xref: /openbmc/linux/arch/x86/mm/fault.c (revision 643d1f7f)
1 /*
2  *  Copyright (C) 1995  Linus Torvalds
3  *  Copyright (C) 2001,2002 Andi Kleen, SuSE Labs.
4  */
5 
6 #include <linux/signal.h>
7 #include <linux/sched.h>
8 #include <linux/kernel.h>
9 #include <linux/errno.h>
10 #include <linux/string.h>
11 #include <linux/types.h>
12 #include <linux/ptrace.h>
13 #include <linux/mman.h>
14 #include <linux/mm.h>
15 #include <linux/smp.h>
16 #include <linux/interrupt.h>
17 #include <linux/init.h>
18 #include <linux/tty.h>
19 #include <linux/vt_kern.h>		/* For unblank_screen() */
20 #include <linux/compiler.h>
21 #include <linux/highmem.h>
22 #include <linux/bootmem.h>		/* for max_low_pfn */
23 #include <linux/vmalloc.h>
24 #include <linux/module.h>
25 #include <linux/kprobes.h>
26 #include <linux/uaccess.h>
27 #include <linux/kdebug.h>
28 
29 #include <asm/system.h>
30 #include <asm/desc.h>
31 #include <asm/segment.h>
32 #include <asm/pgalloc.h>
33 #include <asm/smp.h>
34 #include <asm/tlbflush.h>
35 #include <asm/proto.h>
36 #include <asm-generic/sections.h>
37 
38 /*
39  * Page fault error code bits
40  *	bit 0 == 0 means no page found, 1 means protection fault
41  *	bit 1 == 0 means read, 1 means write
42  *	bit 2 == 0 means kernel, 1 means user-mode
43  *	bit 3 == 1 means use of reserved bit detected
44  *	bit 4 == 1 means fault was an instruction fetch
45  */
46 #define PF_PROT		(1<<0)
47 #define PF_WRITE	(1<<1)
48 #define PF_USER		(1<<2)
49 #define PF_RSVD		(1<<3)
50 #define PF_INSTR	(1<<4)
51 
52 static inline int notify_page_fault(struct pt_regs *regs)
53 {
54 #ifdef CONFIG_KPROBES
55 	int ret = 0;
56 
57 	/* kprobe_running() needs smp_processor_id() */
58 #ifdef CONFIG_X86_32
59 	if (!user_mode_vm(regs)) {
60 #else
61 	if (!user_mode(regs)) {
62 #endif
63 		preempt_disable();
64 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
65 			ret = 1;
66 		preempt_enable();
67 	}
68 
69 	return ret;
70 #else
71 	return 0;
72 #endif
73 }
74 
75 /*
76  * X86_32
77  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
78  * Check that here and ignore it.
79  *
80  * X86_64
81  * Sometimes the CPU reports invalid exceptions on prefetch.
82  * Check that here and ignore it.
83  *
84  * Opcode checker based on code by Richard Brunner
85  */
86 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
87 		       unsigned long error_code)
88 {
89 	unsigned char *instr;
90 	int scan_more = 1;
91 	int prefetch = 0;
92 	unsigned char *max_instr;
93 
94 #ifdef CONFIG_X86_32
95 	if (!(__supported_pte_mask & _PAGE_NX))
96 		return 0;
97 #endif
98 
99 	/* If it was a exec fault on NX page, ignore */
100 	if (error_code & PF_INSTR)
101 		return 0;
102 
103 	instr = (unsigned char *)convert_ip_to_linear(current, regs);
104 	max_instr = instr + 15;
105 
106 	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
107 		return 0;
108 
109 	while (scan_more && instr < max_instr) {
110 		unsigned char opcode;
111 		unsigned char instr_hi;
112 		unsigned char instr_lo;
113 
114 		if (probe_kernel_address(instr, opcode))
115 			break;
116 
117 		instr_hi = opcode & 0xf0;
118 		instr_lo = opcode & 0x0f;
119 		instr++;
120 
121 		switch (instr_hi) {
122 		case 0x20:
123 		case 0x30:
124 			/*
125 			 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
126 			 * In X86_64 long mode, the CPU will signal invalid
127 			 * opcode if some of these prefixes are present so
128 			 * X86_64 will never get here anyway
129 			 */
130 			scan_more = ((instr_lo & 7) == 0x6);
131 			break;
132 #ifdef CONFIG_X86_64
133 		case 0x40:
134 			/*
135 			 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
136 			 * Need to figure out under what instruction mode the
137 			 * instruction was issued. Could check the LDT for lm,
138 			 * but for now it's good enough to assume that long
139 			 * mode only uses well known segments or kernel.
140 			 */
141 			scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
142 			break;
143 #endif
144 		case 0x60:
145 			/* 0x64 thru 0x67 are valid prefixes in all modes. */
146 			scan_more = (instr_lo & 0xC) == 0x4;
147 			break;
148 		case 0xF0:
149 			/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
150 			scan_more = !instr_lo || (instr_lo>>1) == 1;
151 			break;
152 		case 0x00:
153 			/* Prefetch instruction is 0x0F0D or 0x0F18 */
154 			scan_more = 0;
155 
156 			if (probe_kernel_address(instr, opcode))
157 				break;
158 			prefetch = (instr_lo == 0xF) &&
159 				(opcode == 0x0D || opcode == 0x18);
160 			break;
161 		default:
162 			scan_more = 0;
163 			break;
164 		}
165 	}
166 	return prefetch;
167 }
168 
169 static void force_sig_info_fault(int si_signo, int si_code,
170 	unsigned long address, struct task_struct *tsk)
171 {
172 	siginfo_t info;
173 
174 	info.si_signo = si_signo;
175 	info.si_errno = 0;
176 	info.si_code = si_code;
177 	info.si_addr = (void __user *)address;
178 	force_sig_info(si_signo, &info, tsk);
179 }
180 
181 #ifdef CONFIG_X86_64
182 static int bad_address(void *p)
183 {
184 	unsigned long dummy;
185 	return probe_kernel_address((unsigned long *)p, dummy);
186 }
187 #endif
188 
189 void dump_pagetable(unsigned long address)
190 {
191 #ifdef CONFIG_X86_32
192 	__typeof__(pte_val(__pte(0))) page;
193 
194 	page = read_cr3();
195 	page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
196 #ifdef CONFIG_X86_PAE
197 	printk("*pdpt = %016Lx ", page);
198 	if ((page >> PAGE_SHIFT) < max_low_pfn
199 	    && page & _PAGE_PRESENT) {
200 		page &= PAGE_MASK;
201 		page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
202 		                                         & (PTRS_PER_PMD - 1)];
203 		printk(KERN_CONT "*pde = %016Lx ", page);
204 		page &= ~_PAGE_NX;
205 	}
206 #else
207 	printk("*pde = %08lx ", page);
208 #endif
209 
210 	/*
211 	 * We must not directly access the pte in the highpte
212 	 * case if the page table is located in highmem.
213 	 * And let's rather not kmap-atomic the pte, just in case
214 	 * it's allocated already.
215 	 */
216 	if ((page >> PAGE_SHIFT) < max_low_pfn
217 	    && (page & _PAGE_PRESENT)
218 	    && !(page & _PAGE_PSE)) {
219 		page &= PAGE_MASK;
220 		page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
221 		                                         & (PTRS_PER_PTE - 1)];
222 		printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
223 	}
224 
225 	printk("\n");
226 #else /* CONFIG_X86_64 */
227 	pgd_t *pgd;
228 	pud_t *pud;
229 	pmd_t *pmd;
230 	pte_t *pte;
231 
232 	pgd = (pgd_t *)read_cr3();
233 
234 	pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
235 	pgd += pgd_index(address);
236 	if (bad_address(pgd)) goto bad;
237 	printk("PGD %lx ", pgd_val(*pgd));
238 	if (!pgd_present(*pgd)) goto ret;
239 
240 	pud = pud_offset(pgd, address);
241 	if (bad_address(pud)) goto bad;
242 	printk("PUD %lx ", pud_val(*pud));
243 	if (!pud_present(*pud))	goto ret;
244 
245 	pmd = pmd_offset(pud, address);
246 	if (bad_address(pmd)) goto bad;
247 	printk("PMD %lx ", pmd_val(*pmd));
248 	if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
249 
250 	pte = pte_offset_kernel(pmd, address);
251 	if (bad_address(pte)) goto bad;
252 	printk("PTE %lx", pte_val(*pte));
253 ret:
254 	printk("\n");
255 	return;
256 bad:
257 	printk("BAD\n");
258 #endif
259 }
260 
261 #ifdef CONFIG_X86_32
262 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
263 {
264 	unsigned index = pgd_index(address);
265 	pgd_t *pgd_k;
266 	pud_t *pud, *pud_k;
267 	pmd_t *pmd, *pmd_k;
268 
269 	pgd += index;
270 	pgd_k = init_mm.pgd + index;
271 
272 	if (!pgd_present(*pgd_k))
273 		return NULL;
274 
275 	/*
276 	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
277 	 * and redundant with the set_pmd() on non-PAE. As would
278 	 * set_pud.
279 	 */
280 
281 	pud = pud_offset(pgd, address);
282 	pud_k = pud_offset(pgd_k, address);
283 	if (!pud_present(*pud_k))
284 		return NULL;
285 
286 	pmd = pmd_offset(pud, address);
287 	pmd_k = pmd_offset(pud_k, address);
288 	if (!pmd_present(*pmd_k))
289 		return NULL;
290 	if (!pmd_present(*pmd)) {
291 		set_pmd(pmd, *pmd_k);
292 		arch_flush_lazy_mmu_mode();
293 	} else
294 		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
295 	return pmd_k;
296 }
297 #endif
298 
299 #ifdef CONFIG_X86_64
300 static const char errata93_warning[] =
301 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
302 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
303 KERN_ERR "******* Please consider a BIOS update.\n"
304 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
305 #endif
306 
307 /* Workaround for K8 erratum #93 & buggy BIOS.
308    BIOS SMM functions are required to use a specific workaround
309    to avoid corruption of the 64bit RIP register on C stepping K8.
310    A lot of BIOS that didn't get tested properly miss this.
311    The OS sees this as a page fault with the upper 32bits of RIP cleared.
312    Try to work around it here.
313    Note we only handle faults in kernel here.
314    Does nothing for X86_32
315  */
316 static int is_errata93(struct pt_regs *regs, unsigned long address)
317 {
318 #ifdef CONFIG_X86_64
319 	static int warned;
320 	if (address != regs->ip)
321 		return 0;
322 	if ((address >> 32) != 0)
323 		return 0;
324 	address |= 0xffffffffUL << 32;
325 	if ((address >= (u64)_stext && address <= (u64)_etext) ||
326 	    (address >= MODULES_VADDR && address <= MODULES_END)) {
327 		if (!warned) {
328 			printk(errata93_warning);
329 			warned = 1;
330 		}
331 		regs->ip = address;
332 		return 1;
333 	}
334 #endif
335 	return 0;
336 }
337 
338 /*
339  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
340  * addresses >4GB.  We catch this in the page fault handler because these
341  * addresses are not reachable. Just detect this case and return.  Any code
342  * segment in LDT is compatibility mode.
343  */
344 static int is_errata100(struct pt_regs *regs, unsigned long address)
345 {
346 #ifdef CONFIG_X86_64
347 	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
348 	    (address >> 32))
349 		return 1;
350 #endif
351 	return 0;
352 }
353 
354 void do_invalid_op(struct pt_regs *, unsigned long);
355 
356 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
357 {
358 #ifdef CONFIG_X86_F00F_BUG
359 	unsigned long nr;
360 	/*
361 	 * Pentium F0 0F C7 C8 bug workaround.
362 	 */
363 	if (boot_cpu_data.f00f_bug) {
364 		nr = (address - idt_descr.address) >> 3;
365 
366 		if (nr == 6) {
367 			do_invalid_op(regs, 0);
368 			return 1;
369 		}
370 	}
371 #endif
372 	return 0;
373 }
374 
375 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
376 			    unsigned long address)
377 {
378 #ifdef CONFIG_X86_32
379 	if (!oops_may_print())
380 		return;
381 #endif
382 
383 #ifdef CONFIG_X86_PAE
384 	if (error_code & PF_INSTR) {
385 		unsigned int level;
386 		pte_t *pte = lookup_address(address, &level);
387 
388 		if (pte && pte_present(*pte) && !pte_exec(*pte))
389 			printk(KERN_CRIT "kernel tried to execute "
390 				"NX-protected page - exploit attempt? "
391 				"(uid: %d)\n", current->uid);
392 	}
393 #endif
394 
395 	printk(KERN_ALERT "BUG: unable to handle kernel ");
396 	if (address < PAGE_SIZE)
397 		printk(KERN_CONT "NULL pointer dereference");
398 	else
399 		printk(KERN_CONT "paging request");
400 #ifdef CONFIG_X86_32
401 	printk(KERN_CONT " at %08lx\n", address);
402 #else
403 	printk(KERN_CONT " at %016lx\n", address);
404 #endif
405 	printk(KERN_ALERT "IP:");
406 	printk_address(regs->ip, 1);
407 	dump_pagetable(address);
408 }
409 
410 #ifdef CONFIG_X86_64
411 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
412 				 unsigned long error_code)
413 {
414 	unsigned long flags = oops_begin();
415 	struct task_struct *tsk;
416 
417 	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
418 	       current->comm, address);
419 	dump_pagetable(address);
420 	tsk = current;
421 	tsk->thread.cr2 = address;
422 	tsk->thread.trap_no = 14;
423 	tsk->thread.error_code = error_code;
424 	if (__die("Bad pagetable", regs, error_code))
425 		regs = NULL;
426 	oops_end(flags, regs, SIGKILL);
427 }
428 #endif
429 
430 /*
431  * Handle a spurious fault caused by a stale TLB entry.  This allows
432  * us to lazily refresh the TLB when increasing the permissions of a
433  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
434  * expensive since that implies doing a full cross-processor TLB
435  * flush, even if no stale TLB entries exist on other processors.
436  * There are no security implications to leaving a stale TLB when
437  * increasing the permissions on a page.
438  */
439 static int spurious_fault(unsigned long address,
440 			  unsigned long error_code)
441 {
442 	pgd_t *pgd;
443 	pud_t *pud;
444 	pmd_t *pmd;
445 	pte_t *pte;
446 
447 	/* Reserved-bit violation or user access to kernel space? */
448 	if (error_code & (PF_USER | PF_RSVD))
449 		return 0;
450 
451 	pgd = init_mm.pgd + pgd_index(address);
452 	if (!pgd_present(*pgd))
453 		return 0;
454 
455 	pud = pud_offset(pgd, address);
456 	if (!pud_present(*pud))
457 		return 0;
458 
459 	pmd = pmd_offset(pud, address);
460 	if (!pmd_present(*pmd))
461 		return 0;
462 
463 	pte = pte_offset_kernel(pmd, address);
464 	if (!pte_present(*pte))
465 		return 0;
466 
467 	if ((error_code & PF_WRITE) && !pte_write(*pte))
468 		return 0;
469 	if ((error_code & PF_INSTR) && !pte_exec(*pte))
470 		return 0;
471 
472 	return 1;
473 }
474 
475 /*
476  * X86_32
477  * Handle a fault on the vmalloc or module mapping area
478  *
479  * X86_64
480  * Handle a fault on the vmalloc area
481  *
482  * This assumes no large pages in there.
483  */
484 static int vmalloc_fault(unsigned long address)
485 {
486 #ifdef CONFIG_X86_32
487 	unsigned long pgd_paddr;
488 	pmd_t *pmd_k;
489 	pte_t *pte_k;
490 	/*
491 	 * Synchronize this task's top level page-table
492 	 * with the 'reference' page table.
493 	 *
494 	 * Do _not_ use "current" here. We might be inside
495 	 * an interrupt in the middle of a task switch..
496 	 */
497 	pgd_paddr = read_cr3();
498 	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
499 	if (!pmd_k)
500 		return -1;
501 	pte_k = pte_offset_kernel(pmd_k, address);
502 	if (!pte_present(*pte_k))
503 		return -1;
504 	return 0;
505 #else
506 	pgd_t *pgd, *pgd_ref;
507 	pud_t *pud, *pud_ref;
508 	pmd_t *pmd, *pmd_ref;
509 	pte_t *pte, *pte_ref;
510 
511 	/* Copy kernel mappings over when needed. This can also
512 	   happen within a race in page table update. In the later
513 	   case just flush. */
514 
515 	pgd = pgd_offset(current->mm ?: &init_mm, address);
516 	pgd_ref = pgd_offset_k(address);
517 	if (pgd_none(*pgd_ref))
518 		return -1;
519 	if (pgd_none(*pgd))
520 		set_pgd(pgd, *pgd_ref);
521 	else
522 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
523 
524 	/* Below here mismatches are bugs because these lower tables
525 	   are shared */
526 
527 	pud = pud_offset(pgd, address);
528 	pud_ref = pud_offset(pgd_ref, address);
529 	if (pud_none(*pud_ref))
530 		return -1;
531 	if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
532 		BUG();
533 	pmd = pmd_offset(pud, address);
534 	pmd_ref = pmd_offset(pud_ref, address);
535 	if (pmd_none(*pmd_ref))
536 		return -1;
537 	if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
538 		BUG();
539 	pte_ref = pte_offset_kernel(pmd_ref, address);
540 	if (!pte_present(*pte_ref))
541 		return -1;
542 	pte = pte_offset_kernel(pmd, address);
543 	/* Don't use pte_page here, because the mappings can point
544 	   outside mem_map, and the NUMA hash lookup cannot handle
545 	   that. */
546 	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
547 		BUG();
548 	return 0;
549 #endif
550 }
551 
552 int show_unhandled_signals = 1;
553 
554 /*
555  * This routine handles page faults.  It determines the address,
556  * and the problem, and then passes it off to one of the appropriate
557  * routines.
558  */
559 #ifdef CONFIG_X86_64
560 asmlinkage
561 #endif
562 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
563 {
564 	struct task_struct *tsk;
565 	struct mm_struct *mm;
566 	struct vm_area_struct *vma;
567 	unsigned long address;
568 	int write, si_code;
569 	int fault;
570 #ifdef CONFIG_X86_64
571 	unsigned long flags;
572 #endif
573 
574 	/*
575 	 * We can fault from pretty much anywhere, with unknown IRQ state.
576 	 */
577 	trace_hardirqs_fixup();
578 
579 	tsk = current;
580 	mm = tsk->mm;
581 	prefetchw(&mm->mmap_sem);
582 
583 	/* get the address */
584 	address = read_cr2();
585 
586 	si_code = SEGV_MAPERR;
587 
588 	if (notify_page_fault(regs))
589 		return;
590 
591 	/*
592 	 * We fault-in kernel-space virtual memory on-demand. The
593 	 * 'reference' page table is init_mm.pgd.
594 	 *
595 	 * NOTE! We MUST NOT take any locks for this case. We may
596 	 * be in an interrupt or a critical region, and should
597 	 * only copy the information from the master page table,
598 	 * nothing more.
599 	 *
600 	 * This verifies that the fault happens in kernel space
601 	 * (error_code & 4) == 0, and that the fault was not a
602 	 * protection error (error_code & 9) == 0.
603 	 */
604 #ifdef CONFIG_X86_32
605 	if (unlikely(address >= TASK_SIZE)) {
606 		if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
607 		    vmalloc_fault(address) >= 0)
608 			return;
609 
610 		/* Can handle a stale RO->RW TLB */
611 		if (spurious_fault(address, error_code))
612 			return;
613 
614 		/*
615 		 * Don't take the mm semaphore here. If we fixup a prefetch
616 		 * fault we could otherwise deadlock.
617 		 */
618 		goto bad_area_nosemaphore;
619 	}
620 
621 	/* It's safe to allow irq's after cr2 has been saved and the vmalloc
622 	   fault has been handled. */
623 	if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
624 		local_irq_enable();
625 
626 	/*
627 	 * If we're in an interrupt, have no user context or are running in an
628 	 * atomic region then we must not take the fault.
629 	 */
630 	if (in_atomic() || !mm)
631 		goto bad_area_nosemaphore;
632 #else /* CONFIG_X86_64 */
633 	if (unlikely(address >= TASK_SIZE64)) {
634 		/*
635 		 * Don't check for the module range here: its PML4
636 		 * is always initialized because it's shared with the main
637 		 * kernel text. Only vmalloc may need PML4 syncups.
638 		 */
639 		if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
640 		      ((address >= VMALLOC_START && address < VMALLOC_END))) {
641 			if (vmalloc_fault(address) >= 0)
642 				return;
643 		}
644 
645 		/* Can handle a stale RO->RW TLB */
646 		if (spurious_fault(address, error_code))
647 			return;
648 
649 		/*
650 		 * Don't take the mm semaphore here. If we fixup a prefetch
651 		 * fault we could otherwise deadlock.
652 		 */
653 		goto bad_area_nosemaphore;
654 	}
655 	if (likely(regs->flags & X86_EFLAGS_IF))
656 		local_irq_enable();
657 
658 	if (unlikely(error_code & PF_RSVD))
659 		pgtable_bad(address, regs, error_code);
660 
661 	/*
662 	 * If we're in an interrupt, have no user context or are running in an
663 	 * atomic region then we must not take the fault.
664 	 */
665 	if (unlikely(in_atomic() || !mm))
666 		goto bad_area_nosemaphore;
667 
668 	/*
669 	 * User-mode registers count as a user access even for any
670 	 * potential system fault or CPU buglet.
671 	 */
672 	if (user_mode_vm(regs))
673 		error_code |= PF_USER;
674 again:
675 #endif
676 	/* When running in the kernel we expect faults to occur only to
677 	 * addresses in user space.  All other faults represent errors in the
678 	 * kernel and should generate an OOPS.  Unfortunately, in the case of an
679 	 * erroneous fault occurring in a code path which already holds mmap_sem
680 	 * we will deadlock attempting to validate the fault against the
681 	 * address space.  Luckily the kernel only validly references user
682 	 * space from well defined areas of code, which are listed in the
683 	 * exceptions table.
684 	 *
685 	 * As the vast majority of faults will be valid we will only perform
686 	 * the source reference check when there is a possibility of a deadlock.
687 	 * Attempt to lock the address space, if we cannot we then validate the
688 	 * source.  If this is invalid we can skip the address space check,
689 	 * thus avoiding the deadlock.
690 	 */
691 	if (!down_read_trylock(&mm->mmap_sem)) {
692 		if ((error_code & PF_USER) == 0 &&
693 		    !search_exception_tables(regs->ip))
694 			goto bad_area_nosemaphore;
695 		down_read(&mm->mmap_sem);
696 	}
697 
698 	vma = find_vma(mm, address);
699 	if (!vma)
700 		goto bad_area;
701 	if (vma->vm_start <= address)
702 		goto good_area;
703 	if (!(vma->vm_flags & VM_GROWSDOWN))
704 		goto bad_area;
705 	if (error_code & PF_USER) {
706 		/*
707 		 * Accessing the stack below %sp is always a bug.
708 		 * The large cushion allows instructions like enter
709 		 * and pusha to work.  ("enter $65535,$31" pushes
710 		 * 32 pointers and then decrements %sp by 65535.)
711 		 */
712 		if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
713 			goto bad_area;
714 	}
715 	if (expand_stack(vma, address))
716 		goto bad_area;
717 /*
718  * Ok, we have a good vm_area for this memory access, so
719  * we can handle it..
720  */
721 good_area:
722 	si_code = SEGV_ACCERR;
723 	write = 0;
724 	switch (error_code & (PF_PROT|PF_WRITE)) {
725 	default:	/* 3: write, present */
726 		/* fall through */
727 	case PF_WRITE:		/* write, not present */
728 		if (!(vma->vm_flags & VM_WRITE))
729 			goto bad_area;
730 		write++;
731 		break;
732 	case PF_PROT:		/* read, present */
733 		goto bad_area;
734 	case 0:			/* read, not present */
735 		if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
736 			goto bad_area;
737 	}
738 
739 #ifdef CONFIG_X86_32
740 survive:
741 #endif
742 	/*
743 	 * If for any reason at all we couldn't handle the fault,
744 	 * make sure we exit gracefully rather than endlessly redo
745 	 * the fault.
746 	 */
747 	fault = handle_mm_fault(mm, vma, address, write);
748 	if (unlikely(fault & VM_FAULT_ERROR)) {
749 		if (fault & VM_FAULT_OOM)
750 			goto out_of_memory;
751 		else if (fault & VM_FAULT_SIGBUS)
752 			goto do_sigbus;
753 		BUG();
754 	}
755 	if (fault & VM_FAULT_MAJOR)
756 		tsk->maj_flt++;
757 	else
758 		tsk->min_flt++;
759 
760 #ifdef CONFIG_X86_32
761 	/*
762 	 * Did it hit the DOS screen memory VA from vm86 mode?
763 	 */
764 	if (v8086_mode(regs)) {
765 		unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
766 		if (bit < 32)
767 			tsk->thread.screen_bitmap |= 1 << bit;
768 	}
769 #endif
770 	up_read(&mm->mmap_sem);
771 	return;
772 
773 /*
774  * Something tried to access memory that isn't in our memory map..
775  * Fix it, but check if it's kernel or user first..
776  */
777 bad_area:
778 	up_read(&mm->mmap_sem);
779 
780 bad_area_nosemaphore:
781 	/* User mode accesses just cause a SIGSEGV */
782 	if (error_code & PF_USER) {
783 		/*
784 		 * It's possible to have interrupts off here.
785 		 */
786 		local_irq_enable();
787 
788 		/*
789 		 * Valid to do another page fault here because this one came
790 		 * from user space.
791 		 */
792 		if (is_prefetch(regs, address, error_code))
793 			return;
794 
795 		if (is_errata100(regs, address))
796 			return;
797 
798 		if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
799 		    printk_ratelimit()) {
800 			printk(
801 #ifdef CONFIG_X86_32
802 			"%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
803 #else
804 			"%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx",
805 #endif
806 			task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
807 			tsk->comm, task_pid_nr(tsk), address, regs->ip,
808 			regs->sp, error_code);
809 			print_vma_addr(" in ", regs->ip);
810 			printk("\n");
811 		}
812 
813 		tsk->thread.cr2 = address;
814 		/* Kernel addresses are always protection faults */
815 		tsk->thread.error_code = error_code | (address >= TASK_SIZE);
816 		tsk->thread.trap_no = 14;
817 		force_sig_info_fault(SIGSEGV, si_code, address, tsk);
818 		return;
819 	}
820 
821 	if (is_f00f_bug(regs, address))
822 		return;
823 
824 no_context:
825 	/* Are we prepared to handle this kernel fault?  */
826 	if (fixup_exception(regs))
827 		return;
828 
829 	/*
830 	 * X86_32
831 	 * Valid to do another page fault here, because if this fault
832 	 * had been triggered by is_prefetch fixup_exception would have
833 	 * handled it.
834 	 *
835 	 * X86_64
836 	 * Hall of shame of CPU/BIOS bugs.
837 	 */
838 	if (is_prefetch(regs, address, error_code))
839 		return;
840 
841 	if (is_errata93(regs, address))
842 		return;
843 
844 /*
845  * Oops. The kernel tried to access some bad page. We'll have to
846  * terminate things with extreme prejudice.
847  */
848 #ifdef CONFIG_X86_32
849 	bust_spinlocks(1);
850 #else
851 	flags = oops_begin();
852 #endif
853 
854 	show_fault_oops(regs, error_code, address);
855 
856 	tsk->thread.cr2 = address;
857 	tsk->thread.trap_no = 14;
858 	tsk->thread.error_code = error_code;
859 
860 #ifdef CONFIG_X86_32
861 	die("Oops", regs, error_code);
862 	bust_spinlocks(0);
863 	do_exit(SIGKILL);
864 #else
865 	if (__die("Oops", regs, error_code))
866 		regs = NULL;
867 	/* Executive summary in case the body of the oops scrolled away */
868 	printk(KERN_EMERG "CR2: %016lx\n", address);
869 	oops_end(flags, regs, SIGKILL);
870 #endif
871 
872 /*
873  * We ran out of memory, or some other thing happened to us that made
874  * us unable to handle the page fault gracefully.
875  */
876 out_of_memory:
877 	up_read(&mm->mmap_sem);
878 	if (is_global_init(tsk)) {
879 		yield();
880 #ifdef CONFIG_X86_32
881 		down_read(&mm->mmap_sem);
882 		goto survive;
883 #else
884 		goto again;
885 #endif
886 	}
887 
888 	printk("VM: killing process %s\n", tsk->comm);
889 	if (error_code & PF_USER)
890 		do_group_exit(SIGKILL);
891 	goto no_context;
892 
893 do_sigbus:
894 	up_read(&mm->mmap_sem);
895 
896 	/* Kernel mode? Handle exceptions or die */
897 	if (!(error_code & PF_USER))
898 		goto no_context;
899 #ifdef CONFIG_X86_32
900 	/* User space => ok to do another page fault */
901 	if (is_prefetch(regs, address, error_code))
902 		return;
903 #endif
904 	tsk->thread.cr2 = address;
905 	tsk->thread.error_code = error_code;
906 	tsk->thread.trap_no = 14;
907 	force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
908 }
909 
910 DEFINE_SPINLOCK(pgd_lock);
911 LIST_HEAD(pgd_list);
912 
913 void vmalloc_sync_all(void)
914 {
915 #ifdef CONFIG_X86_32
916 	/*
917 	 * Note that races in the updates of insync and start aren't
918 	 * problematic: insync can only get set bits added, and updates to
919 	 * start are only improving performance (without affecting correctness
920 	 * if undone).
921 	 */
922 	static DECLARE_BITMAP(insync, PTRS_PER_PGD);
923 	static unsigned long start = TASK_SIZE;
924 	unsigned long address;
925 
926 	if (SHARED_KERNEL_PMD)
927 		return;
928 
929 	BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
930 	for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
931 		if (!test_bit(pgd_index(address), insync)) {
932 			unsigned long flags;
933 			struct page *page;
934 
935 			spin_lock_irqsave(&pgd_lock, flags);
936 			list_for_each_entry(page, &pgd_list, lru) {
937 				if (!vmalloc_sync_one(page_address(page),
938 						      address))
939 					break;
940 			}
941 			spin_unlock_irqrestore(&pgd_lock, flags);
942 			if (!page)
943 				set_bit(pgd_index(address), insync);
944 		}
945 		if (address == start && test_bit(pgd_index(address), insync))
946 			start = address + PGDIR_SIZE;
947 	}
948 #else /* CONFIG_X86_64 */
949 	/*
950 	 * Note that races in the updates of insync and start aren't
951 	 * problematic: insync can only get set bits added, and updates to
952 	 * start are only improving performance (without affecting correctness
953 	 * if undone).
954 	 */
955 	static DECLARE_BITMAP(insync, PTRS_PER_PGD);
956 	static unsigned long start = VMALLOC_START & PGDIR_MASK;
957 	unsigned long address;
958 
959 	for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
960 		if (!test_bit(pgd_index(address), insync)) {
961 			const pgd_t *pgd_ref = pgd_offset_k(address);
962 			struct page *page;
963 
964 			if (pgd_none(*pgd_ref))
965 				continue;
966 			spin_lock(&pgd_lock);
967 			list_for_each_entry(page, &pgd_list, lru) {
968 				pgd_t *pgd;
969 				pgd = (pgd_t *)page_address(page) + pgd_index(address);
970 				if (pgd_none(*pgd))
971 					set_pgd(pgd, *pgd_ref);
972 				else
973 					BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
974 			}
975 			spin_unlock(&pgd_lock);
976 			set_bit(pgd_index(address), insync);
977 		}
978 		if (address == start)
979 			start = address + PGDIR_SIZE;
980 	}
981 	/* Check that there is no need to do the same for the modules area. */
982 	BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
983 	BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
984 				(__START_KERNEL & PGDIR_MASK)));
985 #endif
986 }
987