) 1325 flags |= FAULT_FLAG_INSTRUCTION; 1326 1327 #ifdef CONFIG_X86_64 1328 /* 1329 * Faults in the vsyscall page might need emulation. The 1330 * vsyscall page is at a high address (>PAGE_OFFSET), but is 1331 * considered to be part of the user address space. 1332 * 1333 * The vsyscall page does not have a "real" VMA, so do this 1334 * emulation before we go searching for VMAs. 1335 * 1336 * PKRU never rejects instruction fetches, so we don't need 1337 * to consider the PF_PK bit. 1338 */ 1339 if (is_vsyscall_vaddr(address)) { 1340 if (emulate_vsyscall(error_code, regs, address)) 1341 return; 1342 } 1343 #endif 1344 1345 if (!(flags & FAULT_FLAG_USER)) 1346 goto lock_mmap; 1347 1348 vma = lock_vma_under_rcu(mm, address); 1349 if (!vma) 1350 goto lock_mmap; 1351 1352 if (unlikely(access_error(error_code, vma))) { 1353 vma_end_read(vma); 1354 goto lock_mmap; 1355 } 1356 fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs); 1357 if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED))) 1358 vma_end_read(vma); 1359 1360 if (!(fault & VM_FAULT_RETRY)) { 1361 count_vm_vma_lock_event(VMA_LOCK_SUCCESS); 1362 goto done; 1363 } 1364 count_vm_vma_lock_event(VMA_LOCK_RETRY); 1365 if (fault & VM_FAULT_MAJOR) 1366 flags |= FAULT_FLAG_TRIED; 1367 1368 /* Quick path to respond to signals */ 1369 if (fault_signal_pending(fault, regs)) { 1370 if (!user_mode(regs)) 1371 kernelmode_fixup_or_oops(regs, error_code, address, 1372 SIGBUS, BUS_ADRERR, 1373 ARCH_DEFAULT_PKEY); 1374 return; 1375 } 1376 lock_mmap: 1377 1378 retry: 1379 vma = lock_mm_and_find_vma(mm, address, regs); 1380 if (unlikely(!vma)) { 1381 bad_area_nosemaphore(regs, error_code, address); 1382 return; 1383 } 1384 1385 /* 1386 * Ok, we have a good vm_area for this memory access, so 1387 * we can handle it.. 1388 */ 1389 if (unlikely(access_error(error_code, vma))) { 1390 bad_area_access_error(regs, error_code, address, vma); 1391 return; 1392 } 1393 1394 /* 1395 * If for any reason at all we couldn't handle the fault, 1396 * make sure we exit gracefully rather than endlessly redo 1397 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if 1398 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked. 1399 * 1400 * Note that handle_userfault() may also release and reacquire mmap_lock 1401 * (and not return with VM_FAULT_RETRY), when returning to userland to 1402 * repeat the page fault later with a VM_FAULT_NOPAGE retval 1403 * (potentially after handling any pending signal during the return to 1404 * userland). The return to userland is identified whenever 1405 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags. 1406 */ 1407 fault = handle_mm_fault(vma, address, flags, regs); 1408 1409 if (fault_signal_pending(fault, regs)) { 1410 /* 1411 * Quick path to respond to signals. The core mm code 1412 * has unlocked the mm for us if we get here. 1413 */ 1414 if (!user_mode(regs)) 1415 kernelmode_fixup_or_oops(regs, error_code, address, 1416 SIGBUS, BUS_ADRERR, 1417 ARCH_DEFAULT_PKEY); 1418 return; 1419 } 1420 1421 /* The fault is fully completed (including releasing mmap lock) */ 1422 if (fault & VM_FAULT_COMPLETED) 1423 return; 1424 1425 /* 1426 * If we need to retry the mmap_lock has already been released, 1427 * and if there is a fatal signal pending there is no guarantee 1428 * that we made any progress. Handle this case first. 1429 */ 1430 if (unlikely(fault & VM_FAULT_RETRY)) { 1431 flags |= FAULT_FLAG_TRIED; 1432 goto retry; 1433 } 1434 1435 mmap_read_unlock(mm); 1436 done: 1437 if (likely(!(fault & VM_FAULT_ERROR))) 1438 return; 1439 1440 if (fatal_signal_pending(current) && !user_mode(regs)) { 1441 kernelmode_fixup_or_oops(regs, error_code, address, 1442 0, 0, ARCH_DEFAULT_PKEY); 1443 return; 1444 } 1445 1446 if (fault & VM_FAULT_OOM) { 1447 /* Kernel mode? Handle exceptions or die: */ 1448 if (!user_mode(regs)) { 1449 kernelmode_fixup_or_oops(regs, error_code, address, 1450 SIGSEGV, SEGV_MAPERR, 1451 ARCH_DEFAULT_PKEY); 1452 return; 1453 } 1454 1455 /* 1456 * We ran out of memory, call the OOM killer, and return the 1457 * userspace (which will retry the fault, or kill us if we got 1458 * oom-killed): 1459 */ 1460 pagefault_out_of_memory(); 1461 } else { 1462 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON| 1463 VM_FAULT_HWPOISON_LARGE)) 1464 do_sigbus(regs, error_code, address, fault); 1465 else if (fault & VM_FAULT_SIGSEGV) 1466 bad_area_nosemaphore(regs, error_code, address); 1467 else 1468 BUG(); 1469 } 1470 } 1471 NOKPROBE_SYMBOL(do_user_addr_fault); 1472 1473 static __always_inline void 1474 trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code, 1475 unsigned long address) 1476 { 1477 if (!trace_pagefault_enabled()) 1478 return; 1479 1480 if (user_mode(regs)) 1481 trace_page_fault_user(address, regs, error_code); 1482 else 1483 trace_page_fault_kernel(address, regs, error_code); 1484 } 1485 1486 static __always_inline void 1487 handle_page_fault(struct pt_regs *regs, unsigned long error_code, 1488 unsigned long address) 1489 { 1490 trace_page_fault_entries(regs, error_code, address); 1491 1492 if (unlikely(kmmio_fault(regs, address))) 1493 return; 1494 1495 /* Was the fault on kernel-controlled part of the address space? */ 1496 if (unlikely(fault_in_kernel_space(address))) { 1497 do_kern_addr_fault(regs, error_code, address); 1498 } else { 1499 do_user_addr_fault(regs, error_code, address); 1500 /* 1501 * User address page fault handling might have reenabled 1502 * interrupts. Fixing up all potential exit points of 1503 * do_user_addr_fault() and its leaf functions is just not 1504 * doable w/o creating an unholy mess or turning the code 1505 * upside down. 1506 */ 1507 local_irq_disable(); 1508 } 1509 } 1510 1511 DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault) 1512 { 1513 irqentry_state_t state; 1514 unsigned long address; 1515 1516 address = cpu_feature_enabled(X86_FEATURE_FRED) ? fred_event_data(regs) : read_cr2(); 1517 1518 prefetchw(¤t->mm->mmap_lock); 1519 1520 /* 1521 * KVM uses #PF vector to deliver 'page not present' events to guests 1522 * (asynchronous page fault mechanism). The event happens when a 1523 * userspace task is trying to access some valid (from guest's point of 1524 * view) memory which is not currently mapped by the host (e.g. the 1525 * memory is swapped out). Note, the corresponding "page ready" event 1526 * which is injected when the memory becomes available, is delivered via 1527 * an interrupt mechanism and not a #PF exception 1528 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()). 1529 * 1530 * We are relying on the interrupted context being sane (valid RSP, 1531 * relevant locks not held, etc.), which is fine as long as the 1532 * interrupted context had IF=1. We are also relying on the KVM 1533 * async pf type field and CR2 being read consistently instead of 1534 * getting values from real and async page faults mixed up. 1535 * 1536 * Fingers crossed. 1537 * 1538 * The async #PF handling code takes care of idtentry handling 1539 * itself. 1540 */ 1541 if (kvm_handle_async_pf(regs, (u32)address)) 1542 return; 1543 1544 /* 1545 * Entry handling for valid #PF from kernel mode is slightly 1546 * different: RCU is already watching and ct_irq_enter() must not 1547 * be invoked because a kernel fault on a user space address might 1548 * sleep. 1549 * 1550 * In case the fault hit a RCU idle region the conditional entry 1551 * code reenabled RCU to avoid subsequent wreckage which helps 1552 * debuggability. 1553 */ 1554 state = irqentry_enter(regs); 1555 1556 instrumentation_begin(); 1557 handle_page_fault(regs, error_code, address); 1558 instrumentation_end(); 1559 1560 irqentry_exit(regs, state); 1561 } 1562