1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  Copyright (C) 1995  Linus Torvalds
   4 *  Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
   5 *  Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
   6 */
   7#include <linux/sched.h>		/* test_thread_flag(), ...	*/
   8#include <linux/sched/task_stack.h>	/* task_stack_*(), ...		*/
   9#include <linux/kdebug.h>		/* oops_begin/end, ...		*/
  10#include <linux/extable.h>		/* search_exception_tables	*/
  11#include <linux/memblock.h>		/* max_low_pfn			*/
  12#include <linux/kfence.h>		/* kfence_handle_page_fault	*/
  13#include <linux/kprobes.h>		/* NOKPROBE_SYMBOL, ...		*/
  14#include <linux/mmiotrace.h>		/* kmmio_handler, ...		*/
  15#include <linux/perf_event.h>		/* perf_sw_event		*/
  16#include <linux/hugetlb.h>		/* hstate_index_to_shift	*/
  17#include <linux/prefetch.h>		/* prefetchw			*/
  18#include <linux/context_tracking.h>	/* exception_enter(), ...	*/
  19#include <linux/uaccess.h>		/* faulthandler_disabled()	*/
  20#include <linux/efi.h>			/* efi_crash_gracefully_on_page_fault()*/
  21#include <linux/mm_types.h>
  22#include <linux/mm.h>			/* find_and_lock_vma() */
  23#include <linux/vmalloc.h>
  24
  25#include <asm/cpufeature.h>		/* boot_cpu_has, ...		*/
  26#include <asm/traps.h>			/* dotraplinkage, ...		*/
  27#include <asm/fixmap.h>			/* VSYSCALL_ADDR		*/
  28#include <asm/vsyscall.h>		/* emulate_vsyscall		*/
  29#include <asm/vm86.h>			/* struct vm86			*/
  30#include <asm/mmu_context.h>		/* vma_pkey()			*/
  31#include <asm/efi.h>			/* efi_crash_gracefully_on_page_fault()*/
  32#include <asm/desc.h>			/* store_idt(), ...		*/
  33#include <asm/cpu_entry_area.h>		/* exception stack		*/
  34#include <asm/pgtable_areas.h>		/* VMALLOC_START, ...		*/
  35#include <asm/kvm_para.h>		/* kvm_handle_async_pf		*/
  36#include <asm/vdso.h>			/* fixup_vdso_exception()	*/
  37#include <asm/irq_stack.h>
  38#include <asm/fred.h>
  39#include <asm/sev.h>			/* snp_dump_hva_rmpentry()	*/
  40
  41#define CREATE_TRACE_POINTS
  42#include <asm/trace/exceptions.h>
  43
  44/*
  45 * Returns 0 if mmiotrace is disabled, or if the fault is not
  46 * handled by mmiotrace:
  47 */
  48static nokprobe_inline int
  49kmmio_fault(struct pt_regs *regs, unsigned long addr)
  50{
  51	if (unlikely(is_kmmio_active()))
  52		if (kmmio_handler(regs, addr) == 1)
  53			return -1;
  54	return 0;
  55}
  56
  57/*
  58 * Prefetch quirks:
  59 *
  60 * 32-bit mode:
  61 *
  62 *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  63 *   Check that here and ignore it.  This is AMD erratum #91.
  64 *
  65 * 64-bit mode:
  66 *
  67 *   Sometimes the CPU reports invalid exceptions on prefetch.
  68 *   Check that here and ignore it.
  69 *
  70 * Opcode checker based on code by Richard Brunner.
  71 */
  72static inline int
  73check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  74		      unsigned char opcode, int *prefetch)
  75{
  76	unsigned char instr_hi = opcode & 0xf0;
  77	unsigned char instr_lo = opcode & 0x0f;
  78
  79	switch (instr_hi) {
  80	case 0x20:
  81	case 0x30:
  82		/*
  83		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  84		 * In X86_64 long mode, the CPU will signal invalid
  85		 * opcode if some of these prefixes are present so
  86		 * X86_64 will never get here anyway
  87		 */
  88		return ((instr_lo & 7) == 0x6);
  89#ifdef CONFIG_X86_64
  90	case 0x40:
  91		/*
  92		 * In 64-bit mode 0x40..0x4F are valid REX prefixes
  93		 */
  94		return (!user_mode(regs) || user_64bit_mode(regs));
  95#endif
  96	case 0x60:
  97		/* 0x64 thru 0x67 are valid prefixes in all modes. */
  98		return (instr_lo & 0xC) == 0x4;
  99	case 0xF0:
 100		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 101		return !instr_lo || (instr_lo>>1) == 1;
 102	case 0x00:
 103		/* Prefetch instruction is 0x0F0D or 0x0F18 */
 104		if (get_kernel_nofault(opcode, instr))
 105			return 0;
 106
 107		*prefetch = (instr_lo == 0xF) &&
 108			(opcode == 0x0D || opcode == 0x18);
 109		return 0;
 110	default:
 111		return 0;
 112	}
 113}
 114
 115static bool is_amd_k8_pre_npt(void)
 116{
 117	struct cpuinfo_x86 *c = &boot_cpu_data;
 118
 119	return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
 120			c->x86_vendor == X86_VENDOR_AMD &&
 121			c->x86 == 0xf && c->x86_model < 0x40);
 122}
 123
 124static int
 125is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 126{
 127	unsigned char *max_instr;
 128	unsigned char *instr;
 129	int prefetch = 0;
 130
 131	/* Erratum #91 affects AMD K8, pre-NPT CPUs */
 132	if (!is_amd_k8_pre_npt())
 133		return 0;
 134
 135	/*
 136	 * If it was a exec (instruction fetch) fault on NX page, then
 137	 * do not ignore the fault:
 138	 */
 139	if (error_code & X86_PF_INSTR)
 140		return 0;
 141
 142	instr = (void *)convert_ip_to_linear(current, regs);
 143	max_instr = instr + 15;
 144
 145	/*
 146	 * This code has historically always bailed out if IP points to a
 147	 * not-present page (e.g. due to a race).  No one has ever
 148	 * complained about this.
 149	 */
 150	pagefault_disable();
 151
 152	while (instr < max_instr) {
 153		unsigned char opcode;
 154
 155		if (user_mode(regs)) {
 156			if (get_user(opcode, (unsigned char __user *) instr))
 157				break;
 158		} else {
 159			if (get_kernel_nofault(opcode, instr))
 160				break;
 161		}
 162
 163		instr++;
 164
 165		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 166			break;
 167	}
 168
 169	pagefault_enable();
 170	return prefetch;
 171}
 172
 173DEFINE_SPINLOCK(pgd_lock);
 174LIST_HEAD(pgd_list);
 175
 176#ifdef CONFIG_X86_32
 177static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 178{
 179	unsigned index = pgd_index(address);
 180	pgd_t *pgd_k;
 181	p4d_t *p4d, *p4d_k;
 182	pud_t *pud, *pud_k;
 183	pmd_t *pmd, *pmd_k;
 184
 185	pgd += index;
 186	pgd_k = init_mm.pgd + index;
 187
 188	if (!pgd_present(*pgd_k))
 189		return NULL;
 190
 191	/*
 192	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
 193	 * and redundant with the set_pmd() on non-PAE. As would
 194	 * set_p4d/set_pud.
 195	 */
 196	p4d = p4d_offset(pgd, address);
 197	p4d_k = p4d_offset(pgd_k, address);
 198	if (!p4d_present(*p4d_k))
 199		return NULL;
 200
 201	pud = pud_offset(p4d, address);
 202	pud_k = pud_offset(p4d_k, address);
 203	if (!pud_present(*pud_k))
 204		return NULL;
 205
 206	pmd = pmd_offset(pud, address);
 207	pmd_k = pmd_offset(pud_k, address);
 208
 209	if (pmd_present(*pmd) != pmd_present(*pmd_k))
 210		set_pmd(pmd, *pmd_k);
 211
 212	if (!pmd_present(*pmd_k))
 213		return NULL;
 214	else
 215		BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
 216
 217	return pmd_k;
 218}
 219
 220/*
 221 *   Handle a fault on the vmalloc or module mapping area
 222 *
 223 *   This is needed because there is a race condition between the time
 224 *   when the vmalloc mapping code updates the PMD to the point in time
 225 *   where it synchronizes this update with the other page-tables in the
 226 *   system.
 227 *
 228 *   In this race window another thread/CPU can map an area on the same
 229 *   PMD, finds it already present and does not synchronize it with the
 230 *   rest of the system yet. As a result v[mz]alloc might return areas
 231 *   which are not mapped in every page-table in the system, causing an
 232 *   unhandled page-fault when they are accessed.
 233 */
 234static noinline int vmalloc_fault(unsigned long address)
 235{
 236	unsigned long pgd_paddr;
 237	pmd_t *pmd_k;
 238	pte_t *pte_k;
 239
 240	/* Make sure we are in vmalloc area: */
 241	if (!(address >= VMALLOC_START && address < VMALLOC_END))
 242		return -1;
 243
 244	/*
 245	 * Synchronize this task's top level page-table
 246	 * with the 'reference' page table.
 247	 *
 248	 * Do _not_ use "current" here. We might be inside
 249	 * an interrupt in the middle of a task switch..
 250	 */
 251	pgd_paddr = read_cr3_pa();
 252	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 253	if (!pmd_k)
 254		return -1;
 255
 256	if (pmd_leaf(*pmd_k))
 257		return 0;
 258
 259	pte_k = pte_offset_kernel(pmd_k, address);
 260	if (!pte_present(*pte_k))
 261		return -1;
 262
 263	return 0;
 264}
 265NOKPROBE_SYMBOL(vmalloc_fault);
 266
 267void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
 268{
 269	unsigned long addr;
 270
 271	for (addr = start & PMD_MASK;
 272	     addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
 273	     addr += PMD_SIZE) {
 274		struct page *page;
 275
 276		spin_lock(&pgd_lock);
 277		list_for_each_entry(page, &pgd_list, lru) {
 278			spinlock_t *pgt_lock;
 279
 280			/* the pgt_lock only for Xen */
 281			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 282
 283			spin_lock(pgt_lock);
 284			vmalloc_sync_one(page_address(page), addr);
 285			spin_unlock(pgt_lock);
 286		}
 287		spin_unlock(&pgd_lock);
 288	}
 289}
 290
 291static bool low_pfn(unsigned long pfn)
 292{
 293	return pfn < max_low_pfn;
 294}
 295
 296static void dump_pagetable(unsigned long address)
 297{
 298	pgd_t *base = __va(read_cr3_pa());
 299	pgd_t *pgd = &base[pgd_index(address)];
 300	p4d_t *p4d;
 301	pud_t *pud;
 302	pmd_t *pmd;
 303	pte_t *pte;
 304
 305#ifdef CONFIG_X86_PAE
 306	pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
 307	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 308		goto out;
 309#define pr_pde pr_cont
 310#else
 311#define pr_pde pr_info
 312#endif
 313	p4d = p4d_offset(pgd, address);
 314	pud = pud_offset(p4d, address);
 315	pmd = pmd_offset(pud, address);
 316	pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 317#undef pr_pde
 318
 319	/*
 320	 * We must not directly access the pte in the highpte
 321	 * case if the page table is located in highmem.
 322	 * And let's rather not kmap-atomic the pte, just in case
 323	 * it's allocated already:
 324	 */
 325	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_leaf(*pmd))
 326		goto out;
 327
 328	pte = pte_offset_kernel(pmd, address);
 329	pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 330out:
 331	pr_cont("\n");
 332}
 333
 334#else /* CONFIG_X86_64: */
 335
 336#ifdef CONFIG_CPU_SUP_AMD
 337static const char errata93_warning[] =
 338KERN_ERR 
 339"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 340"******* Working around it, but it may cause SEGVs or burn power.\n"
 341"******* Please consider a BIOS update.\n"
 342"******* Disabling USB legacy in the BIOS may also help.\n";
 343#endif
 344
 345static int bad_address(void *p)
 346{
 347	unsigned long dummy;
 348
 349	return get_kernel_nofault(dummy, (unsigned long *)p);
 350}
 351
 352static void dump_pagetable(unsigned long address)
 353{
 354	pgd_t *base = __va(read_cr3_pa());
 355	pgd_t *pgd = base + pgd_index(address);
 356	p4d_t *p4d;
 357	pud_t *pud;
 358	pmd_t *pmd;
 359	pte_t *pte;
 360
 361	if (bad_address(pgd))
 362		goto bad;
 363
 364	pr_info("PGD %lx ", pgd_val(*pgd));
 365
 366	if (!pgd_present(*pgd))
 367		goto out;
 368
 369	p4d = p4d_offset(pgd, address);
 370	if (bad_address(p4d))
 371		goto bad;
 372
 373	pr_cont("P4D %lx ", p4d_val(*p4d));
 374	if (!p4d_present(*p4d) || p4d_leaf(*p4d))
 375		goto out;
 376
 377	pud = pud_offset(p4d, address);
 378	if (bad_address(pud))
 379		goto bad;
 380
 381	pr_cont("PUD %lx ", pud_val(*pud));
 382	if (!pud_present(*pud) || pud_leaf(*pud))
 383		goto out;
 384
 385	pmd = pmd_offset(pud, address);
 386	if (bad_address(pmd))
 387		goto bad;
 388
 389	pr_cont("PMD %lx ", pmd_val(*pmd));
 390	if (!pmd_present(*pmd) || pmd_leaf(*pmd))
 391		goto out;
 392
 393	pte = pte_offset_kernel(pmd, address);
 394	if (bad_address(pte))
 395		goto bad;
 396
 397	pr_cont("PTE %lx", pte_val(*pte));
 398out:
 399	pr_cont("\n");
 400	return;
 401bad:
 402	pr_info("BAD\n");
 403}
 404
 405#endif /* CONFIG_X86_64 */
 406
 407/*
 408 * Workaround for K8 erratum #93 & buggy BIOS.
 409 *
 410 * BIOS SMM functions are required to use a specific workaround
 411 * to avoid corruption of the 64bit RIP register on C stepping K8.
 412 *
 413 * A lot of BIOS that didn't get tested properly miss this.
 414 *
 415 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 416 * Try to work around it here.
 417 *
 418 * Note we only handle faults in kernel here.
 419 * Does nothing on 32-bit.
 420 */
 421static int is_errata93(struct pt_regs *regs, unsigned long address)
 422{
 423#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
 424	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
 425	    || boot_cpu_data.x86 != 0xf)
 426		return 0;
 427
 428	if (user_mode(regs))
 429		return 0;
 430
 431	if (address != regs->ip)
 432		return 0;
 433
 434	if ((address >> 32) != 0)
 435		return 0;
 436
 437	address |= 0xffffffffUL << 32;
 438	if ((address >= (u64)_stext && address <= (u64)_etext) ||
 439	    (address >= MODULES_VADDR && address <= MODULES_END)) {
 440		printk_once(errata93_warning);
 441		regs->ip = address;
 442		return 1;
 443	}
 444#endif
 445	return 0;
 446}
 447
 448/*
 449 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 450 * to illegal addresses >4GB.
 451 *
 452 * We catch this in the page fault handler because these addresses
 453 * are not reachable. Just detect this case and return.  Any code
 454 * segment in LDT is compatibility mode.
 455 */
 456static int is_errata100(struct pt_regs *regs, unsigned long address)
 457{
 458#ifdef CONFIG_X86_64
 459	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 460		return 1;
 461#endif
 462	return 0;
 463}
 464
 465/* Pentium F0 0F C7 C8 bug workaround: */
 466static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
 467		       unsigned long address)
 468{
 469#ifdef CONFIG_X86_F00F_BUG
 470	if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
 471	    idt_is_f00f_address(address)) {
 472		handle_invalid_op(regs);
 473		return 1;
 474	}
 475#endif
 476	return 0;
 477}
 478
 479static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
 480{
 481	u32 offset = (index >> 3) * sizeof(struct desc_struct);
 482	unsigned long addr;
 483	struct ldttss_desc desc;
 484
 485	if (index == 0) {
 486		pr_alert("%s: NULL\n", name);
 487		return;
 488	}
 489
 490	if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
 491		pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
 492		return;
 493	}
 494
 495	if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
 496			      sizeof(struct ldttss_desc))) {
 497		pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
 498			 name, index);
 499		return;
 500	}
 501
 502	addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
 503#ifdef CONFIG_X86_64
 504	addr |= ((u64)desc.base3 << 32);
 505#endif
 506	pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
 507		 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
 508}
 509
 510static void
 511show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 512{
 513	if (!oops_may_print())
 514		return;
 515
 516	if (error_code & X86_PF_INSTR) {
 517		unsigned int level;
 518		bool nx, rw;
 519		pgd_t *pgd;
 520		pte_t *pte;
 521
 522		pgd = __va(read_cr3_pa());
 523		pgd += pgd_index(address);
 524
 525		pte = lookup_address_in_pgd_attr(pgd, address, &level, &nx, &rw);
 526
 527		if (pte && pte_present(*pte) && (!pte_exec(*pte) || nx))
 528			pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
 529				from_kuid(&init_user_ns, current_uid()));
 530		if (pte && pte_present(*pte) && pte_exec(*pte) && !nx &&
 531				(pgd_flags(*pgd) & _PAGE_USER) &&
 532				(__read_cr4() & X86_CR4_SMEP))
 533			pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
 534				from_kuid(&init_user_ns, current_uid()));
 535	}
 536
 537	if (address < PAGE_SIZE && !user_mode(regs))
 538		pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
 539			(void *)address);
 540	else
 541		pr_alert("BUG: unable to handle page fault for address: %px\n",
 542			(void *)address);
 543
 544	pr_alert("#PF: %s %s in %s mode\n",
 545		 (error_code & X86_PF_USER)  ? "user" : "supervisor",
 546		 (error_code & X86_PF_INSTR) ? "instruction fetch" :
 547		 (error_code & X86_PF_WRITE) ? "write access" :
 548					       "read access",
 549			     user_mode(regs) ? "user" : "kernel");
 550	pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
 551		 !(error_code & X86_PF_PROT) ? "not-present page" :
 552		 (error_code & X86_PF_RSVD)  ? "reserved bit violation" :
 553		 (error_code & X86_PF_PK)    ? "protection keys violation" :
 554		 (error_code & X86_PF_RMP)   ? "RMP violation" :
 555					       "permissions violation");
 556
 557	if (!(error_code & X86_PF_USER) && user_mode(regs)) {
 558		struct desc_ptr idt, gdt;
 559		u16 ldtr, tr;
 560
 561		/*
 562		 * This can happen for quite a few reasons.  The more obvious
 563		 * ones are faults accessing the GDT, or LDT.  Perhaps
 564		 * surprisingly, if the CPU tries to deliver a benign or
 565		 * contributory exception from user code and gets a page fault
 566		 * during delivery, the page fault can be delivered as though
 567		 * it originated directly from user code.  This could happen
 568		 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
 569		 * kernel or IST stack.
 570		 */
 571		store_idt(&idt);
 572
 573		/* Usable even on Xen PV -- it's just slow. */
 574		native_store_gdt(&gdt);
 575
 576		pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
 577			 idt.address, idt.size, gdt.address, gdt.size);
 578
 579		store_ldt(ldtr);
 580		show_ldttss(&gdt, "LDTR", ldtr);
 581
 582		store_tr(tr);
 583		show_ldttss(&gdt, "TR", tr);
 584	}
 585
 586	dump_pagetable(address);
 587
 588	if (error_code & X86_PF_RMP)
 589		snp_dump_hva_rmpentry(address);
 590}
 591
 592static noinline void
 593pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 594	    unsigned long address)
 595{
 596	struct task_struct *tsk;
 597	unsigned long flags;
 598	int sig;
 599
 600	flags = oops_begin();
 601	tsk = current;
 602	sig = SIGKILL;
 603
 604	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 605	       tsk->comm, address);
 606	dump_pagetable(address);
 607
 608	if (__die("Bad pagetable", regs, error_code))
 609		sig = 0;
 610
 611	oops_end(flags, regs, sig);
 612}
 613
 614static void sanitize_error_code(unsigned long address,
 615				unsigned long *error_code)
 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
 630static void set_signal_archinfo(unsigned long address,
 631				unsigned long error_code)
 632{
 633	struct task_struct *tsk = current;
 634
 635	tsk->thread.trap_nr = X86_TRAP_PF;
 636	tsk->thread.error_code = error_code | X86_PF_USER;
 637	tsk->thread.cr2 = address;
 638}
 639
 640static noinline void
 641page_fault_oops(struct pt_regs *regs, unsigned long error_code,
 642		unsigned long address)
 643{
 644#ifdef CONFIG_VMAP_STACK
 645	struct stack_info info;
 646#endif
 647	unsigned long flags;
 648	int sig;
 649
 650	if (user_mode(regs)) {
 651		/*
 652		 * Implicit kernel access from user mode?  Skip the stack
 653		 * overflow and EFI special cases.
 654		 */
 655		goto oops;
 656	}
 657
 658#ifdef CONFIG_VMAP_STACK
 659	/*
 660	 * Stack overflow?  During boot, we can fault near the initial
 661	 * stack in the direct map, but that's not an overflow -- check
 662	 * that we're in vmalloc space to avoid this.
 663	 */
 664	if (is_vmalloc_addr((void *)address) &&
 665	    get_stack_guard_info((void *)address, &info)) {
 666		/*
 667		 * We're likely to be running with very little stack space
 668		 * left.  It's plausible that we'd hit this condition but
 669		 * double-fault even before we get this far, in which case
 670		 * we're fine: the double-fault handler will deal with it.
 671		 *
 672		 * We don't want to make it all the way into the oops code
 673		 * and then double-fault, though, because we're likely to
 674		 * break the console driver and lose most of the stack dump.
 675		 */
 676		call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
 677			      handle_stack_overflow,
 678			      ASM_CALL_ARG3,
 679			      , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
 680
 681		BUG();
 682	}
 683#endif
 684
 685	/*
 686	 * Buggy firmware could access regions which might page fault.  If
 687	 * this happens, EFI has a special OOPS path that will try to
 688	 * avoid hanging the system.
 689	 */
 690	if (IS_ENABLED(CONFIG_EFI))
 691		efi_crash_gracefully_on_page_fault(address);
 692
 693	/* Only not-present faults should be handled by KFENCE. */
 694	if (!(error_code & X86_PF_PROT) &&
 695	    kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
 696		return;
 697
 698oops:
 699	/*
 700	 * Oops. The kernel tried to access some bad page. We'll have to
 701	 * terminate things with extreme prejudice:
 702	 */
 703	flags = oops_begin();
 704
 705	show_fault_oops(regs, error_code, address);
 706
 707	if (task_stack_end_corrupted(current))
 708		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
 709
 710	sig = SIGKILL;
 711	if (__die("Oops", regs, error_code))
 712		sig = 0;
 713
 714	/* Executive summary in case the body of the oops scrolled away */
 715	printk(KERN_DEFAULT "CR2: %016lx\n", address);
 716
 717	oops_end(flags, regs, sig);
 718}
 719
 720static noinline void
 721kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
 722			 unsigned long address, int signal, int si_code,
 723			 u32 pkey)
 724{
 725	WARN_ON_ONCE(user_mode(regs));
 726
 727	/* Are we prepared to handle this kernel fault? */
 728	if (fixup_exception(regs, X86_TRAP_PF, error_code, address))
 729		return;
 730
 731	/*
 732	 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
 733	 * instruction.
 734	 */
 735	if (is_prefetch(regs, error_code, address))
 736		return;
 737
 738	page_fault_oops(regs, error_code, address);
 739}
 740
 741/*
 742 * Print out info about fatal segfaults, if the show_unhandled_signals
 743 * sysctl is set:
 744 */
 745static inline void
 746show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 747		unsigned long address, struct task_struct *tsk)
 748{
 749	const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
 750	/* This is a racy snapshot, but it's better than nothing. */
 751	int cpu = raw_smp_processor_id();
 752
 753	if (!unhandled_signal(tsk, SIGSEGV))
 754		return;
 755
 756	if (!printk_ratelimit())
 757		return;
 758
 759	printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
 760		loglvl, tsk->comm, task_pid_nr(tsk), address,
 761		(void *)regs->ip, (void *)regs->sp, error_code);
 762
 763	print_vma_addr(KERN_CONT " in ", regs->ip);
 764
 765	/*
 766	 * Dump the likely CPU where the fatal segfault happened.
 767	 * This can help identify faulty hardware.
 768	 */
 769	printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu,
 770	       topology_core_id(cpu), topology_physical_package_id(cpu));
 771
 772
 773	printk(KERN_CONT "\n");
 774
 775	show_opcodes(regs, loglvl);
 776}
 777
 778static void
 779__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 780		       unsigned long address, u32 pkey, int si_code)
 781{
 782	struct task_struct *tsk = current;
 783
 784	if (!user_mode(regs)) {
 785		kernelmode_fixup_or_oops(regs, error_code, address,
 786					 SIGSEGV, si_code, pkey);
 787		return;
 788	}
 789
 790	if (!(error_code & X86_PF_USER)) {
 791		/* Implicit user access to kernel memory -- just oops */
 792		page_fault_oops(regs, error_code, address);
 793		return;
 794	}
 795
 796	/*
 797	 * User mode accesses just cause a SIGSEGV.
 798	 * It's possible to have interrupts off here:
 799	 */
 800	local_irq_enable();
 801
 802	/*
 803	 * Valid to do another page fault here because this one came
 804	 * from user space:
 805	 */
 806	if (is_prefetch(regs, error_code, address))
 807		return;
 808
 809	if (is_errata100(regs, address))
 810		return;
 811
 812	sanitize_error_code(address, &error_code);
 813
 814	if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
 815		return;
 816
 817	if (likely(show_unhandled_signals))
 818		show_signal_msg(regs, error_code, address, tsk);
 819
 820	set_signal_archinfo(address, error_code);
 821
 822	if (si_code == SEGV_PKUERR)
 823		force_sig_pkuerr((void __user *)address, pkey);
 824	else
 825		force_sig_fault(SIGSEGV, si_code, (void __user *)address);
 826
 827	local_irq_disable();
 828}
 829
 830static noinline void
 831bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 832		     unsigned long address)
 833{
 834	__bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
 835}
 836
 837static void
 838__bad_area(struct pt_regs *regs, unsigned long error_code,
 839	   unsigned long address, struct mm_struct *mm,
 840	   struct vm_area_struct *vma, u32 pkey, int si_code)
 841{
 
 842	/*
 843	 * Something tried to access memory that isn't in our memory map..
 844	 * Fix it, but check if it's kernel or user first..
 845	 */
 846	if (mm)
 847		mmap_read_unlock(mm);
 848	else
 849		vma_end_read(vma);
 850
 851	__bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
 852}
 853
 854static inline bool bad_area_access_from_pkeys(unsigned long error_code,
 855		struct vm_area_struct *vma)
 856{
 857	/* This code is always called on the current mm */
 858	bool foreign = false;
 859
 860	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
 861		return false;
 862	if (error_code & X86_PF_PK)
 863		return true;
 864	/* this checks permission keys on the VMA: */
 865	if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
 866				       (error_code & X86_PF_INSTR), foreign))
 867		return true;
 868	return false;
 869}
 870
 871static noinline void
 872bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 873		      unsigned long address, struct mm_struct *mm,
 874		      struct vm_area_struct *vma)
 875{
 876	/*
 877	 * This OSPKE check is not strictly necessary at runtime.
 878	 * But, doing it this way allows compiler optimizations
 879	 * if pkeys are compiled out.
 880	 */
 881	if (bad_area_access_from_pkeys(error_code, vma)) {
 882		/*
 883		 * A protection key fault means that the PKRU value did not allow
 884		 * access to some PTE.  Userspace can figure out what PKRU was
 885		 * from the XSAVE state.  This function captures the pkey from
 886		 * the vma and passes it to userspace so userspace can discover
 887		 * which protection key was set on the PTE.
 888		 *
 889		 * If we get here, we know that the hardware signaled a X86_PF_PK
 890		 * fault and that there was a VMA once we got in the fault
 891		 * handler.  It does *not* guarantee that the VMA we find here
 892		 * was the one that we faulted on.
 893		 *
 894		 * 1. T1   : mprotect_key(foo, PAGE_SIZE, pkey=4);
 895		 * 2. T1   : set PKRU to deny access to pkey=4, touches page
 896		 * 3. T1   : faults...
 897		 * 4.    T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
 898		 * 5. T1   : enters fault handler, takes mmap_lock, etc...
 899		 * 6. T1   : reaches here, sees vma_pkey(vma)=5, when we really
 900		 *	     faulted on a pte with its pkey=4.
 901		 */
 902		u32 pkey = vma_pkey(vma);
 903
 904		__bad_area(regs, error_code, address, mm, vma, pkey, SEGV_PKUERR);
 905	} else {
 906		__bad_area(regs, error_code, address, mm, vma, 0, SEGV_ACCERR);
 907	}
 908}
 909
 910static void
 911do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 912	  vm_fault_t fault)
 913{
 914	/* Kernel mode? Handle exceptions or die: */
 915	if (!user_mode(regs)) {
 916		kernelmode_fixup_or_oops(regs, error_code, address,
 917					 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
 918		return;
 919	}
 920
 921	/* User-space => ok to do another page fault: */
 922	if (is_prefetch(regs, error_code, address))
 923		return;
 924
 925	sanitize_error_code(address, &error_code);
 926
 927	if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
 928		return;
 929
 930	set_signal_archinfo(address, error_code);
 931
 932#ifdef CONFIG_MEMORY_FAILURE
 933	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
 934		struct task_struct *tsk = current;
 935		unsigned lsb = 0;
 936
 937		pr_err(
 938	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
 939			tsk->comm, tsk->pid, address);
 940		if (fault & VM_FAULT_HWPOISON_LARGE)
 941			lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
 942		if (fault & VM_FAULT_HWPOISON)
 943			lsb = PAGE_SHIFT;
 944		force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
 945		return;
 946	}
 947#endif
 948	force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
 949}
 950
 951static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
 952{
 953	if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
 954		return 0;
 955
 956	if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
 957		return 0;
 958
 959	return 1;
 960}
 961
 962/*
 963 * Handle a spurious fault caused by a stale TLB entry.
 964 *
 965 * This allows us to lazily refresh the TLB when increasing the
 966 * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 967 * eagerly is very expensive since that implies doing a full
 968 * cross-processor TLB flush, even if no stale TLB entries exist
 969 * on other processors.
 970 *
 971 * Spurious faults may only occur if the TLB contains an entry with
 972 * fewer permission than the page table entry.  Non-present (P = 0)
 973 * and reserved bit (R = 1) faults are never spurious.
 974 *
 975 * There are no security implications to leaving a stale TLB when
 976 * increasing the permissions on a page.
 977 *
 978 * Returns non-zero if a spurious fault was handled, zero otherwise.
 979 *
 980 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
 981 * (Optional Invalidation).
 982 */
 983static noinline int
 984spurious_kernel_fault(unsigned long error_code, unsigned long address)
 985{
 986	pgd_t *pgd;
 987	p4d_t *p4d;
 988	pud_t *pud;
 989	pmd_t *pmd;
 990	pte_t *pte;
 991	int ret;
 992
 993	/*
 994	 * Only writes to RO or instruction fetches from NX may cause
 995	 * spurious faults.
 996	 *
 997	 * These could be from user or supervisor accesses but the TLB
 998	 * is only lazily flushed after a kernel mapping protection
 999	 * change, so user accesses are not expected to cause spurious
1000	 * faults.
1001	 */
1002	if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1003	    error_code != (X86_PF_INSTR | X86_PF_PROT))
1004		return 0;
1005
1006	pgd = init_mm.pgd + pgd_index(address);
1007	if (!pgd_present(*pgd))
1008		return 0;
1009
1010	p4d = p4d_offset(pgd, address);
1011	if (!p4d_present(*p4d))
1012		return 0;
1013
1014	if (p4d_leaf(*p4d))
1015		return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1016
1017	pud = pud_offset(p4d, address);
1018	if (!pud_present(*pud))
1019		return 0;
1020
1021	if (pud_leaf(*pud))
1022		return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1023
1024	pmd = pmd_offset(pud, address);
1025	if (!pmd_present(*pmd))
1026		return 0;
1027
1028	if (pmd_leaf(*pmd))
1029		return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1030
1031	pte = pte_offset_kernel(pmd, address);
1032	if (!pte_present(*pte))
1033		return 0;
1034
1035	ret = spurious_kernel_fault_check(error_code, pte);
1036	if (!ret)
1037		return 0;
1038
1039	/*
1040	 * Make sure we have permissions in PMD.
1041	 * If not, then there's a bug in the page tables:
1042	 */
1043	ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1044	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1045
1046	return ret;
1047}
1048NOKPROBE_SYMBOL(spurious_kernel_fault);
1049
1050int show_unhandled_signals = 1;
1051
1052static inline int
1053access_error(unsigned long error_code, struct vm_area_struct *vma)
1054{
1055	/* This is only called for the current mm, so: */
1056	bool foreign = false;
1057
1058	/*
1059	 * Read or write was blocked by protection keys.  This is
1060	 * always an unconditional error and can never result in
1061	 * a follow-up action to resolve the fault, like a COW.
1062	 */
1063	if (error_code & X86_PF_PK)
1064		return 1;
1065
1066	/*
1067	 * SGX hardware blocked the access.  This usually happens
1068	 * when the enclave memory contents have been destroyed, like
1069	 * after a suspend/resume cycle. In any case, the kernel can't
1070	 * fix the cause of the fault.  Handle the fault as an access
1071	 * error even in cases where no actual access violation
1072	 * occurred.  This allows userspace to rebuild the enclave in
1073	 * response to the signal.
1074	 */
1075	if (unlikely(error_code & X86_PF_SGX))
1076		return 1;
1077
1078	/*
1079	 * Make sure to check the VMA so that we do not perform
1080	 * faults just to hit a X86_PF_PK as soon as we fill in a
1081	 * page.
1082	 */
1083	if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1084				       (error_code & X86_PF_INSTR), foreign))
1085		return 1;
1086
1087	/*
1088	 * Shadow stack accesses (PF_SHSTK=1) are only permitted to
1089	 * shadow stack VMAs. All other accesses result in an error.
1090	 */
1091	if (error_code & X86_PF_SHSTK) {
1092		if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK)))
1093			return 1;
1094		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1095			return 1;
1096		return 0;
1097	}
1098
1099	if (error_code & X86_PF_WRITE) {
1100		/* write, present and write, not present: */
1101		if (unlikely(vma->vm_flags & VM_SHADOW_STACK))
1102			return 1;
1103		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1104			return 1;
1105		return 0;
1106	}
1107
1108	/* read, present: */
1109	if (unlikely(error_code & X86_PF_PROT))
1110		return 1;
1111
1112	/* read, not present: */
1113	if (unlikely(!vma_is_accessible(vma)))
1114		return 1;
1115
1116	return 0;
1117}
1118
1119bool fault_in_kernel_space(unsigned long address)
1120{
1121	/*
1122	 * On 64-bit systems, the vsyscall page is at an address above
1123	 * TASK_SIZE_MAX, but is not considered part of the kernel
1124	 * address space.
1125	 */
1126	if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1127		return false;
1128
1129	return address >= TASK_SIZE_MAX;
1130}
1131
1132/*
1133 * Called for all faults where 'address' is part of the kernel address
1134 * space.  Might get called for faults that originate from *code* that
1135 * ran in userspace or the kernel.
1136 */
1137static void
1138do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1139		   unsigned long address)
1140{
1141	/*
1142	 * Protection keys exceptions only happen on user pages.  We
1143	 * have no user pages in the kernel portion of the address
1144	 * space, so do not expect them here.
1145	 */
1146	WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1147
1148#ifdef CONFIG_X86_32
1149	/*
1150	 * We can fault-in kernel-space virtual memory on-demand. The
1151	 * 'reference' page table is init_mm.pgd.
1152	 *
1153	 * NOTE! We MUST NOT take any locks for this case. We may
1154	 * be in an interrupt or a critical region, and should
1155	 * only copy the information from the master page table,
1156	 * nothing more.
1157	 *
1158	 * Before doing this on-demand faulting, ensure that the
1159	 * fault is not any of the following:
1160	 * 1. A fault on a PTE with a reserved bit set.
1161	 * 2. A fault caused by a user-mode access.  (Do not demand-
1162	 *    fault kernel memory due to user-mode accesses).
1163	 * 3. A fault caused by a page-level protection violation.
1164	 *    (A demand fault would be on a non-present page which
1165	 *     would have X86_PF_PROT==0).
1166	 *
1167	 * This is only needed to close a race condition on x86-32 in
1168	 * the vmalloc mapping/unmapping code. See the comment above
1169	 * vmalloc_fault() for details. On x86-64 the race does not
1170	 * exist as the vmalloc mappings don't need to be synchronized
1171	 * there.
1172	 */
1173	if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1174		if (vmalloc_fault(address) >= 0)
1175			return;
1176	}
1177#endif
1178
1179	if (is_f00f_bug(regs, hw_error_code, address))
1180		return;
1181
1182	/* Was the fault spurious, caused by lazy TLB invalidation? */
1183	if (spurious_kernel_fault(hw_error_code, address))
1184		return;
1185
1186	/* kprobes don't want to hook the spurious faults: */
1187	if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1188		return;
1189
1190	/*
1191	 * Note, despite being a "bad area", there are quite a few
1192	 * acceptable reasons to get here, such as erratum fixups
1193	 * and handling kernel code that can fault, like get_user().
1194	 *
1195	 * Don't take the mm semaphore here. If we fixup a prefetch
1196	 * fault we could otherwise deadlock:
1197	 */
1198	bad_area_nosemaphore(regs, hw_error_code, address);
1199}
1200NOKPROBE_SYMBOL(do_kern_addr_fault);
1201
1202/*
1203 * Handle faults in the user portion of the address space.  Nothing in here
1204 * should check X86_PF_USER without a specific justification: for almost
1205 * all purposes, we should treat a normal kernel access to user memory
1206 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1207 * The one exception is AC flag handling, which is, per the x86
1208 * architecture, special for WRUSS.
1209 */
1210static inline
1211void do_user_addr_fault(struct pt_regs *regs,
1212			unsigned long error_code,
1213			unsigned long address)
1214{
1215	struct vm_area_struct *vma;
1216	struct task_struct *tsk;
1217	struct mm_struct *mm;
1218	vm_fault_t fault;
1219	unsigned int flags = FAULT_FLAG_DEFAULT;
1220
1221	tsk = current;
1222	mm = tsk->mm;
1223
1224	if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1225		/*
1226		 * Whoops, this is kernel mode code trying to execute from
1227		 * user memory.  Unless this is AMD erratum #93, which
1228		 * corrupts RIP such that it looks like a user address,
1229		 * this is unrecoverable.  Don't even try to look up the
1230		 * VMA or look for extable entries.
1231		 */
1232		if (is_errata93(regs, address))
1233			return;
1234
1235		page_fault_oops(regs, error_code, address);
1236		return;
1237	}
1238
1239	/* kprobes don't want to hook the spurious faults: */
1240	if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1241		return;
1242
1243	/*
1244	 * Reserved bits are never expected to be set on
1245	 * entries in the user portion of the page tables.
1246	 */
1247	if (unlikely(error_code & X86_PF_RSVD))
1248		pgtable_bad(regs, error_code, address);
1249
1250	/*
1251	 * If SMAP is on, check for invalid kernel (supervisor) access to user
1252	 * pages in the user address space.  The odd case here is WRUSS,
1253	 * which, according to the preliminary documentation, does not respect
1254	 * SMAP and will have the USER bit set so, in all cases, SMAP
1255	 * enforcement appears to be consistent with the USER bit.
1256	 */
1257	if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1258		     !(error_code & X86_PF_USER) &&
1259		     !(regs->flags & X86_EFLAGS_AC))) {
1260		/*
1261		 * No extable entry here.  This was a kernel access to an
1262		 * invalid pointer.  get_kernel_nofault() will not get here.
1263		 */
1264		page_fault_oops(regs, error_code, address);
1265		return;
1266	}
1267
1268	/*
1269	 * If we're in an interrupt, have no user context or are running
1270	 * in a region with pagefaults disabled then we must not take the fault
1271	 */
1272	if (unlikely(faulthandler_disabled() || !mm)) {
1273		bad_area_nosemaphore(regs, error_code, address);
1274		return;
1275	}
1276
1277	/* Legacy check - remove this after verifying that it doesn't trigger */
1278	if (WARN_ON_ONCE(!(regs->flags & X86_EFLAGS_IF))) {
1279		bad_area_nosemaphore(regs, error_code, address);
1280		return;
1281	}
1282
1283	local_irq_enable();
1284
1285	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1286
1287	/*
1288	 * Read-only permissions can not be expressed in shadow stack PTEs.
1289	 * Treat all shadow stack accesses as WRITE faults. This ensures
1290	 * that the MM will prepare everything (e.g., break COW) such that
1291	 * maybe_mkwrite() can create a proper shadow stack PTE.
1292	 */
1293	if (error_code & X86_PF_SHSTK)
1294		flags |= FAULT_FLAG_WRITE;
1295	if (error_code & X86_PF_WRITE)
1296		flags |= FAULT_FLAG_WRITE;
1297	if (error_code & X86_PF_INSTR)
1298		flags |= FAULT_FLAG_INSTRUCTION;
1299
1300	/*
1301	 * We set FAULT_FLAG_USER based on the register state, not
1302	 * based on X86_PF_USER. User space accesses that cause
1303	 * system page faults are still user accesses.
1304	 */
1305	if (user_mode(regs))
1306		flags |= FAULT_FLAG_USER;
1307
1308#ifdef CONFIG_X86_64
1309	/*
1310	 * Faults in the vsyscall page might need emulation.  The
1311	 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1312	 * considered to be part of the user address space.
1313	 *
1314	 * The vsyscall page does not have a "real" VMA, so do this
1315	 * emulation before we go searching for VMAs.
1316	 *
1317	 * PKRU never rejects instruction fetches, so we don't need
1318	 * to consider the PF_PK bit.
1319	 */
1320	if (is_vsyscall_vaddr(address)) {
1321		if (emulate_vsyscall(error_code, regs, address))
1322			return;
1323	}
1324#endif
1325
1326	if (!(flags & FAULT_FLAG_USER))
1327		goto lock_mmap;
1328
1329	vma = lock_vma_under_rcu(mm, address);
1330	if (!vma)
1331		goto lock_mmap;
1332
1333	if (unlikely(access_error(error_code, vma))) {
1334		bad_area_access_error(regs, error_code, address, NULL, vma);
1335		count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1336		return;
1337	}
1338	fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
1339	if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
1340		vma_end_read(vma);
1341
1342	if (!(fault & VM_FAULT_RETRY)) {
1343		count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1344		goto done;
1345	}
1346	count_vm_vma_lock_event(VMA_LOCK_RETRY);
1347	if (fault & VM_FAULT_MAJOR)
1348		flags |= FAULT_FLAG_TRIED;
1349
1350	/* Quick path to respond to signals */
1351	if (fault_signal_pending(fault, regs)) {
1352		if (!user_mode(regs))
1353			kernelmode_fixup_or_oops(regs, error_code, address,
1354						 SIGBUS, BUS_ADRERR,
1355						 ARCH_DEFAULT_PKEY);
1356		return;
1357	}
1358lock_mmap:
1359
1360retry:
1361	vma = lock_mm_and_find_vma(mm, address, regs);
1362	if (unlikely(!vma)) {
1363		bad_area_nosemaphore(regs, error_code, address);
1364		return;
1365	}
1366
1367	/*
1368	 * Ok, we have a good vm_area for this memory access, so
1369	 * we can handle it..
1370	 */
1371	if (unlikely(access_error(error_code, vma))) {
1372		bad_area_access_error(regs, error_code, address, mm, vma);
1373		return;
1374	}
1375
1376	/*
1377	 * If for any reason at all we couldn't handle the fault,
1378	 * make sure we exit gracefully rather than endlessly redo
1379	 * the fault.  Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1380	 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1381	 *
1382	 * Note that handle_userfault() may also release and reacquire mmap_lock
1383	 * (and not return with VM_FAULT_RETRY), when returning to userland to
1384	 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1385	 * (potentially after handling any pending signal during the return to
1386	 * userland). The return to userland is identified whenever
1387	 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1388	 */
1389	fault = handle_mm_fault(vma, address, flags, regs);
1390
1391	if (fault_signal_pending(fault, regs)) {
1392		/*
1393		 * Quick path to respond to signals.  The core mm code
1394		 * has unlocked the mm for us if we get here.
1395		 */
1396		if (!user_mode(regs))
1397			kernelmode_fixup_or_oops(regs, error_code, address,
1398						 SIGBUS, BUS_ADRERR,
1399						 ARCH_DEFAULT_PKEY);
1400		return;
1401	}
1402
1403	/* The fault is fully completed (including releasing mmap lock) */
1404	if (fault & VM_FAULT_COMPLETED)
1405		return;
1406
1407	/*
1408	 * If we need to retry the mmap_lock has already been released,
1409	 * and if there is a fatal signal pending there is no guarantee
1410	 * that we made any progress. Handle this case first.
1411	 */
1412	if (unlikely(fault & VM_FAULT_RETRY)) {
1413		flags |= FAULT_FLAG_TRIED;
1414		goto retry;
1415	}
1416
1417	mmap_read_unlock(mm);
1418done:
1419	if (likely(!(fault & VM_FAULT_ERROR)))
1420		return;
1421
1422	if (fatal_signal_pending(current) && !user_mode(regs)) {
1423		kernelmode_fixup_or_oops(regs, error_code, address,
1424					 0, 0, ARCH_DEFAULT_PKEY);
1425		return;
1426	}
1427
1428	if (fault & VM_FAULT_OOM) {
1429		/* Kernel mode? Handle exceptions or die: */
1430		if (!user_mode(regs)) {
1431			kernelmode_fixup_or_oops(regs, error_code, address,
1432						 SIGSEGV, SEGV_MAPERR,
1433						 ARCH_DEFAULT_PKEY);
1434			return;
1435		}
1436
1437		/*
1438		 * We ran out of memory, call the OOM killer, and return the
1439		 * userspace (which will retry the fault, or kill us if we got
1440		 * oom-killed):
1441		 */
1442		pagefault_out_of_memory();
1443	} else {
1444		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1445			     VM_FAULT_HWPOISON_LARGE))
1446			do_sigbus(regs, error_code, address, fault);
1447		else if (fault & VM_FAULT_SIGSEGV)
1448			bad_area_nosemaphore(regs, error_code, address);
1449		else
1450			BUG();
1451	}
1452}
1453NOKPROBE_SYMBOL(do_user_addr_fault);
1454
1455static __always_inline void
1456trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1457			 unsigned long address)
1458{
1459	if (!trace_pagefault_enabled())
1460		return;
1461
1462	if (user_mode(regs))
1463		trace_page_fault_user(address, regs, error_code);
1464	else
1465		trace_page_fault_kernel(address, regs, error_code);
1466}
1467
1468static __always_inline void
1469handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1470			      unsigned long address)
1471{
1472	trace_page_fault_entries(regs, error_code, address);
1473
1474	if (unlikely(kmmio_fault(regs, address)))
1475		return;
1476
1477	/* Was the fault on kernel-controlled part of the address space? */
1478	if (unlikely(fault_in_kernel_space(address))) {
1479		do_kern_addr_fault(regs, error_code, address);
1480	} else {
1481		do_user_addr_fault(regs, error_code, address);
1482		/*
1483		 * User address page fault handling might have reenabled
1484		 * interrupts. Fixing up all potential exit points of
1485		 * do_user_addr_fault() and its leaf functions is just not
1486		 * doable w/o creating an unholy mess or turning the code
1487		 * upside down.
1488		 */
1489		local_irq_disable();
1490	}
1491}
1492
1493DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1494{
1495	irqentry_state_t state;
1496	unsigned long address;
1497
1498	address = cpu_feature_enabled(X86_FEATURE_FRED) ? fred_event_data(regs) : read_cr2();
1499
1500	prefetchw(&current->mm->mmap_lock);
1501
1502	/*
1503	 * KVM uses #PF vector to deliver 'page not present' events to guests
1504	 * (asynchronous page fault mechanism). The event happens when a
1505	 * userspace task is trying to access some valid (from guest's point of
1506	 * view) memory which is not currently mapped by the host (e.g. the
1507	 * memory is swapped out). Note, the corresponding "page ready" event
1508	 * which is injected when the memory becomes available, is delivered via
1509	 * an interrupt mechanism and not a #PF exception
1510	 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1511	 *
1512	 * We are relying on the interrupted context being sane (valid RSP,
1513	 * relevant locks not held, etc.), which is fine as long as the
1514	 * interrupted context had IF=1.  We are also relying on the KVM
1515	 * async pf type field and CR2 being read consistently instead of
1516	 * getting values from real and async page faults mixed up.
1517	 *
1518	 * Fingers crossed.
1519	 *
1520	 * The async #PF handling code takes care of idtentry handling
1521	 * itself.
1522	 */
1523	if (kvm_handle_async_pf(regs, (u32)address))
1524		return;
1525
1526	/*
1527	 * Entry handling for valid #PF from kernel mode is slightly
1528	 * different: RCU is already watching and ct_irq_enter() must not
1529	 * be invoked because a kernel fault on a user space address might
1530	 * sleep.
1531	 *
1532	 * In case the fault hit a RCU idle region the conditional entry
1533	 * code reenabled RCU to avoid subsequent wreckage which helps
1534	 * debuggability.
1535	 */
1536	state = irqentry_enter(regs);
1537
1538	instrumentation_begin();
1539	handle_page_fault(regs, error_code, address);
1540	instrumentation_end();
1541
1542	irqentry_exit(regs, state);
1543}