Linux Audio

Check our new training course

Loading...
v3.15
   1/*
   2 *  Copyright (C) 1995  Linus Torvalds
   3 *  Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
   4 *  Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
   5 */
   6#include <linux/magic.h>		/* STACK_END_MAGIC		*/
   7#include <linux/sched.h>		/* test_thread_flag(), ...	*/
   8#include <linux/kdebug.h>		/* oops_begin/end, ...		*/
   9#include <linux/module.h>		/* search_exception_table	*/
  10#include <linux/bootmem.h>		/* max_low_pfn			*/
  11#include <linux/kprobes.h>		/* __kprobes, ...		*/
  12#include <linux/mmiotrace.h>		/* kmmio_handler, ...		*/
  13#include <linux/perf_event.h>		/* perf_sw_event		*/
  14#include <linux/hugetlb.h>		/* hstate_index_to_shift	*/
  15#include <linux/prefetch.h>		/* prefetchw			*/
  16#include <linux/context_tracking.h>	/* exception_enter(), ...	*/
 
  17
 
  18#include <asm/traps.h>			/* dotraplinkage, ...		*/
  19#include <asm/pgalloc.h>		/* pgd_*(), ...			*/
  20#include <asm/kmemcheck.h>		/* kmemcheck_*(), ...		*/
  21#include <asm/fixmap.h>			/* VSYSCALL_START		*/
 
 
 
  22
  23#define CREATE_TRACE_POINTS
  24#include <asm/trace/exceptions.h>
  25
  26/*
  27 * Page fault error code bits:
  28 *
  29 *   bit 0 ==	 0: no page found	1: protection fault
  30 *   bit 1 ==	 0: read access		1: write access
  31 *   bit 2 ==	 0: kernel-mode access	1: user-mode access
  32 *   bit 3 ==				1: use of reserved bit detected
  33 *   bit 4 ==				1: fault was an instruction fetch
 
  34 */
  35enum x86_pf_error_code {
  36
  37	PF_PROT		=		1 << 0,
  38	PF_WRITE	=		1 << 1,
  39	PF_USER		=		1 << 2,
  40	PF_RSVD		=		1 << 3,
  41	PF_INSTR	=		1 << 4,
 
  42};
  43
  44/*
  45 * Returns 0 if mmiotrace is disabled, or if the fault is not
  46 * handled by mmiotrace:
  47 */
  48static inline int __kprobes
  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
  57static inline int __kprobes kprobes_fault(struct pt_regs *regs)
  58{
  59	int ret = 0;
  60
  61	/* kprobe_running() needs smp_processor_id() */
  62	if (kprobes_built_in() && !user_mode_vm(regs)) {
  63		preempt_disable();
  64		if (kprobe_running() && kprobe_fault_handler(regs, 14))
  65			ret = 1;
  66		preempt_enable();
  67	}
  68
  69	return ret;
  70}
  71
  72/*
  73 * Prefetch quirks:
  74 *
  75 * 32-bit mode:
  76 *
  77 *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  78 *   Check that here and ignore it.
  79 *
  80 * 64-bit mode:
  81 *
  82 *   Sometimes the CPU reports invalid exceptions on prefetch.
  83 *   Check that here and ignore it.
  84 *
  85 * Opcode checker based on code by Richard Brunner.
  86 */
  87static inline int
  88check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  89		      unsigned char opcode, int *prefetch)
  90{
  91	unsigned char instr_hi = opcode & 0xf0;
  92	unsigned char instr_lo = opcode & 0x0f;
  93
  94	switch (instr_hi) {
  95	case 0x20:
  96	case 0x30:
  97		/*
  98		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  99		 * In X86_64 long mode, the CPU will signal invalid
 100		 * opcode if some of these prefixes are present so
 101		 * X86_64 will never get here anyway
 102		 */
 103		return ((instr_lo & 7) == 0x6);
 104#ifdef CONFIG_X86_64
 105	case 0x40:
 106		/*
 107		 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 108		 * Need to figure out under what instruction mode the
 109		 * instruction was issued. Could check the LDT for lm,
 110		 * but for now it's good enough to assume that long
 111		 * mode only uses well known segments or kernel.
 112		 */
 113		return (!user_mode(regs) || user_64bit_mode(regs));
 114#endif
 115	case 0x60:
 116		/* 0x64 thru 0x67 are valid prefixes in all modes. */
 117		return (instr_lo & 0xC) == 0x4;
 118	case 0xF0:
 119		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 120		return !instr_lo || (instr_lo>>1) == 1;
 121	case 0x00:
 122		/* Prefetch instruction is 0x0F0D or 0x0F18 */
 123		if (probe_kernel_address(instr, opcode))
 124			return 0;
 125
 126		*prefetch = (instr_lo == 0xF) &&
 127			(opcode == 0x0D || opcode == 0x18);
 128		return 0;
 129	default:
 130		return 0;
 131	}
 132}
 133
 134static int
 135is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 136{
 137	unsigned char *max_instr;
 138	unsigned char *instr;
 139	int prefetch = 0;
 140
 141	/*
 142	 * If it was a exec (instruction fetch) fault on NX page, then
 143	 * do not ignore the fault:
 144	 */
 145	if (error_code & PF_INSTR)
 146		return 0;
 147
 148	instr = (void *)convert_ip_to_linear(current, regs);
 149	max_instr = instr + 15;
 150
 151	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 152		return 0;
 153
 154	while (instr < max_instr) {
 155		unsigned char opcode;
 156
 157		if (probe_kernel_address(instr, opcode))
 158			break;
 159
 160		instr++;
 161
 162		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 163			break;
 164	}
 165	return prefetch;
 166}
 167
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 168static void
 169force_sig_info_fault(int si_signo, int si_code, unsigned long address,
 170		     struct task_struct *tsk, int fault)
 
 171{
 172	unsigned lsb = 0;
 173	siginfo_t info;
 174
 175	info.si_signo	= si_signo;
 176	info.si_errno	= 0;
 177	info.si_code	= si_code;
 178	info.si_addr	= (void __user *)address;
 179	if (fault & VM_FAULT_HWPOISON_LARGE)
 180		lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 
 181	if (fault & VM_FAULT_HWPOISON)
 182		lsb = PAGE_SHIFT;
 183	info.si_addr_lsb = lsb;
 184
 
 
 185	force_sig_info(si_signo, &info, tsk);
 186}
 187
 188DEFINE_SPINLOCK(pgd_lock);
 189LIST_HEAD(pgd_list);
 190
 191#ifdef CONFIG_X86_32
 192static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 193{
 194	unsigned index = pgd_index(address);
 195	pgd_t *pgd_k;
 196	pud_t *pud, *pud_k;
 197	pmd_t *pmd, *pmd_k;
 198
 199	pgd += index;
 200	pgd_k = init_mm.pgd + index;
 201
 202	if (!pgd_present(*pgd_k))
 203		return NULL;
 204
 205	/*
 206	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
 207	 * and redundant with the set_pmd() on non-PAE. As would
 208	 * set_pud.
 209	 */
 210	pud = pud_offset(pgd, address);
 211	pud_k = pud_offset(pgd_k, address);
 212	if (!pud_present(*pud_k))
 213		return NULL;
 214
 215	pmd = pmd_offset(pud, address);
 216	pmd_k = pmd_offset(pud_k, address);
 217	if (!pmd_present(*pmd_k))
 218		return NULL;
 219
 220	if (!pmd_present(*pmd))
 221		set_pmd(pmd, *pmd_k);
 222	else
 223		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 224
 225	return pmd_k;
 226}
 227
 228void vmalloc_sync_all(void)
 229{
 230	unsigned long address;
 231
 232	if (SHARED_KERNEL_PMD)
 233		return;
 234
 235	for (address = VMALLOC_START & PMD_MASK;
 236	     address >= TASK_SIZE && address < FIXADDR_TOP;
 237	     address += PMD_SIZE) {
 238		struct page *page;
 239
 240		spin_lock(&pgd_lock);
 241		list_for_each_entry(page, &pgd_list, lru) {
 242			spinlock_t *pgt_lock;
 243			pmd_t *ret;
 244
 245			/* the pgt_lock only for Xen */
 246			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 247
 248			spin_lock(pgt_lock);
 249			ret = vmalloc_sync_one(page_address(page), address);
 250			spin_unlock(pgt_lock);
 251
 252			if (!ret)
 253				break;
 254		}
 255		spin_unlock(&pgd_lock);
 256	}
 257}
 258
 259/*
 260 * 32-bit:
 261 *
 262 *   Handle a fault on the vmalloc or module mapping area
 263 */
 264static noinline __kprobes int vmalloc_fault(unsigned long address)
 265{
 266	unsigned long pgd_paddr;
 267	pmd_t *pmd_k;
 268	pte_t *pte_k;
 269
 270	/* Make sure we are in vmalloc area: */
 271	if (!(address >= VMALLOC_START && address < VMALLOC_END))
 272		return -1;
 273
 274	WARN_ON_ONCE(in_nmi());
 275
 276	/*
 277	 * Synchronize this task's top level page-table
 278	 * with the 'reference' page table.
 279	 *
 280	 * Do _not_ use "current" here. We might be inside
 281	 * an interrupt in the middle of a task switch..
 282	 */
 283	pgd_paddr = read_cr3();
 284	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 285	if (!pmd_k)
 286		return -1;
 287
 
 
 
 288	pte_k = pte_offset_kernel(pmd_k, address);
 289	if (!pte_present(*pte_k))
 290		return -1;
 291
 292	return 0;
 293}
 
 294
 295/*
 296 * Did it hit the DOS screen memory VA from vm86 mode?
 297 */
 298static inline void
 299check_v8086_mode(struct pt_regs *regs, unsigned long address,
 300		 struct task_struct *tsk)
 301{
 
 302	unsigned long bit;
 303
 304	if (!v8086_mode(regs))
 305		return;
 306
 307	bit = (address - 0xA0000) >> PAGE_SHIFT;
 308	if (bit < 32)
 309		tsk->thread.screen_bitmap |= 1 << bit;
 
 310}
 311
 312static bool low_pfn(unsigned long pfn)
 313{
 314	return pfn < max_low_pfn;
 315}
 316
 317static void dump_pagetable(unsigned long address)
 318{
 319	pgd_t *base = __va(read_cr3());
 320	pgd_t *pgd = &base[pgd_index(address)];
 321	pmd_t *pmd;
 322	pte_t *pte;
 323
 324#ifdef CONFIG_X86_PAE
 325	printk("*pdpt = %016Lx ", pgd_val(*pgd));
 326	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 327		goto out;
 328#endif
 329	pmd = pmd_offset(pud_offset(pgd, address), address);
 330	printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 331
 332	/*
 333	 * We must not directly access the pte in the highpte
 334	 * case if the page table is located in highmem.
 335	 * And let's rather not kmap-atomic the pte, just in case
 336	 * it's allocated already:
 337	 */
 338	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
 339		goto out;
 340
 341	pte = pte_offset_kernel(pmd, address);
 342	printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 343out:
 344	printk("\n");
 345}
 346
 347#else /* CONFIG_X86_64: */
 348
 349void vmalloc_sync_all(void)
 350{
 351	sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
 352}
 353
 354/*
 355 * 64-bit:
 356 *
 357 *   Handle a fault on the vmalloc area
 358 *
 359 * This assumes no large pages in there.
 360 */
 361static noinline __kprobes int vmalloc_fault(unsigned long address)
 362{
 363	pgd_t *pgd, *pgd_ref;
 364	pud_t *pud, *pud_ref;
 365	pmd_t *pmd, *pmd_ref;
 366	pte_t *pte, *pte_ref;
 367
 368	/* Make sure we are in vmalloc area: */
 369	if (!(address >= VMALLOC_START && address < VMALLOC_END))
 370		return -1;
 371
 372	WARN_ON_ONCE(in_nmi());
 373
 374	/*
 375	 * Copy kernel mappings over when needed. This can also
 376	 * happen within a race in page table update. In the later
 377	 * case just flush:
 378	 */
 379	pgd = pgd_offset(current->active_mm, address);
 380	pgd_ref = pgd_offset_k(address);
 381	if (pgd_none(*pgd_ref))
 382		return -1;
 383
 384	if (pgd_none(*pgd)) {
 385		set_pgd(pgd, *pgd_ref);
 386		arch_flush_lazy_mmu_mode();
 387	} else {
 388		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 389	}
 390
 391	/*
 392	 * Below here mismatches are bugs because these lower tables
 393	 * are shared:
 394	 */
 395
 396	pud = pud_offset(pgd, address);
 397	pud_ref = pud_offset(pgd_ref, address);
 398	if (pud_none(*pud_ref))
 399		return -1;
 400
 401	if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 402		BUG();
 403
 
 
 
 404	pmd = pmd_offset(pud, address);
 405	pmd_ref = pmd_offset(pud_ref, address);
 406	if (pmd_none(*pmd_ref))
 407		return -1;
 408
 409	if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 410		BUG();
 411
 
 
 
 412	pte_ref = pte_offset_kernel(pmd_ref, address);
 413	if (!pte_present(*pte_ref))
 414		return -1;
 415
 416	pte = pte_offset_kernel(pmd, address);
 417
 418	/*
 419	 * Don't use pte_page here, because the mappings can point
 420	 * outside mem_map, and the NUMA hash lookup cannot handle
 421	 * that:
 422	 */
 423	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 424		BUG();
 425
 426	return 0;
 427}
 
 428
 429#ifdef CONFIG_CPU_SUP_AMD
 430static const char errata93_warning[] =
 431KERN_ERR 
 432"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 433"******* Working around it, but it may cause SEGVs or burn power.\n"
 434"******* Please consider a BIOS update.\n"
 435"******* Disabling USB legacy in the BIOS may also help.\n";
 436#endif
 437
 438/*
 439 * No vm86 mode in 64-bit mode:
 440 */
 441static inline void
 442check_v8086_mode(struct pt_regs *regs, unsigned long address,
 443		 struct task_struct *tsk)
 444{
 445}
 446
 447static int bad_address(void *p)
 448{
 449	unsigned long dummy;
 450
 451	return probe_kernel_address((unsigned long *)p, dummy);
 452}
 453
 454static void dump_pagetable(unsigned long address)
 455{
 456	pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 457	pgd_t *pgd = base + pgd_index(address);
 458	pud_t *pud;
 459	pmd_t *pmd;
 460	pte_t *pte;
 461
 462	if (bad_address(pgd))
 463		goto bad;
 464
 465	printk("PGD %lx ", pgd_val(*pgd));
 466
 467	if (!pgd_present(*pgd))
 468		goto out;
 469
 470	pud = pud_offset(pgd, address);
 471	if (bad_address(pud))
 472		goto bad;
 473
 474	printk("PUD %lx ", pud_val(*pud));
 475	if (!pud_present(*pud) || pud_large(*pud))
 476		goto out;
 477
 478	pmd = pmd_offset(pud, address);
 479	if (bad_address(pmd))
 480		goto bad;
 481
 482	printk("PMD %lx ", pmd_val(*pmd));
 483	if (!pmd_present(*pmd) || pmd_large(*pmd))
 484		goto out;
 485
 486	pte = pte_offset_kernel(pmd, address);
 487	if (bad_address(pte))
 488		goto bad;
 489
 490	printk("PTE %lx", pte_val(*pte));
 491out:
 492	printk("\n");
 493	return;
 494bad:
 495	printk("BAD\n");
 496}
 497
 498#endif /* CONFIG_X86_64 */
 499
 500/*
 501 * Workaround for K8 erratum #93 & buggy BIOS.
 502 *
 503 * BIOS SMM functions are required to use a specific workaround
 504 * to avoid corruption of the 64bit RIP register on C stepping K8.
 505 *
 506 * A lot of BIOS that didn't get tested properly miss this.
 507 *
 508 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 509 * Try to work around it here.
 510 *
 511 * Note we only handle faults in kernel here.
 512 * Does nothing on 32-bit.
 513 */
 514static int is_errata93(struct pt_regs *regs, unsigned long address)
 515{
 516#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
 517	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
 518	    || boot_cpu_data.x86 != 0xf)
 519		return 0;
 520
 521	if (address != regs->ip)
 522		return 0;
 523
 524	if ((address >> 32) != 0)
 525		return 0;
 526
 527	address |= 0xffffffffUL << 32;
 528	if ((address >= (u64)_stext && address <= (u64)_etext) ||
 529	    (address >= MODULES_VADDR && address <= MODULES_END)) {
 530		printk_once(errata93_warning);
 531		regs->ip = address;
 532		return 1;
 533	}
 534#endif
 535	return 0;
 536}
 537
 538/*
 539 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 540 * to illegal addresses >4GB.
 541 *
 542 * We catch this in the page fault handler because these addresses
 543 * are not reachable. Just detect this case and return.  Any code
 544 * segment in LDT is compatibility mode.
 545 */
 546static int is_errata100(struct pt_regs *regs, unsigned long address)
 547{
 548#ifdef CONFIG_X86_64
 549	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 550		return 1;
 551#endif
 552	return 0;
 553}
 554
 555static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 556{
 557#ifdef CONFIG_X86_F00F_BUG
 558	unsigned long nr;
 559
 560	/*
 561	 * Pentium F0 0F C7 C8 bug workaround:
 562	 */
 563	if (boot_cpu_has_bug(X86_BUG_F00F)) {
 564		nr = (address - idt_descr.address) >> 3;
 565
 566		if (nr == 6) {
 567			do_invalid_op(regs, 0);
 568			return 1;
 569		}
 570	}
 571#endif
 572	return 0;
 573}
 574
 575static const char nx_warning[] = KERN_CRIT
 576"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 
 
 577
 578static void
 579show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 580		unsigned long address)
 581{
 582	if (!oops_may_print())
 583		return;
 584
 585	if (error_code & PF_INSTR) {
 586		unsigned int level;
 587		pgd_t *pgd;
 588		pte_t *pte;
 589
 590		pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 591		pgd += pgd_index(address);
 592
 593		pte = lookup_address_in_pgd(pgd, address, &level);
 594
 595		if (pte && pte_present(*pte) && !pte_exec(*pte))
 596			printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
 
 
 
 
 597	}
 598
 599	printk(KERN_ALERT "BUG: unable to handle kernel ");
 600	if (address < PAGE_SIZE)
 601		printk(KERN_CONT "NULL pointer dereference");
 602	else
 603		printk(KERN_CONT "paging request");
 604
 605	printk(KERN_CONT " at %p\n", (void *) address);
 606	printk(KERN_ALERT "IP:");
 607	printk_address(regs->ip);
 608
 609	dump_pagetable(address);
 610}
 611
 612static noinline void
 613pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 614	    unsigned long address)
 615{
 616	struct task_struct *tsk;
 617	unsigned long flags;
 618	int sig;
 619
 620	flags = oops_begin();
 621	tsk = current;
 622	sig = SIGKILL;
 623
 624	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 625	       tsk->comm, address);
 626	dump_pagetable(address);
 627
 628	tsk->thread.cr2		= address;
 629	tsk->thread.trap_nr	= X86_TRAP_PF;
 630	tsk->thread.error_code	= error_code;
 631
 632	if (__die("Bad pagetable", regs, error_code))
 633		sig = 0;
 634
 635	oops_end(flags, regs, sig);
 636}
 637
 638static noinline void
 639no_context(struct pt_regs *regs, unsigned long error_code,
 640	   unsigned long address, int signal, int si_code)
 641{
 642	struct task_struct *tsk = current;
 643	unsigned long *stackend;
 644	unsigned long flags;
 645	int sig;
 
 
 646
 647	/* Are we prepared to handle this kernel fault? */
 648	if (fixup_exception(regs)) {
 649		/*
 650		 * Any interrupt that takes a fault gets the fixup. This makes
 651		 * the below recursive fault logic only apply to a faults from
 652		 * task context.
 653		 */
 654		if (in_interrupt())
 655			return;
 656
 657		/*
 658		 * Per the above we're !in_interrupt(), aka. task context.
 659		 *
 660		 * In this case we need to make sure we're not recursively
 661		 * faulting through the emulate_vsyscall() logic.
 662		 */
 663		if (current_thread_info()->sig_on_uaccess_error && signal) {
 664			tsk->thread.trap_nr = X86_TRAP_PF;
 665			tsk->thread.error_code = error_code | PF_USER;
 666			tsk->thread.cr2 = address;
 667
 668			/* XXX: hwpoison faults will set the wrong code. */
 669			force_sig_info_fault(signal, si_code, address, tsk, 0);
 
 670		}
 671
 672		/*
 673		 * Barring that, we can do the fixup and be happy.
 674		 */
 675		return;
 676	}
 677
 678	/*
 679	 * 32-bit:
 680	 *
 681	 *   Valid to do another page fault here, because if this fault
 682	 *   had been triggered by is_prefetch fixup_exception would have
 683	 *   handled it.
 684	 *
 685	 * 64-bit:
 686	 *
 687	 *   Hall of shame of CPU/BIOS bugs.
 688	 */
 689	if (is_prefetch(regs, error_code, address))
 690		return;
 691
 692	if (is_errata93(regs, address))
 693		return;
 694
 695	/*
 696	 * Oops. The kernel tried to access some bad page. We'll have to
 697	 * terminate things with extreme prejudice:
 698	 */
 699	flags = oops_begin();
 700
 701	show_fault_oops(regs, error_code, address);
 702
 703	stackend = end_of_stack(tsk);
 704	if (tsk != &init_task && *stackend != STACK_END_MAGIC)
 705		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
 706
 707	tsk->thread.cr2		= address;
 708	tsk->thread.trap_nr	= X86_TRAP_PF;
 709	tsk->thread.error_code	= error_code;
 710
 711	sig = SIGKILL;
 712	if (__die("Oops", regs, error_code))
 713		sig = 0;
 714
 715	/* Executive summary in case the body of the oops scrolled away */
 716	printk(KERN_DEFAULT "CR2: %016lx\n", address);
 717
 718	oops_end(flags, regs, sig);
 719}
 720
 721/*
 722 * Print out info about fatal segfaults, if the show_unhandled_signals
 723 * sysctl is set:
 724 */
 725static inline void
 726show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 727		unsigned long address, struct task_struct *tsk)
 728{
 729	if (!unhandled_signal(tsk, SIGSEGV))
 730		return;
 731
 732	if (!printk_ratelimit())
 733		return;
 734
 735	printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 736		task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 737		tsk->comm, task_pid_nr(tsk), address,
 738		(void *)regs->ip, (void *)regs->sp, error_code);
 739
 740	print_vma_addr(KERN_CONT " in ", regs->ip);
 741
 742	printk(KERN_CONT "\n");
 743}
 744
 745static void
 746__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 747		       unsigned long address, int si_code)
 
 748{
 749	struct task_struct *tsk = current;
 750
 751	/* User mode accesses just cause a SIGSEGV */
 752	if (error_code & PF_USER) {
 753		/*
 754		 * It's possible to have interrupts off here:
 755		 */
 756		local_irq_enable();
 757
 758		/*
 759		 * Valid to do another page fault here because this one came
 760		 * from user space:
 761		 */
 762		if (is_prefetch(regs, error_code, address))
 763			return;
 764
 765		if (is_errata100(regs, address))
 766			return;
 767
 768#ifdef CONFIG_X86_64
 769		/*
 770		 * Instruction fetch faults in the vsyscall page might need
 771		 * emulation.
 772		 */
 773		if (unlikely((error_code & PF_INSTR) &&
 774			     ((address & ~0xfff) == VSYSCALL_START))) {
 775			if (emulate_vsyscall(regs, address))
 776				return;
 777		}
 778#endif
 779		/* Kernel addresses are always protection faults: */
 780		if (address >= TASK_SIZE)
 781			error_code |= PF_PROT;
 782
 783		if (likely(show_unhandled_signals))
 784			show_signal_msg(regs, error_code, address, tsk);
 785
 786		tsk->thread.cr2		= address;
 787		tsk->thread.error_code	= error_code;
 788		tsk->thread.trap_nr	= X86_TRAP_PF;
 789
 790		force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
 791
 792		return;
 793	}
 794
 795	if (is_f00f_bug(regs, address))
 796		return;
 797
 798	no_context(regs, error_code, address, SIGSEGV, si_code);
 799}
 800
 801static noinline void
 802bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 803		     unsigned long address)
 804{
 805	__bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 806}
 807
 808static void
 809__bad_area(struct pt_regs *regs, unsigned long error_code,
 810	   unsigned long address, int si_code)
 811{
 812	struct mm_struct *mm = current->mm;
 813
 814	/*
 815	 * Something tried to access memory that isn't in our memory map..
 816	 * Fix it, but check if it's kernel or user first..
 817	 */
 818	up_read(&mm->mmap_sem);
 819
 820	__bad_area_nosemaphore(regs, error_code, address, si_code);
 821}
 822
 823static noinline void
 824bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 825{
 826	__bad_area(regs, error_code, address, SEGV_MAPERR);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 827}
 828
 829static noinline void
 830bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 831		      unsigned long address)
 832{
 833	__bad_area(regs, error_code, address, SEGV_ACCERR);
 
 
 
 
 
 
 
 
 834}
 835
 836static void
 837do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 838	  unsigned int fault)
 839{
 840	struct task_struct *tsk = current;
 841	struct mm_struct *mm = tsk->mm;
 842	int code = BUS_ADRERR;
 843
 844	up_read(&mm->mmap_sem);
 845
 846	/* Kernel mode? Handle exceptions or die: */
 847	if (!(error_code & PF_USER)) {
 848		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
 849		return;
 850	}
 851
 852	/* User-space => ok to do another page fault: */
 853	if (is_prefetch(regs, error_code, address))
 854		return;
 855
 856	tsk->thread.cr2		= address;
 857	tsk->thread.error_code	= error_code;
 858	tsk->thread.trap_nr	= X86_TRAP_PF;
 859
 860#ifdef CONFIG_MEMORY_FAILURE
 861	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
 862		printk(KERN_ERR
 863	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
 864			tsk->comm, tsk->pid, address);
 865		code = BUS_MCEERR_AR;
 866	}
 867#endif
 868	force_sig_info_fault(SIGBUS, code, address, tsk, fault);
 869}
 870
 871static noinline void
 872mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 873	       unsigned long address, unsigned int fault)
 
 874{
 875	if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
 876		up_read(&current->mm->mmap_sem);
 877		no_context(regs, error_code, address, 0, 0);
 878		return;
 879	}
 880
 881	if (fault & VM_FAULT_OOM) {
 882		/* Kernel mode? Handle exceptions or die: */
 883		if (!(error_code & PF_USER)) {
 884			up_read(&current->mm->mmap_sem);
 885			no_context(regs, error_code, address,
 886				   SIGSEGV, SEGV_MAPERR);
 887			return;
 888		}
 889
 890		up_read(&current->mm->mmap_sem);
 891
 892		/*
 893		 * We ran out of memory, call the OOM killer, and return the
 894		 * userspace (which will retry the fault, or kill us if we got
 895		 * oom-killed):
 896		 */
 897		pagefault_out_of_memory();
 898	} else {
 899		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
 900			     VM_FAULT_HWPOISON_LARGE))
 901			do_sigbus(regs, error_code, address, fault);
 
 
 902		else
 903			BUG();
 904	}
 905}
 906
 907static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 908{
 909	if ((error_code & PF_WRITE) && !pte_write(*pte))
 910		return 0;
 911
 912	if ((error_code & PF_INSTR) && !pte_exec(*pte))
 913		return 0;
 
 
 
 
 
 
 914
 915	return 1;
 916}
 917
 918/*
 919 * Handle a spurious fault caused by a stale TLB entry.
 920 *
 921 * This allows us to lazily refresh the TLB when increasing the
 922 * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 923 * eagerly is very expensive since that implies doing a full
 924 * cross-processor TLB flush, even if no stale TLB entries exist
 925 * on other processors.
 926 *
 
 
 
 
 927 * There are no security implications to leaving a stale TLB when
 928 * increasing the permissions on a page.
 
 
 
 
 
 929 */
 930static noinline __kprobes int
 931spurious_fault(unsigned long error_code, unsigned long address)
 932{
 933	pgd_t *pgd;
 934	pud_t *pud;
 935	pmd_t *pmd;
 936	pte_t *pte;
 937	int ret;
 938
 939	/* Reserved-bit violation or user access to kernel space? */
 940	if (error_code & (PF_USER | PF_RSVD))
 
 
 
 
 
 
 
 
 
 941		return 0;
 942
 943	pgd = init_mm.pgd + pgd_index(address);
 944	if (!pgd_present(*pgd))
 945		return 0;
 946
 947	pud = pud_offset(pgd, address);
 948	if (!pud_present(*pud))
 949		return 0;
 950
 951	if (pud_large(*pud))
 952		return spurious_fault_check(error_code, (pte_t *) pud);
 953
 954	pmd = pmd_offset(pud, address);
 955	if (!pmd_present(*pmd))
 956		return 0;
 957
 958	if (pmd_large(*pmd))
 959		return spurious_fault_check(error_code, (pte_t *) pmd);
 960
 961	pte = pte_offset_kernel(pmd, address);
 962	if (!pte_present(*pte))
 963		return 0;
 964
 965	ret = spurious_fault_check(error_code, pte);
 966	if (!ret)
 967		return 0;
 968
 969	/*
 970	 * Make sure we have permissions in PMD.
 971	 * If not, then there's a bug in the page tables:
 972	 */
 973	ret = spurious_fault_check(error_code, (pte_t *) pmd);
 974	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 975
 976	return ret;
 977}
 
 978
 979int show_unhandled_signals = 1;
 980
 981static inline int
 982access_error(unsigned long error_code, struct vm_area_struct *vma)
 983{
 
 
 
 
 
 
 
 
 
 
 
 984	if (error_code & PF_WRITE) {
 985		/* write, present and write, not present: */
 986		if (unlikely(!(vma->vm_flags & VM_WRITE)))
 987			return 1;
 988		return 0;
 989	}
 990
 991	/* read, present: */
 992	if (unlikely(error_code & PF_PROT))
 993		return 1;
 994
 995	/* read, not present: */
 996	if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 997		return 1;
 998
 999	return 0;
1000}
1001
1002static int fault_in_kernel_space(unsigned long address)
1003{
1004	return address >= TASK_SIZE_MAX;
1005}
1006
1007static inline bool smap_violation(int error_code, struct pt_regs *regs)
1008{
1009	if (!IS_ENABLED(CONFIG_X86_SMAP))
1010		return false;
1011
1012	if (!static_cpu_has(X86_FEATURE_SMAP))
1013		return false;
1014
1015	if (error_code & PF_USER)
1016		return false;
1017
1018	if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
1019		return false;
1020
1021	return true;
1022}
1023
1024/*
1025 * This routine handles page faults.  It determines the address,
1026 * and the problem, and then passes it off to one of the appropriate
1027 * routines.
1028 *
1029 * This function must have noinline because both callers
1030 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1031 * guarantees there's a function trace entry.
1032 */
1033static void __kprobes noinline
1034__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1035		unsigned long address)
1036{
1037	struct vm_area_struct *vma;
1038	struct task_struct *tsk;
1039	struct mm_struct *mm;
1040	int fault;
1041	unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1042
1043	tsk = current;
1044	mm = tsk->mm;
1045
1046	/*
1047	 * Detect and handle instructions that would cause a page fault for
1048	 * both a tracked kernel page and a userspace page.
1049	 */
1050	if (kmemcheck_active(regs))
1051		kmemcheck_hide(regs);
1052	prefetchw(&mm->mmap_sem);
1053
1054	if (unlikely(kmmio_fault(regs, address)))
1055		return;
1056
1057	/*
1058	 * We fault-in kernel-space virtual memory on-demand. The
1059	 * 'reference' page table is init_mm.pgd.
1060	 *
1061	 * NOTE! We MUST NOT take any locks for this case. We may
1062	 * be in an interrupt or a critical region, and should
1063	 * only copy the information from the master page table,
1064	 * nothing more.
1065	 *
1066	 * This verifies that the fault happens in kernel space
1067	 * (error_code & 4) == 0, and that the fault was not a
1068	 * protection error (error_code & 9) == 0.
1069	 */
1070	if (unlikely(fault_in_kernel_space(address))) {
1071		if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1072			if (vmalloc_fault(address) >= 0)
1073				return;
1074
1075			if (kmemcheck_fault(regs, address, error_code))
1076				return;
1077		}
1078
1079		/* Can handle a stale RO->RW TLB: */
1080		if (spurious_fault(error_code, address))
1081			return;
1082
1083		/* kprobes don't want to hook the spurious faults: */
1084		if (kprobes_fault(regs))
1085			return;
1086		/*
1087		 * Don't take the mm semaphore here. If we fixup a prefetch
1088		 * fault we could otherwise deadlock:
1089		 */
1090		bad_area_nosemaphore(regs, error_code, address);
1091
1092		return;
1093	}
1094
1095	/* kprobes don't want to hook the spurious faults: */
1096	if (unlikely(kprobes_fault(regs)))
1097		return;
1098
1099	if (unlikely(error_code & PF_RSVD))
1100		pgtable_bad(regs, error_code, address);
1101
1102	if (unlikely(smap_violation(error_code, regs))) {
1103		bad_area_nosemaphore(regs, error_code, address);
1104		return;
1105	}
1106
1107	/*
1108	 * If we're in an interrupt, have no user context or are running
1109	 * in an atomic region then we must not take the fault:
1110	 */
1111	if (unlikely(in_atomic() || !mm)) {
1112		bad_area_nosemaphore(regs, error_code, address);
1113		return;
1114	}
1115
1116	/*
1117	 * It's safe to allow irq's after cr2 has been saved and the
1118	 * vmalloc fault has been handled.
1119	 *
1120	 * User-mode registers count as a user access even for any
1121	 * potential system fault or CPU buglet:
1122	 */
1123	if (user_mode_vm(regs)) {
1124		local_irq_enable();
1125		error_code |= PF_USER;
1126		flags |= FAULT_FLAG_USER;
1127	} else {
1128		if (regs->flags & X86_EFLAGS_IF)
1129			local_irq_enable();
1130	}
1131
1132	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1133
1134	if (error_code & PF_WRITE)
1135		flags |= FAULT_FLAG_WRITE;
 
 
1136
1137	/*
1138	 * When running in the kernel we expect faults to occur only to
1139	 * addresses in user space.  All other faults represent errors in
1140	 * the kernel and should generate an OOPS.  Unfortunately, in the
1141	 * case of an erroneous fault occurring in a code path which already
1142	 * holds mmap_sem we will deadlock attempting to validate the fault
1143	 * against the address space.  Luckily the kernel only validly
1144	 * references user space from well defined areas of code, which are
1145	 * listed in the exceptions table.
1146	 *
1147	 * As the vast majority of faults will be valid we will only perform
1148	 * the source reference check when there is a possibility of a
1149	 * deadlock. Attempt to lock the address space, if we cannot we then
1150	 * validate the source. If this is invalid we can skip the address
1151	 * space check, thus avoiding the deadlock:
1152	 */
1153	if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1154		if ((error_code & PF_USER) == 0 &&
1155		    !search_exception_tables(regs->ip)) {
1156			bad_area_nosemaphore(regs, error_code, address);
1157			return;
1158		}
1159retry:
1160		down_read(&mm->mmap_sem);
1161	} else {
1162		/*
1163		 * The above down_read_trylock() might have succeeded in
1164		 * which case we'll have missed the might_sleep() from
1165		 * down_read():
1166		 */
1167		might_sleep();
1168	}
1169
1170	vma = find_vma(mm, address);
1171	if (unlikely(!vma)) {
1172		bad_area(regs, error_code, address);
1173		return;
1174	}
1175	if (likely(vma->vm_start <= address))
1176		goto good_area;
1177	if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1178		bad_area(regs, error_code, address);
1179		return;
1180	}
1181	if (error_code & PF_USER) {
1182		/*
1183		 * Accessing the stack below %sp is always a bug.
1184		 * The large cushion allows instructions like enter
1185		 * and pusha to work. ("enter $65535, $31" pushes
1186		 * 32 pointers and then decrements %sp by 65535.)
1187		 */
1188		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1189			bad_area(regs, error_code, address);
1190			return;
1191		}
1192	}
1193	if (unlikely(expand_stack(vma, address))) {
1194		bad_area(regs, error_code, address);
1195		return;
1196	}
1197
1198	/*
1199	 * Ok, we have a good vm_area for this memory access, so
1200	 * we can handle it..
1201	 */
1202good_area:
1203	if (unlikely(access_error(error_code, vma))) {
1204		bad_area_access_error(regs, error_code, address);
1205		return;
1206	}
1207
1208	/*
1209	 * If for any reason at all we couldn't handle the fault,
1210	 * make sure we exit gracefully rather than endlessly redo
1211	 * the fault:
 
1212	 */
1213	fault = handle_mm_fault(mm, vma, address, flags);
 
1214
1215	/*
1216	 * If we need to retry but a fatal signal is pending, handle the
1217	 * signal first. We do not need to release the mmap_sem because it
1218	 * would already be released in __lock_page_or_retry in mm/filemap.c.
1219	 */
1220	if (unlikely((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1221		return;
 
1222
 
1223	if (unlikely(fault & VM_FAULT_ERROR)) {
1224		mm_fault_error(regs, error_code, address, fault);
1225		return;
1226	}
1227
1228	/*
1229	 * Major/minor page fault accounting is only done on the
1230	 * initial attempt. If we go through a retry, it is extremely
1231	 * likely that the page will be found in page cache at that point.
1232	 */
1233	if (flags & FAULT_FLAG_ALLOW_RETRY) {
1234		if (fault & VM_FAULT_MAJOR) {
1235			tsk->maj_flt++;
1236			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1237				      regs, address);
1238		} else {
1239			tsk->min_flt++;
1240			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1241				      regs, address);
1242		}
1243		if (fault & VM_FAULT_RETRY) {
1244			/* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1245			 * of starvation. */
1246			flags &= ~FAULT_FLAG_ALLOW_RETRY;
1247			flags |= FAULT_FLAG_TRIED;
1248			goto retry;
1249		}
1250	}
1251
1252	check_v8086_mode(regs, address, tsk);
1253
1254	up_read(&mm->mmap_sem);
1255}
 
1256
1257dotraplinkage void __kprobes notrace
1258do_page_fault(struct pt_regs *regs, unsigned long error_code)
1259{
1260	unsigned long address = read_cr2(); /* Get the faulting address */
1261	enum ctx_state prev_state;
1262
1263	/*
1264	 * We must have this function tagged with __kprobes, notrace and call
1265	 * read_cr2() before calling anything else. To avoid calling any kind
1266	 * of tracing machinery before we've observed the CR2 value.
1267	 *
1268	 * exception_{enter,exit}() contain all sorts of tracepoints.
1269	 */
1270
1271	prev_state = exception_enter();
1272	__do_page_fault(regs, error_code, address);
1273	exception_exit(prev_state);
1274}
 
1275
1276#ifdef CONFIG_TRACING
1277static void trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1278				     unsigned long error_code)
 
1279{
1280	if (user_mode(regs))
1281		trace_page_fault_user(address, regs, error_code);
1282	else
1283		trace_page_fault_kernel(address, regs, error_code);
1284}
1285
1286dotraplinkage void __kprobes notrace
1287trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1288{
1289	/*
1290	 * The exception_enter and tracepoint processing could
1291	 * trigger another page faults (user space callchain
1292	 * reading) and destroy the original cr2 value, so read
1293	 * the faulting address now.
1294	 */
1295	unsigned long address = read_cr2();
1296	enum ctx_state prev_state;
1297
1298	prev_state = exception_enter();
1299	trace_page_fault_entries(address, regs, error_code);
1300	__do_page_fault(regs, error_code, address);
1301	exception_exit(prev_state);
1302}
 
1303#endif /* CONFIG_TRACING */
v4.6
   1/*
   2 *  Copyright (C) 1995  Linus Torvalds
   3 *  Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
   4 *  Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
   5 */
 
   6#include <linux/sched.h>		/* test_thread_flag(), ...	*/
   7#include <linux/kdebug.h>		/* oops_begin/end, ...		*/
   8#include <linux/module.h>		/* search_exception_table	*/
   9#include <linux/bootmem.h>		/* max_low_pfn			*/
  10#include <linux/kprobes.h>		/* NOKPROBE_SYMBOL, ...		*/
  11#include <linux/mmiotrace.h>		/* kmmio_handler, ...		*/
  12#include <linux/perf_event.h>		/* perf_sw_event		*/
  13#include <linux/hugetlb.h>		/* hstate_index_to_shift	*/
  14#include <linux/prefetch.h>		/* prefetchw			*/
  15#include <linux/context_tracking.h>	/* exception_enter(), ...	*/
  16#include <linux/uaccess.h>		/* faulthandler_disabled()	*/
  17
  18#include <asm/cpufeature.h>		/* boot_cpu_has, ...		*/
  19#include <asm/traps.h>			/* dotraplinkage, ...		*/
  20#include <asm/pgalloc.h>		/* pgd_*(), ...			*/
  21#include <asm/kmemcheck.h>		/* kmemcheck_*(), ...		*/
  22#include <asm/fixmap.h>			/* VSYSCALL_ADDR		*/
  23#include <asm/vsyscall.h>		/* emulate_vsyscall		*/
  24#include <asm/vm86.h>			/* struct vm86			*/
  25#include <asm/mmu_context.h>		/* vma_pkey()			*/
  26
  27#define CREATE_TRACE_POINTS
  28#include <asm/trace/exceptions.h>
  29
  30/*
  31 * Page fault error code bits:
  32 *
  33 *   bit 0 ==	 0: no page found	1: protection fault
  34 *   bit 1 ==	 0: read access		1: write access
  35 *   bit 2 ==	 0: kernel-mode access	1: user-mode access
  36 *   bit 3 ==				1: use of reserved bit detected
  37 *   bit 4 ==				1: fault was an instruction fetch
  38 *   bit 5 ==				1: protection keys block access
  39 */
  40enum x86_pf_error_code {
  41
  42	PF_PROT		=		1 << 0,
  43	PF_WRITE	=		1 << 1,
  44	PF_USER		=		1 << 2,
  45	PF_RSVD		=		1 << 3,
  46	PF_INSTR	=		1 << 4,
  47	PF_PK		=		1 << 5,
  48};
  49
  50/*
  51 * Returns 0 if mmiotrace is disabled, or if the fault is not
  52 * handled by mmiotrace:
  53 */
  54static nokprobe_inline int
  55kmmio_fault(struct pt_regs *regs, unsigned long addr)
  56{
  57	if (unlikely(is_kmmio_active()))
  58		if (kmmio_handler(regs, addr) == 1)
  59			return -1;
  60	return 0;
  61}
  62
  63static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
  64{
  65	int ret = 0;
  66
  67	/* kprobe_running() needs smp_processor_id() */
  68	if (kprobes_built_in() && !user_mode(regs)) {
  69		preempt_disable();
  70		if (kprobe_running() && kprobe_fault_handler(regs, 14))
  71			ret = 1;
  72		preempt_enable();
  73	}
  74
  75	return ret;
  76}
  77
  78/*
  79 * Prefetch quirks:
  80 *
  81 * 32-bit mode:
  82 *
  83 *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  84 *   Check that here and ignore it.
  85 *
  86 * 64-bit mode:
  87 *
  88 *   Sometimes the CPU reports invalid exceptions on prefetch.
  89 *   Check that here and ignore it.
  90 *
  91 * Opcode checker based on code by Richard Brunner.
  92 */
  93static inline int
  94check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  95		      unsigned char opcode, int *prefetch)
  96{
  97	unsigned char instr_hi = opcode & 0xf0;
  98	unsigned char instr_lo = opcode & 0x0f;
  99
 100	switch (instr_hi) {
 101	case 0x20:
 102	case 0x30:
 103		/*
 104		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 105		 * In X86_64 long mode, the CPU will signal invalid
 106		 * opcode if some of these prefixes are present so
 107		 * X86_64 will never get here anyway
 108		 */
 109		return ((instr_lo & 7) == 0x6);
 110#ifdef CONFIG_X86_64
 111	case 0x40:
 112		/*
 113		 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 114		 * Need to figure out under what instruction mode the
 115		 * instruction was issued. Could check the LDT for lm,
 116		 * but for now it's good enough to assume that long
 117		 * mode only uses well known segments or kernel.
 118		 */
 119		return (!user_mode(regs) || user_64bit_mode(regs));
 120#endif
 121	case 0x60:
 122		/* 0x64 thru 0x67 are valid prefixes in all modes. */
 123		return (instr_lo & 0xC) == 0x4;
 124	case 0xF0:
 125		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 126		return !instr_lo || (instr_lo>>1) == 1;
 127	case 0x00:
 128		/* Prefetch instruction is 0x0F0D or 0x0F18 */
 129		if (probe_kernel_address(instr, opcode))
 130			return 0;
 131
 132		*prefetch = (instr_lo == 0xF) &&
 133			(opcode == 0x0D || opcode == 0x18);
 134		return 0;
 135	default:
 136		return 0;
 137	}
 138}
 139
 140static int
 141is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 142{
 143	unsigned char *max_instr;
 144	unsigned char *instr;
 145	int prefetch = 0;
 146
 147	/*
 148	 * If it was a exec (instruction fetch) fault on NX page, then
 149	 * do not ignore the fault:
 150	 */
 151	if (error_code & PF_INSTR)
 152		return 0;
 153
 154	instr = (void *)convert_ip_to_linear(current, regs);
 155	max_instr = instr + 15;
 156
 157	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
 158		return 0;
 159
 160	while (instr < max_instr) {
 161		unsigned char opcode;
 162
 163		if (probe_kernel_address(instr, opcode))
 164			break;
 165
 166		instr++;
 167
 168		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 169			break;
 170	}
 171	return prefetch;
 172}
 173
 174/*
 175 * A protection key fault means that the PKRU value did not allow
 176 * access to some PTE.  Userspace can figure out what PKRU was
 177 * from the XSAVE state, and this function fills out a field in
 178 * siginfo so userspace can discover which protection key was set
 179 * on the PTE.
 180 *
 181 * If we get here, we know that the hardware signaled a PF_PK
 182 * fault and that there was a VMA once we got in the fault
 183 * handler.  It does *not* guarantee that the VMA we find here
 184 * was the one that we faulted on.
 185 *
 186 * 1. T1   : mprotect_key(foo, PAGE_SIZE, pkey=4);
 187 * 2. T1   : set PKRU to deny access to pkey=4, touches page
 188 * 3. T1   : faults...
 189 * 4.    T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
 190 * 5. T1   : enters fault handler, takes mmap_sem, etc...
 191 * 6. T1   : reaches here, sees vma_pkey(vma)=5, when we really
 192 *	     faulted on a pte with its pkey=4.
 193 */
 194static void fill_sig_info_pkey(int si_code, siginfo_t *info,
 195		struct vm_area_struct *vma)
 196{
 197	/* This is effectively an #ifdef */
 198	if (!boot_cpu_has(X86_FEATURE_OSPKE))
 199		return;
 200
 201	/* Fault not from Protection Keys: nothing to do */
 202	if (si_code != SEGV_PKUERR)
 203		return;
 204	/*
 205	 * force_sig_info_fault() is called from a number of
 206	 * contexts, some of which have a VMA and some of which
 207	 * do not.  The PF_PK handing happens after we have a
 208	 * valid VMA, so we should never reach this without a
 209	 * valid VMA.
 210	 */
 211	if (!vma) {
 212		WARN_ONCE(1, "PKU fault with no VMA passed in");
 213		info->si_pkey = 0;
 214		return;
 215	}
 216	/*
 217	 * si_pkey should be thought of as a strong hint, but not
 218	 * absolutely guranteed to be 100% accurate because of
 219	 * the race explained above.
 220	 */
 221	info->si_pkey = vma_pkey(vma);
 222}
 223
 224static void
 225force_sig_info_fault(int si_signo, int si_code, unsigned long address,
 226		     struct task_struct *tsk, struct vm_area_struct *vma,
 227		     int fault)
 228{
 229	unsigned lsb = 0;
 230	siginfo_t info;
 231
 232	info.si_signo	= si_signo;
 233	info.si_errno	= 0;
 234	info.si_code	= si_code;
 235	info.si_addr	= (void __user *)address;
 236	if (fault & VM_FAULT_HWPOISON_LARGE)
 237		lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); 
 238	if (fault & VM_FAULT_HWPOISON)
 239		lsb = PAGE_SHIFT;
 240	info.si_addr_lsb = lsb;
 241
 242	fill_sig_info_pkey(si_code, &info, vma);
 243
 244	force_sig_info(si_signo, &info, tsk);
 245}
 246
 247DEFINE_SPINLOCK(pgd_lock);
 248LIST_HEAD(pgd_list);
 249
 250#ifdef CONFIG_X86_32
 251static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 252{
 253	unsigned index = pgd_index(address);
 254	pgd_t *pgd_k;
 255	pud_t *pud, *pud_k;
 256	pmd_t *pmd, *pmd_k;
 257
 258	pgd += index;
 259	pgd_k = init_mm.pgd + index;
 260
 261	if (!pgd_present(*pgd_k))
 262		return NULL;
 263
 264	/*
 265	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
 266	 * and redundant with the set_pmd() on non-PAE. As would
 267	 * set_pud.
 268	 */
 269	pud = pud_offset(pgd, address);
 270	pud_k = pud_offset(pgd_k, address);
 271	if (!pud_present(*pud_k))
 272		return NULL;
 273
 274	pmd = pmd_offset(pud, address);
 275	pmd_k = pmd_offset(pud_k, address);
 276	if (!pmd_present(*pmd_k))
 277		return NULL;
 278
 279	if (!pmd_present(*pmd))
 280		set_pmd(pmd, *pmd_k);
 281	else
 282		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 283
 284	return pmd_k;
 285}
 286
 287void vmalloc_sync_all(void)
 288{
 289	unsigned long address;
 290
 291	if (SHARED_KERNEL_PMD)
 292		return;
 293
 294	for (address = VMALLOC_START & PMD_MASK;
 295	     address >= TASK_SIZE && address < FIXADDR_TOP;
 296	     address += PMD_SIZE) {
 297		struct page *page;
 298
 299		spin_lock(&pgd_lock);
 300		list_for_each_entry(page, &pgd_list, lru) {
 301			spinlock_t *pgt_lock;
 302			pmd_t *ret;
 303
 304			/* the pgt_lock only for Xen */
 305			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 306
 307			spin_lock(pgt_lock);
 308			ret = vmalloc_sync_one(page_address(page), address);
 309			spin_unlock(pgt_lock);
 310
 311			if (!ret)
 312				break;
 313		}
 314		spin_unlock(&pgd_lock);
 315	}
 316}
 317
 318/*
 319 * 32-bit:
 320 *
 321 *   Handle a fault on the vmalloc or module mapping area
 322 */
 323static noinline int vmalloc_fault(unsigned long address)
 324{
 325	unsigned long pgd_paddr;
 326	pmd_t *pmd_k;
 327	pte_t *pte_k;
 328
 329	/* Make sure we are in vmalloc area: */
 330	if (!(address >= VMALLOC_START && address < VMALLOC_END))
 331		return -1;
 332
 333	WARN_ON_ONCE(in_nmi());
 334
 335	/*
 336	 * Synchronize this task's top level page-table
 337	 * with the 'reference' page table.
 338	 *
 339	 * Do _not_ use "current" here. We might be inside
 340	 * an interrupt in the middle of a task switch..
 341	 */
 342	pgd_paddr = read_cr3();
 343	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 344	if (!pmd_k)
 345		return -1;
 346
 347	if (pmd_huge(*pmd_k))
 348		return 0;
 349
 350	pte_k = pte_offset_kernel(pmd_k, address);
 351	if (!pte_present(*pte_k))
 352		return -1;
 353
 354	return 0;
 355}
 356NOKPROBE_SYMBOL(vmalloc_fault);
 357
 358/*
 359 * Did it hit the DOS screen memory VA from vm86 mode?
 360 */
 361static inline void
 362check_v8086_mode(struct pt_regs *regs, unsigned long address,
 363		 struct task_struct *tsk)
 364{
 365#ifdef CONFIG_VM86
 366	unsigned long bit;
 367
 368	if (!v8086_mode(regs) || !tsk->thread.vm86)
 369		return;
 370
 371	bit = (address - 0xA0000) >> PAGE_SHIFT;
 372	if (bit < 32)
 373		tsk->thread.vm86->screen_bitmap |= 1 << bit;
 374#endif
 375}
 376
 377static bool low_pfn(unsigned long pfn)
 378{
 379	return pfn < max_low_pfn;
 380}
 381
 382static void dump_pagetable(unsigned long address)
 383{
 384	pgd_t *base = __va(read_cr3());
 385	pgd_t *pgd = &base[pgd_index(address)];
 386	pmd_t *pmd;
 387	pte_t *pte;
 388
 389#ifdef CONFIG_X86_PAE
 390	printk("*pdpt = %016Lx ", pgd_val(*pgd));
 391	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 392		goto out;
 393#endif
 394	pmd = pmd_offset(pud_offset(pgd, address), address);
 395	printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 396
 397	/*
 398	 * We must not directly access the pte in the highpte
 399	 * case if the page table is located in highmem.
 400	 * And let's rather not kmap-atomic the pte, just in case
 401	 * it's allocated already:
 402	 */
 403	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
 404		goto out;
 405
 406	pte = pte_offset_kernel(pmd, address);
 407	printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 408out:
 409	printk("\n");
 410}
 411
 412#else /* CONFIG_X86_64: */
 413
 414void vmalloc_sync_all(void)
 415{
 416	sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0);
 417}
 418
 419/*
 420 * 64-bit:
 421 *
 422 *   Handle a fault on the vmalloc area
 
 
 423 */
 424static noinline int vmalloc_fault(unsigned long address)
 425{
 426	pgd_t *pgd, *pgd_ref;
 427	pud_t *pud, *pud_ref;
 428	pmd_t *pmd, *pmd_ref;
 429	pte_t *pte, *pte_ref;
 430
 431	/* Make sure we are in vmalloc area: */
 432	if (!(address >= VMALLOC_START && address < VMALLOC_END))
 433		return -1;
 434
 435	WARN_ON_ONCE(in_nmi());
 436
 437	/*
 438	 * Copy kernel mappings over when needed. This can also
 439	 * happen within a race in page table update. In the later
 440	 * case just flush:
 441	 */
 442	pgd = pgd_offset(current->active_mm, address);
 443	pgd_ref = pgd_offset_k(address);
 444	if (pgd_none(*pgd_ref))
 445		return -1;
 446
 447	if (pgd_none(*pgd)) {
 448		set_pgd(pgd, *pgd_ref);
 449		arch_flush_lazy_mmu_mode();
 450	} else {
 451		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 452	}
 453
 454	/*
 455	 * Below here mismatches are bugs because these lower tables
 456	 * are shared:
 457	 */
 458
 459	pud = pud_offset(pgd, address);
 460	pud_ref = pud_offset(pgd_ref, address);
 461	if (pud_none(*pud_ref))
 462		return -1;
 463
 464	if (pud_none(*pud) || pud_pfn(*pud) != pud_pfn(*pud_ref))
 465		BUG();
 466
 467	if (pud_huge(*pud))
 468		return 0;
 469
 470	pmd = pmd_offset(pud, address);
 471	pmd_ref = pmd_offset(pud_ref, address);
 472	if (pmd_none(*pmd_ref))
 473		return -1;
 474
 475	if (pmd_none(*pmd) || pmd_pfn(*pmd) != pmd_pfn(*pmd_ref))
 476		BUG();
 477
 478	if (pmd_huge(*pmd))
 479		return 0;
 480
 481	pte_ref = pte_offset_kernel(pmd_ref, address);
 482	if (!pte_present(*pte_ref))
 483		return -1;
 484
 485	pte = pte_offset_kernel(pmd, address);
 486
 487	/*
 488	 * Don't use pte_page here, because the mappings can point
 489	 * outside mem_map, and the NUMA hash lookup cannot handle
 490	 * that:
 491	 */
 492	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 493		BUG();
 494
 495	return 0;
 496}
 497NOKPROBE_SYMBOL(vmalloc_fault);
 498
 499#ifdef CONFIG_CPU_SUP_AMD
 500static const char errata93_warning[] =
 501KERN_ERR 
 502"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 503"******* Working around it, but it may cause SEGVs or burn power.\n"
 504"******* Please consider a BIOS update.\n"
 505"******* Disabling USB legacy in the BIOS may also help.\n";
 506#endif
 507
 508/*
 509 * No vm86 mode in 64-bit mode:
 510 */
 511static inline void
 512check_v8086_mode(struct pt_regs *regs, unsigned long address,
 513		 struct task_struct *tsk)
 514{
 515}
 516
 517static int bad_address(void *p)
 518{
 519	unsigned long dummy;
 520
 521	return probe_kernel_address((unsigned long *)p, dummy);
 522}
 523
 524static void dump_pagetable(unsigned long address)
 525{
 526	pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 527	pgd_t *pgd = base + pgd_index(address);
 528	pud_t *pud;
 529	pmd_t *pmd;
 530	pte_t *pte;
 531
 532	if (bad_address(pgd))
 533		goto bad;
 534
 535	printk("PGD %lx ", pgd_val(*pgd));
 536
 537	if (!pgd_present(*pgd))
 538		goto out;
 539
 540	pud = pud_offset(pgd, address);
 541	if (bad_address(pud))
 542		goto bad;
 543
 544	printk("PUD %lx ", pud_val(*pud));
 545	if (!pud_present(*pud) || pud_large(*pud))
 546		goto out;
 547
 548	pmd = pmd_offset(pud, address);
 549	if (bad_address(pmd))
 550		goto bad;
 551
 552	printk("PMD %lx ", pmd_val(*pmd));
 553	if (!pmd_present(*pmd) || pmd_large(*pmd))
 554		goto out;
 555
 556	pte = pte_offset_kernel(pmd, address);
 557	if (bad_address(pte))
 558		goto bad;
 559
 560	printk("PTE %lx", pte_val(*pte));
 561out:
 562	printk("\n");
 563	return;
 564bad:
 565	printk("BAD\n");
 566}
 567
 568#endif /* CONFIG_X86_64 */
 569
 570/*
 571 * Workaround for K8 erratum #93 & buggy BIOS.
 572 *
 573 * BIOS SMM functions are required to use a specific workaround
 574 * to avoid corruption of the 64bit RIP register on C stepping K8.
 575 *
 576 * A lot of BIOS that didn't get tested properly miss this.
 577 *
 578 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 579 * Try to work around it here.
 580 *
 581 * Note we only handle faults in kernel here.
 582 * Does nothing on 32-bit.
 583 */
 584static int is_errata93(struct pt_regs *regs, unsigned long address)
 585{
 586#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
 587	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
 588	    || boot_cpu_data.x86 != 0xf)
 589		return 0;
 590
 591	if (address != regs->ip)
 592		return 0;
 593
 594	if ((address >> 32) != 0)
 595		return 0;
 596
 597	address |= 0xffffffffUL << 32;
 598	if ((address >= (u64)_stext && address <= (u64)_etext) ||
 599	    (address >= MODULES_VADDR && address <= MODULES_END)) {
 600		printk_once(errata93_warning);
 601		regs->ip = address;
 602		return 1;
 603	}
 604#endif
 605	return 0;
 606}
 607
 608/*
 609 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 610 * to illegal addresses >4GB.
 611 *
 612 * We catch this in the page fault handler because these addresses
 613 * are not reachable. Just detect this case and return.  Any code
 614 * segment in LDT is compatibility mode.
 615 */
 616static int is_errata100(struct pt_regs *regs, unsigned long address)
 617{
 618#ifdef CONFIG_X86_64
 619	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 620		return 1;
 621#endif
 622	return 0;
 623}
 624
 625static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 626{
 627#ifdef CONFIG_X86_F00F_BUG
 628	unsigned long nr;
 629
 630	/*
 631	 * Pentium F0 0F C7 C8 bug workaround:
 632	 */
 633	if (boot_cpu_has_bug(X86_BUG_F00F)) {
 634		nr = (address - idt_descr.address) >> 3;
 635
 636		if (nr == 6) {
 637			do_invalid_op(regs, 0);
 638			return 1;
 639		}
 640	}
 641#endif
 642	return 0;
 643}
 644
 645static const char nx_warning[] = KERN_CRIT
 646"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 647static const char smep_warning[] = KERN_CRIT
 648"unable to execute userspace code (SMEP?) (uid: %d)\n";
 649
 650static void
 651show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 652		unsigned long address)
 653{
 654	if (!oops_may_print())
 655		return;
 656
 657	if (error_code & PF_INSTR) {
 658		unsigned int level;
 659		pgd_t *pgd;
 660		pte_t *pte;
 661
 662		pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 663		pgd += pgd_index(address);
 664
 665		pte = lookup_address_in_pgd(pgd, address, &level);
 666
 667		if (pte && pte_present(*pte) && !pte_exec(*pte))
 668			printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
 669		if (pte && pte_present(*pte) && pte_exec(*pte) &&
 670				(pgd_flags(*pgd) & _PAGE_USER) &&
 671				(__read_cr4() & X86_CR4_SMEP))
 672			printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
 673	}
 674
 675	printk(KERN_ALERT "BUG: unable to handle kernel ");
 676	if (address < PAGE_SIZE)
 677		printk(KERN_CONT "NULL pointer dereference");
 678	else
 679		printk(KERN_CONT "paging request");
 680
 681	printk(KERN_CONT " at %p\n", (void *) address);
 682	printk(KERN_ALERT "IP:");
 683	printk_address(regs->ip);
 684
 685	dump_pagetable(address);
 686}
 687
 688static noinline void
 689pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 690	    unsigned long address)
 691{
 692	struct task_struct *tsk;
 693	unsigned long flags;
 694	int sig;
 695
 696	flags = oops_begin();
 697	tsk = current;
 698	sig = SIGKILL;
 699
 700	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 701	       tsk->comm, address);
 702	dump_pagetable(address);
 703
 704	tsk->thread.cr2		= address;
 705	tsk->thread.trap_nr	= X86_TRAP_PF;
 706	tsk->thread.error_code	= error_code;
 707
 708	if (__die("Bad pagetable", regs, error_code))
 709		sig = 0;
 710
 711	oops_end(flags, regs, sig);
 712}
 713
 714static noinline void
 715no_context(struct pt_regs *regs, unsigned long error_code,
 716	   unsigned long address, int signal, int si_code)
 717{
 718	struct task_struct *tsk = current;
 
 719	unsigned long flags;
 720	int sig;
 721	/* No context means no VMA to pass down */
 722	struct vm_area_struct *vma = NULL;
 723
 724	/* Are we prepared to handle this kernel fault? */
 725	if (fixup_exception(regs, X86_TRAP_PF)) {
 726		/*
 727		 * Any interrupt that takes a fault gets the fixup. This makes
 728		 * the below recursive fault logic only apply to a faults from
 729		 * task context.
 730		 */
 731		if (in_interrupt())
 732			return;
 733
 734		/*
 735		 * Per the above we're !in_interrupt(), aka. task context.
 736		 *
 737		 * In this case we need to make sure we're not recursively
 738		 * faulting through the emulate_vsyscall() logic.
 739		 */
 740		if (current_thread_info()->sig_on_uaccess_error && signal) {
 741			tsk->thread.trap_nr = X86_TRAP_PF;
 742			tsk->thread.error_code = error_code | PF_USER;
 743			tsk->thread.cr2 = address;
 744
 745			/* XXX: hwpoison faults will set the wrong code. */
 746			force_sig_info_fault(signal, si_code, address,
 747					     tsk, vma, 0);
 748		}
 749
 750		/*
 751		 * Barring that, we can do the fixup and be happy.
 752		 */
 753		return;
 754	}
 755
 756	/*
 757	 * 32-bit:
 758	 *
 759	 *   Valid to do another page fault here, because if this fault
 760	 *   had been triggered by is_prefetch fixup_exception would have
 761	 *   handled it.
 762	 *
 763	 * 64-bit:
 764	 *
 765	 *   Hall of shame of CPU/BIOS bugs.
 766	 */
 767	if (is_prefetch(regs, error_code, address))
 768		return;
 769
 770	if (is_errata93(regs, address))
 771		return;
 772
 773	/*
 774	 * Oops. The kernel tried to access some bad page. We'll have to
 775	 * terminate things with extreme prejudice:
 776	 */
 777	flags = oops_begin();
 778
 779	show_fault_oops(regs, error_code, address);
 780
 781	if (task_stack_end_corrupted(tsk))
 
 782		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
 783
 784	tsk->thread.cr2		= address;
 785	tsk->thread.trap_nr	= X86_TRAP_PF;
 786	tsk->thread.error_code	= error_code;
 787
 788	sig = SIGKILL;
 789	if (__die("Oops", regs, error_code))
 790		sig = 0;
 791
 792	/* Executive summary in case the body of the oops scrolled away */
 793	printk(KERN_DEFAULT "CR2: %016lx\n", address);
 794
 795	oops_end(flags, regs, sig);
 796}
 797
 798/*
 799 * Print out info about fatal segfaults, if the show_unhandled_signals
 800 * sysctl is set:
 801 */
 802static inline void
 803show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 804		unsigned long address, struct task_struct *tsk)
 805{
 806	if (!unhandled_signal(tsk, SIGSEGV))
 807		return;
 808
 809	if (!printk_ratelimit())
 810		return;
 811
 812	printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 813		task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 814		tsk->comm, task_pid_nr(tsk), address,
 815		(void *)regs->ip, (void *)regs->sp, error_code);
 816
 817	print_vma_addr(KERN_CONT " in ", regs->ip);
 818
 819	printk(KERN_CONT "\n");
 820}
 821
 822static void
 823__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 824		       unsigned long address, struct vm_area_struct *vma,
 825		       int si_code)
 826{
 827	struct task_struct *tsk = current;
 828
 829	/* User mode accesses just cause a SIGSEGV */
 830	if (error_code & PF_USER) {
 831		/*
 832		 * It's possible to have interrupts off here:
 833		 */
 834		local_irq_enable();
 835
 836		/*
 837		 * Valid to do another page fault here because this one came
 838		 * from user space:
 839		 */
 840		if (is_prefetch(regs, error_code, address))
 841			return;
 842
 843		if (is_errata100(regs, address))
 844			return;
 845
 846#ifdef CONFIG_X86_64
 847		/*
 848		 * Instruction fetch faults in the vsyscall page might need
 849		 * emulation.
 850		 */
 851		if (unlikely((error_code & PF_INSTR) &&
 852			     ((address & ~0xfff) == VSYSCALL_ADDR))) {
 853			if (emulate_vsyscall(regs, address))
 854				return;
 855		}
 856#endif
 857		/* Kernel addresses are always protection faults: */
 858		if (address >= TASK_SIZE)
 859			error_code |= PF_PROT;
 860
 861		if (likely(show_unhandled_signals))
 862			show_signal_msg(regs, error_code, address, tsk);
 863
 864		tsk->thread.cr2		= address;
 865		tsk->thread.error_code	= error_code;
 866		tsk->thread.trap_nr	= X86_TRAP_PF;
 867
 868		force_sig_info_fault(SIGSEGV, si_code, address, tsk, vma, 0);
 869
 870		return;
 871	}
 872
 873	if (is_f00f_bug(regs, address))
 874		return;
 875
 876	no_context(regs, error_code, address, SIGSEGV, si_code);
 877}
 878
 879static noinline void
 880bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 881		     unsigned long address, struct vm_area_struct *vma)
 882{
 883	__bad_area_nosemaphore(regs, error_code, address, vma, SEGV_MAPERR);
 884}
 885
 886static void
 887__bad_area(struct pt_regs *regs, unsigned long error_code,
 888	   unsigned long address,  struct vm_area_struct *vma, int si_code)
 889{
 890	struct mm_struct *mm = current->mm;
 891
 892	/*
 893	 * Something tried to access memory that isn't in our memory map..
 894	 * Fix it, but check if it's kernel or user first..
 895	 */
 896	up_read(&mm->mmap_sem);
 897
 898	__bad_area_nosemaphore(regs, error_code, address, vma, si_code);
 899}
 900
 901static noinline void
 902bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 903{
 904	__bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
 905}
 906
 907static inline bool bad_area_access_from_pkeys(unsigned long error_code,
 908		struct vm_area_struct *vma)
 909{
 910	/* This code is always called on the current mm */
 911	bool foreign = false;
 912
 913	if (!boot_cpu_has(X86_FEATURE_OSPKE))
 914		return false;
 915	if (error_code & PF_PK)
 916		return true;
 917	/* this checks permission keys on the VMA: */
 918	if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
 919				(error_code & PF_INSTR), foreign))
 920		return true;
 921	return false;
 922}
 923
 924static noinline void
 925bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 926		      unsigned long address, struct vm_area_struct *vma)
 927{
 928	/*
 929	 * This OSPKE check is not strictly necessary at runtime.
 930	 * But, doing it this way allows compiler optimizations
 931	 * if pkeys are compiled out.
 932	 */
 933	if (bad_area_access_from_pkeys(error_code, vma))
 934		__bad_area(regs, error_code, address, vma, SEGV_PKUERR);
 935	else
 936		__bad_area(regs, error_code, address, vma, SEGV_ACCERR);
 937}
 938
 939static void
 940do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 941	  struct vm_area_struct *vma, unsigned int fault)
 942{
 943	struct task_struct *tsk = current;
 
 944	int code = BUS_ADRERR;
 945
 
 
 946	/* Kernel mode? Handle exceptions or die: */
 947	if (!(error_code & PF_USER)) {
 948		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
 949		return;
 950	}
 951
 952	/* User-space => ok to do another page fault: */
 953	if (is_prefetch(regs, error_code, address))
 954		return;
 955
 956	tsk->thread.cr2		= address;
 957	tsk->thread.error_code	= error_code;
 958	tsk->thread.trap_nr	= X86_TRAP_PF;
 959
 960#ifdef CONFIG_MEMORY_FAILURE
 961	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
 962		printk(KERN_ERR
 963	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
 964			tsk->comm, tsk->pid, address);
 965		code = BUS_MCEERR_AR;
 966	}
 967#endif
 968	force_sig_info_fault(SIGBUS, code, address, tsk, vma, fault);
 969}
 970
 971static noinline void
 972mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 973	       unsigned long address, struct vm_area_struct *vma,
 974	       unsigned int fault)
 975{
 976	if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
 
 977		no_context(regs, error_code, address, 0, 0);
 978		return;
 979	}
 980
 981	if (fault & VM_FAULT_OOM) {
 982		/* Kernel mode? Handle exceptions or die: */
 983		if (!(error_code & PF_USER)) {
 
 984			no_context(regs, error_code, address,
 985				   SIGSEGV, SEGV_MAPERR);
 986			return;
 987		}
 988
 
 
 989		/*
 990		 * We ran out of memory, call the OOM killer, and return the
 991		 * userspace (which will retry the fault, or kill us if we got
 992		 * oom-killed):
 993		 */
 994		pagefault_out_of_memory();
 995	} else {
 996		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
 997			     VM_FAULT_HWPOISON_LARGE))
 998			do_sigbus(regs, error_code, address, vma, fault);
 999		else if (fault & VM_FAULT_SIGSEGV)
1000			bad_area_nosemaphore(regs, error_code, address, vma);
1001		else
1002			BUG();
1003	}
1004}
1005
1006static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1007{
1008	if ((error_code & PF_WRITE) && !pte_write(*pte))
1009		return 0;
1010
1011	if ((error_code & PF_INSTR) && !pte_exec(*pte))
1012		return 0;
1013	/*
1014	 * Note: We do not do lazy flushing on protection key
1015	 * changes, so no spurious fault will ever set PF_PK.
1016	 */
1017	if ((error_code & PF_PK))
1018		return 1;
1019
1020	return 1;
1021}
1022
1023/*
1024 * Handle a spurious fault caused by a stale TLB entry.
1025 *
1026 * This allows us to lazily refresh the TLB when increasing the
1027 * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
1028 * eagerly is very expensive since that implies doing a full
1029 * cross-processor TLB flush, even if no stale TLB entries exist
1030 * on other processors.
1031 *
1032 * Spurious faults may only occur if the TLB contains an entry with
1033 * fewer permission than the page table entry.  Non-present (P = 0)
1034 * and reserved bit (R = 1) faults are never spurious.
1035 *
1036 * There are no security implications to leaving a stale TLB when
1037 * increasing the permissions on a page.
1038 *
1039 * Returns non-zero if a spurious fault was handled, zero otherwise.
1040 *
1041 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1042 * (Optional Invalidation).
1043 */
1044static noinline int
1045spurious_fault(unsigned long error_code, unsigned long address)
1046{
1047	pgd_t *pgd;
1048	pud_t *pud;
1049	pmd_t *pmd;
1050	pte_t *pte;
1051	int ret;
1052
1053	/*
1054	 * Only writes to RO or instruction fetches from NX may cause
1055	 * spurious faults.
1056	 *
1057	 * These could be from user or supervisor accesses but the TLB
1058	 * is only lazily flushed after a kernel mapping protection
1059	 * change, so user accesses are not expected to cause spurious
1060	 * faults.
1061	 */
1062	if (error_code != (PF_WRITE | PF_PROT)
1063	    && error_code != (PF_INSTR | PF_PROT))
1064		return 0;
1065
1066	pgd = init_mm.pgd + pgd_index(address);
1067	if (!pgd_present(*pgd))
1068		return 0;
1069
1070	pud = pud_offset(pgd, address);
1071	if (!pud_present(*pud))
1072		return 0;
1073
1074	if (pud_large(*pud))
1075		return spurious_fault_check(error_code, (pte_t *) pud);
1076
1077	pmd = pmd_offset(pud, address);
1078	if (!pmd_present(*pmd))
1079		return 0;
1080
1081	if (pmd_large(*pmd))
1082		return spurious_fault_check(error_code, (pte_t *) pmd);
1083
1084	pte = pte_offset_kernel(pmd, address);
1085	if (!pte_present(*pte))
1086		return 0;
1087
1088	ret = spurious_fault_check(error_code, pte);
1089	if (!ret)
1090		return 0;
1091
1092	/*
1093	 * Make sure we have permissions in PMD.
1094	 * If not, then there's a bug in the page tables:
1095	 */
1096	ret = spurious_fault_check(error_code, (pte_t *) pmd);
1097	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1098
1099	return ret;
1100}
1101NOKPROBE_SYMBOL(spurious_fault);
1102
1103int show_unhandled_signals = 1;
1104
1105static inline int
1106access_error(unsigned long error_code, struct vm_area_struct *vma)
1107{
1108	/* This is only called for the current mm, so: */
1109	bool foreign = false;
1110	/*
1111	 * Make sure to check the VMA so that we do not perform
1112	 * faults just to hit a PF_PK as soon as we fill in a
1113	 * page.
1114	 */
1115	if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
1116				(error_code & PF_INSTR), foreign))
1117		return 1;
1118
1119	if (error_code & PF_WRITE) {
1120		/* write, present and write, not present: */
1121		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1122			return 1;
1123		return 0;
1124	}
1125
1126	/* read, present: */
1127	if (unlikely(error_code & PF_PROT))
1128		return 1;
1129
1130	/* read, not present: */
1131	if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1132		return 1;
1133
1134	return 0;
1135}
1136
1137static int fault_in_kernel_space(unsigned long address)
1138{
1139	return address >= TASK_SIZE_MAX;
1140}
1141
1142static inline bool smap_violation(int error_code, struct pt_regs *regs)
1143{
1144	if (!IS_ENABLED(CONFIG_X86_SMAP))
1145		return false;
1146
1147	if (!static_cpu_has(X86_FEATURE_SMAP))
1148		return false;
1149
1150	if (error_code & PF_USER)
1151		return false;
1152
1153	if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1154		return false;
1155
1156	return true;
1157}
1158
1159/*
1160 * This routine handles page faults.  It determines the address,
1161 * and the problem, and then passes it off to one of the appropriate
1162 * routines.
1163 *
1164 * This function must have noinline because both callers
1165 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1166 * guarantees there's a function trace entry.
1167 */
1168static noinline void
1169__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1170		unsigned long address)
1171{
1172	struct vm_area_struct *vma;
1173	struct task_struct *tsk;
1174	struct mm_struct *mm;
1175	int fault, major = 0;
1176	unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1177
1178	tsk = current;
1179	mm = tsk->mm;
1180
1181	/*
1182	 * Detect and handle instructions that would cause a page fault for
1183	 * both a tracked kernel page and a userspace page.
1184	 */
1185	if (kmemcheck_active(regs))
1186		kmemcheck_hide(regs);
1187	prefetchw(&mm->mmap_sem);
1188
1189	if (unlikely(kmmio_fault(regs, address)))
1190		return;
1191
1192	/*
1193	 * We fault-in kernel-space virtual memory on-demand. The
1194	 * 'reference' page table is init_mm.pgd.
1195	 *
1196	 * NOTE! We MUST NOT take any locks for this case. We may
1197	 * be in an interrupt or a critical region, and should
1198	 * only copy the information from the master page table,
1199	 * nothing more.
1200	 *
1201	 * This verifies that the fault happens in kernel space
1202	 * (error_code & 4) == 0, and that the fault was not a
1203	 * protection error (error_code & 9) == 0.
1204	 */
1205	if (unlikely(fault_in_kernel_space(address))) {
1206		if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1207			if (vmalloc_fault(address) >= 0)
1208				return;
1209
1210			if (kmemcheck_fault(regs, address, error_code))
1211				return;
1212		}
1213
1214		/* Can handle a stale RO->RW TLB: */
1215		if (spurious_fault(error_code, address))
1216			return;
1217
1218		/* kprobes don't want to hook the spurious faults: */
1219		if (kprobes_fault(regs))
1220			return;
1221		/*
1222		 * Don't take the mm semaphore here. If we fixup a prefetch
1223		 * fault we could otherwise deadlock:
1224		 */
1225		bad_area_nosemaphore(regs, error_code, address, NULL);
1226
1227		return;
1228	}
1229
1230	/* kprobes don't want to hook the spurious faults: */
1231	if (unlikely(kprobes_fault(regs)))
1232		return;
1233
1234	if (unlikely(error_code & PF_RSVD))
1235		pgtable_bad(regs, error_code, address);
1236
1237	if (unlikely(smap_violation(error_code, regs))) {
1238		bad_area_nosemaphore(regs, error_code, address, NULL);
1239		return;
1240	}
1241
1242	/*
1243	 * If we're in an interrupt, have no user context or are running
1244	 * in a region with pagefaults disabled then we must not take the fault
1245	 */
1246	if (unlikely(faulthandler_disabled() || !mm)) {
1247		bad_area_nosemaphore(regs, error_code, address, NULL);
1248		return;
1249	}
1250
1251	/*
1252	 * It's safe to allow irq's after cr2 has been saved and the
1253	 * vmalloc fault has been handled.
1254	 *
1255	 * User-mode registers count as a user access even for any
1256	 * potential system fault or CPU buglet:
1257	 */
1258	if (user_mode(regs)) {
1259		local_irq_enable();
1260		error_code |= PF_USER;
1261		flags |= FAULT_FLAG_USER;
1262	} else {
1263		if (regs->flags & X86_EFLAGS_IF)
1264			local_irq_enable();
1265	}
1266
1267	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1268
1269	if (error_code & PF_WRITE)
1270		flags |= FAULT_FLAG_WRITE;
1271	if (error_code & PF_INSTR)
1272		flags |= FAULT_FLAG_INSTRUCTION;
1273
1274	/*
1275	 * When running in the kernel we expect faults to occur only to
1276	 * addresses in user space.  All other faults represent errors in
1277	 * the kernel and should generate an OOPS.  Unfortunately, in the
1278	 * case of an erroneous fault occurring in a code path which already
1279	 * holds mmap_sem we will deadlock attempting to validate the fault
1280	 * against the address space.  Luckily the kernel only validly
1281	 * references user space from well defined areas of code, which are
1282	 * listed in the exceptions table.
1283	 *
1284	 * As the vast majority of faults will be valid we will only perform
1285	 * the source reference check when there is a possibility of a
1286	 * deadlock. Attempt to lock the address space, if we cannot we then
1287	 * validate the source. If this is invalid we can skip the address
1288	 * space check, thus avoiding the deadlock:
1289	 */
1290	if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1291		if ((error_code & PF_USER) == 0 &&
1292		    !search_exception_tables(regs->ip)) {
1293			bad_area_nosemaphore(regs, error_code, address, NULL);
1294			return;
1295		}
1296retry:
1297		down_read(&mm->mmap_sem);
1298	} else {
1299		/*
1300		 * The above down_read_trylock() might have succeeded in
1301		 * which case we'll have missed the might_sleep() from
1302		 * down_read():
1303		 */
1304		might_sleep();
1305	}
1306
1307	vma = find_vma(mm, address);
1308	if (unlikely(!vma)) {
1309		bad_area(regs, error_code, address);
1310		return;
1311	}
1312	if (likely(vma->vm_start <= address))
1313		goto good_area;
1314	if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1315		bad_area(regs, error_code, address);
1316		return;
1317	}
1318	if (error_code & PF_USER) {
1319		/*
1320		 * Accessing the stack below %sp is always a bug.
1321		 * The large cushion allows instructions like enter
1322		 * and pusha to work. ("enter $65535, $31" pushes
1323		 * 32 pointers and then decrements %sp by 65535.)
1324		 */
1325		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1326			bad_area(regs, error_code, address);
1327			return;
1328		}
1329	}
1330	if (unlikely(expand_stack(vma, address))) {
1331		bad_area(regs, error_code, address);
1332		return;
1333	}
1334
1335	/*
1336	 * Ok, we have a good vm_area for this memory access, so
1337	 * we can handle it..
1338	 */
1339good_area:
1340	if (unlikely(access_error(error_code, vma))) {
1341		bad_area_access_error(regs, error_code, address, vma);
1342		return;
1343	}
1344
1345	/*
1346	 * If for any reason at all we couldn't handle the fault,
1347	 * make sure we exit gracefully rather than endlessly redo
1348	 * the fault.  Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1349	 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1350	 */
1351	fault = handle_mm_fault(mm, vma, address, flags);
1352	major |= fault & VM_FAULT_MAJOR;
1353
1354	/*
1355	 * If we need to retry the mmap_sem has already been released,
1356	 * and if there is a fatal signal pending there is no guarantee
1357	 * that we made any progress. Handle this case first.
1358	 */
1359	if (unlikely(fault & VM_FAULT_RETRY)) {
1360		/* Retry at most once */
1361		if (flags & FAULT_FLAG_ALLOW_RETRY) {
1362			flags &= ~FAULT_FLAG_ALLOW_RETRY;
1363			flags |= FAULT_FLAG_TRIED;
1364			if (!fatal_signal_pending(tsk))
1365				goto retry;
1366		}
1367
1368		/* User mode? Just return to handle the fatal exception */
1369		if (flags & FAULT_FLAG_USER)
1370			return;
1371
1372		/* Not returning to user mode? Handle exceptions or die: */
1373		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1374		return;
1375	}
1376
1377	up_read(&mm->mmap_sem);
1378	if (unlikely(fault & VM_FAULT_ERROR)) {
1379		mm_fault_error(regs, error_code, address, vma, fault);
1380		return;
1381	}
1382
1383	/*
1384	 * Major/minor page fault accounting. If any of the events
1385	 * returned VM_FAULT_MAJOR, we account it as a major fault.
 
1386	 */
1387	if (major) {
1388		tsk->maj_flt++;
1389		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1390	} else {
1391		tsk->min_flt++;
1392		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
 
 
 
 
 
 
 
 
 
 
 
1393	}
1394
1395	check_v8086_mode(regs, address, tsk);
 
 
1396}
1397NOKPROBE_SYMBOL(__do_page_fault);
1398
1399dotraplinkage void notrace
1400do_page_fault(struct pt_regs *regs, unsigned long error_code)
1401{
1402	unsigned long address = read_cr2(); /* Get the faulting address */
1403	enum ctx_state prev_state;
1404
1405	/*
1406	 * We must have this function tagged with __kprobes, notrace and call
1407	 * read_cr2() before calling anything else. To avoid calling any kind
1408	 * of tracing machinery before we've observed the CR2 value.
1409	 *
1410	 * exception_{enter,exit}() contain all sorts of tracepoints.
1411	 */
1412
1413	prev_state = exception_enter();
1414	__do_page_fault(regs, error_code, address);
1415	exception_exit(prev_state);
1416}
1417NOKPROBE_SYMBOL(do_page_fault);
1418
1419#ifdef CONFIG_TRACING
1420static nokprobe_inline void
1421trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1422			 unsigned long error_code)
1423{
1424	if (user_mode(regs))
1425		trace_page_fault_user(address, regs, error_code);
1426	else
1427		trace_page_fault_kernel(address, regs, error_code);
1428}
1429
1430dotraplinkage void notrace
1431trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1432{
1433	/*
1434	 * The exception_enter and tracepoint processing could
1435	 * trigger another page faults (user space callchain
1436	 * reading) and destroy the original cr2 value, so read
1437	 * the faulting address now.
1438	 */
1439	unsigned long address = read_cr2();
1440	enum ctx_state prev_state;
1441
1442	prev_state = exception_enter();
1443	trace_page_fault_entries(address, regs, error_code);
1444	__do_page_fault(regs, error_code, address);
1445	exception_exit(prev_state);
1446}
1447NOKPROBE_SYMBOL(trace_do_page_fault);
1448#endif /* CONFIG_TRACING */