xref: /openbmc/linux/arch/x86/mm/fault.c (revision 545e4006)
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/mmiotrace.h>
14 #include <linux/mman.h>
15 #include <linux/mm.h>
16 #include <linux/smp.h>
17 #include <linux/interrupt.h>
18 #include <linux/init.h>
19 #include <linux/tty.h>
20 #include <linux/vt_kern.h>		/* For unblank_screen() */
21 #include <linux/compiler.h>
22 #include <linux/highmem.h>
23 #include <linux/bootmem.h>		/* for max_low_pfn */
24 #include <linux/vmalloc.h>
25 #include <linux/module.h>
26 #include <linux/kprobes.h>
27 #include <linux/uaccess.h>
28 #include <linux/kdebug.h>
29 
30 #include <asm/system.h>
31 #include <asm/desc.h>
32 #include <asm/segment.h>
33 #include <asm/pgalloc.h>
34 #include <asm/smp.h>
35 #include <asm/tlbflush.h>
36 #include <asm/proto.h>
37 #include <asm-generic/sections.h>
38 
39 /*
40  * Page fault error code bits
41  *	bit 0 == 0 means no page found, 1 means protection fault
42  *	bit 1 == 0 means read, 1 means write
43  *	bit 2 == 0 means kernel, 1 means user-mode
44  *	bit 3 == 1 means use of reserved bit detected
45  *	bit 4 == 1 means fault was an instruction fetch
46  */
47 #define PF_PROT		(1<<0)
48 #define PF_WRITE	(1<<1)
49 #define PF_USER		(1<<2)
50 #define PF_RSVD		(1<<3)
51 #define PF_INSTR	(1<<4)
52 
53 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
54 {
55 #ifdef CONFIG_MMIOTRACE_HOOKS
56 	if (unlikely(is_kmmio_active()))
57 		if (kmmio_handler(regs, addr) == 1)
58 			return -1;
59 #endif
60 	return 0;
61 }
62 
63 static inline int notify_page_fault(struct pt_regs *regs)
64 {
65 #ifdef CONFIG_KPROBES
66 	int ret = 0;
67 
68 	/* kprobe_running() needs smp_processor_id() */
69 	if (!user_mode_vm(regs)) {
70 		preempt_disable();
71 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
72 			ret = 1;
73 		preempt_enable();
74 	}
75 
76 	return ret;
77 #else
78 	return 0;
79 #endif
80 }
81 
82 /*
83  * X86_32
84  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
85  * Check that here and ignore it.
86  *
87  * X86_64
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  */
93 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
94 		       unsigned long error_code)
95 {
96 	unsigned char *instr;
97 	int scan_more = 1;
98 	int prefetch = 0;
99 	unsigned char *max_instr;
100 
101 	/*
102 	 * If it was a exec (instruction fetch) fault on NX page, then
103 	 * do not ignore the fault:
104 	 */
105 	if (error_code & PF_INSTR)
106 		return 0;
107 
108 	instr = (unsigned char *)convert_ip_to_linear(current, regs);
109 	max_instr = instr + 15;
110 
111 	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
112 		return 0;
113 
114 	while (scan_more && instr < max_instr) {
115 		unsigned char opcode;
116 		unsigned char instr_hi;
117 		unsigned char instr_lo;
118 
119 		if (probe_kernel_address(instr, opcode))
120 			break;
121 
122 		instr_hi = opcode & 0xf0;
123 		instr_lo = opcode & 0x0f;
124 		instr++;
125 
126 		switch (instr_hi) {
127 		case 0x20:
128 		case 0x30:
129 			/*
130 			 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
131 			 * In X86_64 long mode, the CPU will signal invalid
132 			 * opcode if some of these prefixes are present so
133 			 * X86_64 will never get here anyway
134 			 */
135 			scan_more = ((instr_lo & 7) == 0x6);
136 			break;
137 #ifdef CONFIG_X86_64
138 		case 0x40:
139 			/*
140 			 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
141 			 * Need to figure out under what instruction mode the
142 			 * instruction was issued. Could check the LDT for lm,
143 			 * but for now it's good enough to assume that long
144 			 * mode only uses well known segments or kernel.
145 			 */
146 			scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
147 			break;
148 #endif
149 		case 0x60:
150 			/* 0x64 thru 0x67 are valid prefixes in all modes. */
151 			scan_more = (instr_lo & 0xC) == 0x4;
152 			break;
153 		case 0xF0:
154 			/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
155 			scan_more = !instr_lo || (instr_lo>>1) == 1;
156 			break;
157 		case 0x00:
158 			/* Prefetch instruction is 0x0F0D or 0x0F18 */
159 			scan_more = 0;
160 
161 			if (probe_kernel_address(instr, opcode))
162 				break;
163 			prefetch = (instr_lo == 0xF) &&
164 				(opcode == 0x0D || opcode == 0x18);
165 			break;
166 		default:
167 			scan_more = 0;
168 			break;
169 		}
170 	}
171 	return prefetch;
172 }
173 
174 static void force_sig_info_fault(int si_signo, int si_code,
175 	unsigned long address, struct task_struct *tsk)
176 {
177 	siginfo_t info;
178 
179 	info.si_signo = si_signo;
180 	info.si_errno = 0;
181 	info.si_code = si_code;
182 	info.si_addr = (void __user *)address;
183 	force_sig_info(si_signo, &info, tsk);
184 }
185 
186 #ifdef CONFIG_X86_64
187 static int bad_address(void *p)
188 {
189 	unsigned long dummy;
190 	return probe_kernel_address((unsigned long *)p, dummy);
191 }
192 #endif
193 
194 static void dump_pagetable(unsigned long address)
195 {
196 #ifdef CONFIG_X86_32
197 	__typeof__(pte_val(__pte(0))) page;
198 
199 	page = read_cr3();
200 	page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
201 #ifdef CONFIG_X86_PAE
202 	printk("*pdpt = %016Lx ", page);
203 	if ((page >> PAGE_SHIFT) < max_low_pfn
204 	    && page & _PAGE_PRESENT) {
205 		page &= PAGE_MASK;
206 		page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
207 		                                         & (PTRS_PER_PMD - 1)];
208 		printk(KERN_CONT "*pde = %016Lx ", page);
209 		page &= ~_PAGE_NX;
210 	}
211 #else
212 	printk("*pde = %08lx ", page);
213 #endif
214 
215 	/*
216 	 * We must not directly access the pte in the highpte
217 	 * case if the page table is located in highmem.
218 	 * And let's rather not kmap-atomic the pte, just in case
219 	 * it's allocated already.
220 	 */
221 	if ((page >> PAGE_SHIFT) < max_low_pfn
222 	    && (page & _PAGE_PRESENT)
223 	    && !(page & _PAGE_PSE)) {
224 		page &= PAGE_MASK;
225 		page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
226 		                                         & (PTRS_PER_PTE - 1)];
227 		printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
228 	}
229 
230 	printk("\n");
231 #else /* CONFIG_X86_64 */
232 	pgd_t *pgd;
233 	pud_t *pud;
234 	pmd_t *pmd;
235 	pte_t *pte;
236 
237 	pgd = (pgd_t *)read_cr3();
238 
239 	pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
240 	pgd += pgd_index(address);
241 	if (bad_address(pgd)) goto bad;
242 	printk("PGD %lx ", pgd_val(*pgd));
243 	if (!pgd_present(*pgd)) goto ret;
244 
245 	pud = pud_offset(pgd, address);
246 	if (bad_address(pud)) goto bad;
247 	printk("PUD %lx ", pud_val(*pud));
248 	if (!pud_present(*pud) || pud_large(*pud))
249 		goto ret;
250 
251 	pmd = pmd_offset(pud, address);
252 	if (bad_address(pmd)) goto bad;
253 	printk("PMD %lx ", pmd_val(*pmd));
254 	if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
255 
256 	pte = pte_offset_kernel(pmd, address);
257 	if (bad_address(pte)) goto bad;
258 	printk("PTE %lx", pte_val(*pte));
259 ret:
260 	printk("\n");
261 	return;
262 bad:
263 	printk("BAD\n");
264 #endif
265 }
266 
267 #ifdef CONFIG_X86_32
268 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
269 {
270 	unsigned index = pgd_index(address);
271 	pgd_t *pgd_k;
272 	pud_t *pud, *pud_k;
273 	pmd_t *pmd, *pmd_k;
274 
275 	pgd += index;
276 	pgd_k = init_mm.pgd + index;
277 
278 	if (!pgd_present(*pgd_k))
279 		return NULL;
280 
281 	/*
282 	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
283 	 * and redundant with the set_pmd() on non-PAE. As would
284 	 * set_pud.
285 	 */
286 
287 	pud = pud_offset(pgd, address);
288 	pud_k = pud_offset(pgd_k, address);
289 	if (!pud_present(*pud_k))
290 		return NULL;
291 
292 	pmd = pmd_offset(pud, address);
293 	pmd_k = pmd_offset(pud_k, address);
294 	if (!pmd_present(*pmd_k))
295 		return NULL;
296 	if (!pmd_present(*pmd)) {
297 		set_pmd(pmd, *pmd_k);
298 		arch_flush_lazy_mmu_mode();
299 	} else
300 		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
301 	return pmd_k;
302 }
303 #endif
304 
305 #ifdef CONFIG_X86_64
306 static const char errata93_warning[] =
307 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
308 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
309 KERN_ERR "******* Please consider a BIOS update.\n"
310 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
311 #endif
312 
313 /* Workaround for K8 erratum #93 & buggy BIOS.
314    BIOS SMM functions are required to use a specific workaround
315    to avoid corruption of the 64bit RIP register on C stepping K8.
316    A lot of BIOS that didn't get tested properly miss this.
317    The OS sees this as a page fault with the upper 32bits of RIP cleared.
318    Try to work around it here.
319    Note we only handle faults in kernel here.
320    Does nothing for X86_32
321  */
322 static int is_errata93(struct pt_regs *regs, unsigned long address)
323 {
324 #ifdef CONFIG_X86_64
325 	static int warned;
326 	if (address != regs->ip)
327 		return 0;
328 	if ((address >> 32) != 0)
329 		return 0;
330 	address |= 0xffffffffUL << 32;
331 	if ((address >= (u64)_stext && address <= (u64)_etext) ||
332 	    (address >= MODULES_VADDR && address <= MODULES_END)) {
333 		if (!warned) {
334 			printk(errata93_warning);
335 			warned = 1;
336 		}
337 		regs->ip = address;
338 		return 1;
339 	}
340 #endif
341 	return 0;
342 }
343 
344 /*
345  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
346  * addresses >4GB.  We catch this in the page fault handler because these
347  * addresses are not reachable. Just detect this case and return.  Any code
348  * segment in LDT is compatibility mode.
349  */
350 static int is_errata100(struct pt_regs *regs, unsigned long address)
351 {
352 #ifdef CONFIG_X86_64
353 	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
354 	    (address >> 32))
355 		return 1;
356 #endif
357 	return 0;
358 }
359 
360 void do_invalid_op(struct pt_regs *, unsigned long);
361 
362 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
363 {
364 #ifdef CONFIG_X86_F00F_BUG
365 	unsigned long nr;
366 	/*
367 	 * Pentium F0 0F C7 C8 bug workaround.
368 	 */
369 	if (boot_cpu_data.f00f_bug) {
370 		nr = (address - idt_descr.address) >> 3;
371 
372 		if (nr == 6) {
373 			do_invalid_op(regs, 0);
374 			return 1;
375 		}
376 	}
377 #endif
378 	return 0;
379 }
380 
381 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
382 			    unsigned long address)
383 {
384 #ifdef CONFIG_X86_32
385 	if (!oops_may_print())
386 		return;
387 #endif
388 
389 #ifdef CONFIG_X86_PAE
390 	if (error_code & PF_INSTR) {
391 		unsigned int level;
392 		pte_t *pte = lookup_address(address, &level);
393 
394 		if (pte && pte_present(*pte) && !pte_exec(*pte))
395 			printk(KERN_CRIT "kernel tried to execute "
396 				"NX-protected page - exploit attempt? "
397 				"(uid: %d)\n", current->uid);
398 	}
399 #endif
400 
401 	printk(KERN_ALERT "BUG: unable to handle kernel ");
402 	if (address < PAGE_SIZE)
403 		printk(KERN_CONT "NULL pointer dereference");
404 	else
405 		printk(KERN_CONT "paging request");
406 	printk(KERN_CONT " at %p\n", (void *) address);
407 	printk(KERN_ALERT "IP:");
408 	printk_address(regs->ip, 1);
409 	dump_pagetable(address);
410 }
411 
412 #ifdef CONFIG_X86_64
413 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
414 				 unsigned long error_code)
415 {
416 	unsigned long flags = oops_begin();
417 	struct task_struct *tsk;
418 
419 	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
420 	       current->comm, address);
421 	dump_pagetable(address);
422 	tsk = current;
423 	tsk->thread.cr2 = address;
424 	tsk->thread.trap_no = 14;
425 	tsk->thread.error_code = error_code;
426 	if (__die("Bad pagetable", regs, error_code))
427 		regs = NULL;
428 	oops_end(flags, regs, SIGKILL);
429 }
430 #endif
431 
432 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
433 {
434 	if ((error_code & PF_WRITE) && !pte_write(*pte))
435 		return 0;
436 	if ((error_code & PF_INSTR) && !pte_exec(*pte))
437 		return 0;
438 
439 	return 1;
440 }
441 
442 /*
443  * Handle a spurious fault caused by a stale TLB entry.  This allows
444  * us to lazily refresh the TLB when increasing the permissions of a
445  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
446  * expensive since that implies doing a full cross-processor TLB
447  * flush, even if no stale TLB entries exist on other processors.
448  * There are no security implications to leaving a stale TLB when
449  * increasing the permissions on a page.
450  */
451 static int spurious_fault(unsigned long address,
452 			  unsigned long error_code)
453 {
454 	pgd_t *pgd;
455 	pud_t *pud;
456 	pmd_t *pmd;
457 	pte_t *pte;
458 
459 	/* Reserved-bit violation or user access to kernel space? */
460 	if (error_code & (PF_USER | PF_RSVD))
461 		return 0;
462 
463 	pgd = init_mm.pgd + pgd_index(address);
464 	if (!pgd_present(*pgd))
465 		return 0;
466 
467 	pud = pud_offset(pgd, address);
468 	if (!pud_present(*pud))
469 		return 0;
470 
471 	if (pud_large(*pud))
472 		return spurious_fault_check(error_code, (pte_t *) pud);
473 
474 	pmd = pmd_offset(pud, address);
475 	if (!pmd_present(*pmd))
476 		return 0;
477 
478 	if (pmd_large(*pmd))
479 		return spurious_fault_check(error_code, (pte_t *) pmd);
480 
481 	pte = pte_offset_kernel(pmd, address);
482 	if (!pte_present(*pte))
483 		return 0;
484 
485 	return spurious_fault_check(error_code, pte);
486 }
487 
488 /*
489  * X86_32
490  * Handle a fault on the vmalloc or module mapping area
491  *
492  * X86_64
493  * Handle a fault on the vmalloc area
494  *
495  * This assumes no large pages in there.
496  */
497 static int vmalloc_fault(unsigned long address)
498 {
499 #ifdef CONFIG_X86_32
500 	unsigned long pgd_paddr;
501 	pmd_t *pmd_k;
502 	pte_t *pte_k;
503 
504 	/* Make sure we are in vmalloc area */
505 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
506 		return -1;
507 
508 	/*
509 	 * Synchronize this task's top level page-table
510 	 * with the 'reference' page table.
511 	 *
512 	 * Do _not_ use "current" here. We might be inside
513 	 * an interrupt in the middle of a task switch..
514 	 */
515 	pgd_paddr = read_cr3();
516 	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
517 	if (!pmd_k)
518 		return -1;
519 	pte_k = pte_offset_kernel(pmd_k, address);
520 	if (!pte_present(*pte_k))
521 		return -1;
522 	return 0;
523 #else
524 	pgd_t *pgd, *pgd_ref;
525 	pud_t *pud, *pud_ref;
526 	pmd_t *pmd, *pmd_ref;
527 	pte_t *pte, *pte_ref;
528 
529 	/* Make sure we are in vmalloc area */
530 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
531 		return -1;
532 
533 	/* Copy kernel mappings over when needed. This can also
534 	   happen within a race in page table update. In the later
535 	   case just flush. */
536 
537 	pgd = pgd_offset(current->mm ?: &init_mm, address);
538 	pgd_ref = pgd_offset_k(address);
539 	if (pgd_none(*pgd_ref))
540 		return -1;
541 	if (pgd_none(*pgd))
542 		set_pgd(pgd, *pgd_ref);
543 	else
544 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
545 
546 	/* Below here mismatches are bugs because these lower tables
547 	   are shared */
548 
549 	pud = pud_offset(pgd, address);
550 	pud_ref = pud_offset(pgd_ref, address);
551 	if (pud_none(*pud_ref))
552 		return -1;
553 	if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
554 		BUG();
555 	pmd = pmd_offset(pud, address);
556 	pmd_ref = pmd_offset(pud_ref, address);
557 	if (pmd_none(*pmd_ref))
558 		return -1;
559 	if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
560 		BUG();
561 	pte_ref = pte_offset_kernel(pmd_ref, address);
562 	if (!pte_present(*pte_ref))
563 		return -1;
564 	pte = pte_offset_kernel(pmd, address);
565 	/* Don't use pte_page here, because the mappings can point
566 	   outside mem_map, and the NUMA hash lookup cannot handle
567 	   that. */
568 	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
569 		BUG();
570 	return 0;
571 #endif
572 }
573 
574 int show_unhandled_signals = 1;
575 
576 /*
577  * This routine handles page faults.  It determines the address,
578  * and the problem, and then passes it off to one of the appropriate
579  * routines.
580  */
581 #ifdef CONFIG_X86_64
582 asmlinkage
583 #endif
584 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
585 {
586 	struct task_struct *tsk;
587 	struct mm_struct *mm;
588 	struct vm_area_struct *vma;
589 	unsigned long address;
590 	int write, si_code;
591 	int fault;
592 #ifdef CONFIG_X86_64
593 	unsigned long flags;
594 #endif
595 
596 	/*
597 	 * We can fault from pretty much anywhere, with unknown IRQ state.
598 	 */
599 	trace_hardirqs_fixup();
600 
601 	tsk = current;
602 	mm = tsk->mm;
603 	prefetchw(&mm->mmap_sem);
604 
605 	/* get the address */
606 	address = read_cr2();
607 
608 	si_code = SEGV_MAPERR;
609 
610 	if (notify_page_fault(regs))
611 		return;
612 	if (unlikely(kmmio_fault(regs, address)))
613 		return;
614 
615 	/*
616 	 * We fault-in kernel-space virtual memory on-demand. The
617 	 * 'reference' page table is init_mm.pgd.
618 	 *
619 	 * NOTE! We MUST NOT take any locks for this case. We may
620 	 * be in an interrupt or a critical region, and should
621 	 * only copy the information from the master page table,
622 	 * nothing more.
623 	 *
624 	 * This verifies that the fault happens in kernel space
625 	 * (error_code & 4) == 0, and that the fault was not a
626 	 * protection error (error_code & 9) == 0.
627 	 */
628 #ifdef CONFIG_X86_32
629 	if (unlikely(address >= TASK_SIZE)) {
630 #else
631 	if (unlikely(address >= TASK_SIZE64)) {
632 #endif
633 		if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
634 		    vmalloc_fault(address) >= 0)
635 			return;
636 
637 		/* Can handle a stale RO->RW TLB */
638 		if (spurious_fault(address, error_code))
639 			return;
640 
641 		/*
642 		 * Don't take the mm semaphore here. If we fixup a prefetch
643 		 * fault we could otherwise deadlock.
644 		 */
645 		goto bad_area_nosemaphore;
646 	}
647 
648 
649 #ifdef CONFIG_X86_32
650 	/* It's safe to allow irq's after cr2 has been saved and the vmalloc
651 	   fault has been handled. */
652 	if (regs->flags & (X86_EFLAGS_IF | X86_VM_MASK))
653 		local_irq_enable();
654 
655 	/*
656 	 * If we're in an interrupt, have no user context or are running in an
657 	 * atomic region then we must not take the fault.
658 	 */
659 	if (in_atomic() || !mm)
660 		goto bad_area_nosemaphore;
661 #else /* CONFIG_X86_64 */
662 	if (likely(regs->flags & X86_EFLAGS_IF))
663 		local_irq_enable();
664 
665 	if (unlikely(error_code & PF_RSVD))
666 		pgtable_bad(address, regs, error_code);
667 
668 	/*
669 	 * If we're in an interrupt, have no user context or are running in an
670 	 * atomic region then we must not take the fault.
671 	 */
672 	if (unlikely(in_atomic() || !mm))
673 		goto bad_area_nosemaphore;
674 
675 	/*
676 	 * User-mode registers count as a user access even for any
677 	 * potential system fault or CPU buglet.
678 	 */
679 	if (user_mode_vm(regs))
680 		error_code |= PF_USER;
681 again:
682 #endif
683 	/* When running in the kernel we expect faults to occur only to
684 	 * addresses in user space.  All other faults represent errors in the
685 	 * kernel and should generate an OOPS.  Unfortunately, in the case of an
686 	 * erroneous fault occurring in a code path which already holds mmap_sem
687 	 * we will deadlock attempting to validate the fault against the
688 	 * address space.  Luckily the kernel only validly references user
689 	 * space from well defined areas of code, which are listed in the
690 	 * exceptions table.
691 	 *
692 	 * As the vast majority of faults will be valid we will only perform
693 	 * the source reference check when there is a possibility of a deadlock.
694 	 * Attempt to lock the address space, if we cannot we then validate the
695 	 * source.  If this is invalid we can skip the address space check,
696 	 * thus avoiding the deadlock.
697 	 */
698 	if (!down_read_trylock(&mm->mmap_sem)) {
699 		if ((error_code & PF_USER) == 0 &&
700 		    !search_exception_tables(regs->ip))
701 			goto bad_area_nosemaphore;
702 		down_read(&mm->mmap_sem);
703 	}
704 
705 	vma = find_vma(mm, address);
706 	if (!vma)
707 		goto bad_area;
708 	if (vma->vm_start <= address)
709 		goto good_area;
710 	if (!(vma->vm_flags & VM_GROWSDOWN))
711 		goto bad_area;
712 	if (error_code & PF_USER) {
713 		/*
714 		 * Accessing the stack below %sp is always a bug.
715 		 * The large cushion allows instructions like enter
716 		 * and pusha to work.  ("enter $65535,$31" pushes
717 		 * 32 pointers and then decrements %sp by 65535.)
718 		 */
719 		if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
720 			goto bad_area;
721 	}
722 	if (expand_stack(vma, address))
723 		goto bad_area;
724 /*
725  * Ok, we have a good vm_area for this memory access, so
726  * we can handle it..
727  */
728 good_area:
729 	si_code = SEGV_ACCERR;
730 	write = 0;
731 	switch (error_code & (PF_PROT|PF_WRITE)) {
732 	default:	/* 3: write, present */
733 		/* fall through */
734 	case PF_WRITE:		/* write, not present */
735 		if (!(vma->vm_flags & VM_WRITE))
736 			goto bad_area;
737 		write++;
738 		break;
739 	case PF_PROT:		/* read, present */
740 		goto bad_area;
741 	case 0:			/* read, not present */
742 		if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
743 			goto bad_area;
744 	}
745 
746 #ifdef CONFIG_X86_32
747 survive:
748 #endif
749 	/*
750 	 * If for any reason at all we couldn't handle the fault,
751 	 * make sure we exit gracefully rather than endlessly redo
752 	 * the fault.
753 	 */
754 	fault = handle_mm_fault(mm, vma, address, write);
755 	if (unlikely(fault & VM_FAULT_ERROR)) {
756 		if (fault & VM_FAULT_OOM)
757 			goto out_of_memory;
758 		else if (fault & VM_FAULT_SIGBUS)
759 			goto do_sigbus;
760 		BUG();
761 	}
762 	if (fault & VM_FAULT_MAJOR)
763 		tsk->maj_flt++;
764 	else
765 		tsk->min_flt++;
766 
767 #ifdef CONFIG_X86_32
768 	/*
769 	 * Did it hit the DOS screen memory VA from vm86 mode?
770 	 */
771 	if (v8086_mode(regs)) {
772 		unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
773 		if (bit < 32)
774 			tsk->thread.screen_bitmap |= 1 << bit;
775 	}
776 #endif
777 	up_read(&mm->mmap_sem);
778 	return;
779 
780 /*
781  * Something tried to access memory that isn't in our memory map..
782  * Fix it, but check if it's kernel or user first..
783  */
784 bad_area:
785 	up_read(&mm->mmap_sem);
786 
787 bad_area_nosemaphore:
788 	/* User mode accesses just cause a SIGSEGV */
789 	if (error_code & PF_USER) {
790 		/*
791 		 * It's possible to have interrupts off here.
792 		 */
793 		local_irq_enable();
794 
795 		/*
796 		 * Valid to do another page fault here because this one came
797 		 * from user space.
798 		 */
799 		if (is_prefetch(regs, address, error_code))
800 			return;
801 
802 		if (is_errata100(regs, address))
803 			return;
804 
805 		if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
806 		    printk_ratelimit()) {
807 			printk(
808 			"%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
809 			task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
810 			tsk->comm, task_pid_nr(tsk), address,
811 			(void *) regs->ip, (void *) regs->sp, error_code);
812 			print_vma_addr(" in ", regs->ip);
813 			printk("\n");
814 		}
815 
816 		tsk->thread.cr2 = address;
817 		/* Kernel addresses are always protection faults */
818 		tsk->thread.error_code = error_code | (address >= TASK_SIZE);
819 		tsk->thread.trap_no = 14;
820 		force_sig_info_fault(SIGSEGV, si_code, address, tsk);
821 		return;
822 	}
823 
824 	if (is_f00f_bug(regs, address))
825 		return;
826 
827 no_context:
828 	/* Are we prepared to handle this kernel fault?  */
829 	if (fixup_exception(regs))
830 		return;
831 
832 	/*
833 	 * X86_32
834 	 * Valid to do another page fault here, because if this fault
835 	 * had been triggered by is_prefetch fixup_exception would have
836 	 * handled it.
837 	 *
838 	 * X86_64
839 	 * Hall of shame of CPU/BIOS bugs.
840 	 */
841 	if (is_prefetch(regs, address, error_code))
842 		return;
843 
844 	if (is_errata93(regs, address))
845 		return;
846 
847 /*
848  * Oops. The kernel tried to access some bad page. We'll have to
849  * terminate things with extreme prejudice.
850  */
851 #ifdef CONFIG_X86_32
852 	bust_spinlocks(1);
853 #else
854 	flags = oops_begin();
855 #endif
856 
857 	show_fault_oops(regs, error_code, address);
858 
859 	tsk->thread.cr2 = address;
860 	tsk->thread.trap_no = 14;
861 	tsk->thread.error_code = error_code;
862 
863 #ifdef CONFIG_X86_32
864 	die("Oops", regs, error_code);
865 	bust_spinlocks(0);
866 	do_exit(SIGKILL);
867 #else
868 	if (__die("Oops", regs, error_code))
869 		regs = NULL;
870 	/* Executive summary in case the body of the oops scrolled away */
871 	printk(KERN_EMERG "CR2: %016lx\n", address);
872 	oops_end(flags, regs, SIGKILL);
873 #endif
874 
875 /*
876  * We ran out of memory, or some other thing happened to us that made
877  * us unable to handle the page fault gracefully.
878  */
879 out_of_memory:
880 	up_read(&mm->mmap_sem);
881 	if (is_global_init(tsk)) {
882 		yield();
883 #ifdef CONFIG_X86_32
884 		down_read(&mm->mmap_sem);
885 		goto survive;
886 #else
887 		goto again;
888 #endif
889 	}
890 
891 	printk("VM: killing process %s\n", tsk->comm);
892 	if (error_code & PF_USER)
893 		do_group_exit(SIGKILL);
894 	goto no_context;
895 
896 do_sigbus:
897 	up_read(&mm->mmap_sem);
898 
899 	/* Kernel mode? Handle exceptions or die */
900 	if (!(error_code & PF_USER))
901 		goto no_context;
902 #ifdef CONFIG_X86_32
903 	/* User space => ok to do another page fault */
904 	if (is_prefetch(regs, address, error_code))
905 		return;
906 #endif
907 	tsk->thread.cr2 = address;
908 	tsk->thread.error_code = error_code;
909 	tsk->thread.trap_no = 14;
910 	force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
911 }
912 
913 DEFINE_SPINLOCK(pgd_lock);
914 LIST_HEAD(pgd_list);
915 
916 void vmalloc_sync_all(void)
917 {
918 #ifdef CONFIG_X86_32
919 	unsigned long start = VMALLOC_START & PGDIR_MASK;
920 	unsigned long address;
921 
922 	if (SHARED_KERNEL_PMD)
923 		return;
924 
925 	BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
926 	for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
927 		unsigned long flags;
928 		struct page *page;
929 
930 		spin_lock_irqsave(&pgd_lock, flags);
931 		list_for_each_entry(page, &pgd_list, lru) {
932 			if (!vmalloc_sync_one(page_address(page),
933 					      address))
934 				break;
935 		}
936 		spin_unlock_irqrestore(&pgd_lock, flags);
937 	}
938 #else /* CONFIG_X86_64 */
939 	unsigned long start = VMALLOC_START & PGDIR_MASK;
940 	unsigned long address;
941 
942 	for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
943 		const pgd_t *pgd_ref = pgd_offset_k(address);
944 		unsigned long flags;
945 		struct page *page;
946 
947 		if (pgd_none(*pgd_ref))
948 			continue;
949 		spin_lock_irqsave(&pgd_lock, flags);
950 		list_for_each_entry(page, &pgd_list, lru) {
951 			pgd_t *pgd;
952 			pgd = (pgd_t *)page_address(page) + pgd_index(address);
953 			if (pgd_none(*pgd))
954 				set_pgd(pgd, *pgd_ref);
955 			else
956 				BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
957 		}
958 		spin_unlock_irqrestore(&pgd_lock, flags);
959 	}
960 #endif
961 }
962