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/kprobes.h> /* NOKPROBE_SYMBOL, ... */ 13#include <linux/mmiotrace.h> /* kmmio_handler, ... */ 14#include <linux/perf_event.h> /* perf_sw_event */ 15#include <linux/hugetlb.h> /* hstate_index_to_shift */ 16#include <linux/prefetch.h> /* prefetchw */ 17#include <linux/context_tracking.h> /* exception_enter(), ... */ 18#include <linux/uaccess.h> /* faulthandler_disabled() */ 19#include <linux/efi.h> /* efi_recover_from_page_fault()*/ 20#include <linux/mm_types.h> 21 22#include <asm/cpufeature.h> /* boot_cpu_has, ... */ 23#include <asm/traps.h> /* dotraplinkage, ... */ 24#include <asm/fixmap.h> /* VSYSCALL_ADDR */ 25#include <asm/vsyscall.h> /* emulate_vsyscall */ 26#include <asm/vm86.h> /* struct vm86 */ 27#include <asm/mmu_context.h> /* vma_pkey() */ 28#include <asm/efi.h> /* efi_recover_from_page_fault()*/ 29#include <asm/desc.h> /* store_idt(), ... */ 30#include <asm/cpu_entry_area.h> /* exception stack */ 31#include <asm/pgtable_areas.h> /* VMALLOC_START, ... */ 32#include <asm/kvm_para.h> /* kvm_handle_async_pf */ 33 34#define CREATE_TRACE_POINTS 35#include <asm/trace/exceptions.h> 36 37/* 38 * Returns 0 if mmiotrace is disabled, or if the fault is not 39 * handled by mmiotrace: 40 */ 41static nokprobe_inline int 42kmmio_fault(struct pt_regs *regs, unsigned long addr) 43{ 44 if (unlikely(is_kmmio_active())) 45 if (kmmio_handler(regs, addr) == 1) 46 return -1; 47 return 0; 48} 49 50/* 51 * Prefetch quirks: 52 * 53 * 32-bit mode: 54 * 55 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch. 56 * Check that here and ignore it. This is AMD erratum #91. 57 * 58 * 64-bit mode: 59 * 60 * Sometimes the CPU reports invalid exceptions on prefetch. 61 * Check that here and ignore it. 62 * 63 * Opcode checker based on code by Richard Brunner. 64 */ 65static inline int 66check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr, 67 unsigned char opcode, int *prefetch) 68{ 69 unsigned char instr_hi = opcode & 0xf0; 70 unsigned char instr_lo = opcode & 0x0f; 71 72 switch (instr_hi) { 73 case 0x20: 74 case 0x30: 75 /* 76 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. 77 * In X86_64 long mode, the CPU will signal invalid 78 * opcode if some of these prefixes are present so 79 * X86_64 will never get here anyway 80 */ 81 return ((instr_lo & 7) == 0x6); 82#ifdef CONFIG_X86_64 83 case 0x40: 84 /* 85 * In 64-bit mode 0x40..0x4F are valid REX prefixes 86 */ 87 return (!user_mode(regs) || user_64bit_mode(regs)); 88#endif 89 case 0x60: 90 /* 0x64 thru 0x67 are valid prefixes in all modes. */ 91 return (instr_lo & 0xC) == 0x4; 92 case 0xF0: 93 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */ 94 return !instr_lo || (instr_lo>>1) == 1; 95 case 0x00: 96 /* Prefetch instruction is 0x0F0D or 0x0F18 */ 97 if (get_kernel_nofault(opcode, instr)) 98 return 0; 99 100 *prefetch = (instr_lo == 0xF) && 101 (opcode == 0x0D || opcode == 0x18); 102 return 0; 103 default: 104 return 0; 105 } 106} 107 108static int 109is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr) 110{ 111 unsigned char *max_instr; 112 unsigned char *instr; 113 int prefetch = 0; 114 115 /* 116 * If it was a exec (instruction fetch) fault on NX page, then 117 * do not ignore the fault: 118 */ 119 if (error_code & X86_PF_INSTR) 120 return 0; 121 122 instr = (void *)convert_ip_to_linear(current, regs); 123 max_instr = instr + 15; 124 125 /* 126 * This code has historically always bailed out if IP points to a 127 * not-present page (e.g. due to a race). No one has ever 128 * complained about this. 129 */ 130 pagefault_disable(); 131 132 while (instr < max_instr) { 133 unsigned char opcode; 134 135 if (user_mode(regs)) { 136 if (get_user(opcode, instr)) 137 break; 138 } else { 139 if (get_kernel_nofault(opcode, instr)) 140 break; 141 } 142 143 instr++; 144 145 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch)) 146 break; 147 } 148 149 pagefault_enable(); 150 return prefetch; 151} 152 153DEFINE_SPINLOCK(pgd_lock); 154LIST_HEAD(pgd_list); 155 156#ifdef CONFIG_X86_32 157static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) 158{ 159 unsigned index = pgd_index(address); 160 pgd_t *pgd_k; 161 p4d_t *p4d, *p4d_k; 162 pud_t *pud, *pud_k; 163 pmd_t *pmd, *pmd_k; 164 165 pgd += index; 166 pgd_k = init_mm.pgd + index; 167 168 if (!pgd_present(*pgd_k)) 169 return NULL; 170 171 /* 172 * set_pgd(pgd, *pgd_k); here would be useless on PAE 173 * and redundant with the set_pmd() on non-PAE. As would 174 * set_p4d/set_pud. 175 */ 176 p4d = p4d_offset(pgd, address); 177 p4d_k = p4d_offset(pgd_k, address); 178 if (!p4d_present(*p4d_k)) 179 return NULL; 180 181 pud = pud_offset(p4d, address); 182 pud_k = pud_offset(p4d_k, address); 183 if (!pud_present(*pud_k)) 184 return NULL; 185 186 pmd = pmd_offset(pud, address); 187 pmd_k = pmd_offset(pud_k, address); 188 189 if (pmd_present(*pmd) != pmd_present(*pmd_k)) 190 set_pmd(pmd, *pmd_k); 191 192 if (!pmd_present(*pmd_k)) 193 return NULL; 194 else 195 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k)); 196 197 return pmd_k; 198} 199 200/* 201 * Handle a fault on the vmalloc or module mapping area 202 * 203 * This is needed because there is a race condition between the time 204 * when the vmalloc mapping code updates the PMD to the point in time 205 * where it synchronizes this update with the other page-tables in the 206 * system. 207 * 208 * In this race window another thread/CPU can map an area on the same 209 * PMD, finds it already present and does not synchronize it with the 210 * rest of the system yet. As a result v[mz]alloc might return areas 211 * which are not mapped in every page-table in the system, causing an 212 * unhandled page-fault when they are accessed. 213 */ 214static noinline int vmalloc_fault(unsigned long address) 215{ 216 unsigned long pgd_paddr; 217 pmd_t *pmd_k; 218 pte_t *pte_k; 219 220 /* Make sure we are in vmalloc area: */ 221 if (!(address >= VMALLOC_START && address < VMALLOC_END)) 222 return -1; 223 224 /* 225 * Synchronize this task's top level page-table 226 * with the 'reference' page table. 227 * 228 * Do _not_ use "current" here. We might be inside 229 * an interrupt in the middle of a task switch.. 230 */ 231 pgd_paddr = read_cr3_pa(); 232 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address); 233 if (!pmd_k) 234 return -1; 235 236 if (pmd_large(*pmd_k)) 237 return 0; 238 239 pte_k = pte_offset_kernel(pmd_k, address); 240 if (!pte_present(*pte_k)) 241 return -1; 242 243 return 0; 244} 245NOKPROBE_SYMBOL(vmalloc_fault); 246 247void arch_sync_kernel_mappings(unsigned long start, unsigned long end) 248{ 249 unsigned long addr; 250 251 for (addr = start & PMD_MASK; 252 addr >= TASK_SIZE_MAX && addr < VMALLOC_END; 253 addr += PMD_SIZE) { 254 struct page *page; 255 256 spin_lock(&pgd_lock); 257 list_for_each_entry(page, &pgd_list, lru) { 258 spinlock_t *pgt_lock; 259 260 /* the pgt_lock only for Xen */ 261 pgt_lock = &pgd_page_get_mm(page)->page_table_lock; 262 263 spin_lock(pgt_lock); 264 vmalloc_sync_one(page_address(page), addr); 265 spin_unlock(pgt_lock); 266 } 267 spin_unlock(&pgd_lock); 268 } 269} 270 271/* 272 * Did it hit the DOS screen memory VA from vm86 mode? 273 */ 274static inline void 275check_v8086_mode(struct pt_regs *regs, unsigned long address, 276 struct task_struct *tsk) 277{ 278#ifdef CONFIG_VM86 279 unsigned long bit; 280 281 if (!v8086_mode(regs) || !tsk->thread.vm86) 282 return; 283 284 bit = (address - 0xA0000) >> PAGE_SHIFT; 285 if (bit < 32) 286 tsk->thread.vm86->screen_bitmap |= 1 << bit; 287#endif 288} 289 290static bool low_pfn(unsigned long pfn) 291{ 292 return pfn < max_low_pfn; 293} 294 295static 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_large(*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)); 329out: 330 pr_cont("\n"); 331} 332 333#else /* CONFIG_X86_64: */ 334 335#ifdef CONFIG_CPU_SUP_AMD 336static const char errata93_warning[] = 337KERN_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/* 345 * No vm86 mode in 64-bit mode: 346 */ 347static inline void 348check_v8086_mode(struct pt_regs *regs, unsigned long address, 349 struct task_struct *tsk) 350{ 351} 352 353static int bad_address(void *p) 354{ 355 unsigned long dummy; 356 357 return get_kernel_nofault(dummy, (unsigned long *)p); 358} 359 360static void dump_pagetable(unsigned long address) 361{ 362 pgd_t *base = __va(read_cr3_pa()); 363 pgd_t *pgd = base + pgd_index(address); 364 p4d_t *p4d; 365 pud_t *pud; 366 pmd_t *pmd; 367 pte_t *pte; 368 369 if (bad_address(pgd)) 370 goto bad; 371 372 pr_info("PGD %lx ", pgd_val(*pgd)); 373 374 if (!pgd_present(*pgd)) 375 goto out; 376 377 p4d = p4d_offset(pgd, address); 378 if (bad_address(p4d)) 379 goto bad; 380 381 pr_cont("P4D %lx ", p4d_val(*p4d)); 382 if (!p4d_present(*p4d) || p4d_large(*p4d)) 383 goto out; 384 385 pud = pud_offset(p4d, address); 386 if (bad_address(pud)) 387 goto bad; 388 389 pr_cont("PUD %lx ", pud_val(*pud)); 390 if (!pud_present(*pud) || pud_large(*pud)) 391 goto out; 392 393 pmd = pmd_offset(pud, address); 394 if (bad_address(pmd)) 395 goto bad; 396 397 pr_cont("PMD %lx ", pmd_val(*pmd)); 398 if (!pmd_present(*pmd) || pmd_large(*pmd)) 399 goto out; 400 401 pte = pte_offset_kernel(pmd, address); 402 if (bad_address(pte)) 403 goto bad; 404 405 pr_cont("PTE %lx", pte_val(*pte)); 406out: 407 pr_cont("\n"); 408 return; 409bad: 410 pr_info("BAD\n"); 411} 412 413#endif /* CONFIG_X86_64 */ 414 415/* 416 * Workaround for K8 erratum #93 & buggy BIOS. 417 * 418 * BIOS SMM functions are required to use a specific workaround 419 * to avoid corruption of the 64bit RIP register on C stepping K8. 420 * 421 * A lot of BIOS that didn't get tested properly miss this. 422 * 423 * The OS sees this as a page fault with the upper 32bits of RIP cleared. 424 * Try to work around it here. 425 * 426 * Note we only handle faults in kernel here. 427 * Does nothing on 32-bit. 428 */ 429static int is_errata93(struct pt_regs *regs, unsigned long address) 430{ 431#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD) 432 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD 433 || boot_cpu_data.x86 != 0xf) 434 return 0; 435 436 if (address != regs->ip) 437 return 0; 438 439 if ((address >> 32) != 0) 440 return 0; 441 442 address |= 0xffffffffUL << 32; 443 if ((address >= (u64)_stext && address <= (u64)_etext) || 444 (address >= MODULES_VADDR && address <= MODULES_END)) { 445 printk_once(errata93_warning); 446 regs->ip = address; 447 return 1; 448 } 449#endif 450 return 0; 451} 452 453/* 454 * Work around K8 erratum #100 K8 in compat mode occasionally jumps 455 * to illegal addresses >4GB. 456 * 457 * We catch this in the page fault handler because these addresses 458 * are not reachable. Just detect this case and return. Any code 459 * segment in LDT is compatibility mode. 460 */ 461static int is_errata100(struct pt_regs *regs, unsigned long address) 462{ 463#ifdef CONFIG_X86_64 464 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32)) 465 return 1; 466#endif 467 return 0; 468} 469 470/* Pentium F0 0F C7 C8 bug workaround: */ 471static int is_f00f_bug(struct pt_regs *regs, unsigned long address) 472{ 473#ifdef CONFIG_X86_F00F_BUG 474 if (boot_cpu_has_bug(X86_BUG_F00F) && idt_is_f00f_address(address)) { 475 handle_invalid_op(regs); 476 return 1; 477 } 478#endif 479 return 0; 480} 481 482static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index) 483{ 484 u32 offset = (index >> 3) * sizeof(struct desc_struct); 485 unsigned long addr; 486 struct ldttss_desc desc; 487 488 if (index == 0) { 489 pr_alert("%s: NULL\n", name); 490 return; 491 } 492 493 if (offset + sizeof(struct ldttss_desc) >= gdt->size) { 494 pr_alert("%s: 0x%hx -- out of bounds\n", name, index); 495 return; 496 } 497 498 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset), 499 sizeof(struct ldttss_desc))) { 500 pr_alert("%s: 0x%hx -- GDT entry is not readable\n", 501 name, index); 502 return; 503 } 504 505 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24); 506#ifdef CONFIG_X86_64 507 addr |= ((u64)desc.base3 << 32); 508#endif 509 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n", 510 name, index, addr, (desc.limit0 | (desc.limit1 << 16))); 511} 512 513static void 514show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address) 515{ 516 if (!oops_may_print()) 517 return; 518 519 if (error_code & X86_PF_INSTR) { 520 unsigned int level; 521 pgd_t *pgd; 522 pte_t *pte; 523 524 pgd = __va(read_cr3_pa()); 525 pgd += pgd_index(address); 526 527 pte = lookup_address_in_pgd(pgd, address, &level); 528 529 if (pte && pte_present(*pte) && !pte_exec(*pte)) 530 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", 531 from_kuid(&init_user_ns, current_uid())); 532 if (pte && pte_present(*pte) && pte_exec(*pte) && 533 (pgd_flags(*pgd) & _PAGE_USER) && 534 (__read_cr4() & X86_CR4_SMEP)) 535 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n", 536 from_kuid(&init_user_ns, current_uid())); 537 } 538 539 if (address < PAGE_SIZE && !user_mode(regs)) 540 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n", 541 (void *)address); 542 else 543 pr_alert("BUG: unable to handle page fault for address: %px\n", 544 (void *)address); 545 546 pr_alert("#PF: %s %s in %s mode\n", 547 (error_code & X86_PF_USER) ? "user" : "supervisor", 548 (error_code & X86_PF_INSTR) ? "instruction fetch" : 549 (error_code & X86_PF_WRITE) ? "write access" : 550 "read access", 551 user_mode(regs) ? "user" : "kernel"); 552 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code, 553 !(error_code & X86_PF_PROT) ? "not-present page" : 554 (error_code & X86_PF_RSVD) ? "reserved bit violation" : 555 (error_code & X86_PF_PK) ? "protection keys violation" : 556 "permissions violation"); 557 558 if (!(error_code & X86_PF_USER) && user_mode(regs)) { 559 struct desc_ptr idt, gdt; 560 u16 ldtr, tr; 561 562 /* 563 * This can happen for quite a few reasons. The more obvious 564 * ones are faults accessing the GDT, or LDT. Perhaps 565 * surprisingly, if the CPU tries to deliver a benign or 566 * contributory exception from user code and gets a page fault 567 * during delivery, the page fault can be delivered as though 568 * it originated directly from user code. This could happen 569 * due to wrong permissions on the IDT, GDT, LDT, TSS, or 570 * kernel or IST stack. 571 */ 572 store_idt(&idt); 573 574 /* Usable even on Xen PV -- it's just slow. */ 575 native_store_gdt(&gdt); 576 577 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n", 578 idt.address, idt.size, gdt.address, gdt.size); 579 580 store_ldt(ldtr); 581 show_ldttss(&gdt, "LDTR", ldtr); 582 583 store_tr(tr); 584 show_ldttss(&gdt, "TR", tr); 585 } 586 587 dump_pagetable(address); 588} 589 590static noinline void 591pgtable_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 612static void set_signal_archinfo(unsigned long address, 613 unsigned long error_code) 614{ 615 struct task_struct *tsk = current; 616 617 /* 618 * To avoid leaking information about the kernel page 619 * table layout, pretend that user-mode accesses to 620 * kernel addresses are always protection faults. 621 * 622 * NB: This means that failed vsyscalls with vsyscall=none 623 * will have the PROT bit. This doesn't leak any 624 * information and does not appear to cause any problems. 625 */ 626 if (address >= TASK_SIZE_MAX) 627 error_code |= X86_PF_PROT; 628 629 tsk->thread.trap_nr = X86_TRAP_PF; 630 tsk->thread.error_code = error_code | X86_PF_USER; 631 tsk->thread.cr2 = address; 632} 633 634static noinline void 635no_context(struct pt_regs *regs, unsigned long error_code, 636 unsigned long address, int signal, int si_code) 637{ 638 struct task_struct *tsk = current; 639 unsigned long flags; 640 int sig; 641 642 if (user_mode(regs)) { 643 /* 644 * This is an implicit supervisor-mode access from user 645 * mode. Bypass all the kernel-mode recovery code and just 646 * OOPS. 647 */ 648 goto oops; 649 } 650 651 /* Are we prepared to handle this kernel fault? */ 652 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) { 653 /* 654 * Any interrupt that takes a fault gets the fixup. This makes 655 * the below recursive fault logic only apply to a faults from 656 * task context. 657 */ 658 if (in_interrupt()) 659 return; 660 661 /* 662 * Per the above we're !in_interrupt(), aka. task context. 663 * 664 * In this case we need to make sure we're not recursively 665 * faulting through the emulate_vsyscall() logic. 666 */ 667 if (current->thread.sig_on_uaccess_err && signal) { 668 set_signal_archinfo(address, error_code); 669 670 /* XXX: hwpoison faults will set the wrong code. */ 671 force_sig_fault(signal, si_code, (void __user *)address); 672 } 673 674 /* 675 * Barring that, we can do the fixup and be happy. 676 */ 677 return; 678 } 679 680#ifdef CONFIG_VMAP_STACK 681 /* 682 * Stack overflow? During boot, we can fault near the initial 683 * stack in the direct map, but that's not an overflow -- check 684 * that we're in vmalloc space to avoid this. 685 */ 686 if (is_vmalloc_addr((void *)address) && 687 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) || 688 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) { 689 unsigned long stack = __this_cpu_ist_top_va(DF) - sizeof(void *); 690 /* 691 * We're likely to be running with very little stack space 692 * left. It's plausible that we'd hit this condition but 693 * double-fault even before we get this far, in which case 694 * we're fine: the double-fault handler will deal with it. 695 * 696 * We don't want to make it all the way into the oops code 697 * and then double-fault, though, because we're likely to 698 * break the console driver and lose most of the stack dump. 699 */ 700 asm volatile ("movq %[stack], %%rsp\n\t" 701 "call handle_stack_overflow\n\t" 702 "1: jmp 1b" 703 : ASM_CALL_CONSTRAINT 704 : "D" ("kernel stack overflow (page fault)"), 705 "S" (regs), "d" (address), 706 [stack] "rm" (stack)); 707 unreachable(); 708 } 709#endif 710 711 /* 712 * 32-bit: 713 * 714 * Valid to do another page fault here, because if this fault 715 * had been triggered by is_prefetch fixup_exception would have 716 * handled it. 717 * 718 * 64-bit: 719 * 720 * Hall of shame of CPU/BIOS bugs. 721 */ 722 if (is_prefetch(regs, error_code, address)) 723 return; 724 725 if (is_errata93(regs, address)) 726 return; 727 728 /* 729 * Buggy firmware could access regions which might page fault, try to 730 * recover from such faults. 731 */ 732 if (IS_ENABLED(CONFIG_EFI)) 733 efi_recover_from_page_fault(address); 734 735oops: 736 /* 737 * Oops. The kernel tried to access some bad page. We'll have to 738 * terminate things with extreme prejudice: 739 */ 740 flags = oops_begin(); 741 742 show_fault_oops(regs, error_code, address); 743 744 if (task_stack_end_corrupted(tsk)) 745 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 746 747 sig = SIGKILL; 748 if (__die("Oops", regs, error_code)) 749 sig = 0; 750 751 /* Executive summary in case the body of the oops scrolled away */ 752 printk(KERN_DEFAULT "CR2: %016lx\n", address); 753 754 oops_end(flags, regs, sig); 755} 756 757/* 758 * Print out info about fatal segfaults, if the show_unhandled_signals 759 * sysctl is set: 760 */ 761static inline void 762show_signal_msg(struct pt_regs *regs, unsigned long error_code, 763 unsigned long address, struct task_struct *tsk) 764{ 765 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG; 766 767 if (!unhandled_signal(tsk, SIGSEGV)) 768 return; 769 770 if (!printk_ratelimit()) 771 return; 772 773 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx", 774 loglvl, tsk->comm, task_pid_nr(tsk), address, 775 (void *)regs->ip, (void *)regs->sp, error_code); 776 777 print_vma_addr(KERN_CONT " in ", regs->ip); 778 779 printk(KERN_CONT "\n"); 780 781 show_opcodes(regs, loglvl); 782} 783 784/* 785 * The (legacy) vsyscall page is the long page in the kernel portion 786 * of the address space that has user-accessible permissions. 787 */ 788static bool is_vsyscall_vaddr(unsigned long vaddr) 789{ 790 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR); 791} 792 793static void 794__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 795 unsigned long address, u32 pkey, int si_code) 796{ 797 struct task_struct *tsk = current; 798 799 /* User mode accesses just cause a SIGSEGV */ 800 if (user_mode(regs) && (error_code & X86_PF_USER)) { 801 /* 802 * It's possible to have interrupts off here: 803 */ 804 local_irq_enable(); 805 806 /* 807 * Valid to do another page fault here because this one came 808 * from user space: 809 */ 810 if (is_prefetch(regs, error_code, address)) 811 return; 812 813 if (is_errata100(regs, address)) 814 return; 815 816 /* 817 * To avoid leaking information about the kernel page table 818 * layout, pretend that user-mode accesses to kernel addresses 819 * are always protection faults. 820 */ 821 if (address >= TASK_SIZE_MAX) 822 error_code |= X86_PF_PROT; 823 824 if (likely(show_unhandled_signals)) 825 show_signal_msg(regs, error_code, address, tsk); 826 827 set_signal_archinfo(address, error_code); 828 829 if (si_code == SEGV_PKUERR) 830 force_sig_pkuerr((void __user *)address, pkey); 831 832 force_sig_fault(SIGSEGV, si_code, (void __user *)address); 833 834 local_irq_disable(); 835 836 return; 837 } 838 839 if (is_f00f_bug(regs, address)) 840 return; 841 842 no_context(regs, error_code, address, SIGSEGV, si_code); 843} 844 845static noinline void 846bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code, 847 unsigned long address) 848{ 849 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR); 850} 851 852static void 853__bad_area(struct pt_regs *regs, unsigned long error_code, 854 unsigned long address, u32 pkey, int si_code) 855{ 856 struct mm_struct *mm = current->mm; 857 /* 858 * Something tried to access memory that isn't in our memory map.. 859 * Fix it, but check if it's kernel or user first.. 860 */ 861 mmap_read_unlock(mm); 862 863 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code); 864} 865 866static noinline void 867bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address) 868{ 869 __bad_area(regs, error_code, address, 0, SEGV_MAPERR); 870} 871 872static inline bool bad_area_access_from_pkeys(unsigned long error_code, 873 struct vm_area_struct *vma) 874{ 875 /* This code is always called on the current mm */ 876 bool foreign = false; 877 878 if (!boot_cpu_has(X86_FEATURE_OSPKE)) 879 return false; 880 if (error_code & X86_PF_PK) 881 return true; 882 /* this checks permission keys on the VMA: */ 883 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 884 (error_code & X86_PF_INSTR), foreign)) 885 return true; 886 return false; 887} 888 889static noinline void 890bad_area_access_error(struct pt_regs *regs, unsigned long error_code, 891 unsigned long address, struct vm_area_struct *vma) 892{ 893 /* 894 * This OSPKE check is not strictly necessary at runtime. 895 * But, doing it this way allows compiler optimizations 896 * if pkeys are compiled out. 897 */ 898 if (bad_area_access_from_pkeys(error_code, vma)) { 899 /* 900 * A protection key fault means that the PKRU value did not allow 901 * access to some PTE. Userspace can figure out what PKRU was 902 * from the XSAVE state. This function captures the pkey from 903 * the vma and passes it to userspace so userspace can discover 904 * which protection key was set on the PTE. 905 * 906 * If we get here, we know that the hardware signaled a X86_PF_PK 907 * fault and that there was a VMA once we got in the fault 908 * handler. It does *not* guarantee that the VMA we find here 909 * was the one that we faulted on. 910 * 911 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4); 912 * 2. T1 : set PKRU to deny access to pkey=4, touches page 913 * 3. T1 : faults... 914 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5); 915 * 5. T1 : enters fault handler, takes mmap_lock, etc... 916 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really 917 * faulted on a pte with its pkey=4. 918 */ 919 u32 pkey = vma_pkey(vma); 920 921 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR); 922 } else { 923 __bad_area(regs, error_code, address, 0, SEGV_ACCERR); 924 } 925} 926 927static void 928do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address, 929 vm_fault_t fault) 930{ 931 /* Kernel mode? Handle exceptions or die: */ 932 if (!(error_code & X86_PF_USER)) { 933 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR); 934 return; 935 } 936 937 /* User-space => ok to do another page fault: */ 938 if (is_prefetch(regs, error_code, address)) 939 return; 940 941 set_signal_archinfo(address, error_code); 942 943#ifdef CONFIG_MEMORY_FAILURE 944 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) { 945 struct task_struct *tsk = current; 946 unsigned lsb = 0; 947 948 pr_err( 949 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n", 950 tsk->comm, tsk->pid, address); 951 if (fault & VM_FAULT_HWPOISON_LARGE) 952 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 953 if (fault & VM_FAULT_HWPOISON) 954 lsb = PAGE_SHIFT; 955 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb); 956 return; 957 } 958#endif 959 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address); 960} 961 962static noinline void 963mm_fault_error(struct pt_regs *regs, unsigned long error_code, 964 unsigned long address, vm_fault_t fault) 965{ 966 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) { 967 no_context(regs, error_code, address, 0, 0); 968 return; 969 } 970 971 if (fault & VM_FAULT_OOM) { 972 /* Kernel mode? Handle exceptions or die: */ 973 if (!(error_code & X86_PF_USER)) { 974 no_context(regs, error_code, address, 975 SIGSEGV, SEGV_MAPERR); 976 return; 977 } 978 979 /* 980 * We ran out of memory, call the OOM killer, and return the 981 * userspace (which will retry the fault, or kill us if we got 982 * oom-killed): 983 */ 984 pagefault_out_of_memory(); 985 } else { 986 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON| 987 VM_FAULT_HWPOISON_LARGE)) 988 do_sigbus(regs, error_code, address, fault); 989 else if (fault & VM_FAULT_SIGSEGV) 990 bad_area_nosemaphore(regs, error_code, address); 991 else 992 BUG(); 993 } 994} 995 996static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte) 997{ 998 if ((error_code & X86_PF_WRITE) && !pte_write(*pte)) 999 return 0; 1000 1001 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte)) 1002 return 0; 1003 1004 return 1; 1005} 1006 1007/* 1008 * Handle a spurious fault caused by a stale TLB entry. 1009 * 1010 * This allows us to lazily refresh the TLB when increasing the 1011 * permissions of a kernel page (RO -> RW or NX -> X). Doing it 1012 * eagerly is very expensive since that implies doing a full 1013 * cross-processor TLB flush, even if no stale TLB entries exist 1014 * on other processors. 1015 * 1016 * Spurious faults may only occur if the TLB contains an entry with 1017 * fewer permission than the page table entry. Non-present (P = 0) 1018 * and reserved bit (R = 1) faults are never spurious. 1019 * 1020 * There are no security implications to leaving a stale TLB when 1021 * increasing the permissions on a page. 1022 * 1023 * Returns non-zero if a spurious fault was handled, zero otherwise. 1024 * 1025 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3 1026 * (Optional Invalidation). 1027 */ 1028static noinline int 1029spurious_kernel_fault(unsigned long error_code, unsigned long address) 1030{ 1031 pgd_t *pgd; 1032 p4d_t *p4d; 1033 pud_t *pud; 1034 pmd_t *pmd; 1035 pte_t *pte; 1036 int ret; 1037 1038 /* 1039 * Only writes to RO or instruction fetches from NX may cause 1040 * spurious faults. 1041 * 1042 * These could be from user or supervisor accesses but the TLB 1043 * is only lazily flushed after a kernel mapping protection 1044 * change, so user accesses are not expected to cause spurious 1045 * faults. 1046 */ 1047 if (error_code != (X86_PF_WRITE | X86_PF_PROT) && 1048 error_code != (X86_PF_INSTR | X86_PF_PROT)) 1049 return 0; 1050 1051 pgd = init_mm.pgd + pgd_index(address); 1052 if (!pgd_present(*pgd)) 1053 return 0; 1054 1055 p4d = p4d_offset(pgd, address); 1056 if (!p4d_present(*p4d)) 1057 return 0; 1058 1059 if (p4d_large(*p4d)) 1060 return spurious_kernel_fault_check(error_code, (pte_t *) p4d); 1061 1062 pud = pud_offset(p4d, address); 1063 if (!pud_present(*pud)) 1064 return 0; 1065 1066 if (pud_large(*pud)) 1067 return spurious_kernel_fault_check(error_code, (pte_t *) pud); 1068 1069 pmd = pmd_offset(pud, address); 1070 if (!pmd_present(*pmd)) 1071 return 0; 1072 1073 if (pmd_large(*pmd)) 1074 return spurious_kernel_fault_check(error_code, (pte_t *) pmd); 1075 1076 pte = pte_offset_kernel(pmd, address); 1077 if (!pte_present(*pte)) 1078 return 0; 1079 1080 ret = spurious_kernel_fault_check(error_code, pte); 1081 if (!ret) 1082 return 0; 1083 1084 /* 1085 * Make sure we have permissions in PMD. 1086 * If not, then there's a bug in the page tables: 1087 */ 1088 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd); 1089 WARN_ONCE(!ret, "PMD has incorrect permission bits\n"); 1090 1091 return ret; 1092} 1093NOKPROBE_SYMBOL(spurious_kernel_fault); 1094 1095int show_unhandled_signals = 1; 1096 1097static inline int 1098access_error(unsigned long error_code, struct vm_area_struct *vma) 1099{ 1100 /* This is only called for the current mm, so: */ 1101 bool foreign = false; 1102 1103 /* 1104 * Read or write was blocked by protection keys. This is 1105 * always an unconditional error and can never result in 1106 * a follow-up action to resolve the fault, like a COW. 1107 */ 1108 if (error_code & X86_PF_PK) 1109 return 1; 1110 1111 /* 1112 * Make sure to check the VMA so that we do not perform 1113 * faults just to hit a X86_PF_PK as soon as we fill in a 1114 * page. 1115 */ 1116 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE), 1117 (error_code & X86_PF_INSTR), foreign)) 1118 return 1; 1119 1120 if (error_code & X86_PF_WRITE) { 1121 /* write, present and write, not present: */ 1122 if (unlikely(!(vma->vm_flags & VM_WRITE))) 1123 return 1; 1124 return 0; 1125 } 1126 1127 /* read, present: */ 1128 if (unlikely(error_code & X86_PF_PROT)) 1129 return 1; 1130 1131 /* read, not present: */ 1132 if (unlikely(!vma_is_accessible(vma))) 1133 return 1; 1134 1135 return 0; 1136} 1137 1138bool fault_in_kernel_space(unsigned long address) 1139{ 1140 /* 1141 * On 64-bit systems, the vsyscall page is at an address above 1142 * TASK_SIZE_MAX, but is not considered part of the kernel 1143 * address space. 1144 */ 1145 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address)) 1146 return false; 1147 1148 return address >= TASK_SIZE_MAX; 1149} 1150 1151/* 1152 * Called for all faults where 'address' is part of the kernel address 1153 * space. Might get called for faults that originate from *code* that 1154 * ran in userspace or the kernel. 1155 */ 1156static void 1157do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code, 1158 unsigned long address) 1159{ 1160 /* 1161 * Protection keys exceptions only happen on user pages. We 1162 * have no user pages in the kernel portion of the address 1163 * space, so do not expect them here. 1164 */ 1165 WARN_ON_ONCE(hw_error_code & X86_PF_PK); 1166 1167#ifdef CONFIG_X86_32 1168 /* 1169 * We can fault-in kernel-space virtual memory on-demand. The 1170 * 'reference' page table is init_mm.pgd. 1171 * 1172 * NOTE! We MUST NOT take any locks for this case. We may 1173 * be in an interrupt or a critical region, and should 1174 * only copy the information from the master page table, 1175 * nothing more. 1176 * 1177 * Before doing this on-demand faulting, ensure that the 1178 * fault is not any of the following: 1179 * 1. A fault on a PTE with a reserved bit set. 1180 * 2. A fault caused by a user-mode access. (Do not demand- 1181 * fault kernel memory due to user-mode accesses). 1182 * 3. A fault caused by a page-level protection violation. 1183 * (A demand fault would be on a non-present page which 1184 * would have X86_PF_PROT==0). 1185 * 1186 * This is only needed to close a race condition on x86-32 in 1187 * the vmalloc mapping/unmapping code. See the comment above 1188 * vmalloc_fault() for details. On x86-64 the race does not 1189 * exist as the vmalloc mappings don't need to be synchronized 1190 * there. 1191 */ 1192 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) { 1193 if (vmalloc_fault(address) >= 0) 1194 return; 1195 } 1196#endif 1197 1198 /* Was the fault spurious, caused by lazy TLB invalidation? */ 1199 if (spurious_kernel_fault(hw_error_code, address)) 1200 return; 1201 1202 /* kprobes don't want to hook the spurious faults: */ 1203 if (kprobe_page_fault(regs, X86_TRAP_PF)) 1204 return; 1205 1206 /* 1207 * Note, despite being a "bad area", there are quite a few 1208 * acceptable reasons to get here, such as erratum fixups 1209 * and handling kernel code that can fault, like get_user(). 1210 * 1211 * Don't take the mm semaphore here. If we fixup a prefetch 1212 * fault we could otherwise deadlock: 1213 */ 1214 bad_area_nosemaphore(regs, hw_error_code, address); 1215} 1216NOKPROBE_SYMBOL(do_kern_addr_fault); 1217 1218/* Handle faults in the user portion of the address space */ 1219static inline 1220void do_user_addr_fault(struct pt_regs *regs, 1221 unsigned long hw_error_code, 1222 unsigned long address) 1223{ 1224 struct vm_area_struct *vma; 1225 struct task_struct *tsk; 1226 struct mm_struct *mm; 1227 vm_fault_t fault; 1228 unsigned int flags = FAULT_FLAG_DEFAULT; 1229 1230 tsk = current; 1231 mm = tsk->mm; 1232 1233 /* kprobes don't want to hook the spurious faults: */ 1234 if (unlikely(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(hw_error_code & X86_PF_RSVD)) 1242 pgtable_bad(regs, hw_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 !(hw_error_code & X86_PF_USER) && 1253 !(regs->flags & X86_EFLAGS_AC))) 1254 { 1255 bad_area_nosemaphore(regs, hw_error_code, address); 1256 return; 1257 } 1258 1259 /* 1260 * If we're in an interrupt, have no user context or are running 1261 * in a region with pagefaults disabled then we must not take the fault 1262 */ 1263 if (unlikely(faulthandler_disabled() || !mm)) { 1264 bad_area_nosemaphore(regs, hw_error_code, address); 1265 return; 1266 } 1267 1268 /* 1269 * It's safe to allow irq's after cr2 has been saved and the 1270 * vmalloc fault has been handled. 1271 * 1272 * User-mode registers count as a user access even for any 1273 * potential system fault or CPU buglet: 1274 */ 1275 if (user_mode(regs)) { 1276 local_irq_enable(); 1277 flags |= FAULT_FLAG_USER; 1278 } else { 1279 if (regs->flags & X86_EFLAGS_IF) 1280 local_irq_enable(); 1281 } 1282 1283 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); 1284 1285 if (hw_error_code & X86_PF_WRITE) 1286 flags |= FAULT_FLAG_WRITE; 1287 if (hw_error_code & X86_PF_INSTR) 1288 flags |= FAULT_FLAG_INSTRUCTION; 1289 1290#ifdef CONFIG_X86_64 1291 /* 1292 * Faults in the vsyscall page might need emulation. The 1293 * vsyscall page is at a high address (>PAGE_OFFSET), but is 1294 * considered to be part of the user address space. 1295 * 1296 * The vsyscall page does not have a "real" VMA, so do this 1297 * emulation before we go searching for VMAs. 1298 * 1299 * PKRU never rejects instruction fetches, so we don't need 1300 * to consider the PF_PK bit. 1301 */ 1302 if (is_vsyscall_vaddr(address)) { 1303 if (emulate_vsyscall(hw_error_code, regs, address)) 1304 return; 1305 } 1306#endif 1307 1308 /* 1309 * Kernel-mode access to the user address space should only occur 1310 * on well-defined single instructions listed in the exception 1311 * tables. But, an erroneous kernel fault occurring outside one of 1312 * those areas which also holds mmap_lock might deadlock attempting 1313 * to validate the fault against the address space. 1314 * 1315 * Only do the expensive exception table search when we might be at 1316 * risk of a deadlock. This happens if we 1317 * 1. Failed to acquire mmap_lock, and 1318 * 2. The access did not originate in userspace. 1319 */ 1320 if (unlikely(!mmap_read_trylock(mm))) { 1321 if (!user_mode(regs) && !search_exception_tables(regs->ip)) { 1322 /* 1323 * Fault from code in kernel from 1324 * which we do not expect faults. 1325 */ 1326 bad_area_nosemaphore(regs, hw_error_code, address); 1327 return; 1328 } 1329retry: 1330 mmap_read_lock(mm); 1331 } else { 1332 /* 1333 * The above down_read_trylock() might have succeeded in 1334 * which case we'll have missed the might_sleep() from 1335 * down_read(): 1336 */ 1337 might_sleep(); 1338 } 1339 1340 vma = find_vma(mm, address); 1341 if (unlikely(!vma)) { 1342 bad_area(regs, hw_error_code, address); 1343 return; 1344 } 1345 if (likely(vma->vm_start <= address)) 1346 goto good_area; 1347 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) { 1348 bad_area(regs, hw_error_code, address); 1349 return; 1350 } 1351 if (unlikely(expand_stack(vma, address))) { 1352 bad_area(regs, hw_error_code, address); 1353 return; 1354 } 1355 1356 /* 1357 * Ok, we have a good vm_area for this memory access, so 1358 * we can handle it.. 1359 */ 1360good_area: 1361 if (unlikely(access_error(hw_error_code, vma))) { 1362 bad_area_access_error(regs, hw_error_code, address, vma); 1363 return; 1364 } 1365 1366 /* 1367 * If for any reason at all we couldn't handle the fault, 1368 * make sure we exit gracefully rather than endlessly redo 1369 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if 1370 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked. 1371 * 1372 * Note that handle_userfault() may also release and reacquire mmap_lock 1373 * (and not return with VM_FAULT_RETRY), when returning to userland to 1374 * repeat the page fault later with a VM_FAULT_NOPAGE retval 1375 * (potentially after handling any pending signal during the return to 1376 * userland). The return to userland is identified whenever 1377 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags. 1378 */ 1379 fault = handle_mm_fault(vma, address, flags, regs); 1380 1381 /* Quick path to respond to signals */ 1382 if (fault_signal_pending(fault, regs)) { 1383 if (!user_mode(regs)) 1384 no_context(regs, hw_error_code, address, SIGBUS, 1385 BUS_ADRERR); 1386 return; 1387 } 1388 1389 /* 1390 * If we need to retry the mmap_lock has already been released, 1391 * and if there is a fatal signal pending there is no guarantee 1392 * that we made any progress. Handle this case first. 1393 */ 1394 if (unlikely((fault & VM_FAULT_RETRY) && 1395 (flags & FAULT_FLAG_ALLOW_RETRY))) { 1396 flags |= FAULT_FLAG_TRIED; 1397 goto retry; 1398 } 1399 1400 mmap_read_unlock(mm); 1401 if (unlikely(fault & VM_FAULT_ERROR)) { 1402 mm_fault_error(regs, hw_error_code, address, fault); 1403 return; 1404 } 1405 1406 check_v8086_mode(regs, address, tsk); 1407} 1408NOKPROBE_SYMBOL(do_user_addr_fault); 1409 1410static __always_inline void 1411trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code, 1412 unsigned long address) 1413{ 1414 if (!trace_pagefault_enabled()) 1415 return; 1416 1417 if (user_mode(regs)) 1418 trace_page_fault_user(address, regs, error_code); 1419 else 1420 trace_page_fault_kernel(address, regs, error_code); 1421} 1422 1423static __always_inline void 1424handle_page_fault(struct pt_regs *regs, unsigned long error_code, 1425 unsigned long address) 1426{ 1427 trace_page_fault_entries(regs, error_code, address); 1428 1429 if (unlikely(kmmio_fault(regs, address))) 1430 return; 1431 1432 /* Was the fault on kernel-controlled part of the address space? */ 1433 if (unlikely(fault_in_kernel_space(address))) { 1434 do_kern_addr_fault(regs, error_code, address); 1435 } else { 1436 do_user_addr_fault(regs, error_code, address); 1437 /* 1438 * User address page fault handling might have reenabled 1439 * interrupts. Fixing up all potential exit points of 1440 * do_user_addr_fault() and its leaf functions is just not 1441 * doable w/o creating an unholy mess or turning the code 1442 * upside down. 1443 */ 1444 local_irq_disable(); 1445 } 1446} 1447 1448DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault) 1449{ 1450 unsigned long address = read_cr2(); 1451 irqentry_state_t state; 1452 1453 prefetchw(¤t->mm->mmap_lock); 1454 1455 /* 1456 * KVM uses #PF vector to deliver 'page not present' events to guests 1457 * (asynchronous page fault mechanism). The event happens when a 1458 * userspace task is trying to access some valid (from guest's point of 1459 * view) memory which is not currently mapped by the host (e.g. the 1460 * memory is swapped out). Note, the corresponding "page ready" event 1461 * which is injected when the memory becomes available, is delived via 1462 * an interrupt mechanism and not a #PF exception 1463 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()). 1464 * 1465 * We are relying on the interrupted context being sane (valid RSP, 1466 * relevant locks not held, etc.), which is fine as long as the 1467 * interrupted context had IF=1. We are also relying on the KVM 1468 * async pf type field and CR2 being read consistently instead of 1469 * getting values from real and async page faults mixed up. 1470 * 1471 * Fingers crossed. 1472 * 1473 * The async #PF handling code takes care of idtentry handling 1474 * itself. 1475 */ 1476 if (kvm_handle_async_pf(regs, (u32)address)) 1477 return; 1478 1479 /* 1480 * Entry handling for valid #PF from kernel mode is slightly 1481 * different: RCU is already watching and rcu_irq_enter() must not 1482 * be invoked because a kernel fault on a user space address might 1483 * sleep. 1484 * 1485 * In case the fault hit a RCU idle region the conditional entry 1486 * code reenabled RCU to avoid subsequent wreckage which helps 1487 * debugability. 1488 */ 1489 state = irqentry_enter(regs); 1490 1491 instrumentation_begin(); 1492 handle_page_fault(regs, error_code, address); 1493 instrumentation_end(); 1494 1495 irqentry_exit(regs, state); 1496} 1497