xref: /openbmc/linux/arch/x86/mm/fault.c (revision 367b8112)
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 #include <asm/traps.h>
39 
40 /*
41  * Page fault error code bits
42  *	bit 0 == 0 means no page found, 1 means protection fault
43  *	bit 1 == 0 means read, 1 means write
44  *	bit 2 == 0 means kernel, 1 means user-mode
45  *	bit 3 == 1 means use of reserved bit detected
46  *	bit 4 == 1 means fault was an instruction fetch
47  */
48 #define PF_PROT		(1<<0)
49 #define PF_WRITE	(1<<1)
50 #define PF_USER		(1<<2)
51 #define PF_RSVD		(1<<3)
52 #define PF_INSTR	(1<<4)
53 
54 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
55 {
56 #ifdef CONFIG_MMIOTRACE_HOOKS
57 	if (unlikely(is_kmmio_active()))
58 		if (kmmio_handler(regs, addr) == 1)
59 			return -1;
60 #endif
61 	return 0;
62 }
63 
64 static inline int notify_page_fault(struct pt_regs *regs)
65 {
66 #ifdef CONFIG_KPROBES
67 	int ret = 0;
68 
69 	/* kprobe_running() needs smp_processor_id() */
70 	if (!user_mode_vm(regs)) {
71 		preempt_disable();
72 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
73 			ret = 1;
74 		preempt_enable();
75 	}
76 
77 	return ret;
78 #else
79 	return 0;
80 #endif
81 }
82 
83 /*
84  * X86_32
85  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
86  * Check that here and ignore it.
87  *
88  * X86_64
89  * Sometimes the CPU reports invalid exceptions on prefetch.
90  * Check that here and ignore it.
91  *
92  * Opcode checker based on code by Richard Brunner
93  */
94 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
95 		       unsigned long error_code)
96 {
97 	unsigned char *instr;
98 	int scan_more = 1;
99 	int prefetch = 0;
100 	unsigned char *max_instr;
101 
102 	/*
103 	 * If it was a exec (instruction fetch) fault on NX page, then
104 	 * do not ignore the fault:
105 	 */
106 	if (error_code & PF_INSTR)
107 		return 0;
108 
109 	instr = (unsigned char *)convert_ip_to_linear(current, regs);
110 	max_instr = instr + 15;
111 
112 	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
113 		return 0;
114 
115 	while (scan_more && instr < max_instr) {
116 		unsigned char opcode;
117 		unsigned char instr_hi;
118 		unsigned char instr_lo;
119 
120 		if (probe_kernel_address(instr, opcode))
121 			break;
122 
123 		instr_hi = opcode & 0xf0;
124 		instr_lo = opcode & 0x0f;
125 		instr++;
126 
127 		switch (instr_hi) {
128 		case 0x20:
129 		case 0x30:
130 			/*
131 			 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
132 			 * In X86_64 long mode, the CPU will signal invalid
133 			 * opcode if some of these prefixes are present so
134 			 * X86_64 will never get here anyway
135 			 */
136 			scan_more = ((instr_lo & 7) == 0x6);
137 			break;
138 #ifdef CONFIG_X86_64
139 		case 0x40:
140 			/*
141 			 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
142 			 * Need to figure out under what instruction mode the
143 			 * instruction was issued. Could check the LDT for lm,
144 			 * but for now it's good enough to assume that long
145 			 * mode only uses well known segments or kernel.
146 			 */
147 			scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
148 			break;
149 #endif
150 		case 0x60:
151 			/* 0x64 thru 0x67 are valid prefixes in all modes. */
152 			scan_more = (instr_lo & 0xC) == 0x4;
153 			break;
154 		case 0xF0:
155 			/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
156 			scan_more = !instr_lo || (instr_lo>>1) == 1;
157 			break;
158 		case 0x00:
159 			/* Prefetch instruction is 0x0F0D or 0x0F18 */
160 			scan_more = 0;
161 
162 			if (probe_kernel_address(instr, opcode))
163 				break;
164 			prefetch = (instr_lo == 0xF) &&
165 				(opcode == 0x0D || opcode == 0x18);
166 			break;
167 		default:
168 			scan_more = 0;
169 			break;
170 		}
171 	}
172 	return prefetch;
173 }
174 
175 static void force_sig_info_fault(int si_signo, int si_code,
176 	unsigned long address, struct task_struct *tsk)
177 {
178 	siginfo_t info;
179 
180 	info.si_signo = si_signo;
181 	info.si_errno = 0;
182 	info.si_code = si_code;
183 	info.si_addr = (void __user *)address;
184 	force_sig_info(si_signo, &info, tsk);
185 }
186 
187 #ifdef CONFIG_X86_64
188 static int bad_address(void *p)
189 {
190 	unsigned long dummy;
191 	return probe_kernel_address((unsigned long *)p, dummy);
192 }
193 #endif
194 
195 static void dump_pagetable(unsigned long address)
196 {
197 #ifdef CONFIG_X86_32
198 	__typeof__(pte_val(__pte(0))) page;
199 
200 	page = read_cr3();
201 	page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
202 #ifdef CONFIG_X86_PAE
203 	printk("*pdpt = %016Lx ", page);
204 	if ((page >> PAGE_SHIFT) < max_low_pfn
205 	    && page & _PAGE_PRESENT) {
206 		page &= PAGE_MASK;
207 		page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
208 		                                         & (PTRS_PER_PMD - 1)];
209 		printk(KERN_CONT "*pde = %016Lx ", page);
210 		page &= ~_PAGE_NX;
211 	}
212 #else
213 	printk("*pde = %08lx ", page);
214 #endif
215 
216 	/*
217 	 * We must not directly access the pte in the highpte
218 	 * case if the page table is located in highmem.
219 	 * And let's rather not kmap-atomic the pte, just in case
220 	 * it's allocated already.
221 	 */
222 	if ((page >> PAGE_SHIFT) < max_low_pfn
223 	    && (page & _PAGE_PRESENT)
224 	    && !(page & _PAGE_PSE)) {
225 		page &= PAGE_MASK;
226 		page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
227 		                                         & (PTRS_PER_PTE - 1)];
228 		printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
229 	}
230 
231 	printk("\n");
232 #else /* CONFIG_X86_64 */
233 	pgd_t *pgd;
234 	pud_t *pud;
235 	pmd_t *pmd;
236 	pte_t *pte;
237 
238 	pgd = (pgd_t *)read_cr3();
239 
240 	pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
241 	pgd += pgd_index(address);
242 	if (bad_address(pgd)) goto bad;
243 	printk("PGD %lx ", pgd_val(*pgd));
244 	if (!pgd_present(*pgd)) goto ret;
245 
246 	pud = pud_offset(pgd, address);
247 	if (bad_address(pud)) goto bad;
248 	printk("PUD %lx ", pud_val(*pud));
249 	if (!pud_present(*pud) || pud_large(*pud))
250 		goto ret;
251 
252 	pmd = pmd_offset(pud, address);
253 	if (bad_address(pmd)) goto bad;
254 	printk("PMD %lx ", pmd_val(*pmd));
255 	if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
256 
257 	pte = pte_offset_kernel(pmd, address);
258 	if (bad_address(pte)) goto bad;
259 	printk("PTE %lx", pte_val(*pte));
260 ret:
261 	printk("\n");
262 	return;
263 bad:
264 	printk("BAD\n");
265 #endif
266 }
267 
268 #ifdef CONFIG_X86_32
269 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
270 {
271 	unsigned index = pgd_index(address);
272 	pgd_t *pgd_k;
273 	pud_t *pud, *pud_k;
274 	pmd_t *pmd, *pmd_k;
275 
276 	pgd += index;
277 	pgd_k = init_mm.pgd + index;
278 
279 	if (!pgd_present(*pgd_k))
280 		return NULL;
281 
282 	/*
283 	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
284 	 * and redundant with the set_pmd() on non-PAE. As would
285 	 * set_pud.
286 	 */
287 
288 	pud = pud_offset(pgd, address);
289 	pud_k = pud_offset(pgd_k, address);
290 	if (!pud_present(*pud_k))
291 		return NULL;
292 
293 	pmd = pmd_offset(pud, address);
294 	pmd_k = pmd_offset(pud_k, address);
295 	if (!pmd_present(*pmd_k))
296 		return NULL;
297 	if (!pmd_present(*pmd)) {
298 		set_pmd(pmd, *pmd_k);
299 		arch_flush_lazy_mmu_mode();
300 	} else
301 		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
302 	return pmd_k;
303 }
304 #endif
305 
306 #ifdef CONFIG_X86_64
307 static const char errata93_warning[] =
308 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
309 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
310 KERN_ERR "******* Please consider a BIOS update.\n"
311 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
312 #endif
313 
314 /* Workaround for K8 erratum #93 & buggy BIOS.
315    BIOS SMM functions are required to use a specific workaround
316    to avoid corruption of the 64bit RIP register on C stepping K8.
317    A lot of BIOS that didn't get tested properly miss this.
318    The OS sees this as a page fault with the upper 32bits of RIP cleared.
319    Try to work around it here.
320    Note we only handle faults in kernel here.
321    Does nothing for X86_32
322  */
323 static int is_errata93(struct pt_regs *regs, unsigned long address)
324 {
325 #ifdef CONFIG_X86_64
326 	static int warned;
327 	if (address != regs->ip)
328 		return 0;
329 	if ((address >> 32) != 0)
330 		return 0;
331 	address |= 0xffffffffUL << 32;
332 	if ((address >= (u64)_stext && address <= (u64)_etext) ||
333 	    (address >= MODULES_VADDR && address <= MODULES_END)) {
334 		if (!warned) {
335 			printk(errata93_warning);
336 			warned = 1;
337 		}
338 		regs->ip = address;
339 		return 1;
340 	}
341 #endif
342 	return 0;
343 }
344 
345 /*
346  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
347  * addresses >4GB.  We catch this in the page fault handler because these
348  * addresses are not reachable. Just detect this case and return.  Any code
349  * segment in LDT is compatibility mode.
350  */
351 static int is_errata100(struct pt_regs *regs, unsigned long address)
352 {
353 #ifdef CONFIG_X86_64
354 	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
355 	    (address >> 32))
356 		return 1;
357 #endif
358 	return 0;
359 }
360 
361 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
362 {
363 #ifdef CONFIG_X86_F00F_BUG
364 	unsigned long nr;
365 	/*
366 	 * Pentium F0 0F C7 C8 bug workaround.
367 	 */
368 	if (boot_cpu_data.f00f_bug) {
369 		nr = (address - idt_descr.address) >> 3;
370 
371 		if (nr == 6) {
372 			do_invalid_op(regs, 0);
373 			return 1;
374 		}
375 	}
376 #endif
377 	return 0;
378 }
379 
380 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
381 			    unsigned long address)
382 {
383 #ifdef CONFIG_X86_32
384 	if (!oops_may_print())
385 		return;
386 #endif
387 
388 #ifdef CONFIG_X86_PAE
389 	if (error_code & PF_INSTR) {
390 		unsigned int level;
391 		pte_t *pte = lookup_address(address, &level);
392 
393 		if (pte && pte_present(*pte) && !pte_exec(*pte))
394 			printk(KERN_CRIT "kernel tried to execute "
395 				"NX-protected page - exploit attempt? "
396 				"(uid: %d)\n", current->uid);
397 	}
398 #endif
399 
400 	printk(KERN_ALERT "BUG: unable to handle kernel ");
401 	if (address < PAGE_SIZE)
402 		printk(KERN_CONT "NULL pointer dereference");
403 	else
404 		printk(KERN_CONT "paging request");
405 	printk(KERN_CONT " at %p\n", (void *) address);
406 	printk(KERN_ALERT "IP:");
407 	printk_address(regs->ip, 1);
408 	dump_pagetable(address);
409 }
410 
411 #ifdef CONFIG_X86_64
412 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
413 				 unsigned long error_code)
414 {
415 	unsigned long flags = oops_begin();
416 	struct task_struct *tsk;
417 
418 	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
419 	       current->comm, address);
420 	dump_pagetable(address);
421 	tsk = current;
422 	tsk->thread.cr2 = address;
423 	tsk->thread.trap_no = 14;
424 	tsk->thread.error_code = error_code;
425 	if (__die("Bad pagetable", regs, error_code))
426 		regs = NULL;
427 	oops_end(flags, regs, SIGKILL);
428 }
429 #endif
430 
431 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
432 {
433 	if ((error_code & PF_WRITE) && !pte_write(*pte))
434 		return 0;
435 	if ((error_code & PF_INSTR) && !pte_exec(*pte))
436 		return 0;
437 
438 	return 1;
439 }
440 
441 /*
442  * Handle a spurious fault caused by a stale TLB entry.  This allows
443  * us to lazily refresh the TLB when increasing the permissions of a
444  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
445  * expensive since that implies doing a full cross-processor TLB
446  * flush, even if no stale TLB entries exist on other processors.
447  * There are no security implications to leaving a stale TLB when
448  * increasing the permissions on a page.
449  */
450 static int spurious_fault(unsigned long address,
451 			  unsigned long error_code)
452 {
453 	pgd_t *pgd;
454 	pud_t *pud;
455 	pmd_t *pmd;
456 	pte_t *pte;
457 
458 	/* Reserved-bit violation or user access to kernel space? */
459 	if (error_code & (PF_USER | PF_RSVD))
460 		return 0;
461 
462 	pgd = init_mm.pgd + pgd_index(address);
463 	if (!pgd_present(*pgd))
464 		return 0;
465 
466 	pud = pud_offset(pgd, address);
467 	if (!pud_present(*pud))
468 		return 0;
469 
470 	if (pud_large(*pud))
471 		return spurious_fault_check(error_code, (pte_t *) pud);
472 
473 	pmd = pmd_offset(pud, address);
474 	if (!pmd_present(*pmd))
475 		return 0;
476 
477 	if (pmd_large(*pmd))
478 		return spurious_fault_check(error_code, (pte_t *) pmd);
479 
480 	pte = pte_offset_kernel(pmd, address);
481 	if (!pte_present(*pte))
482 		return 0;
483 
484 	return spurious_fault_check(error_code, pte);
485 }
486 
487 /*
488  * X86_32
489  * Handle a fault on the vmalloc or module mapping area
490  *
491  * X86_64
492  * Handle a fault on the vmalloc area
493  *
494  * This assumes no large pages in there.
495  */
496 static int vmalloc_fault(unsigned long address)
497 {
498 #ifdef CONFIG_X86_32
499 	unsigned long pgd_paddr;
500 	pmd_t *pmd_k;
501 	pte_t *pte_k;
502 
503 	/* Make sure we are in vmalloc area */
504 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
505 		return -1;
506 
507 	/*
508 	 * Synchronize this task's top level page-table
509 	 * with the 'reference' page table.
510 	 *
511 	 * Do _not_ use "current" here. We might be inside
512 	 * an interrupt in the middle of a task switch..
513 	 */
514 	pgd_paddr = read_cr3();
515 	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
516 	if (!pmd_k)
517 		return -1;
518 	pte_k = pte_offset_kernel(pmd_k, address);
519 	if (!pte_present(*pte_k))
520 		return -1;
521 	return 0;
522 #else
523 	pgd_t *pgd, *pgd_ref;
524 	pud_t *pud, *pud_ref;
525 	pmd_t *pmd, *pmd_ref;
526 	pte_t *pte, *pte_ref;
527 
528 	/* Make sure we are in vmalloc area */
529 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
530 		return -1;
531 
532 	/* Copy kernel mappings over when needed. This can also
533 	   happen within a race in page table update. In the later
534 	   case just flush. */
535 
536 	pgd = pgd_offset(current->mm ?: &init_mm, address);
537 	pgd_ref = pgd_offset_k(address);
538 	if (pgd_none(*pgd_ref))
539 		return -1;
540 	if (pgd_none(*pgd))
541 		set_pgd(pgd, *pgd_ref);
542 	else
543 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
544 
545 	/* Below here mismatches are bugs because these lower tables
546 	   are shared */
547 
548 	pud = pud_offset(pgd, address);
549 	pud_ref = pud_offset(pgd_ref, address);
550 	if (pud_none(*pud_ref))
551 		return -1;
552 	if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
553 		BUG();
554 	pmd = pmd_offset(pud, address);
555 	pmd_ref = pmd_offset(pud_ref, address);
556 	if (pmd_none(*pmd_ref))
557 		return -1;
558 	if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
559 		BUG();
560 	pte_ref = pte_offset_kernel(pmd_ref, address);
561 	if (!pte_present(*pte_ref))
562 		return -1;
563 	pte = pte_offset_kernel(pmd, address);
564 	/* Don't use pte_page here, because the mappings can point
565 	   outside mem_map, and the NUMA hash lookup cannot handle
566 	   that. */
567 	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
568 		BUG();
569 	return 0;
570 #endif
571 }
572 
573 int show_unhandled_signals = 1;
574 
575 /*
576  * This routine handles page faults.  It determines the address,
577  * and the problem, and then passes it off to one of the appropriate
578  * routines.
579  */
580 #ifdef CONFIG_X86_64
581 asmlinkage
582 #endif
583 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
584 {
585 	struct task_struct *tsk;
586 	struct mm_struct *mm;
587 	struct vm_area_struct *vma;
588 	unsigned long address;
589 	int write, si_code;
590 	int fault;
591 #ifdef CONFIG_X86_64
592 	unsigned long flags;
593 #endif
594 
595 	tsk = current;
596 	mm = tsk->mm;
597 	prefetchw(&mm->mmap_sem);
598 
599 	/* get the address */
600 	address = read_cr2();
601 
602 	si_code = SEGV_MAPERR;
603 
604 	if (notify_page_fault(regs))
605 		return;
606 	if (unlikely(kmmio_fault(regs, address)))
607 		return;
608 
609 	/*
610 	 * We fault-in kernel-space virtual memory on-demand. The
611 	 * 'reference' page table is init_mm.pgd.
612 	 *
613 	 * NOTE! We MUST NOT take any locks for this case. We may
614 	 * be in an interrupt or a critical region, and should
615 	 * only copy the information from the master page table,
616 	 * nothing more.
617 	 *
618 	 * This verifies that the fault happens in kernel space
619 	 * (error_code & 4) == 0, and that the fault was not a
620 	 * protection error (error_code & 9) == 0.
621 	 */
622 #ifdef CONFIG_X86_32
623 	if (unlikely(address >= TASK_SIZE)) {
624 #else
625 	if (unlikely(address >= TASK_SIZE64)) {
626 #endif
627 		if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
628 		    vmalloc_fault(address) >= 0)
629 			return;
630 
631 		/* Can handle a stale RO->RW TLB */
632 		if (spurious_fault(address, error_code))
633 			return;
634 
635 		/*
636 		 * Don't take the mm semaphore here. If we fixup a prefetch
637 		 * fault we could otherwise deadlock.
638 		 */
639 		goto bad_area_nosemaphore;
640 	}
641 
642 
643 	/*
644 	 * It's safe to allow irq's after cr2 has been saved and the
645 	 * vmalloc fault has been handled.
646 	 *
647 	 * User-mode registers count as a user access even for any
648 	 * potential system fault or CPU buglet.
649 	 */
650 	if (user_mode_vm(regs)) {
651 		local_irq_enable();
652 		error_code |= PF_USER;
653 	} else if (regs->flags & X86_EFLAGS_IF)
654 		local_irq_enable();
655 
656 #ifdef CONFIG_X86_64
657 	if (unlikely(error_code & PF_RSVD))
658 		pgtable_bad(address, regs, error_code);
659 #endif
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 again:
669 	/*
670 	 * When running in the kernel we expect faults to occur only to
671 	 * addresses in user space.  All other faults represent errors in the
672 	 * kernel and should generate an OOPS.  Unfortunately, in the case of an
673 	 * erroneous fault occurring in a code path which already holds mmap_sem
674 	 * we will deadlock attempting to validate the fault against the
675 	 * address space.  Luckily the kernel only validly references user
676 	 * space from well defined areas of code, which are listed in the
677 	 * exceptions table.
678 	 *
679 	 * As the vast majority of faults will be valid we will only perform
680 	 * the source reference check when there is a possibility of a deadlock.
681 	 * Attempt to lock the address space, if we cannot we then validate the
682 	 * source.  If this is invalid we can skip the address space check,
683 	 * thus avoiding the deadlock.
684 	 */
685 	if (!down_read_trylock(&mm->mmap_sem)) {
686 		if ((error_code & PF_USER) == 0 &&
687 		    !search_exception_tables(regs->ip))
688 			goto bad_area_nosemaphore;
689 		down_read(&mm->mmap_sem);
690 	}
691 
692 	vma = find_vma(mm, address);
693 	if (!vma)
694 		goto bad_area;
695 	if (vma->vm_start <= address)
696 		goto good_area;
697 	if (!(vma->vm_flags & VM_GROWSDOWN))
698 		goto bad_area;
699 	if (error_code & PF_USER) {
700 		/*
701 		 * Accessing the stack below %sp is always a bug.
702 		 * The large cushion allows instructions like enter
703 		 * and pusha to work.  ("enter $65535,$31" pushes
704 		 * 32 pointers and then decrements %sp by 65535.)
705 		 */
706 		if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
707 			goto bad_area;
708 	}
709 	if (expand_stack(vma, address))
710 		goto bad_area;
711 /*
712  * Ok, we have a good vm_area for this memory access, so
713  * we can handle it..
714  */
715 good_area:
716 	si_code = SEGV_ACCERR;
717 	write = 0;
718 	switch (error_code & (PF_PROT|PF_WRITE)) {
719 	default:	/* 3: write, present */
720 		/* fall through */
721 	case PF_WRITE:		/* write, not present */
722 		if (!(vma->vm_flags & VM_WRITE))
723 			goto bad_area;
724 		write++;
725 		break;
726 	case PF_PROT:		/* read, present */
727 		goto bad_area;
728 	case 0:			/* read, not present */
729 		if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
730 			goto bad_area;
731 	}
732 
733 	/*
734 	 * If for any reason at all we couldn't handle the fault,
735 	 * make sure we exit gracefully rather than endlessly redo
736 	 * the fault.
737 	 */
738 	fault = handle_mm_fault(mm, vma, address, write);
739 	if (unlikely(fault & VM_FAULT_ERROR)) {
740 		if (fault & VM_FAULT_OOM)
741 			goto out_of_memory;
742 		else if (fault & VM_FAULT_SIGBUS)
743 			goto do_sigbus;
744 		BUG();
745 	}
746 	if (fault & VM_FAULT_MAJOR)
747 		tsk->maj_flt++;
748 	else
749 		tsk->min_flt++;
750 
751 #ifdef CONFIG_X86_32
752 	/*
753 	 * Did it hit the DOS screen memory VA from vm86 mode?
754 	 */
755 	if (v8086_mode(regs)) {
756 		unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
757 		if (bit < 32)
758 			tsk->thread.screen_bitmap |= 1 << bit;
759 	}
760 #endif
761 	up_read(&mm->mmap_sem);
762 	return;
763 
764 /*
765  * Something tried to access memory that isn't in our memory map..
766  * Fix it, but check if it's kernel or user first..
767  */
768 bad_area:
769 	up_read(&mm->mmap_sem);
770 
771 bad_area_nosemaphore:
772 	/* User mode accesses just cause a SIGSEGV */
773 	if (error_code & PF_USER) {
774 		/*
775 		 * It's possible to have interrupts off here.
776 		 */
777 		local_irq_enable();
778 
779 		/*
780 		 * Valid to do another page fault here because this one came
781 		 * from user space.
782 		 */
783 		if (is_prefetch(regs, address, error_code))
784 			return;
785 
786 		if (is_errata100(regs, address))
787 			return;
788 
789 		if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
790 		    printk_ratelimit()) {
791 			printk(
792 			"%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
793 			task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
794 			tsk->comm, task_pid_nr(tsk), address,
795 			(void *) regs->ip, (void *) regs->sp, error_code);
796 			print_vma_addr(" in ", regs->ip);
797 			printk("\n");
798 		}
799 
800 		tsk->thread.cr2 = address;
801 		/* Kernel addresses are always protection faults */
802 		tsk->thread.error_code = error_code | (address >= TASK_SIZE);
803 		tsk->thread.trap_no = 14;
804 		force_sig_info_fault(SIGSEGV, si_code, address, tsk);
805 		return;
806 	}
807 
808 	if (is_f00f_bug(regs, address))
809 		return;
810 
811 no_context:
812 	/* Are we prepared to handle this kernel fault?  */
813 	if (fixup_exception(regs))
814 		return;
815 
816 	/*
817 	 * X86_32
818 	 * Valid to do another page fault here, because if this fault
819 	 * had been triggered by is_prefetch fixup_exception would have
820 	 * handled it.
821 	 *
822 	 * X86_64
823 	 * Hall of shame of CPU/BIOS bugs.
824 	 */
825 	if (is_prefetch(regs, address, error_code))
826 		return;
827 
828 	if (is_errata93(regs, address))
829 		return;
830 
831 /*
832  * Oops. The kernel tried to access some bad page. We'll have to
833  * terminate things with extreme prejudice.
834  */
835 #ifdef CONFIG_X86_32
836 	bust_spinlocks(1);
837 #else
838 	flags = oops_begin();
839 #endif
840 
841 	show_fault_oops(regs, error_code, address);
842 
843 	tsk->thread.cr2 = address;
844 	tsk->thread.trap_no = 14;
845 	tsk->thread.error_code = error_code;
846 
847 #ifdef CONFIG_X86_32
848 	die("Oops", regs, error_code);
849 	bust_spinlocks(0);
850 	do_exit(SIGKILL);
851 #else
852 	if (__die("Oops", regs, error_code))
853 		regs = NULL;
854 	/* Executive summary in case the body of the oops scrolled away */
855 	printk(KERN_EMERG "CR2: %016lx\n", address);
856 	oops_end(flags, regs, SIGKILL);
857 #endif
858 
859 /*
860  * We ran out of memory, or some other thing happened to us that made
861  * us unable to handle the page fault gracefully.
862  */
863 out_of_memory:
864 	up_read(&mm->mmap_sem);
865 	if (is_global_init(tsk)) {
866 		yield();
867 		/*
868 		 * Re-lookup the vma - in theory the vma tree might
869 		 * have changed:
870 		 */
871 		goto again;
872 	}
873 
874 	printk("VM: killing process %s\n", tsk->comm);
875 	if (error_code & PF_USER)
876 		do_group_exit(SIGKILL);
877 	goto no_context;
878 
879 do_sigbus:
880 	up_read(&mm->mmap_sem);
881 
882 	/* Kernel mode? Handle exceptions or die */
883 	if (!(error_code & PF_USER))
884 		goto no_context;
885 #ifdef CONFIG_X86_32
886 	/* User space => ok to do another page fault */
887 	if (is_prefetch(regs, address, error_code))
888 		return;
889 #endif
890 	tsk->thread.cr2 = address;
891 	tsk->thread.error_code = error_code;
892 	tsk->thread.trap_no = 14;
893 	force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
894 }
895 
896 DEFINE_SPINLOCK(pgd_lock);
897 LIST_HEAD(pgd_list);
898 
899 void vmalloc_sync_all(void)
900 {
901 	unsigned long address;
902 
903 #ifdef CONFIG_X86_32
904 	if (SHARED_KERNEL_PMD)
905 		return;
906 
907 	for (address = VMALLOC_START & PMD_MASK;
908 	     address >= TASK_SIZE && address < FIXADDR_TOP;
909 	     address += PMD_SIZE) {
910 		unsigned long flags;
911 		struct page *page;
912 
913 		spin_lock_irqsave(&pgd_lock, flags);
914 		list_for_each_entry(page, &pgd_list, lru) {
915 			if (!vmalloc_sync_one(page_address(page),
916 					      address))
917 				break;
918 		}
919 		spin_unlock_irqrestore(&pgd_lock, flags);
920 	}
921 #else /* CONFIG_X86_64 */
922 	for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
923 	     address += PGDIR_SIZE) {
924 		const pgd_t *pgd_ref = pgd_offset_k(address);
925 		unsigned long flags;
926 		struct page *page;
927 
928 		if (pgd_none(*pgd_ref))
929 			continue;
930 		spin_lock_irqsave(&pgd_lock, flags);
931 		list_for_each_entry(page, &pgd_list, lru) {
932 			pgd_t *pgd;
933 			pgd = (pgd_t *)page_address(page) + pgd_index(address);
934 			if (pgd_none(*pgd))
935 				set_pgd(pgd, *pgd_ref);
936 			else
937 				BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
938 		}
939 		spin_unlock_irqrestore(&pgd_lock, flags);
940 	}
941 #endif
942 }
943