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
  23#include <asm/cpufeature.h>		/* boot_cpu_has, ...		*/
  24#include <asm/traps.h>			/* dotraplinkage, ...		*/
 
  25#include <asm/fixmap.h>			/* VSYSCALL_ADDR		*/
  26#include <asm/vsyscall.h>		/* emulate_vsyscall		*/
  27#include <asm/vm86.h>			/* struct vm86			*/
  28#include <asm/mmu_context.h>		/* vma_pkey()			*/
  29#include <asm/efi.h>			/* efi_crash_gracefully_on_page_fault()*/
  30#include <asm/desc.h>			/* store_idt(), ...		*/
  31#include <asm/cpu_entry_area.h>		/* exception stack		*/
  32#include <asm/pgtable_areas.h>		/* VMALLOC_START, ...		*/
  33#include <asm/kvm_para.h>		/* kvm_handle_async_pf		*/
  34#include <asm/vdso.h>			/* fixup_vdso_exception()	*/
  35#include <asm/irq_stack.h>
  36
  37#define CREATE_TRACE_POINTS
  38#include <asm/trace/exceptions.h>
  39
  40/*
  41 * Returns 0 if mmiotrace is disabled, or if the fault is not
  42 * handled by mmiotrace:
  43 */
  44static nokprobe_inline int
  45kmmio_fault(struct pt_regs *regs, unsigned long addr)
  46{
  47	if (unlikely(is_kmmio_active()))
  48		if (kmmio_handler(regs, addr) == 1)
  49			return -1;
  50	return 0;
  51}
  52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  53/*
  54 * Prefetch quirks:
  55 *
  56 * 32-bit mode:
  57 *
  58 *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  59 *   Check that here and ignore it.  This is AMD erratum #91.
  60 *
  61 * 64-bit mode:
  62 *
  63 *   Sometimes the CPU reports invalid exceptions on prefetch.
  64 *   Check that here and ignore it.
  65 *
  66 * Opcode checker based on code by Richard Brunner.
  67 */
  68static inline int
  69check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  70		      unsigned char opcode, int *prefetch)
  71{
  72	unsigned char instr_hi = opcode & 0xf0;
  73	unsigned char instr_lo = opcode & 0x0f;
  74
  75	switch (instr_hi) {
  76	case 0x20:
  77	case 0x30:
  78		/*
  79		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  80		 * In X86_64 long mode, the CPU will signal invalid
  81		 * opcode if some of these prefixes are present so
  82		 * X86_64 will never get here anyway
  83		 */
  84		return ((instr_lo & 7) == 0x6);
  85#ifdef CONFIG_X86_64
  86	case 0x40:
  87		/*
  88		 * In 64-bit mode 0x40..0x4F are valid REX prefixes
 
 
 
 
  89		 */
  90		return (!user_mode(regs) || user_64bit_mode(regs));
  91#endif
  92	case 0x60:
  93		/* 0x64 thru 0x67 are valid prefixes in all modes. */
  94		return (instr_lo & 0xC) == 0x4;
  95	case 0xF0:
  96		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
  97		return !instr_lo || (instr_lo>>1) == 1;
  98	case 0x00:
  99		/* Prefetch instruction is 0x0F0D or 0x0F18 */
 100		if (get_kernel_nofault(opcode, instr))
 101			return 0;
 102
 103		*prefetch = (instr_lo == 0xF) &&
 104			(opcode == 0x0D || opcode == 0x18);
 105		return 0;
 106	default:
 107		return 0;
 108	}
 109}
 110
 111static bool is_amd_k8_pre_npt(void)
 112{
 113	struct cpuinfo_x86 *c = &boot_cpu_data;
 114
 115	return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
 116			c->x86_vendor == X86_VENDOR_AMD &&
 117			c->x86 == 0xf && c->x86_model < 0x40);
 118}
 119
 120static int
 121is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 122{
 123	unsigned char *max_instr;
 124	unsigned char *instr;
 125	int prefetch = 0;
 126
 127	/* Erratum #91 affects AMD K8, pre-NPT CPUs */
 128	if (!is_amd_k8_pre_npt())
 129		return 0;
 130
 131	/*
 132	 * If it was a exec (instruction fetch) fault on NX page, then
 133	 * do not ignore the fault:
 134	 */
 135	if (error_code & X86_PF_INSTR)
 136		return 0;
 137
 138	instr = (void *)convert_ip_to_linear(current, regs);
 139	max_instr = instr + 15;
 140
 141	/*
 142	 * This code has historically always bailed out if IP points to a
 143	 * not-present page (e.g. due to a race).  No one has ever
 144	 * complained about this.
 145	 */
 146	pagefault_disable();
 147
 148	while (instr < max_instr) {
 149		unsigned char opcode;
 150
 151		if (user_mode(regs)) {
 152			if (get_user(opcode, (unsigned char __user *) instr))
 153				break;
 154		} else {
 155			if (get_kernel_nofault(opcode, instr))
 156				break;
 157		}
 158
 159		instr++;
 160
 161		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 162			break;
 163	}
 164
 165	pagefault_enable();
 166	return prefetch;
 167}
 168
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 169DEFINE_SPINLOCK(pgd_lock);
 170LIST_HEAD(pgd_list);
 171
 172#ifdef CONFIG_X86_32
 173static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 174{
 175	unsigned index = pgd_index(address);
 176	pgd_t *pgd_k;
 177	p4d_t *p4d, *p4d_k;
 178	pud_t *pud, *pud_k;
 179	pmd_t *pmd, *pmd_k;
 180
 181	pgd += index;
 182	pgd_k = init_mm.pgd + index;
 183
 184	if (!pgd_present(*pgd_k))
 185		return NULL;
 186
 187	/*
 188	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
 189	 * and redundant with the set_pmd() on non-PAE. As would
 190	 * set_p4d/set_pud.
 191	 */
 192	p4d = p4d_offset(pgd, address);
 193	p4d_k = p4d_offset(pgd_k, address);
 194	if (!p4d_present(*p4d_k))
 195		return NULL;
 196
 197	pud = pud_offset(p4d, address);
 198	pud_k = pud_offset(p4d_k, address);
 199	if (!pud_present(*pud_k))
 200		return NULL;
 201
 202	pmd = pmd_offset(pud, address);
 203	pmd_k = pmd_offset(pud_k, address);
 204
 205	if (pmd_present(*pmd) != pmd_present(*pmd_k))
 206		set_pmd(pmd, *pmd_k);
 207
 208	if (!pmd_present(*pmd_k))
 209		return NULL;
 
 
 
 210	else
 211		BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
 212
 213	return pmd_k;
 214}
 215
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 216/*
 217 *   Handle a fault on the vmalloc or module mapping area
 218 *
 219 *   This is needed because there is a race condition between the time
 220 *   when the vmalloc mapping code updates the PMD to the point in time
 221 *   where it synchronizes this update with the other page-tables in the
 222 *   system.
 223 *
 224 *   In this race window another thread/CPU can map an area on the same
 225 *   PMD, finds it already present and does not synchronize it with the
 226 *   rest of the system yet. As a result v[mz]alloc might return areas
 227 *   which are not mapped in every page-table in the system, causing an
 228 *   unhandled page-fault when they are accessed.
 229 */
 230static noinline int vmalloc_fault(unsigned long address)
 231{
 232	unsigned long pgd_paddr;
 233	pmd_t *pmd_k;
 234	pte_t *pte_k;
 235
 236	/* Make sure we are in vmalloc area: */
 237	if (!(address >= VMALLOC_START && address < VMALLOC_END))
 238		return -1;
 239
 
 
 240	/*
 241	 * Synchronize this task's top level page-table
 242	 * with the 'reference' page table.
 243	 *
 244	 * Do _not_ use "current" here. We might be inside
 245	 * an interrupt in the middle of a task switch..
 246	 */
 247	pgd_paddr = read_cr3_pa();
 248	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 249	if (!pmd_k)
 250		return -1;
 251
 252	if (pmd_large(*pmd_k))
 253		return 0;
 254
 255	pte_k = pte_offset_kernel(pmd_k, address);
 256	if (!pte_present(*pte_k))
 257		return -1;
 258
 259	return 0;
 260}
 261NOKPROBE_SYMBOL(vmalloc_fault);
 262
 263static void __arch_sync_kernel_mappings(unsigned long start, unsigned long end)
 
 
 
 
 
 264{
 265	unsigned long addr;
 266
 267	for (addr = start & PMD_MASK;
 268	     addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
 269	     addr += PMD_SIZE) {
 270		struct page *page;
 271
 272		spin_lock(&pgd_lock);
 273		list_for_each_entry(page, &pgd_list, lru) {
 274			spinlock_t *pgt_lock;
 275
 276			/* the pgt_lock only for Xen */
 277			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 278
 279			spin_lock(pgt_lock);
 280			vmalloc_sync_one(page_address(page), addr);
 281			spin_unlock(pgt_lock);
 282		}
 283		spin_unlock(&pgd_lock);
 284	}
 285}
 286
 287void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
 288{
 289	__arch_sync_kernel_mappings(start, end);
 290#ifdef CONFIG_KMSAN
 291	/*
 292	 * KMSAN maintains two additional metadata page mappings for the
 293	 * [VMALLOC_START, VMALLOC_END) range. These mappings start at
 294	 * KMSAN_VMALLOC_SHADOW_START and KMSAN_VMALLOC_ORIGIN_START and
 295	 * have to be synced together with the vmalloc memory mapping.
 296	 */
 297	if (start >= VMALLOC_START && end < VMALLOC_END) {
 298		__arch_sync_kernel_mappings(
 299			start - VMALLOC_START + KMSAN_VMALLOC_SHADOW_START,
 300			end - VMALLOC_START + KMSAN_VMALLOC_SHADOW_START);
 301		__arch_sync_kernel_mappings(
 302			start - VMALLOC_START + KMSAN_VMALLOC_ORIGIN_START,
 303			end - VMALLOC_START + KMSAN_VMALLOC_ORIGIN_START);
 304	}
 305#endif
 306}
 307
 308static bool low_pfn(unsigned long pfn)
 309{
 310	return pfn < max_low_pfn;
 311}
 312
 313static void dump_pagetable(unsigned long address)
 314{
 315	pgd_t *base = __va(read_cr3_pa());
 316	pgd_t *pgd = &base[pgd_index(address)];
 317	p4d_t *p4d;
 318	pud_t *pud;
 319	pmd_t *pmd;
 320	pte_t *pte;
 321
 322#ifdef CONFIG_X86_PAE
 323	pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
 324	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 325		goto out;
 326#define pr_pde pr_cont
 327#else
 328#define pr_pde pr_info
 329#endif
 330	p4d = p4d_offset(pgd, address);
 331	pud = pud_offset(p4d, address);
 332	pmd = pmd_offset(pud, address);
 333	pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 334#undef pr_pde
 335
 336	/*
 337	 * We must not directly access the pte in the highpte
 338	 * case if the page table is located in highmem.
 339	 * And let's rather not kmap-atomic the pte, just in case
 340	 * it's allocated already:
 341	 */
 342	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
 343		goto out;
 344
 345	pte = pte_offset_kernel(pmd, address);
 346	pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 347out:
 348	pr_cont("\n");
 349}
 350
 351#else /* CONFIG_X86_64: */
 352
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 353#ifdef CONFIG_CPU_SUP_AMD
 354static const char errata93_warning[] =
 355KERN_ERR 
 356"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 357"******* Working around it, but it may cause SEGVs or burn power.\n"
 358"******* Please consider a BIOS update.\n"
 359"******* Disabling USB legacy in the BIOS may also help.\n";
 360#endif
 361
 
 
 
 
 
 
 
 
 
 362static int bad_address(void *p)
 363{
 364	unsigned long dummy;
 365
 366	return get_kernel_nofault(dummy, (unsigned long *)p);
 367}
 368
 369static void dump_pagetable(unsigned long address)
 370{
 371	pgd_t *base = __va(read_cr3_pa());
 372	pgd_t *pgd = base + pgd_index(address);
 373	p4d_t *p4d;
 374	pud_t *pud;
 375	pmd_t *pmd;
 376	pte_t *pte;
 377
 378	if (bad_address(pgd))
 379		goto bad;
 380
 381	pr_info("PGD %lx ", pgd_val(*pgd));
 382
 383	if (!pgd_present(*pgd))
 384		goto out;
 385
 386	p4d = p4d_offset(pgd, address);
 387	if (bad_address(p4d))
 388		goto bad;
 389
 390	pr_cont("P4D %lx ", p4d_val(*p4d));
 391	if (!p4d_present(*p4d) || p4d_large(*p4d))
 392		goto out;
 393
 394	pud = pud_offset(p4d, address);
 395	if (bad_address(pud))
 396		goto bad;
 397
 398	pr_cont("PUD %lx ", pud_val(*pud));
 399	if (!pud_present(*pud) || pud_large(*pud))
 400		goto out;
 401
 402	pmd = pmd_offset(pud, address);
 403	if (bad_address(pmd))
 404		goto bad;
 405
 406	pr_cont("PMD %lx ", pmd_val(*pmd));
 407	if (!pmd_present(*pmd) || pmd_large(*pmd))
 408		goto out;
 409
 410	pte = pte_offset_kernel(pmd, address);
 411	if (bad_address(pte))
 412		goto bad;
 413
 414	pr_cont("PTE %lx", pte_val(*pte));
 415out:
 416	pr_cont("\n");
 417	return;
 418bad:
 419	pr_info("BAD\n");
 420}
 421
 422#endif /* CONFIG_X86_64 */
 423
 424/*
 425 * Workaround for K8 erratum #93 & buggy BIOS.
 426 *
 427 * BIOS SMM functions are required to use a specific workaround
 428 * to avoid corruption of the 64bit RIP register on C stepping K8.
 429 *
 430 * A lot of BIOS that didn't get tested properly miss this.
 431 *
 432 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 433 * Try to work around it here.
 434 *
 435 * Note we only handle faults in kernel here.
 436 * Does nothing on 32-bit.
 437 */
 438static int is_errata93(struct pt_regs *regs, unsigned long address)
 439{
 440#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
 441	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
 442	    || boot_cpu_data.x86 != 0xf)
 443		return 0;
 444
 445	if (user_mode(regs))
 446		return 0;
 447
 448	if (address != regs->ip)
 449		return 0;
 450
 451	if ((address >> 32) != 0)
 452		return 0;
 453
 454	address |= 0xffffffffUL << 32;
 455	if ((address >= (u64)_stext && address <= (u64)_etext) ||
 456	    (address >= MODULES_VADDR && address <= MODULES_END)) {
 457		printk_once(errata93_warning);
 458		regs->ip = address;
 459		return 1;
 460	}
 461#endif
 462	return 0;
 463}
 464
 465/*
 466 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 467 * to illegal addresses >4GB.
 468 *
 469 * We catch this in the page fault handler because these addresses
 470 * are not reachable. Just detect this case and return.  Any code
 471 * segment in LDT is compatibility mode.
 472 */
 473static int is_errata100(struct pt_regs *regs, unsigned long address)
 474{
 475#ifdef CONFIG_X86_64
 476	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 477		return 1;
 478#endif
 479	return 0;
 480}
 481
 482/* Pentium F0 0F C7 C8 bug workaround: */
 483static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
 484		       unsigned long address)
 485{
 486#ifdef CONFIG_X86_F00F_BUG
 487	if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
 488	    idt_is_f00f_address(address)) {
 489		handle_invalid_op(regs);
 490		return 1;
 491	}
 492#endif
 493	return 0;
 494}
 495
 496static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
 497{
 498	u32 offset = (index >> 3) * sizeof(struct desc_struct);
 499	unsigned long addr;
 500	struct ldttss_desc desc;
 501
 502	if (index == 0) {
 503		pr_alert("%s: NULL\n", name);
 504		return;
 505	}
 506
 507	if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
 508		pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
 509		return;
 510	}
 511
 512	if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
 513			      sizeof(struct ldttss_desc))) {
 514		pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
 515			 name, index);
 516		return;
 517	}
 518
 519	addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
 520#ifdef CONFIG_X86_64
 521	addr |= ((u64)desc.base3 << 32);
 522#endif
 523	pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
 524		 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
 525}
 526
 
 
 
 
 
 527static void
 528show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 
 529{
 530	if (!oops_may_print())
 531		return;
 532
 533	if (error_code & X86_PF_INSTR) {
 534		unsigned int level;
 535		pgd_t *pgd;
 536		pte_t *pte;
 537
 538		pgd = __va(read_cr3_pa());
 539		pgd += pgd_index(address);
 540
 541		pte = lookup_address_in_pgd(pgd, address, &level);
 542
 543		if (pte && pte_present(*pte) && !pte_exec(*pte))
 544			pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
 545				from_kuid(&init_user_ns, current_uid()));
 546		if (pte && pte_present(*pte) && pte_exec(*pte) &&
 547				(pgd_flags(*pgd) & _PAGE_USER) &&
 548				(__read_cr4() & X86_CR4_SMEP))
 549			pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
 550				from_kuid(&init_user_ns, current_uid()));
 551	}
 552
 553	if (address < PAGE_SIZE && !user_mode(regs))
 554		pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
 555			(void *)address);
 556	else
 557		pr_alert("BUG: unable to handle page fault for address: %px\n",
 558			(void *)address);
 559
 560	pr_alert("#PF: %s %s in %s mode\n",
 561		 (error_code & X86_PF_USER)  ? "user" : "supervisor",
 562		 (error_code & X86_PF_INSTR) ? "instruction fetch" :
 563		 (error_code & X86_PF_WRITE) ? "write access" :
 564					       "read access",
 565			     user_mode(regs) ? "user" : "kernel");
 566	pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
 567		 !(error_code & X86_PF_PROT) ? "not-present page" :
 568		 (error_code & X86_PF_RSVD)  ? "reserved bit violation" :
 569		 (error_code & X86_PF_PK)    ? "protection keys violation" :
 570					       "permissions violation");
 571
 572	if (!(error_code & X86_PF_USER) && user_mode(regs)) {
 573		struct desc_ptr idt, gdt;
 574		u16 ldtr, tr;
 575
 576		/*
 577		 * This can happen for quite a few reasons.  The more obvious
 578		 * ones are faults accessing the GDT, or LDT.  Perhaps
 579		 * surprisingly, if the CPU tries to deliver a benign or
 580		 * contributory exception from user code and gets a page fault
 581		 * during delivery, the page fault can be delivered as though
 582		 * it originated directly from user code.  This could happen
 583		 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
 584		 * kernel or IST stack.
 585		 */
 586		store_idt(&idt);
 587
 588		/* Usable even on Xen PV -- it's just slow. */
 589		native_store_gdt(&gdt);
 590
 591		pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
 592			 idt.address, idt.size, gdt.address, gdt.size);
 593
 594		store_ldt(ldtr);
 595		show_ldttss(&gdt, "LDTR", ldtr);
 596
 597		store_tr(tr);
 598		show_ldttss(&gdt, "TR", tr);
 599	}
 600
 601	dump_pagetable(address);
 602}
 603
 604static noinline void
 605pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 606	    unsigned long address)
 607{
 608	struct task_struct *tsk;
 609	unsigned long flags;
 610	int sig;
 611
 612	flags = oops_begin();
 613	tsk = current;
 614	sig = SIGKILL;
 615
 616	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 617	       tsk->comm, address);
 618	dump_pagetable(address);
 619
 
 
 
 
 620	if (__die("Bad pagetable", regs, error_code))
 621		sig = 0;
 622
 623	oops_end(flags, regs, sig);
 624}
 625
 626static void sanitize_error_code(unsigned long address,
 627				unsigned long *error_code)
 628{
 629	/*
 630	 * To avoid leaking information about the kernel page
 631	 * table layout, pretend that user-mode accesses to
 632	 * kernel addresses are always protection faults.
 633	 *
 634	 * NB: This means that failed vsyscalls with vsyscall=none
 635	 * will have the PROT bit.  This doesn't leak any
 636	 * information and does not appear to cause any problems.
 637	 */
 638	if (address >= TASK_SIZE_MAX)
 639		*error_code |= X86_PF_PROT;
 640}
 641
 642static void set_signal_archinfo(unsigned long address,
 643				unsigned long error_code)
 644{
 645	struct task_struct *tsk = current;
 646
 647	tsk->thread.trap_nr = X86_TRAP_PF;
 648	tsk->thread.error_code = error_code | X86_PF_USER;
 649	tsk->thread.cr2 = address;
 650}
 651
 652static noinline void
 653page_fault_oops(struct pt_regs *regs, unsigned long error_code,
 654		unsigned long address)
 655{
 656#ifdef CONFIG_VMAP_STACK
 657	struct stack_info info;
 658#endif
 659	unsigned long flags;
 660	int sig;
 661
 662	if (user_mode(regs)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 663		/*
 664		 * Implicit kernel access from user mode?  Skip the stack
 665		 * overflow and EFI special cases.
 666		 */
 667		goto oops;
 668	}
 669
 670#ifdef CONFIG_VMAP_STACK
 671	/*
 672	 * Stack overflow?  During boot, we can fault near the initial
 673	 * stack in the direct map, but that's not an overflow -- check
 674	 * that we're in vmalloc space to avoid this.
 675	 */
 676	if (is_vmalloc_addr((void *)address) &&
 677	    get_stack_guard_info((void *)address, &info)) {
 
 
 678		/*
 679		 * We're likely to be running with very little stack space
 680		 * left.  It's plausible that we'd hit this condition but
 681		 * double-fault even before we get this far, in which case
 682		 * we're fine: the double-fault handler will deal with it.
 683		 *
 684		 * We don't want to make it all the way into the oops code
 685		 * and then double-fault, though, because we're likely to
 686		 * break the console driver and lose most of the stack dump.
 687		 */
 688		call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
 689			      handle_stack_overflow,
 690			      ASM_CALL_ARG3,
 691			      , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
 692
 
 
 693		unreachable();
 694	}
 695#endif
 696
 697	/*
 698	 * Buggy firmware could access regions which might page fault.  If
 699	 * this happens, EFI has a special OOPS path that will try to
 700	 * avoid hanging the system.
 
 
 
 
 
 
 701	 */
 702	if (IS_ENABLED(CONFIG_EFI))
 703		efi_crash_gracefully_on_page_fault(address);
 704
 705	/* Only not-present faults should be handled by KFENCE. */
 706	if (!(error_code & X86_PF_PROT) &&
 707	    kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
 708		return;
 709
 710oops:
 711	/*
 712	 * Oops. The kernel tried to access some bad page. We'll have to
 713	 * terminate things with extreme prejudice:
 714	 */
 715	flags = oops_begin();
 716
 717	show_fault_oops(regs, error_code, address);
 718
 719	if (task_stack_end_corrupted(current))
 720		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
 721
 
 
 
 
 722	sig = SIGKILL;
 723	if (__die("Oops", regs, error_code))
 724		sig = 0;
 725
 726	/* Executive summary in case the body of the oops scrolled away */
 727	printk(KERN_DEFAULT "CR2: %016lx\n", address);
 728
 729	oops_end(flags, regs, sig);
 730}
 731
 732static noinline void
 733kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
 734			 unsigned long address, int signal, int si_code,
 735			 u32 pkey)
 736{
 737	WARN_ON_ONCE(user_mode(regs));
 738
 739	/* Are we prepared to handle this kernel fault? */
 740	if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
 741		/*
 742		 * Any interrupt that takes a fault gets the fixup. This makes
 743		 * the below recursive fault logic only apply to a faults from
 744		 * task context.
 745		 */
 746		if (in_interrupt())
 747			return;
 748
 749		/*
 750		 * Per the above we're !in_interrupt(), aka. task context.
 751		 *
 752		 * In this case we need to make sure we're not recursively
 753		 * faulting through the emulate_vsyscall() logic.
 754		 */
 755		if (current->thread.sig_on_uaccess_err && signal) {
 756			sanitize_error_code(address, &error_code);
 757
 758			set_signal_archinfo(address, error_code);
 759
 760			if (si_code == SEGV_PKUERR) {
 761				force_sig_pkuerr((void __user *)address, pkey);
 762			} else {
 763				/* XXX: hwpoison faults will set the wrong code. */
 764				force_sig_fault(signal, si_code, (void __user *)address);
 765			}
 766		}
 767
 768		/*
 769		 * Barring that, we can do the fixup and be happy.
 770		 */
 771		return;
 772	}
 773
 774	/*
 775	 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
 776	 * instruction.
 777	 */
 778	if (is_prefetch(regs, error_code, address))
 779		return;
 780
 781	page_fault_oops(regs, error_code, address);
 782}
 783
 784/*
 785 * Print out info about fatal segfaults, if the show_unhandled_signals
 786 * sysctl is set:
 787 */
 788static inline void
 789show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 790		unsigned long address, struct task_struct *tsk)
 791{
 792	const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
 793	/* This is a racy snapshot, but it's better than nothing. */
 794	int cpu = raw_smp_processor_id();
 795
 796	if (!unhandled_signal(tsk, SIGSEGV))
 797		return;
 798
 799	if (!printk_ratelimit())
 800		return;
 801
 802	printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
 803		loglvl, tsk->comm, task_pid_nr(tsk), address,
 
 804		(void *)regs->ip, (void *)regs->sp, error_code);
 805
 806	print_vma_addr(KERN_CONT " in ", regs->ip);
 807
 808	/*
 809	 * Dump the likely CPU where the fatal segfault happened.
 810	 * This can help identify faulty hardware.
 811	 */
 812	printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu,
 813	       topology_core_id(cpu), topology_physical_package_id(cpu));
 814
 815
 816	printk(KERN_CONT "\n");
 817
 818	show_opcodes(regs, loglvl);
 819}
 820
 821/*
 822 * The (legacy) vsyscall page is the long page in the kernel portion
 823 * of the address space that has user-accessible permissions.
 824 */
 825static bool is_vsyscall_vaddr(unsigned long vaddr)
 826{
 827	return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
 828}
 829
 830static void
 831__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 832		       unsigned long address, u32 pkey, int si_code)
 833{
 834	struct task_struct *tsk = current;
 835
 836	if (!user_mode(regs)) {
 837		kernelmode_fixup_or_oops(regs, error_code, address,
 838					 SIGSEGV, si_code, pkey);
 839		return;
 840	}
 
 841
 842	if (!(error_code & X86_PF_USER)) {
 843		/* Implicit user access to kernel memory -- just oops */
 844		page_fault_oops(regs, error_code, address);
 845		return;
 846	}
 
 847
 848	/*
 849	 * User mode accesses just cause a SIGSEGV.
 850	 * It's possible to have interrupts off here:
 851	 */
 852	local_irq_enable();
 853
 854	/*
 855	 * Valid to do another page fault here because this one came
 856	 * from user space:
 857	 */
 858	if (is_prefetch(regs, error_code, address))
 859		return;
 
 
 
 
 
 860
 861	if (is_errata100(regs, address))
 862		return;
 
 
 
 
 
 863
 864	sanitize_error_code(address, &error_code);
 
 865
 866	if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
 867		return;
 
 868
 869	if (likely(show_unhandled_signals))
 870		show_signal_msg(regs, error_code, address, tsk);
 871
 872	set_signal_archinfo(address, error_code);
 
 873
 874	if (si_code == SEGV_PKUERR)
 875		force_sig_pkuerr((void __user *)address, pkey);
 876	else
 877		force_sig_fault(SIGSEGV, si_code, (void __user *)address);
 878
 879	local_irq_disable();
 880}
 881
 882static noinline void
 883bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 884		     unsigned long address)
 885{
 886	__bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
 887}
 888
 889static void
 890__bad_area(struct pt_regs *regs, unsigned long error_code,
 891	   unsigned long address, u32 pkey, int si_code)
 892{
 893	struct mm_struct *mm = current->mm;
 
 
 
 
 
 894	/*
 895	 * Something tried to access memory that isn't in our memory map..
 896	 * Fix it, but check if it's kernel or user first..
 897	 */
 898	mmap_read_unlock(mm);
 899
 900	__bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
 
 901}
 902
 903static noinline void
 904bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 905{
 906	__bad_area(regs, error_code, address, 0, SEGV_MAPERR);
 907}
 908
 909static inline bool bad_area_access_from_pkeys(unsigned long error_code,
 910		struct vm_area_struct *vma)
 911{
 912	/* This code is always called on the current mm */
 913	bool foreign = false;
 914
 915	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
 916		return false;
 917	if (error_code & X86_PF_PK)
 918		return true;
 919	/* this checks permission keys on the VMA: */
 920	if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
 921				       (error_code & X86_PF_INSTR), foreign))
 922		return true;
 923	return false;
 924}
 925
 926static noinline void
 927bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 928		      unsigned long address, struct vm_area_struct *vma)
 929{
 930	/*
 931	 * This OSPKE check is not strictly necessary at runtime.
 932	 * But, doing it this way allows compiler optimizations
 933	 * if pkeys are compiled out.
 934	 */
 935	if (bad_area_access_from_pkeys(error_code, vma)) {
 936		/*
 937		 * A protection key fault means that the PKRU value did not allow
 938		 * access to some PTE.  Userspace can figure out what PKRU was
 939		 * from the XSAVE state.  This function captures the pkey from
 940		 * the vma and passes it to userspace so userspace can discover
 941		 * which protection key was set on the PTE.
 942		 *
 943		 * If we get here, we know that the hardware signaled a X86_PF_PK
 944		 * fault and that there was a VMA once we got in the fault
 945		 * handler.  It does *not* guarantee that the VMA we find here
 946		 * was the one that we faulted on.
 947		 *
 948		 * 1. T1   : mprotect_key(foo, PAGE_SIZE, pkey=4);
 949		 * 2. T1   : set PKRU to deny access to pkey=4, touches page
 950		 * 3. T1   : faults...
 951		 * 4.    T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
 952		 * 5. T1   : enters fault handler, takes mmap_lock, etc...
 953		 * 6. T1   : reaches here, sees vma_pkey(vma)=5, when we really
 954		 *	     faulted on a pte with its pkey=4.
 955		 */
 956		u32 pkey = vma_pkey(vma);
 957
 958		__bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
 959	} else {
 960		__bad_area(regs, error_code, address, 0, SEGV_ACCERR);
 961	}
 962}
 963
 964static void
 965do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 966	  vm_fault_t fault)
 967{
 
 
 
 968	/* Kernel mode? Handle exceptions or die: */
 969	if (!user_mode(regs)) {
 970		kernelmode_fixup_or_oops(regs, error_code, address,
 971					 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
 972		return;
 973	}
 974
 975	/* User-space => ok to do another page fault: */
 976	if (is_prefetch(regs, error_code, address))
 977		return;
 978
 979	sanitize_error_code(address, &error_code);
 980
 981	if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
 982		return;
 983
 984	set_signal_archinfo(address, error_code);
 985
 986#ifdef CONFIG_MEMORY_FAILURE
 987	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
 988		struct task_struct *tsk = current;
 989		unsigned lsb = 0;
 990
 991		pr_err(
 992	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
 993			tsk->comm, tsk->pid, address);
 994		if (fault & VM_FAULT_HWPOISON_LARGE)
 995			lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
 996		if (fault & VM_FAULT_HWPOISON)
 997			lsb = PAGE_SHIFT;
 998		force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
 999		return;
1000	}
1001#endif
1002	force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1003}
1004
1005static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
1006{
1007	if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1008		return 0;
1009
1010	if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1011		return 0;
 
 
 
 
 
 
1012
1013	return 1;
1014}
1015
1016/*
1017 * Handle a spurious fault caused by a stale TLB entry.
1018 *
1019 * This allows us to lazily refresh the TLB when increasing the
1020 * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
1021 * eagerly is very expensive since that implies doing a full
1022 * cross-processor TLB flush, even if no stale TLB entries exist
1023 * on other processors.
1024 *
1025 * Spurious faults may only occur if the TLB contains an entry with
1026 * fewer permission than the page table entry.  Non-present (P = 0)
1027 * and reserved bit (R = 1) faults are never spurious.
1028 *
1029 * There are no security implications to leaving a stale TLB when
1030 * increasing the permissions on a page.
1031 *
1032 * Returns non-zero if a spurious fault was handled, zero otherwise.
1033 *
1034 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1035 * (Optional Invalidation).
1036 */
1037static noinline int
1038spurious_kernel_fault(unsigned long error_code, unsigned long address)
1039{
1040	pgd_t *pgd;
1041	p4d_t *p4d;
1042	pud_t *pud;
1043	pmd_t *pmd;
1044	pte_t *pte;
1045	int ret;
1046
1047	/*
1048	 * Only writes to RO or instruction fetches from NX may cause
1049	 * spurious faults.
1050	 *
1051	 * These could be from user or supervisor accesses but the TLB
1052	 * is only lazily flushed after a kernel mapping protection
1053	 * change, so user accesses are not expected to cause spurious
1054	 * faults.
1055	 */
1056	if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1057	    error_code != (X86_PF_INSTR | X86_PF_PROT))
1058		return 0;
1059
1060	pgd = init_mm.pgd + pgd_index(address);
1061	if (!pgd_present(*pgd))
1062		return 0;
1063
1064	p4d = p4d_offset(pgd, address);
1065	if (!p4d_present(*p4d))
1066		return 0;
1067
1068	if (p4d_large(*p4d))
1069		return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1070
1071	pud = pud_offset(p4d, address);
1072	if (!pud_present(*pud))
1073		return 0;
1074
1075	if (pud_large(*pud))
1076		return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1077
1078	pmd = pmd_offset(pud, address);
1079	if (!pmd_present(*pmd))
1080		return 0;
1081
1082	if (pmd_large(*pmd))
1083		return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1084
1085	pte = pte_offset_kernel(pmd, address);
1086	if (!pte_present(*pte))
1087		return 0;
1088
1089	ret = spurious_kernel_fault_check(error_code, pte);
1090	if (!ret)
1091		return 0;
1092
1093	/*
1094	 * Make sure we have permissions in PMD.
1095	 * If not, then there's a bug in the page tables:
1096	 */
1097	ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1098	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1099
1100	return ret;
1101}
1102NOKPROBE_SYMBOL(spurious_kernel_fault);
1103
1104int show_unhandled_signals = 1;
1105
1106static inline int
1107access_error(unsigned long error_code, struct vm_area_struct *vma)
1108{
1109	/* This is only called for the current mm, so: */
1110	bool foreign = false;
1111
1112	/*
1113	 * Read or write was blocked by protection keys.  This is
1114	 * always an unconditional error and can never result in
1115	 * a follow-up action to resolve the fault, like a COW.
1116	 */
1117	if (error_code & X86_PF_PK)
1118		return 1;
1119
1120	/*
1121	 * SGX hardware blocked the access.  This usually happens
1122	 * when the enclave memory contents have been destroyed, like
1123	 * after a suspend/resume cycle. In any case, the kernel can't
1124	 * fix the cause of the fault.  Handle the fault as an access
1125	 * error even in cases where no actual access violation
1126	 * occurred.  This allows userspace to rebuild the enclave in
1127	 * response to the signal.
1128	 */
1129	if (unlikely(error_code & X86_PF_SGX))
1130		return 1;
1131
1132	/*
1133	 * Make sure to check the VMA so that we do not perform
1134	 * faults just to hit a X86_PF_PK as soon as we fill in a
1135	 * page.
1136	 */
1137	if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1138				       (error_code & X86_PF_INSTR), foreign))
1139		return 1;
1140
1141	if (error_code & X86_PF_WRITE) {
1142		/* write, present and write, not present: */
1143		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1144			return 1;
1145		return 0;
1146	}
1147
1148	/* read, present: */
1149	if (unlikely(error_code & X86_PF_PROT))
1150		return 1;
1151
1152	/* read, not present: */
1153	if (unlikely(!vma_is_accessible(vma)))
1154		return 1;
1155
1156	return 0;
1157}
1158
1159bool fault_in_kernel_space(unsigned long address)
1160{
1161	/*
1162	 * On 64-bit systems, the vsyscall page is at an address above
1163	 * TASK_SIZE_MAX, but is not considered part of the kernel
1164	 * address space.
1165	 */
1166	if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1167		return false;
1168
1169	return address >= TASK_SIZE_MAX;
1170}
1171
1172/*
1173 * Called for all faults where 'address' is part of the kernel address
1174 * space.  Might get called for faults that originate from *code* that
1175 * ran in userspace or the kernel.
1176 */
1177static void
1178do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1179		   unsigned long address)
1180{
1181	/*
1182	 * Protection keys exceptions only happen on user pages.  We
1183	 * have no user pages in the kernel portion of the address
1184	 * space, so do not expect them here.
1185	 */
1186	WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1187
1188#ifdef CONFIG_X86_32
1189	/*
1190	 * We can fault-in kernel-space virtual memory on-demand. The
1191	 * 'reference' page table is init_mm.pgd.
1192	 *
1193	 * NOTE! We MUST NOT take any locks for this case. We may
1194	 * be in an interrupt or a critical region, and should
1195	 * only copy the information from the master page table,
1196	 * nothing more.
1197	 *
1198	 * Before doing this on-demand faulting, ensure that the
1199	 * fault is not any of the following:
1200	 * 1. A fault on a PTE with a reserved bit set.
1201	 * 2. A fault caused by a user-mode access.  (Do not demand-
1202	 *    fault kernel memory due to user-mode accesses).
1203	 * 3. A fault caused by a page-level protection violation.
1204	 *    (A demand fault would be on a non-present page which
1205	 *     would have X86_PF_PROT==0).
1206	 *
1207	 * This is only needed to close a race condition on x86-32 in
1208	 * the vmalloc mapping/unmapping code. See the comment above
1209	 * vmalloc_fault() for details. On x86-64 the race does not
1210	 * exist as the vmalloc mappings don't need to be synchronized
1211	 * there.
1212	 */
1213	if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1214		if (vmalloc_fault(address) >= 0)
1215			return;
1216	}
1217#endif
1218
1219	if (is_f00f_bug(regs, hw_error_code, address))
1220		return;
1221
1222	/* Was the fault spurious, caused by lazy TLB invalidation? */
1223	if (spurious_kernel_fault(hw_error_code, address))
1224		return;
1225
1226	/* kprobes don't want to hook the spurious faults: */
1227	if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1228		return;
1229
1230	/*
1231	 * Note, despite being a "bad area", there are quite a few
1232	 * acceptable reasons to get here, such as erratum fixups
1233	 * and handling kernel code that can fault, like get_user().
1234	 *
1235	 * Don't take the mm semaphore here. If we fixup a prefetch
1236	 * fault we could otherwise deadlock:
1237	 */
1238	bad_area_nosemaphore(regs, hw_error_code, address);
1239}
1240NOKPROBE_SYMBOL(do_kern_addr_fault);
1241
1242/*
1243 * Handle faults in the user portion of the address space.  Nothing in here
1244 * should check X86_PF_USER without a specific justification: for almost
1245 * all purposes, we should treat a normal kernel access to user memory
1246 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1247 * The one exception is AC flag handling, which is, per the x86
1248 * architecture, special for WRUSS.
1249 */
1250static inline
1251void do_user_addr_fault(struct pt_regs *regs,
1252			unsigned long error_code,
1253			unsigned long address)
1254{
1255	struct vm_area_struct *vma;
1256	struct task_struct *tsk;
1257	struct mm_struct *mm;
1258	vm_fault_t fault;
1259	unsigned int flags = FAULT_FLAG_DEFAULT;
 
1260
1261	tsk = current;
1262	mm = tsk->mm;
1263
1264	if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1265		/*
1266		 * Whoops, this is kernel mode code trying to execute from
1267		 * user memory.  Unless this is AMD erratum #93, which
1268		 * corrupts RIP such that it looks like a user address,
1269		 * this is unrecoverable.  Don't even try to look up the
1270		 * VMA or look for extable entries.
1271		 */
1272		if (is_errata93(regs, address))
1273			return;
1274
1275		page_fault_oops(regs, error_code, address);
1276		return;
1277	}
1278
1279	/* kprobes don't want to hook the spurious faults: */
1280	if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1281		return;
1282
1283	/*
1284	 * Reserved bits are never expected to be set on
1285	 * entries in the user portion of the page tables.
1286	 */
1287	if (unlikely(error_code & X86_PF_RSVD))
1288		pgtable_bad(regs, error_code, address);
1289
1290	/*
1291	 * If SMAP is on, check for invalid kernel (supervisor) access to user
1292	 * pages in the user address space.  The odd case here is WRUSS,
1293	 * which, according to the preliminary documentation, does not respect
1294	 * SMAP and will have the USER bit set so, in all cases, SMAP
1295	 * enforcement appears to be consistent with the USER bit.
1296	 */
1297	if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1298		     !(error_code & X86_PF_USER) &&
1299		     !(regs->flags & X86_EFLAGS_AC))) {
1300		/*
1301		 * No extable entry here.  This was a kernel access to an
1302		 * invalid pointer.  get_kernel_nofault() will not get here.
1303		 */
1304		page_fault_oops(regs, error_code, address);
1305		return;
1306	}
1307
1308	/*
1309	 * If we're in an interrupt, have no user context or are running
1310	 * in a region with pagefaults disabled then we must not take the fault
1311	 */
1312	if (unlikely(faulthandler_disabled() || !mm)) {
1313		bad_area_nosemaphore(regs, error_code, address);
1314		return;
1315	}
1316
1317	/*
1318	 * It's safe to allow irq's after cr2 has been saved and the
1319	 * vmalloc fault has been handled.
1320	 *
1321	 * User-mode registers count as a user access even for any
1322	 * potential system fault or CPU buglet:
1323	 */
1324	if (user_mode(regs)) {
1325		local_irq_enable();
 
1326		flags |= FAULT_FLAG_USER;
1327	} else {
1328		if (regs->flags & X86_EFLAGS_IF)
1329			local_irq_enable();
1330	}
1331
1332	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1333
1334	if (error_code & X86_PF_WRITE)
1335		flags |= FAULT_FLAG_WRITE;
1336	if (error_code & X86_PF_INSTR)
1337		flags |= FAULT_FLAG_INSTRUCTION;
1338
1339#ifdef CONFIG_X86_64
1340	/*
1341	 * Faults in the vsyscall page might need emulation.  The
1342	 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1343	 * considered to be part of the user address space.
 
 
 
 
 
1344	 *
1345	 * The vsyscall page does not have a "real" VMA, so do this
1346	 * emulation before we go searching for VMAs.
1347	 *
1348	 * PKRU never rejects instruction fetches, so we don't need
1349	 * to consider the PF_PK bit.
1350	 */
1351	if (is_vsyscall_vaddr(address)) {
1352		if (emulate_vsyscall(error_code, regs, address))
1353			return;
1354	}
1355#endif
1356
1357	/*
1358	 * Kernel-mode access to the user address space should only occur
1359	 * on well-defined single instructions listed in the exception
1360	 * tables.  But, an erroneous kernel fault occurring outside one of
1361	 * those areas which also holds mmap_lock might deadlock attempting
1362	 * to validate the fault against the address space.
1363	 *
1364	 * Only do the expensive exception table search when we might be at
1365	 * risk of a deadlock.  This happens if we
1366	 * 1. Failed to acquire mmap_lock, and
1367	 * 2. The access did not originate in userspace.
1368	 */
1369	if (unlikely(!mmap_read_trylock(mm))) {
1370		if (!user_mode(regs) && !search_exception_tables(regs->ip)) {
1371			/*
1372			 * Fault from code in kernel from
1373			 * which we do not expect faults.
1374			 */
1375			bad_area_nosemaphore(regs, error_code, address);
1376			return;
1377		}
1378retry:
1379		mmap_read_lock(mm);
1380	} else {
1381		/*
1382		 * The above down_read_trylock() might have succeeded in
1383		 * which case we'll have missed the might_sleep() from
1384		 * down_read():
1385		 */
1386		might_sleep();
1387	}
1388
1389	vma = find_vma(mm, address);
1390	if (unlikely(!vma)) {
1391		bad_area(regs, error_code, address);
1392		return;
1393	}
1394	if (likely(vma->vm_start <= address))
1395		goto good_area;
1396	if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1397		bad_area(regs, error_code, address);
1398		return;
1399	}
 
 
 
 
 
 
 
 
 
 
 
 
1400	if (unlikely(expand_stack(vma, address))) {
1401		bad_area(regs, error_code, address);
1402		return;
1403	}
1404
1405	/*
1406	 * Ok, we have a good vm_area for this memory access, so
1407	 * we can handle it..
1408	 */
1409good_area:
1410	if (unlikely(access_error(error_code, vma))) {
1411		bad_area_access_error(regs, error_code, address, vma);
1412		return;
1413	}
1414
1415	/*
1416	 * If for any reason at all we couldn't handle the fault,
1417	 * make sure we exit gracefully rather than endlessly redo
1418	 * the fault.  Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1419	 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1420	 *
1421	 * Note that handle_userfault() may also release and reacquire mmap_lock
1422	 * (and not return with VM_FAULT_RETRY), when returning to userland to
1423	 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1424	 * (potentially after handling any pending signal during the return to
1425	 * userland). The return to userland is identified whenever
1426	 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
 
 
1427	 */
1428	fault = handle_mm_fault(vma, address, flags, regs);
1429
1430	if (fault_signal_pending(fault, regs)) {
1431		/*
1432		 * Quick path to respond to signals.  The core mm code
1433		 * has unlocked the mm for us if we get here.
1434		 */
1435		if (!user_mode(regs))
1436			kernelmode_fixup_or_oops(regs, error_code, address,
1437						 SIGBUS, BUS_ADRERR,
1438						 ARCH_DEFAULT_PKEY);
1439		return;
1440	}
1441
1442	/* The fault is fully completed (including releasing mmap lock) */
1443	if (fault & VM_FAULT_COMPLETED)
1444		return;
1445
1446	/*
1447	 * If we need to retry the mmap_lock has already been released,
1448	 * and if there is a fatal signal pending there is no guarantee
1449	 * that we made any progress. Handle this case first.
1450	 */
1451	if (unlikely(fault & VM_FAULT_RETRY)) {
1452		flags |= FAULT_FLAG_TRIED;
1453		goto retry;
1454	}
 
 
 
 
1455
1456	mmap_read_unlock(mm);
1457	if (likely(!(fault & VM_FAULT_ERROR)))
1458		return;
1459
1460	if (fatal_signal_pending(current) && !user_mode(regs)) {
1461		kernelmode_fixup_or_oops(regs, error_code, address,
1462					 0, 0, ARCH_DEFAULT_PKEY);
1463		return;
1464	}
1465
1466	if (fault & VM_FAULT_OOM) {
1467		/* Kernel mode? Handle exceptions or die: */
1468		if (!user_mode(regs)) {
1469			kernelmode_fixup_or_oops(regs, error_code, address,
1470						 SIGSEGV, SEGV_MAPERR,
1471						 ARCH_DEFAULT_PKEY);
1472			return;
1473		}
1474
1475		/*
1476		 * We ran out of memory, call the OOM killer, and return the
1477		 * userspace (which will retry the fault, or kill us if we got
1478		 * oom-killed):
1479		 */
1480		pagefault_out_of_memory();
 
1481	} else {
1482		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1483			     VM_FAULT_HWPOISON_LARGE))
1484			do_sigbus(regs, error_code, address, fault);
1485		else if (fault & VM_FAULT_SIGSEGV)
1486			bad_area_nosemaphore(regs, error_code, address);
1487		else
1488			BUG();
1489	}
 
 
1490}
1491NOKPROBE_SYMBOL(do_user_addr_fault);
1492
1493static __always_inline void
1494trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1495			 unsigned long address)
1496{
1497	if (!trace_pagefault_enabled())
1498		return;
1499
1500	if (user_mode(regs))
1501		trace_page_fault_user(address, regs, error_code);
1502	else
1503		trace_page_fault_kernel(address, regs, error_code);
1504}
1505
1506static __always_inline void
1507handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1508			      unsigned long address)
1509{
1510	trace_page_fault_entries(regs, error_code, address);
1511
1512	if (unlikely(kmmio_fault(regs, address)))
1513		return;
1514
1515	/* Was the fault on kernel-controlled part of the address space? */
1516	if (unlikely(fault_in_kernel_space(address))) {
1517		do_kern_addr_fault(regs, error_code, address);
1518	} else {
1519		do_user_addr_fault(regs, error_code, address);
1520		/*
1521		 * User address page fault handling might have reenabled
1522		 * interrupts. Fixing up all potential exit points of
1523		 * do_user_addr_fault() and its leaf functions is just not
1524		 * doable w/o creating an unholy mess or turning the code
1525		 * upside down.
1526		 */
1527		local_irq_disable();
1528	}
1529}
1530
1531DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1532{
1533	unsigned long address = read_cr2();
1534	irqentry_state_t state;
1535
1536	prefetchw(&current->mm->mmap_lock);
1537
1538	/*
1539	 * KVM uses #PF vector to deliver 'page not present' events to guests
1540	 * (asynchronous page fault mechanism). The event happens when a
1541	 * userspace task is trying to access some valid (from guest's point of
1542	 * view) memory which is not currently mapped by the host (e.g. the
1543	 * memory is swapped out). Note, the corresponding "page ready" event
1544	 * which is injected when the memory becomes available, is delivered via
1545	 * an interrupt mechanism and not a #PF exception
1546	 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1547	 *
1548	 * We are relying on the interrupted context being sane (valid RSP,
1549	 * relevant locks not held, etc.), which is fine as long as the
1550	 * interrupted context had IF=1.  We are also relying on the KVM
1551	 * async pf type field and CR2 being read consistently instead of
1552	 * getting values from real and async page faults mixed up.
1553	 *
1554	 * Fingers crossed.
1555	 *
1556	 * The async #PF handling code takes care of idtentry handling
1557	 * itself.
1558	 */
1559	if (kvm_handle_async_pf(regs, (u32)address))
1560		return;
1561
1562	/*
1563	 * Entry handling for valid #PF from kernel mode is slightly
1564	 * different: RCU is already watching and ct_irq_enter() must not
1565	 * be invoked because a kernel fault on a user space address might
1566	 * sleep.
1567	 *
1568	 * In case the fault hit a RCU idle region the conditional entry
1569	 * code reenabled RCU to avoid subsequent wreckage which helps
1570	 * debuggability.
1571	 */
1572	state = irqentry_enter(regs);
1573
1574	instrumentation_begin();
1575	handle_page_fault(regs, error_code, address);
1576	instrumentation_end();
1577
1578	irqentry_exit(regs, state);
 
1579}