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 #include <asm/sev.h> /* snp_dump_hva_rmpentry() */ 39 40 #define CREATE_TRACE_POINTS 41 #include <asm/trace/exceptions.h> 42 43 /* 44 * Returns 0 if mmiotrace is disabled, or if the fault is not 45 * handled by mmiotrace: 46 */ 47 static nokprobe_inline int 48 kmmio_fault(struct pt_regs *regs, unsigned long addr) 49 { 50 if (unlikely(is_kmmio_active())) 51 if (kmmio_handler(regs, addr) == 1) 52 return -1; 53 return 0; 54 } 55 56 /* 57 * Prefetch quirks: 58 * 59 * 32-bit mode: 60 * 61 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch. 62 * Check that here and ignore it. This is AMD erratum #91. 63 * 64 * 64-bit mode: 65 * 66 * Sometimes the CPU reports invalid exceptions on prefetch. 67 * Check that here and ignore it. 68 * 69 * Opcode checker based on code by Richard Brunner. 70 */ 71 static inline int 72 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr, 73 unsigned char opcode, int *prefetch) 74 { 75 unsigned char instr_hi = opcode & 0xf0; 76 unsigned char instr_lo = opcode & 0x0f; 77 78 switch (instr_hi) { 79 case 0x20: 80 case 0x30: 81 /* 82 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. 83 * In X86_64 long mode, the CPU will signal invalid 84 * opcode if some of these prefixes are present so 85 * X86_64 will never get here anyway 86 */ 87 return ((instr_lo & 7) == 0x6); 88 #ifdef CONFIG_X86_64 89 case 0x40: 90 /* 91 * In 64-bit mode 0x40..0x4F are valid REX prefixes 92 */ 93 return (!user_mode(regs) || user_64bit_mode(regs)); 94 #endif 95 case 0x60: 96 /* 0x64 thru 0x67 are valid prefixes in all modes. */ 97 return (instr_lo & 0xC) == 0x4; 98 case 0xF0: 99 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */ 100 return !instr_lo || (instr_lo>>1) == 1; 101 case 0x00: 102 /* Prefetch instruction is 0x0F0D or 0x0F18 */ 103 if (get_kernel_nofault(opcode, instr)) 104 return 0; 105 106 *prefetch = (instr_lo == 0xF) && 107 (opcode == 0x0D || opcode == 0x18); 108 return 0; 109 default: 110 return 0; 111 } 112 } 113 114 static bool is_amd_k8_pre_npt(void) 115 { 116 struct cpuinfo_x86 *c = &boot_cpu_data; 117 118 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) && 119 c->x86_vendor == X86_VENDOR_AMD && 120 c->x86 == 0xf && c->x86_model < 0x40); 121 } 122 123 static int 124 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr) 125 { 126 unsigned char *max_instr; 127 unsigned char *instr; 128 int prefetch = 0; 129 130 /* Erratum #91 affects AMD K8, pre-NPT CPUs */ 131 if (!is_amd_k8_pre_npt()) 132 return 0; 133 134 /* 135 * If it was a exec (instruction fetch) fault on NX page, then 136 * do not ignore the fault: 137 */ 138 if (error_code & X86_PF_INSTR) 139 return 0; 140 141 instr = (void *)convert_ip_to_linear(current, regs); 142 max_instr = instr + 15; 143 144 /* 145 * This code has historically always bailed out if IP points to a 146 * not-present page (e.g. due to a race). No one has ever 147 * complained about this. 148 */ 149 pagefault_disable(); 150 151 while (instr < max_instr) { 152 unsigned char opcode; 153 154 if (user_mode(regs)) { 155 if (get_user(opcode, (unsigned char __user *) instr)) 156 break; 157 } else { 158 if (get_kernel_nofault(opcode, instr)) 159 break; 160 } 161 162 instr++; 163 164 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch)) 165 break; 166 } 167 168 pagefault_enable(); 169 return prefetch; 170 } 171 172 DEFINE_SPINLOCK(pgd_lock); 173 LIST_HEAD(pgd_list); 174 175 #ifdef CONFIG_X86_32 176 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) 177 { 178 unsigned index = pgd_index(address); 179 pgd_t *pgd_k; 180 p4d_t *p4d, *p4d_k; 181 pud_t *pud, *pud_k; 182 pmd_t *pmd, *pmd_k; 183 184 pgd += index; 185 pgd_k = init_mm.pgd + index; 186 187 if (!pgd_present(*pgd_k)) 188 return NULL; 189 190 /* 191 * set_pgd(pgd, *pgd_k); here would be useless on PAE 192 * and redundant with the set_pmd() on non-PAE. As would 193 * set_p4d/set_pud. 194 */ 195 p4d = p4d_offset(pgd, address); 196 p4d_k = p4d_offset(pgd_k, address); 197 if (!p4d_present(*p4d_k)) 198 return NULL; 199 200 pud = pud_offset(p4d, address); 201 pud_k = pud_offset(p4d_k, address); 202 if (!pud_present(*pud_k)) 203 return NULL; 204 205 pmd = pmd_offset(pud, address); 206 pmd_k = pmd_offset(pud_k, address); 207 208 if (pmd_present(*pmd) != pmd_present(*pmd_k)) 209 set_pmd(pmd, *pmd_k); 210 211 if (!pmd_present(*pmd_k)) 212 return NULL; 213 else 214 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k)); 215 216 return pmd_k; 217 } 218 219 /* 220 * Handle a fault on the vmalloc or module mapping area 221 * 222 * This is needed because there is a race condition between the time 223 * when the vmalloc mapping code updates the PMD to the point in time 224 * where it synchronizes this update with the other page-tables in the 225 * system. 226 * 227 * In this race window another thread/CPU can map an area on the same 228 * PMD, finds it already present and does not synchronize it with the 229 * rest of the system yet. As a result v[mz]alloc might return areas 230 * which are not mapped in every page-table in the system, causing an 231 * unhandled page-fault when they are accessed. 232 */ 233 static noinline int vmalloc_fault(unsigned long address) 234 { 235 unsigned long pgd_paddr; 236 pmd_t *pmd_k; 237 pte_t *pte_k; 238 239 /* Make sure we are in vmalloc area: */ 240 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 241 return -1; 242 243 /* 244 * Synchronize this task's top level page-table 245 * with the 'reference' page table. 246 * 247 * Do _not_ use "current" here. We might be inside 248 * an interrupt in the middle of a task switch.. 249 */ 250 pgd_paddr = read_cr3_pa(); 251 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address); 252 if (!pmd_k) 253 return -1; 254 255 if (pmd_leaf(*pmd_k)) 256 return 0; 257 258 pte_k = pte_offset_kernel(pmd_k, address); 259 if (!pte_present(*pte_k)) 260 return -1; 261 262 return 0; 263 } 264 NOKPROBE_SYMBOL(vmalloc_fault); 265 266 void arch_sync_kernel_mappings(unsigned long start, unsigned long end) 267 { 268 unsigned long addr; 269 270 for (addr = start & PMD_MASK; 271 addr >= TASK_SIZE_MAX && addr < VMALLOC_END; 272 addr += PMD_SIZE) { 273 struct page *page; 274 275 spin_lock(&pgd_lock); 276 list_for_each_entry(page, &pgd_list, lru) { 277 spinlock_t *pgt_lock; 278 279 /* the pgt_lock only for Xen */ 280 pgt_lock = &pgd_page_get_mm(page)->page_table_lock; 281 282 spin_lock(pgt_lock); 283 vmalloc_sync_one(page_address(page), addr); 284 spin_unlock(pgt_lock); 285 } 286 spin_unlock(&pgd_lock); 287 } 288 } 289 290 static bool low_pfn(unsigned long pfn) 291 { 292 return pfn < max_low_pfn; 293 } 294 295 static void dump_pagetable(unsigned long address) 296 { 297 pgd_t *base = __va(read_cr3_pa()); 298 pgd_t *pgd = &base[pgd_index(address)]; 299 p4d_t *p4d; 300 pud_t *pud; 301 pmd_t *pmd; 302 pte_t *pte; 303 304 #ifdef CONFIG_X86_PAE 305 pr_info("*pdpt = %016Lx ", pgd_val(*pgd)); 306 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd)) 307 goto out; 308 #define pr_pde pr_cont 309 #else 310 #define pr_pde pr_info 311 #endif 312 p4d = p4d_offset(pgd, address); 313 pud = pud_offset(p4d, address); 314 pmd = pmd_offset(pud, address); 315 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd)); 316 #undef pr_pde 317 318 /* 319 * We must not directly access the pte in the highpte 320 * case if the page table is located in highmem. 321 * And let's rather not kmap-atomic the pte, just in case 322 * it's allocated already: 323 */ 324 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_leaf(*pmd)) 325 goto out; 326 327 pte = pte_offset_kernel(pmd, address); 328 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte)); 329 out: 330 pr_cont("\n"); 331 } 332 333 #else /* CONFIG_X86_64: */ 334 335 #ifdef CONFIG_CPU_SUP_AMD 336 static const char errata93_warning[] = 337 KERN_ERR 338 "******* Your BIOS seems to not contain a fix for K8 errata #93\n" 339 "******* Working around it, but it may cause SEGVs or burn power.\n" 340 "******* Please consider a BIOS update.\n" 341 "******* Disabling USB legacy in the BIOS may also help.\n"; 342 #endif 343 344 static int bad_address(void *p) 345 { 346 unsigned long dummy; 347 348 return get_kernel_nofault(dummy, (unsigned long *)p); 349 } 350 351 static void dump_pagetable(unsigned long address) 352 { 353 pgd_t *base = __va(read_cr3_pa()); 354 pgd_t *pgd = base + pgd_index(address); 355 p4d_t *p4d; 356 pud_t *pud; 357 pmd_t *pmd; 358 pte_t *pte; 359 360 if (bad_address(pgd)) 361 goto bad; 362 363 pr_info("PGD %lx ", pgd_val(*pgd)); 364 365 if (!pgd_present(*pgd)) 366 goto out; 367 368 p4d = p4d_offset(pgd, address); 369 if (bad_address(p4d)) 370 goto bad; 371 372 pr_cont("P4D %lx ", p4d_val(*p4d)); 373 if (!p4d_present(*p4d) || p4d_leaf(*p4d)) 374 goto out; 375 376 pud = pud_offset(p4d, address); 377 if (bad_address(pud)) 378 goto bad; 379 380 pr_cont("PUD %lx ", pud_val(*pud)); 381 if (!pud_present(*pud) || pud_leaf(*pud)) 382 goto out; 383 384 pmd = pmd_offset(pud, address); 385 if (bad_address(pmd)) 386 goto bad; 387 388 pr_cont("PMD %lx ", pmd_val(*pmd)); 389 if (!pmd_present(*pmd) || pmd_leaf(*pmd)) 390 goto out; 391 392 pte = pte_offset_kernel(pmd, address); 393 if (bad_address(pte)) 394 goto bad; 395 396 pr_cont("PTE %lx", pte_val(*pte)); 397 out: 398 pr_cont("\n"); 399 return; 400 bad: 401 pr_info("BAD\n"); 402 } 403 404 #endif /* CONFIG_X86_64 */ 405 406 /* 407 * Workaround for K8 erratum #93 & buggy BIOS. 408 * 409 * BIOS SMM functions are required to use a specific workaround 410 * to avoid corruption of the 64bit RIP register on C stepping K8. 411 * 412 * A lot of BIOS that didn't get tested properly miss this. 413 * 414 * The OS sees this as a page fault with the upper 32bits of RIP cleared. 415 * Try to work around it here. 416 * 417 * Note we only handle faults in kernel here. 418 * Does nothing on 32-bit. 419 */ 420 static int is_errata93(struct pt_regs *regs, unsigned long address) 421 { 422 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD) 423 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD 424 || boot_cpu_data.x86 != 0xf) 425 return 0; 426 427 if (user_mode(regs)) 428 return 0; 429 430 if (address != regs->ip) 431 return 0; 432 433 if ((address >> 32) != 0) 434 return 0; 435 436 address |= 0xffffffffUL << 32; 437 if ((address >= (u64)_stext && address <= (u64)_etext) || 438 (address >= MODULES_VADDR && address <= MODULES_END)) { 439 printk_once(errata93_warning); 440 regs->ip = address; 441 return 1; 442 } 443 #endif 444 return 0; 445 } 446 447 /* 448 * Work around K8 erratum #100 K8 in compat mode occasionally jumps 449 * to illegal addresses >4GB. 450 * 451 * We catch this in the page fault handler because these addresses 452 * are not reachable. Just detect this case and return. Any code 453 * segment in LDT is compatibility mode. 454 */ 455 static int is_errata100(struct pt_regs *regs, unsigned long address) 456 { 457 #ifdef CONFIG_X86_64 458 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32)) 459 return 1; 460 #endif 461 return 0; 462 } 463 464 /* Pentium F0 0F C7 C8 bug workaround: */ 465 static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code, 466 unsigned long address) 467 { 468 #ifdef CONFIG_X86_F00F_BUG 469 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) && 470 idt_is_f00f_address(address)) { 471 handle_invalid_op(regs); 472 return 1; 473 } 474 #endif 475 return 0; 476 } 477 478 static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index) 479 { 480 u32 offset = (index >> 3) * sizeof(struct desc_struct); 481 unsigned long addr; 482 struct ldttss_desc desc; 483 484 if (index == 0) { 485 pr_alert("%s: NULL\n", name); 486 return; 487 } 488 489 if (offset + sizeof(struct ldttss_desc) >= gdt->size) { 490 pr_alert("%s: 0x%hx -- out of bounds\n", name, index); 491 return; 492 } 493 494 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset), 495 sizeof(struct ldttss_desc))) { 496 pr_alert("%s: 0x%hx -- GDT entry is not readable\n", 497 name, index); 498 return; 499 } 500 501 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24); 502 #ifdef CONFIG_X86_64 503 addr |= ((u64)desc.base3 << 32); 504 #endif 505 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n", 506 name, index, addr, (desc.limit0 | (desc.limit1 << 16))); 507 } 508 509 static void 510 show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address) 511 { 512 if (!oops_may_print()) 513 return; 514 515 if (error_code & X86_PF_INSTR) { 516 unsigned int level; 517 pgd_t *pgd; 518 pte_t *pte; 519 520 pgd = __va(read_cr3_pa()); 521 pgd += pgd_index(address); 522 523 pte = lookup_address_in_pgd(pgd, address, &level); 524 525 if (pte && pte_present(*pte) && !pte_exec(*pte)) 526 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", 527 from_kuid(&init_user_ns, current_uid())); 528 if (pte && pte_present(*pte) && pte_exec(*pte) && 529 (pgd_flags(*pgd) & _PAGE_USER) && 530 (__read_cr4() & X86_CR4_SMEP)) 531 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n", 532 from_kuid(&init_user_ns, current_uid())); 533 } 534 535 if (address < PAGE_SIZE && !user_mode(regs)) 536 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n", 537 (void *)address); 538 else 539 pr_alert("BUG: unable to handle page fault for address: %px\n", 540 (void *)address); 541 542 pr_alert("#PF: %s %s in %s mode\n", 543 (error_code & X86_PF_USER) ? "user" : "supervisor", 544 (error_code & X86_PF_INSTR) ? "instruction fetch" : 545 (error_code & X86_PF_WRITE) ? "write access" : 546 "read access", 547 user_mode(regs) ? "user" : "kernel"); 548 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code, 549 !(error_code & X86_PF_PROT) ? "not-present page" : 550 (error_code & X86_PF_RSVD) ? "reserved bit violation" : 551 (error_code & X86_PF_PK) ? "protection keys violation" : 552 (error_code & X86_PF_RMP) ? "RMP violation" : 553 "permissions violation"); 554 555 if (!(error_code & X86_PF_USER) && user_mode(regs)) { 556 struct desc_ptr idt, gdt; 557 u16 ldtr, tr; 558 559 /* 560 * This can happen for quite a few reasons. The more obvious 561 * ones are faults accessing the GDT, or LDT. Perhaps 562 * surprisingly, if the CPU tries to deliver a benign or 563 * contributory exception from user code and gets a page fault 564 * during delivery, the page fault can be delivered as though 565 * it originated directly from user code. This could happen 566 * due to wrong permissions on the IDT, GDT, LDT, TSS, or 567 * kernel or IST stack. 568 */ 569 store_idt(&idt); 570 571 /* Usable even on Xen PV -- it's just slow. */ 572 native_store_gdt(&gdt); 573 574 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n", 575 idt.address, idt.size, gdt.address, gdt.size); 576 577 store_ldt(ldtr); 578 show_ldttss(&gdt, "LDTR", ldtr); 579 580 store_tr(tr); 581 show_ldttss(&gdt, "TR", tr); 582 } 583 584 dump_pagetable(address); 585 586 if (error_code & X86_PF_RMP) 587 snp_dump_hva_rmpentry(address); 588 } 589 590 static noinline void 591 pgtable_bad(struct pt_regs *regs, unsigned long error_code, 592 unsigned long address) 593 { 594 struct task_struct *tsk; 595 unsigned long flags; 596 int sig; 597 598 flags = oops_begin(); 599 tsk = current; 600 sig = SIGKILL; 601 602 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n", 603 tsk->comm, address); 604 dump_pagetable(address); 605 606 if (__die("Bad pagetable", regs, error_code)) 607 sig = 0; 608 609 oops_end(flags, regs, sig); 610 } 611 612 static void sanitize_error_code(unsigned long address, 613 unsigned long *error_code) 614 { 615 /* 616 * To avoid leaking information about the kernel page 617 * table layout, pretend that user-mode accesses to 618 * kernel addresses are always protection faults. 619 * 620 * NB: This means that failed vsyscalls with vsyscall=none 621 * will have the PROT bit. This doesn't leak any 622 * information and does not appear to cause any problems. 623 */ 624 if (address >= TASK_SIZE_MAX) 625 *error_code |= X86_PF_PROT; 626 } 627 628 static void set_signal_archinfo(unsigned long address, 629 unsigned long error_code) 630 { 631 struct task_struct *tsk = current; 632 633 tsk->thread.trap_nr = X86_TRAP_PF; 634 tsk->thread.error_code = error_code | X86_PF_USER; 635 tsk->thread.cr2 = address; 636 } 637 638 static noinline void 639 page_fault_oops(struct pt_regs *regs, unsigned long error_code, 640 unsigned long address) 641 { 642 #ifdef CONFIG_VMAP_STACK 643 struct stack_info info; 644 #endif 645 unsigned long flags; 646 int sig; 647 648 if (user_mode(regs)) { 649 /* 650 * Implicit kernel access from user mode? Skip the stack 651 * overflow and EFI special cases. 652 */ 653 goto oops; 654 } 655 656 #ifdef CONFIG_VMAP_STACK 657 /* 658 * Stack overflow? During boot, we can fault near the initial 659 * stack in the direct map, but that's not an overflow -- check 660 * that we're in vmalloc space to avoid this. 661 */ 662 if (is_vmalloc_addr((void *)address) && 663 get_stack_guard_info((void *)address, &info)) { 664 /* 665 * We're likely to be running with very little stack space 666 * left. It's plausible that we'd hit this condition but 667 * double-fault even before we get this far, in which case 668 * we're fine: the double-fault handler will deal with it. 669 * 670 * We don't want to make it all the way into the oops code 671 * and then double-fault, though, because we're likely to 672 * break the console driver and lose most of the stack dump. 673 */ 674 call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*), 675 handle_stack_overflow, 676 ASM_CALL_ARG3, 677 , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info)); 678 679 unreachable(); 680 } 681 #endif 682 683 /* 684 * Buggy firmware could access regions which might page fault. If 685 * this happens, EFI has a special OOPS path that will try to 686 * avoid hanging the system. 687 */ 688 if (IS_ENABLED(CONFIG_EFI)) 689 efi_crash_gracefully_on_page_fault(address); 690 691 /* Only not-present faults should be handled by KFENCE. */ 692 if (!(error_code & X86_PF_PROT) && 693 kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs)) 694 return; 695 696 oops: 697 /* 698 * Oops. The kernel tried to access some bad page. We'll have to 699 * terminate things with extreme prejudice: 700 */ 701 flags = oops_begin(); 702 703 show_fault_oops(regs, error_code, address); 704 705 if (task_stack_end_corrupted(current)) 706 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 707 708 sig = SIGKILL; 709 if (__die("Oops", regs, error_code)) 710 sig = 0; 711 712 /* Executive summary in case the body of the oops scrolled away */ 713 printk(KERN_DEFAULT "CR2: %016lx\n", address); 714 715 oops_end(flags, regs, sig); 716 } 717 718 static noinline void 719 kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code, 720 unsigned long address, int signal, int si_code, 721 u32 pkey) 722 { 723 WARN_ON_ONCE(user_mode(regs)); 724 725 /* Are we prepared to handle this kernel fault? */ 726 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) 727 return; 728 729 /* 730 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH 731 * instruction. 732 */ 733 if (is_prefetch(regs, error_code, address)) 734 return; 735 736 page_fault_oops(regs, error_code, address); 737 } 738 739 /* 740 * Print out info about fatal segfaults, if the show_unhandled_signals 741 * sysctl is set: 742 */ 743 static inline void 744 show_signal_msg(struct pt_regs *regs, unsigned long error_code, 745 unsigned long address, struct task_struct *tsk) 746 { 747 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG; 748 /* This is a racy snapshot, but it's better than nothing. */ 749 int cpu = raw_smp_processor_id(); 750 751 if (!unhandled_signal(tsk, SIGSEGV)) 752 return; 753 754 if (!printk_ratelimit()) 755 return; 756 757 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx", 758 loglvl, tsk->comm, task_pid_nr(tsk), address, 759 (void *)regs->ip, (void *)regs->sp, error_code); 760 761 print_vma_addr(KERN_CONT " in ", regs->ip); 762 763 /* 764 * Dump the likely CPU where the fatal segfault happened. 765 * This can help identify faulty hardware. 766 */ 767 printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu, 768 topology_core_id(cpu), topology_physical_package_id(cpu)); 769 770 771 printk(KERN_CONT "\n"); 772 773 show_opcodes(regs, loglvl); 774 } 775 776 static void 777 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 778 unsigned long address, u32 pkey, int si_code) 779 { 780 struct task_struct *tsk = current; 781 782 if (!user_mode(regs)) { 783 kernelmode_fixup_or_oops(regs, error_code, address, 784 SIGSEGV, si_code, pkey); 785 return; 786 } 787 788 if (!(error_code & X86_PF_USER)) { 789 /* Implicit user access to kernel memory -- just oops */ 790 page_fault_oops(regs, error_code, address); 791 return; 792 } 793 794 /* 795 * User mode accesses just cause a SIGSEGV. 796 * It's possible to have interrupts off here: 797 */ 798 local_irq_enable(); 799 800 /* 801 * Valid to do another page fault here because this one came 802 * from user space: 803 */ 804 if (is_prefetch(regs, error_code, address)) 805 return; 806 807 if (is_errata100(regs, address)) 808 return; 809 810 sanitize_error_code(address, &error_code); 811 812 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address)) 813 return; 814 815 if (likely(show_unhandled_signals)) 816 show_signal_msg(regs, error_code, address, tsk); 817 818 set_signal_archinfo(address, error_code); 819 820 if (si_code == SEGV_PKUERR) 821 force_sig_pkuerr((void __user *)address, pkey); 822 else 823 force_sig_fault(SIGSEGV, si_code, (void __user *)address); 824 825 local_irq_disable(); 826 } 827 828 static noinline void 829 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 830 unsigned long address) 831 { 832 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR); 833 } 834 835 static void 836 __bad_area(struct pt_regs *regs, unsigned long error_code, 837 unsigned long address, u32 pkey, int si_code) 838 { 839 struct mm_struct *mm = current->mm; 840 /* 841 * Something tried to access memory that isn't in our memory map.. 842 * Fix it, but check if it's kernel or user first.. 843 */ 844 mmap_read_unlock(mm); 845 846 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code); 847 } 848 849 static inline bool bad_area_access_from_pkeys(unsigned long error_code, 850 struct vm_area_struct *vma) 851 { 852 /* This code is always called on the current mm */ 853 bool foreign = false; 854 855 if (!cpu_feature_enabled(X86_FEATURE_OSPKE)) 856 return false; 857 if (error_code & X86_PF_PK) 858 return true; 859 /* this checks permission keys on the VMA: */ 860 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 861 (error_code & X86_PF_INSTR), foreign)) 862 return true; 863 return false; 864 } 865 866 static noinline void 867 bad_area_access_error(struct pt_regs *regs, unsigned long error_code, 868 unsigned long address, struct vm_area_struct *vma) 869 { 870 /* 871 * This OSPKE check is not strictly necessary at runtime. 872 * But, doing it this way allows compiler optimizations 873 * if pkeys are compiled out. 874 */ 875 if (bad_area_access_from_pkeys(error_code, vma)) { 876 /* 877 * A protection key fault means that the PKRU value did not allow 878 * access to some PTE. Userspace can figure out what PKRU was 879 * from the XSAVE state. This function captures the pkey from 880 * the vma and passes it to userspace so userspace can discover 881 * which protection key was set on the PTE. 882 * 883 * If we get here, we know that the hardware signaled a X86_PF_PK 884 * fault and that there was a VMA once we got in the fault 885 * handler. It does *not* guarantee that the VMA we find here 886 * was the one that we faulted on. 887 * 888 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4); 889 * 2. T1 : set PKRU to deny access to pkey=4, touches page 890 * 3. T1 : faults... 891 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5); 892 * 5. T1 : enters fault handler, takes mmap_lock, etc... 893 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really 894 * faulted on a pte with its pkey=4. 895 */ 896 u32 pkey = vma_pkey(vma); 897 898 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR); 899 } else { 900 __bad_area(regs, error_code, address, 0, SEGV_ACCERR); 901 } 902 } 903 904 static void 905 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address, 906 vm_fault_t fault) 907 { 908 /* Kernel mode? Handle exceptions or die: */ 909 if (!user_mode(regs)) { 910 kernelmode_fixup_or_oops(regs, error_code, address, 911 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY); 912 return; 913 } 914 915 /* User-space => ok to do another page fault: */ 916 if (is_prefetch(regs, error_code, address)) 917 return; 918 919 sanitize_error_code(address, &error_code); 920 921 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address)) 922 return; 923 924 set_signal_archinfo(address, error_code); 925 926 #ifdef CONFIG_MEMORY_FAILURE 927 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) { 928 struct task_struct *tsk = current; 929 unsigned lsb = 0; 930 931 pr_err( 932 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n", 933 tsk->comm, tsk->pid, address); 934 if (fault & VM_FAULT_HWPOISON_LARGE) 935 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 936 if (fault & VM_FAULT_HWPOISON) 937 lsb = PAGE_SHIFT; 938 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb); 939 return; 940 } 941 #endif 942 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address); 943 } 944 945 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte) 946 { 947 if ((error_code & X86_PF_WRITE) && !pte_write(*pte)) 948 return 0; 949 950 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte)) 951 return 0; 952 953 return 1; 954 } 955 956 /* 957 * Handle a spurious fault caused by a stale TLB entry. 958 * 959 * This allows us to lazily refresh the TLB when increasing the 960 * permissions of a kernel page (RO -> RW or NX -> X). Doing it 961 * eagerly is very expensive since that implies doing a full 962 * cross-processor TLB flush, even if no stale TLB entries exist 963 * on other processors. 964 * 965 * Spurious faults may only occur if the TLB contains an entry with 966 * fewer permission than the page table entry. Non-present (P = 0) 967 * and reserved bit (R = 1) faults are never spurious. 968 * 969 * There are no security implications to leaving a stale TLB when 970 * increasing the permissions on a page. 971 * 972 * Returns non-zero if a spurious fault was handled, zero otherwise. 973 * 974 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 975 * (Optional Invalidation). 976 */ 977 static noinline int 978 spurious_kernel_fault(unsigned long error_code, unsigned long address) 979 { 980 pgd_t *pgd; 981 p4d_t *p4d; 982 pud_t *pud; 983 pmd_t *pmd; 984 pte_t *pte; 985 int ret; 986 987 /* 988 * Only writes to RO or instruction fetches from NX may cause 989 * spurious faults. 990 * 991 * These could be from user or supervisor accesses but the TLB 992 * is only lazily flushed after a kernel mapping protection 993 * change, so user accesses are not expected to cause spurious 994 * faults. 995 */ 996 if (error_code != (X86_PF_WRITE | X86_PF_PROT) && 997 error_code != (X86_PF_INSTR | X86_PF_PROT)) 998 return 0; 999 1000 pgd = init_mm.pgd + pgd_index(address); 1001 if (!pgd_present(*pgd)) 1002 return 0; 1003 1004 p4d = p4d_offset(pgd, address); 1005 if (!p4d_present(*p4d)) 1006 return 0; 1007 1008 if (p4d_leaf(*p4d)) 1009 return spurious_kernel_fault_check(error_code, (pte_t *) p4d); 1010 1011 pud = pud_offset(p4d, address); 1012 if (!pud_present(*pud)) 1013 return 0; 1014 1015 if (pud_leaf(*pud)) 1016 return spurious_kernel_fault_check(error_code, (pte_t *) pud); 1017 1018 pmd = pmd_offset(pud, address); 1019 if (!pmd_present(*pmd)) 1020 return 0; 1021 1022 if (pmd_leaf(*pmd)) 1023 return spurious_kernel_fault_check(error_code, (pte_t *) pmd); 1024 1025 pte = pte_offset_kernel(pmd, address); 1026 if (!pte_present(*pte)) 1027 return 0; 1028 1029 ret = spurious_kernel_fault_check(error_code, pte); 1030 if (!ret) 1031 return 0; 1032 1033 /* 1034 * Make sure we have permissions in PMD. 1035 * If not, then there's a bug in the page tables: 1036 */ 1037 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd); 1038 WARN_ONCE(!ret, "PMD has incorrect permission bits\n"); 1039 1040 return ret; 1041 } 1042 NOKPROBE_SYMBOL(spurious_kernel_fault); 1043 1044 int show_unhandled_signals = 1; 1045 1046 static inline int 1047 access_error(unsigned long error_code, struct vm_area_struct *vma) 1048 { 1049 /* This is only called for the current mm, so: */ 1050 bool foreign = false; 1051 1052 /* 1053 * Read or write was blocked by protection keys. This is 1054 * always an unconditional error and can never result in 1055 * a follow-up action to resolve the fault, like a COW. 1056 */ 1057 if (error_code & X86_PF_PK) 1058 return 1; 1059 1060 /* 1061 * SGX hardware blocked the access. This usually happens 1062 * when the enclave memory contents have been destroyed, like 1063 * after a suspend/resume cycle. In any case, the kernel can't 1064 * fix the cause of the fault. Handle the fault as an access 1065 * error even in cases where no actual access violation 1066 * occurred. This allows userspace to rebuild the enclave in 1067 * response to the signal. 1068 */ 1069 if (unlikely(error_code & X86_PF_SGX)) 1070 return 1; 1071 1072 /* 1073 * Make sure to check the VMA so that we do not perform 1074 * faults just to hit a X86_PF_PK as soon as we fill in a 1075 * page. 1076 */ 1077 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 1078 (error_code & X86_PF_INSTR), foreign)) 1079 return 1; 1080 1081 /* 1082 * Shadow stack accesses (PF_SHSTK=1) are only permitted to 1083 * shadow stack VMAs. All other accesses result in an error. 1084 */ 1085 if (error_code & X86_PF_SHSTK) { 1086 if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK))) 1087 return 1; 1088 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1089 return 1; 1090 return 0; 1091 } 1092 1093 if (error_code & X86_PF_WRITE) { 1094 /* write, present and write, not present: */ 1095 if (unlikely(vma->vm_flags & VM_SHADOW_STACK)) 1096 return 1; 1097 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1098 return 1; 1099 return 0; 1100 } 1101 1102 /* read, present: */ 1103 if (unlikely(error_code & X86_PF_PROT)) 1104 return 1; 1105 1106 /* read, not present: */ 1107 if (unlikely(!vma_is_accessible(vma))) 1108 return 1; 1109 1110 return 0; 1111 } 1112 1113 bool fault_in_kernel_space(unsigned long address) 1114 { 1115 /* 1116 * On 64-bit systems, the vsyscall page is at an address above 1117 * TASK_SIZE_MAX, but is not considered part of the kernel 1118 * address space. 1119 */ 1120 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address)) 1121 return false; 1122 1123 return address >= TASK_SIZE_MAX; 1124 } 1125 1126 /* 1127 * Called for all faults where 'address' is part of the kernel address 1128 * space. Might get called for faults that originate from *code* that 1129 * ran in userspace or the kernel. 1130 */ 1131 static void 1132 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code, 1133 unsigned long address) 1134 { 1135 /* 1136 * Protection keys exceptions only happen on user pages. We 1137 * have no user pages in the kernel portion of the address 1138 * space, so do not expect them here. 1139 */ 1140 WARN_ON_ONCE(hw_error_code & X86_PF_PK); 1141 1142 #ifdef CONFIG_X86_32 1143 /* 1144 * We can fault-in kernel-space virtual memory on-demand. The 1145 * 'reference' page table is init_mm.pgd. 1146 * 1147 * NOTE! We MUST NOT take any locks for this case. We may 1148 * be in an interrupt or a critical region, and should 1149 * only copy the information from the master page table, 1150 * nothing more. 1151 * 1152 * Before doing this on-demand faulting, ensure that the 1153 * fault is not any of the following: 1154 * 1. A fault on a PTE with a reserved bit set. 1155 * 2. A fault caused by a user-mode access. (Do not demand- 1156 * fault kernel memory due to user-mode accesses). 1157 * 3. A fault caused by a page-level protection violation. 1158 * (A demand fault would be on a non-present page which 1159 * would have X86_PF_PROT==0). 1160 * 1161 * This is only needed to close a race condition on x86-32 in 1162 * the vmalloc mapping/unmapping code. See the comment above 1163 * vmalloc_fault() for details. On x86-64 the race does not 1164 * exist as the vmalloc mappings don't need to be synchronized 1165 * there. 1166 */ 1167 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) { 1168 if (vmalloc_fault(address) >= 0) 1169 return; 1170 } 1171 #endif 1172 1173 if (is_f00f_bug(regs, hw_error_code, address)) 1174 return; 1175 1176 /* Was the fault spurious, caused by lazy TLB invalidation? */ 1177 if (spurious_kernel_fault(hw_error_code, address)) 1178 return; 1179 1180 /* kprobes don't want to hook the spurious faults: */ 1181 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF))) 1182 return; 1183 1184 /* 1185 * Note, despite being a "bad area", there are quite a few 1186 * acceptable reasons to get here, such as erratum fixups 1187 * and handling kernel code that can fault, like get_user(). 1188 * 1189 * Don't take the mm semaphore here. If we fixup a prefetch 1190 * fault we could otherwise deadlock: 1191 */ 1192 bad_area_nosemaphore(regs, hw_error_code, address); 1193 } 1194 NOKPROBE_SYMBOL(do_kern_addr_fault); 1195 1196 /* 1197 * Handle faults in the user portion of the address space. Nothing in here 1198 * should check X86_PF_USER without a specific justification: for almost 1199 * all purposes, we should treat a normal kernel access to user memory 1200 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction. 1201 * The one exception is AC flag handling, which is, per the x86 1202 * architecture, special for WRUSS. 1203 */ 1204 static inline 1205 void do_user_addr_fault(struct pt_regs *regs, 1206 unsigned long error_code, 1207 unsigned long address) 1208 { 1209 struct vm_area_struct *vma; 1210 struct task_struct *tsk; 1211 struct mm_struct *mm; 1212 vm_fault_t fault; 1213 unsigned int flags = FAULT_FLAG_DEFAULT; 1214 1215 tsk = current; 1216 mm = tsk->mm; 1217 1218 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) { 1219 /* 1220 * Whoops, this is kernel mode code trying to execute from 1221 * user memory. Unless this is AMD erratum #93, which 1222 * corrupts RIP such that it looks like a user address, 1223 * this is unrecoverable. Don't even try to look up the 1224 * VMA or look for extable entries. 1225 */ 1226 if (is_errata93(regs, address)) 1227 return; 1228 1229 page_fault_oops(regs, error_code, address); 1230 return; 1231 } 1232 1233 /* kprobes don't want to hook the spurious faults: */ 1234 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF))) 1235 return; 1236 1237 /* 1238 * Reserved bits are never expected to be set on 1239 * entries in the user portion of the page tables. 1240 */ 1241 if (unlikely(error_code & X86_PF_RSVD)) 1242 pgtable_bad(regs, error_code, address); 1243 1244 /* 1245 * If SMAP is on, check for invalid kernel (supervisor) access to user 1246 * pages in the user address space. The odd case here is WRUSS, 1247 * which, according to the preliminary documentation, does not respect 1248 * SMAP and will have the USER bit set so, in all cases, SMAP 1249 * enforcement appears to be consistent with the USER bit. 1250 */ 1251 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) && 1252 !(error_code & X86_PF_USER) && 1253 !(regs->flags & X86_EFLAGS_AC))) { 1254 /* 1255 * No extable entry here. This was a kernel access to an 1256 * invalid pointer. get_kernel_nofault() will not get here. 1257 */ 1258 page_fault_oops(regs, error_code, address); 1259 return; 1260 } 1261 1262 /* 1263 * If we're in an interrupt, have no user context or are running 1264 * in a region with pagefaults disabled then we must not take the fault 1265 */ 1266 if (unlikely(faulthandler_disabled() || !mm)) { 1267 bad_area_nosemaphore(regs, error_code, address); 1268 return; 1269 } 1270 1271 /* Legacy check - remove this after verifying that it doesn't trigger */ 1272 if (WARN_ON_ONCE(!(regs->flags & X86_EFLAGS_IF))) { 1273 bad_area_nosemaphore(regs, error_code, address); 1274 return; 1275 } 1276 1277 local_irq_enable(); 1278 1279 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); 1280 1281 /* 1282 * Read-only permissions can not be expressed in shadow stack PTEs. 1283 * Treat all shadow stack accesses as WRITE faults. This ensures 1284 * that the MM will prepare everything (e.g., break COW) such that 1285 * maybe_mkwrite() can create a proper shadow stack PTE. 1286 */ 1287 if (error_code & X86_PF_SHSTK) 1288 flags |= FAULT_FLAG_WRITE; 1289 if (error_code & X86_PF_WRITE) 1290 flags |= FAULT_FLAG_WRITE; 1291 if (error_code & X86_PF_INSTR) 1292 flags |= FAULT_FLAG_INSTRUCTION; 1293 1294 /* 1295 * We set FAULT_FLAG_USER based on the register state, not 1296 * based on X86_PF_USER. User space accesses that cause 1297 * system page faults are still user accesses. 1298 */ 1299 if (user_mode(regs)) 1300 flags |= FAULT_FLAG_USER; 1301 1302 #ifdef CONFIG_X86_64 1303 /* 1304 * Faults in the vsyscall page might need emulation. The 1305 * vsyscall page is at a high address (>PAGE_OFFSET), but is 1306 * considered to be part of the user address space. 1307 * 1308 * The vsyscall page does not have a "real" VMA, so do this 1309 * emulation before we go searching for VMAs. 1310 * 1311 * PKRU never rejects instruction fetches, so we don't need 1312 * to consider the PF_PK bit. 1313 */ 1314 if (is_vsyscall_vaddr(address)) { 1315 if (emulate_vsyscall(error_code, regs, address)) 1316 return; 1317 } 1318 #endif 1319 1320 if (!(flags & FAULT_FLAG_USER)) 1321 goto lock_mmap; 1322 1323 vma = lock_vma_under_rcu(mm, address); 1324 if (!vma) 1325 goto lock_mmap; 1326 1327 if (unlikely(access_error(error_code, vma))) { 1328 vma_end_read(vma); 1329 goto lock_mmap; 1330 } 1331 fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs); 1332 if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED))) 1333 vma_end_read(vma); 1334 1335 if (!(fault & VM_FAULT_RETRY)) { 1336 count_vm_vma_lock_event(VMA_LOCK_SUCCESS); 1337 goto done; 1338 } 1339 count_vm_vma_lock_event(VMA_LOCK_RETRY); 1340 if (fault & VM_FAULT_MAJOR) 1341 flags |= FAULT_FLAG_TRIED; 1342 1343 /* Quick path to respond to signals */ 1344 if (fault_signal_pending(fault, regs)) { 1345 if (!user_mode(regs)) 1346 kernelmode_fixup_or_oops(regs, error_code, address, 1347 SIGBUS, BUS_ADRERR, 1348 ARCH_DEFAULT_PKEY); 1349 return; 1350 } 1351 lock_mmap: 1352 1353 retry: 1354 vma = lock_mm_and_find_vma(mm, address, regs); 1355 if (unlikely(!vma)) { 1356 bad_area_nosemaphore(regs, error_code, address); 1357 return; 1358 } 1359 1360 /* 1361 * Ok, we have a good vm_area for this memory access, so 1362 * we can handle it.. 1363 */ 1364 if (unlikely(access_error(error_code, vma))) { 1365 bad_area_access_error(regs, error_code, address, vma); 1366 return; 1367 } 1368 1369 /* 1370 * If for any reason at all we couldn't handle the fault, 1371 * make sure we exit gracefully rather than endlessly redo 1372 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if 1373 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked. 1374 * 1375 * Note that handle_userfault() may also release and reacquire mmap_lock 1376 * (and not return with VM_FAULT_RETRY), when returning to userland to 1377 * repeat the page fault later with a VM_FAULT_NOPAGE retval 1378 * (potentially after handling any pending signal during the return to 1379 * userland). The return to userland is identified whenever 1380 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags. 1381 */ 1382 fault = handle_mm_fault(vma, address, flags, regs); 1383 1384 if (fault_signal_pending(fault, regs)) { 1385 /* 1386 * Quick path to respond to signals. The core mm code 1387 * has unlocked the mm for us if we get here. 1388 */ 1389 if (!user_mode(regs)) 1390 kernelmode_fixup_or_oops(regs, error_code, address, 1391 SIGBUS, BUS_ADRERR, 1392 ARCH_DEFAULT_PKEY); 1393 return; 1394 } 1395 1396 /* The fault is fully completed (including releasing mmap lock) */ 1397 if (fault & VM_FAULT_COMPLETED) 1398 return; 1399 1400 /* 1401 * If we need to retry the mmap_lock has already been released, 1402 * and if there is a fatal signal pending there is no guarantee 1403 * that we made any progress. Handle this case first. 1404 */ 1405 if (unlikely(fault & VM_FAULT_RETRY)) { 1406 flags |= FAULT_FLAG_TRIED; 1407 goto retry; 1408 } 1409 1410 mmap_read_unlock(mm); 1411 done: 1412 if (likely(!(fault & VM_FAULT_ERROR))) 1413 return; 1414 1415 if (fatal_signal_pending(current) && !user_mode(regs)) { 1416 kernelmode_fixup_or_oops(regs, error_code, address, 1417 0, 0, ARCH_DEFAULT_PKEY); 1418 return; 1419 } 1420 1421 if (fault & VM_FAULT_OOM) { 1422 /* Kernel mode? Handle exceptions or die: */ 1423 if (!user_mode(regs)) { 1424 kernelmode_fixup_or_oops(regs, error_code, address, 1425 SIGSEGV, SEGV_MAPERR, 1426 ARCH_DEFAULT_PKEY); 1427 return; 1428 } 1429 1430 /* 1431 * We ran out of memory, call the OOM killer, and return the 1432 * userspace (which will retry the fault, or kill us if we got 1433 * oom-killed): 1434 */ 1435 pagefault_out_of_memory(); 1436 } else { 1437 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON| 1438 VM_FAULT_HWPOISON_LARGE)) 1439 do_sigbus(regs, error_code, address, fault); 1440 else if (fault & VM_FAULT_SIGSEGV) 1441 bad_area_nosemaphore(regs, error_code, address); 1442 else 1443 BUG(); 1444 } 1445 } 1446 NOKPROBE_SYMBOL(do_user_addr_fault); 1447 1448 static __always_inline void 1449 trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code, 1450 unsigned long address) 1451 { 1452 if (!trace_pagefault_enabled()) 1453 return; 1454 1455 if (user_mode(regs)) 1456 trace_page_fault_user(address, regs, error_code); 1457 else 1458 trace_page_fault_kernel(address, regs, error_code); 1459 } 1460 1461 static __always_inline void 1462 handle_page_fault(struct pt_regs *regs, unsigned long error_code, 1463 unsigned long address) 1464 { 1465 trace_page_fault_entries(regs, error_code, address); 1466 1467 if (unlikely(kmmio_fault(regs, address))) 1468 return; 1469 1470 /* Was the fault on kernel-controlled part of the address space? */ 1471 if (unlikely(fault_in_kernel_space(address))) { 1472 do_kern_addr_fault(regs, error_code, address); 1473 } else { 1474 do_user_addr_fault(regs, error_code, address); 1475 /* 1476 * User address page fault handling might have reenabled 1477 * interrupts. Fixing up all potential exit points of 1478 * do_user_addr_fault() and its leaf functions is just not 1479 * doable w/o creating an unholy mess or turning the code 1480 * upside down. 1481 */ 1482 local_irq_disable(); 1483 } 1484 } 1485 1486 DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault) 1487 { 1488 irqentry_state_t state; 1489 unsigned long address; 1490 1491 address = cpu_feature_enabled(X86_FEATURE_FRED) ? fred_event_data(regs) : read_cr2(); 1492 1493 prefetchw(¤t->mm->mmap_lock); 1494 1495 /* 1496 * KVM uses #PF vector to deliver 'page not present' events to guests 1497 * (asynchronous page fault mechanism). The event happens when a 1498 * userspace task is trying to access some valid (from guest's point of 1499 * view) memory which is not currently mapped by the host (e.g. the 1500 * memory is swapped out). Note, the corresponding "page ready" event 1501 * which is injected when the memory becomes available, is delivered via 1502 * an interrupt mechanism and not a #PF exception 1503 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()). 1504 * 1505 * We are relying on the interrupted context being sane (valid RSP, 1506 * relevant locks not held, etc.), which is fine as long as the 1507 * interrupted context had IF=1. We are also relying on the KVM 1508 * async pf type field and CR2 being read consistently instead of 1509 * getting values from real and async page faults mixed up. 1510 * 1511 * Fingers crossed. 1512 * 1513 * The async #PF handling code takes care of idtentry handling 1514 * itself. 1515 */ 1516 if (kvm_handle_async_pf(regs, (u32)address)) 1517 return; 1518 1519 /* 1520 * Entry handling for valid #PF from kernel mode is slightly 1521 * different: RCU is already watching and ct_irq_enter() must not 1522 * be invoked because a kernel fault on a user space address might 1523 * sleep. 1524 * 1525 * In case the fault hit a RCU idle region the conditional entry 1526 * code reenabled RCU to avoid subsequent wreckage which helps 1527 * debuggability. 1528 */ 1529 state = irqentry_enter(regs); 1530 1531 instrumentation_begin(); 1532 handle_page_fault(regs, error_code, address); 1533 instrumentation_end(); 1534 1535 irqentry_exit(regs, state); 1536 } 1537