xref: /openbmc/linux/mm/memory.c (revision abfbd895)
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6 
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11 
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22 
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *		Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30 
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  * 		Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *		(Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40 
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
62 #include <linux/dma-debug.h>
63 #include <linux/debugfs.h>
64 #include <linux/userfaultfd_k.h>
65 
66 #include <asm/io.h>
67 #include <asm/pgalloc.h>
68 #include <asm/uaccess.h>
69 #include <asm/tlb.h>
70 #include <asm/tlbflush.h>
71 #include <asm/pgtable.h>
72 
73 #include "internal.h"
74 
75 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
76 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
77 #endif
78 
79 #ifndef CONFIG_NEED_MULTIPLE_NODES
80 /* use the per-pgdat data instead for discontigmem - mbligh */
81 unsigned long max_mapnr;
82 struct page *mem_map;
83 
84 EXPORT_SYMBOL(max_mapnr);
85 EXPORT_SYMBOL(mem_map);
86 #endif
87 
88 /*
89  * A number of key systems in x86 including ioremap() rely on the assumption
90  * that high_memory defines the upper bound on direct map memory, then end
91  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
92  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
93  * and ZONE_HIGHMEM.
94  */
95 void * high_memory;
96 
97 EXPORT_SYMBOL(high_memory);
98 
99 /*
100  * Randomize the address space (stacks, mmaps, brk, etc.).
101  *
102  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
103  *   as ancient (libc5 based) binaries can segfault. )
104  */
105 int randomize_va_space __read_mostly =
106 #ifdef CONFIG_COMPAT_BRK
107 					1;
108 #else
109 					2;
110 #endif
111 
112 static int __init disable_randmaps(char *s)
113 {
114 	randomize_va_space = 0;
115 	return 1;
116 }
117 __setup("norandmaps", disable_randmaps);
118 
119 unsigned long zero_pfn __read_mostly;
120 unsigned long highest_memmap_pfn __read_mostly;
121 
122 EXPORT_SYMBOL(zero_pfn);
123 
124 /*
125  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
126  */
127 static int __init init_zero_pfn(void)
128 {
129 	zero_pfn = page_to_pfn(ZERO_PAGE(0));
130 	return 0;
131 }
132 core_initcall(init_zero_pfn);
133 
134 
135 #if defined(SPLIT_RSS_COUNTING)
136 
137 void sync_mm_rss(struct mm_struct *mm)
138 {
139 	int i;
140 
141 	for (i = 0; i < NR_MM_COUNTERS; i++) {
142 		if (current->rss_stat.count[i]) {
143 			add_mm_counter(mm, i, current->rss_stat.count[i]);
144 			current->rss_stat.count[i] = 0;
145 		}
146 	}
147 	current->rss_stat.events = 0;
148 }
149 
150 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
151 {
152 	struct task_struct *task = current;
153 
154 	if (likely(task->mm == mm))
155 		task->rss_stat.count[member] += val;
156 	else
157 		add_mm_counter(mm, member, val);
158 }
159 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
160 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
161 
162 /* sync counter once per 64 page faults */
163 #define TASK_RSS_EVENTS_THRESH	(64)
164 static void check_sync_rss_stat(struct task_struct *task)
165 {
166 	if (unlikely(task != current))
167 		return;
168 	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
169 		sync_mm_rss(task->mm);
170 }
171 #else /* SPLIT_RSS_COUNTING */
172 
173 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
174 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
175 
176 static void check_sync_rss_stat(struct task_struct *task)
177 {
178 }
179 
180 #endif /* SPLIT_RSS_COUNTING */
181 
182 #ifdef HAVE_GENERIC_MMU_GATHER
183 
184 static bool tlb_next_batch(struct mmu_gather *tlb)
185 {
186 	struct mmu_gather_batch *batch;
187 
188 	batch = tlb->active;
189 	if (batch->next) {
190 		tlb->active = batch->next;
191 		return true;
192 	}
193 
194 	if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
195 		return false;
196 
197 	batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
198 	if (!batch)
199 		return false;
200 
201 	tlb->batch_count++;
202 	batch->next = NULL;
203 	batch->nr   = 0;
204 	batch->max  = MAX_GATHER_BATCH;
205 
206 	tlb->active->next = batch;
207 	tlb->active = batch;
208 
209 	return true;
210 }
211 
212 /* tlb_gather_mmu
213  *	Called to initialize an (on-stack) mmu_gather structure for page-table
214  *	tear-down from @mm. The @fullmm argument is used when @mm is without
215  *	users and we're going to destroy the full address space (exit/execve).
216  */
217 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
218 {
219 	tlb->mm = mm;
220 
221 	/* Is it from 0 to ~0? */
222 	tlb->fullmm     = !(start | (end+1));
223 	tlb->need_flush_all = 0;
224 	tlb->local.next = NULL;
225 	tlb->local.nr   = 0;
226 	tlb->local.max  = ARRAY_SIZE(tlb->__pages);
227 	tlb->active     = &tlb->local;
228 	tlb->batch_count = 0;
229 
230 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 	tlb->batch = NULL;
232 #endif
233 
234 	__tlb_reset_range(tlb);
235 }
236 
237 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
238 {
239 	if (!tlb->end)
240 		return;
241 
242 	tlb_flush(tlb);
243 	mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
244 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
245 	tlb_table_flush(tlb);
246 #endif
247 	__tlb_reset_range(tlb);
248 }
249 
250 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
251 {
252 	struct mmu_gather_batch *batch;
253 
254 	for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
255 		free_pages_and_swap_cache(batch->pages, batch->nr);
256 		batch->nr = 0;
257 	}
258 	tlb->active = &tlb->local;
259 }
260 
261 void tlb_flush_mmu(struct mmu_gather *tlb)
262 {
263 	tlb_flush_mmu_tlbonly(tlb);
264 	tlb_flush_mmu_free(tlb);
265 }
266 
267 /* tlb_finish_mmu
268  *	Called at the end of the shootdown operation to free up any resources
269  *	that were required.
270  */
271 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
272 {
273 	struct mmu_gather_batch *batch, *next;
274 
275 	tlb_flush_mmu(tlb);
276 
277 	/* keep the page table cache within bounds */
278 	check_pgt_cache();
279 
280 	for (batch = tlb->local.next; batch; batch = next) {
281 		next = batch->next;
282 		free_pages((unsigned long)batch, 0);
283 	}
284 	tlb->local.next = NULL;
285 }
286 
287 /* __tlb_remove_page
288  *	Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
289  *	handling the additional races in SMP caused by other CPUs caching valid
290  *	mappings in their TLBs. Returns the number of free page slots left.
291  *	When out of page slots we must call tlb_flush_mmu().
292  */
293 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
294 {
295 	struct mmu_gather_batch *batch;
296 
297 	VM_BUG_ON(!tlb->end);
298 
299 	batch = tlb->active;
300 	batch->pages[batch->nr++] = page;
301 	if (batch->nr == batch->max) {
302 		if (!tlb_next_batch(tlb))
303 			return 0;
304 		batch = tlb->active;
305 	}
306 	VM_BUG_ON_PAGE(batch->nr > batch->max, page);
307 
308 	return batch->max - batch->nr;
309 }
310 
311 #endif /* HAVE_GENERIC_MMU_GATHER */
312 
313 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
314 
315 /*
316  * See the comment near struct mmu_table_batch.
317  */
318 
319 static void tlb_remove_table_smp_sync(void *arg)
320 {
321 	/* Simply deliver the interrupt */
322 }
323 
324 static void tlb_remove_table_one(void *table)
325 {
326 	/*
327 	 * This isn't an RCU grace period and hence the page-tables cannot be
328 	 * assumed to be actually RCU-freed.
329 	 *
330 	 * It is however sufficient for software page-table walkers that rely on
331 	 * IRQ disabling. See the comment near struct mmu_table_batch.
332 	 */
333 	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
334 	__tlb_remove_table(table);
335 }
336 
337 static void tlb_remove_table_rcu(struct rcu_head *head)
338 {
339 	struct mmu_table_batch *batch;
340 	int i;
341 
342 	batch = container_of(head, struct mmu_table_batch, rcu);
343 
344 	for (i = 0; i < batch->nr; i++)
345 		__tlb_remove_table(batch->tables[i]);
346 
347 	free_page((unsigned long)batch);
348 }
349 
350 void tlb_table_flush(struct mmu_gather *tlb)
351 {
352 	struct mmu_table_batch **batch = &tlb->batch;
353 
354 	if (*batch) {
355 		call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
356 		*batch = NULL;
357 	}
358 }
359 
360 void tlb_remove_table(struct mmu_gather *tlb, void *table)
361 {
362 	struct mmu_table_batch **batch = &tlb->batch;
363 
364 	/*
365 	 * When there's less then two users of this mm there cannot be a
366 	 * concurrent page-table walk.
367 	 */
368 	if (atomic_read(&tlb->mm->mm_users) < 2) {
369 		__tlb_remove_table(table);
370 		return;
371 	}
372 
373 	if (*batch == NULL) {
374 		*batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
375 		if (*batch == NULL) {
376 			tlb_remove_table_one(table);
377 			return;
378 		}
379 		(*batch)->nr = 0;
380 	}
381 	(*batch)->tables[(*batch)->nr++] = table;
382 	if ((*batch)->nr == MAX_TABLE_BATCH)
383 		tlb_table_flush(tlb);
384 }
385 
386 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
387 
388 /*
389  * Note: this doesn't free the actual pages themselves. That
390  * has been handled earlier when unmapping all the memory regions.
391  */
392 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
393 			   unsigned long addr)
394 {
395 	pgtable_t token = pmd_pgtable(*pmd);
396 	pmd_clear(pmd);
397 	pte_free_tlb(tlb, token, addr);
398 	atomic_long_dec(&tlb->mm->nr_ptes);
399 }
400 
401 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
402 				unsigned long addr, unsigned long end,
403 				unsigned long floor, unsigned long ceiling)
404 {
405 	pmd_t *pmd;
406 	unsigned long next;
407 	unsigned long start;
408 
409 	start = addr;
410 	pmd = pmd_offset(pud, addr);
411 	do {
412 		next = pmd_addr_end(addr, end);
413 		if (pmd_none_or_clear_bad(pmd))
414 			continue;
415 		free_pte_range(tlb, pmd, addr);
416 	} while (pmd++, addr = next, addr != end);
417 
418 	start &= PUD_MASK;
419 	if (start < floor)
420 		return;
421 	if (ceiling) {
422 		ceiling &= PUD_MASK;
423 		if (!ceiling)
424 			return;
425 	}
426 	if (end - 1 > ceiling - 1)
427 		return;
428 
429 	pmd = pmd_offset(pud, start);
430 	pud_clear(pud);
431 	pmd_free_tlb(tlb, pmd, start);
432 	mm_dec_nr_pmds(tlb->mm);
433 }
434 
435 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
436 				unsigned long addr, unsigned long end,
437 				unsigned long floor, unsigned long ceiling)
438 {
439 	pud_t *pud;
440 	unsigned long next;
441 	unsigned long start;
442 
443 	start = addr;
444 	pud = pud_offset(pgd, addr);
445 	do {
446 		next = pud_addr_end(addr, end);
447 		if (pud_none_or_clear_bad(pud))
448 			continue;
449 		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
450 	} while (pud++, addr = next, addr != end);
451 
452 	start &= PGDIR_MASK;
453 	if (start < floor)
454 		return;
455 	if (ceiling) {
456 		ceiling &= PGDIR_MASK;
457 		if (!ceiling)
458 			return;
459 	}
460 	if (end - 1 > ceiling - 1)
461 		return;
462 
463 	pud = pud_offset(pgd, start);
464 	pgd_clear(pgd);
465 	pud_free_tlb(tlb, pud, start);
466 }
467 
468 /*
469  * This function frees user-level page tables of a process.
470  */
471 void free_pgd_range(struct mmu_gather *tlb,
472 			unsigned long addr, unsigned long end,
473 			unsigned long floor, unsigned long ceiling)
474 {
475 	pgd_t *pgd;
476 	unsigned long next;
477 
478 	/*
479 	 * The next few lines have given us lots of grief...
480 	 *
481 	 * Why are we testing PMD* at this top level?  Because often
482 	 * there will be no work to do at all, and we'd prefer not to
483 	 * go all the way down to the bottom just to discover that.
484 	 *
485 	 * Why all these "- 1"s?  Because 0 represents both the bottom
486 	 * of the address space and the top of it (using -1 for the
487 	 * top wouldn't help much: the masks would do the wrong thing).
488 	 * The rule is that addr 0 and floor 0 refer to the bottom of
489 	 * the address space, but end 0 and ceiling 0 refer to the top
490 	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
491 	 * that end 0 case should be mythical).
492 	 *
493 	 * Wherever addr is brought up or ceiling brought down, we must
494 	 * be careful to reject "the opposite 0" before it confuses the
495 	 * subsequent tests.  But what about where end is brought down
496 	 * by PMD_SIZE below? no, end can't go down to 0 there.
497 	 *
498 	 * Whereas we round start (addr) and ceiling down, by different
499 	 * masks at different levels, in order to test whether a table
500 	 * now has no other vmas using it, so can be freed, we don't
501 	 * bother to round floor or end up - the tests don't need that.
502 	 */
503 
504 	addr &= PMD_MASK;
505 	if (addr < floor) {
506 		addr += PMD_SIZE;
507 		if (!addr)
508 			return;
509 	}
510 	if (ceiling) {
511 		ceiling &= PMD_MASK;
512 		if (!ceiling)
513 			return;
514 	}
515 	if (end - 1 > ceiling - 1)
516 		end -= PMD_SIZE;
517 	if (addr > end - 1)
518 		return;
519 
520 	pgd = pgd_offset(tlb->mm, addr);
521 	do {
522 		next = pgd_addr_end(addr, end);
523 		if (pgd_none_or_clear_bad(pgd))
524 			continue;
525 		free_pud_range(tlb, pgd, addr, next, floor, ceiling);
526 	} while (pgd++, addr = next, addr != end);
527 }
528 
529 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
530 		unsigned long floor, unsigned long ceiling)
531 {
532 	while (vma) {
533 		struct vm_area_struct *next = vma->vm_next;
534 		unsigned long addr = vma->vm_start;
535 
536 		/*
537 		 * Hide vma from rmap and truncate_pagecache before freeing
538 		 * pgtables
539 		 */
540 		unlink_anon_vmas(vma);
541 		unlink_file_vma(vma);
542 
543 		if (is_vm_hugetlb_page(vma)) {
544 			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
545 				floor, next? next->vm_start: ceiling);
546 		} else {
547 			/*
548 			 * Optimization: gather nearby vmas into one call down
549 			 */
550 			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
551 			       && !is_vm_hugetlb_page(next)) {
552 				vma = next;
553 				next = vma->vm_next;
554 				unlink_anon_vmas(vma);
555 				unlink_file_vma(vma);
556 			}
557 			free_pgd_range(tlb, addr, vma->vm_end,
558 				floor, next? next->vm_start: ceiling);
559 		}
560 		vma = next;
561 	}
562 }
563 
564 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
565 		pmd_t *pmd, unsigned long address)
566 {
567 	spinlock_t *ptl;
568 	pgtable_t new = pte_alloc_one(mm, address);
569 	int wait_split_huge_page;
570 	if (!new)
571 		return -ENOMEM;
572 
573 	/*
574 	 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 	 * visible before the pte is made visible to other CPUs by being
576 	 * put into page tables.
577 	 *
578 	 * The other side of the story is the pointer chasing in the page
579 	 * table walking code (when walking the page table without locking;
580 	 * ie. most of the time). Fortunately, these data accesses consist
581 	 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 	 * being the notable exception) will already guarantee loads are
583 	 * seen in-order. See the alpha page table accessors for the
584 	 * smp_read_barrier_depends() barriers in page table walking code.
585 	 */
586 	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587 
588 	ptl = pmd_lock(mm, pmd);
589 	wait_split_huge_page = 0;
590 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
591 		atomic_long_inc(&mm->nr_ptes);
592 		pmd_populate(mm, pmd, new);
593 		new = NULL;
594 	} else if (unlikely(pmd_trans_splitting(*pmd)))
595 		wait_split_huge_page = 1;
596 	spin_unlock(ptl);
597 	if (new)
598 		pte_free(mm, new);
599 	if (wait_split_huge_page)
600 		wait_split_huge_page(vma->anon_vma, pmd);
601 	return 0;
602 }
603 
604 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
605 {
606 	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
607 	if (!new)
608 		return -ENOMEM;
609 
610 	smp_wmb(); /* See comment in __pte_alloc */
611 
612 	spin_lock(&init_mm.page_table_lock);
613 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
614 		pmd_populate_kernel(&init_mm, pmd, new);
615 		new = NULL;
616 	} else
617 		VM_BUG_ON(pmd_trans_splitting(*pmd));
618 	spin_unlock(&init_mm.page_table_lock);
619 	if (new)
620 		pte_free_kernel(&init_mm, new);
621 	return 0;
622 }
623 
624 static inline void init_rss_vec(int *rss)
625 {
626 	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
627 }
628 
629 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
630 {
631 	int i;
632 
633 	if (current->mm == mm)
634 		sync_mm_rss(mm);
635 	for (i = 0; i < NR_MM_COUNTERS; i++)
636 		if (rss[i])
637 			add_mm_counter(mm, i, rss[i]);
638 }
639 
640 /*
641  * This function is called to print an error when a bad pte
642  * is found. For example, we might have a PFN-mapped pte in
643  * a region that doesn't allow it.
644  *
645  * The calling function must still handle the error.
646  */
647 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
648 			  pte_t pte, struct page *page)
649 {
650 	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
651 	pud_t *pud = pud_offset(pgd, addr);
652 	pmd_t *pmd = pmd_offset(pud, addr);
653 	struct address_space *mapping;
654 	pgoff_t index;
655 	static unsigned long resume;
656 	static unsigned long nr_shown;
657 	static unsigned long nr_unshown;
658 
659 	/*
660 	 * Allow a burst of 60 reports, then keep quiet for that minute;
661 	 * or allow a steady drip of one report per second.
662 	 */
663 	if (nr_shown == 60) {
664 		if (time_before(jiffies, resume)) {
665 			nr_unshown++;
666 			return;
667 		}
668 		if (nr_unshown) {
669 			printk(KERN_ALERT
670 				"BUG: Bad page map: %lu messages suppressed\n",
671 				nr_unshown);
672 			nr_unshown = 0;
673 		}
674 		nr_shown = 0;
675 	}
676 	if (nr_shown++ == 0)
677 		resume = jiffies + 60 * HZ;
678 
679 	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
680 	index = linear_page_index(vma, addr);
681 
682 	printk(KERN_ALERT
683 		"BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
684 		current->comm,
685 		(long long)pte_val(pte), (long long)pmd_val(*pmd));
686 	if (page)
687 		dump_page(page, "bad pte");
688 	printk(KERN_ALERT
689 		"addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
690 		(void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
691 	/*
692 	 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
693 	 */
694 	pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
695 		 vma->vm_file,
696 		 vma->vm_ops ? vma->vm_ops->fault : NULL,
697 		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
698 		 mapping ? mapping->a_ops->readpage : NULL);
699 	dump_stack();
700 	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
701 }
702 
703 /*
704  * vm_normal_page -- This function gets the "struct page" associated with a pte.
705  *
706  * "Special" mappings do not wish to be associated with a "struct page" (either
707  * it doesn't exist, or it exists but they don't want to touch it). In this
708  * case, NULL is returned here. "Normal" mappings do have a struct page.
709  *
710  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
711  * pte bit, in which case this function is trivial. Secondly, an architecture
712  * may not have a spare pte bit, which requires a more complicated scheme,
713  * described below.
714  *
715  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
716  * special mapping (even if there are underlying and valid "struct pages").
717  * COWed pages of a VM_PFNMAP are always normal.
718  *
719  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
720  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
721  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
722  * mapping will always honor the rule
723  *
724  *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
725  *
726  * And for normal mappings this is false.
727  *
728  * This restricts such mappings to be a linear translation from virtual address
729  * to pfn. To get around this restriction, we allow arbitrary mappings so long
730  * as the vma is not a COW mapping; in that case, we know that all ptes are
731  * special (because none can have been COWed).
732  *
733  *
734  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
735  *
736  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
737  * page" backing, however the difference is that _all_ pages with a struct
738  * page (that is, those where pfn_valid is true) are refcounted and considered
739  * normal pages by the VM. The disadvantage is that pages are refcounted
740  * (which can be slower and simply not an option for some PFNMAP users). The
741  * advantage is that we don't have to follow the strict linearity rule of
742  * PFNMAP mappings in order to support COWable mappings.
743  *
744  */
745 #ifdef __HAVE_ARCH_PTE_SPECIAL
746 # define HAVE_PTE_SPECIAL 1
747 #else
748 # define HAVE_PTE_SPECIAL 0
749 #endif
750 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
751 				pte_t pte)
752 {
753 	unsigned long pfn = pte_pfn(pte);
754 
755 	if (HAVE_PTE_SPECIAL) {
756 		if (likely(!pte_special(pte)))
757 			goto check_pfn;
758 		if (vma->vm_ops && vma->vm_ops->find_special_page)
759 			return vma->vm_ops->find_special_page(vma, addr);
760 		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
761 			return NULL;
762 		if (!is_zero_pfn(pfn))
763 			print_bad_pte(vma, addr, pte, NULL);
764 		return NULL;
765 	}
766 
767 	/* !HAVE_PTE_SPECIAL case follows: */
768 
769 	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
770 		if (vma->vm_flags & VM_MIXEDMAP) {
771 			if (!pfn_valid(pfn))
772 				return NULL;
773 			goto out;
774 		} else {
775 			unsigned long off;
776 			off = (addr - vma->vm_start) >> PAGE_SHIFT;
777 			if (pfn == vma->vm_pgoff + off)
778 				return NULL;
779 			if (!is_cow_mapping(vma->vm_flags))
780 				return NULL;
781 		}
782 	}
783 
784 	if (is_zero_pfn(pfn))
785 		return NULL;
786 check_pfn:
787 	if (unlikely(pfn > highest_memmap_pfn)) {
788 		print_bad_pte(vma, addr, pte, NULL);
789 		return NULL;
790 	}
791 
792 	/*
793 	 * NOTE! We still have PageReserved() pages in the page tables.
794 	 * eg. VDSO mappings can cause them to exist.
795 	 */
796 out:
797 	return pfn_to_page(pfn);
798 }
799 
800 /*
801  * copy one vm_area from one task to the other. Assumes the page tables
802  * already present in the new task to be cleared in the whole range
803  * covered by this vma.
804  */
805 
806 static inline unsigned long
807 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
808 		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
809 		unsigned long addr, int *rss)
810 {
811 	unsigned long vm_flags = vma->vm_flags;
812 	pte_t pte = *src_pte;
813 	struct page *page;
814 
815 	/* pte contains position in swap or file, so copy. */
816 	if (unlikely(!pte_present(pte))) {
817 		swp_entry_t entry = pte_to_swp_entry(pte);
818 
819 		if (likely(!non_swap_entry(entry))) {
820 			if (swap_duplicate(entry) < 0)
821 				return entry.val;
822 
823 			/* make sure dst_mm is on swapoff's mmlist. */
824 			if (unlikely(list_empty(&dst_mm->mmlist))) {
825 				spin_lock(&mmlist_lock);
826 				if (list_empty(&dst_mm->mmlist))
827 					list_add(&dst_mm->mmlist,
828 							&src_mm->mmlist);
829 				spin_unlock(&mmlist_lock);
830 			}
831 			rss[MM_SWAPENTS]++;
832 		} else if (is_migration_entry(entry)) {
833 			page = migration_entry_to_page(entry);
834 
835 			if (PageAnon(page))
836 				rss[MM_ANONPAGES]++;
837 			else
838 				rss[MM_FILEPAGES]++;
839 
840 			if (is_write_migration_entry(entry) &&
841 					is_cow_mapping(vm_flags)) {
842 				/*
843 				 * COW mappings require pages in both
844 				 * parent and child to be set to read.
845 				 */
846 				make_migration_entry_read(&entry);
847 				pte = swp_entry_to_pte(entry);
848 				if (pte_swp_soft_dirty(*src_pte))
849 					pte = pte_swp_mksoft_dirty(pte);
850 				set_pte_at(src_mm, addr, src_pte, pte);
851 			}
852 		}
853 		goto out_set_pte;
854 	}
855 
856 	/*
857 	 * If it's a COW mapping, write protect it both
858 	 * in the parent and the child
859 	 */
860 	if (is_cow_mapping(vm_flags)) {
861 		ptep_set_wrprotect(src_mm, addr, src_pte);
862 		pte = pte_wrprotect(pte);
863 	}
864 
865 	/*
866 	 * If it's a shared mapping, mark it clean in
867 	 * the child
868 	 */
869 	if (vm_flags & VM_SHARED)
870 		pte = pte_mkclean(pte);
871 	pte = pte_mkold(pte);
872 
873 	page = vm_normal_page(vma, addr, pte);
874 	if (page) {
875 		get_page(page);
876 		page_dup_rmap(page);
877 		if (PageAnon(page))
878 			rss[MM_ANONPAGES]++;
879 		else
880 			rss[MM_FILEPAGES]++;
881 	}
882 
883 out_set_pte:
884 	set_pte_at(dst_mm, addr, dst_pte, pte);
885 	return 0;
886 }
887 
888 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
889 		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
890 		   unsigned long addr, unsigned long end)
891 {
892 	pte_t *orig_src_pte, *orig_dst_pte;
893 	pte_t *src_pte, *dst_pte;
894 	spinlock_t *src_ptl, *dst_ptl;
895 	int progress = 0;
896 	int rss[NR_MM_COUNTERS];
897 	swp_entry_t entry = (swp_entry_t){0};
898 
899 again:
900 	init_rss_vec(rss);
901 
902 	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
903 	if (!dst_pte)
904 		return -ENOMEM;
905 	src_pte = pte_offset_map(src_pmd, addr);
906 	src_ptl = pte_lockptr(src_mm, src_pmd);
907 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
908 	orig_src_pte = src_pte;
909 	orig_dst_pte = dst_pte;
910 	arch_enter_lazy_mmu_mode();
911 
912 	do {
913 		/*
914 		 * We are holding two locks at this point - either of them
915 		 * could generate latencies in another task on another CPU.
916 		 */
917 		if (progress >= 32) {
918 			progress = 0;
919 			if (need_resched() ||
920 			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
921 				break;
922 		}
923 		if (pte_none(*src_pte)) {
924 			progress++;
925 			continue;
926 		}
927 		entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
928 							vma, addr, rss);
929 		if (entry.val)
930 			break;
931 		progress += 8;
932 	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
933 
934 	arch_leave_lazy_mmu_mode();
935 	spin_unlock(src_ptl);
936 	pte_unmap(orig_src_pte);
937 	add_mm_rss_vec(dst_mm, rss);
938 	pte_unmap_unlock(orig_dst_pte, dst_ptl);
939 	cond_resched();
940 
941 	if (entry.val) {
942 		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
943 			return -ENOMEM;
944 		progress = 0;
945 	}
946 	if (addr != end)
947 		goto again;
948 	return 0;
949 }
950 
951 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
952 		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
953 		unsigned long addr, unsigned long end)
954 {
955 	pmd_t *src_pmd, *dst_pmd;
956 	unsigned long next;
957 
958 	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
959 	if (!dst_pmd)
960 		return -ENOMEM;
961 	src_pmd = pmd_offset(src_pud, addr);
962 	do {
963 		next = pmd_addr_end(addr, end);
964 		if (pmd_trans_huge(*src_pmd)) {
965 			int err;
966 			VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
967 			err = copy_huge_pmd(dst_mm, src_mm,
968 					    dst_pmd, src_pmd, addr, vma);
969 			if (err == -ENOMEM)
970 				return -ENOMEM;
971 			if (!err)
972 				continue;
973 			/* fall through */
974 		}
975 		if (pmd_none_or_clear_bad(src_pmd))
976 			continue;
977 		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
978 						vma, addr, next))
979 			return -ENOMEM;
980 	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
981 	return 0;
982 }
983 
984 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
985 		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
986 		unsigned long addr, unsigned long end)
987 {
988 	pud_t *src_pud, *dst_pud;
989 	unsigned long next;
990 
991 	dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
992 	if (!dst_pud)
993 		return -ENOMEM;
994 	src_pud = pud_offset(src_pgd, addr);
995 	do {
996 		next = pud_addr_end(addr, end);
997 		if (pud_none_or_clear_bad(src_pud))
998 			continue;
999 		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1000 						vma, addr, next))
1001 			return -ENOMEM;
1002 	} while (dst_pud++, src_pud++, addr = next, addr != end);
1003 	return 0;
1004 }
1005 
1006 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1007 		struct vm_area_struct *vma)
1008 {
1009 	pgd_t *src_pgd, *dst_pgd;
1010 	unsigned long next;
1011 	unsigned long addr = vma->vm_start;
1012 	unsigned long end = vma->vm_end;
1013 	unsigned long mmun_start;	/* For mmu_notifiers */
1014 	unsigned long mmun_end;		/* For mmu_notifiers */
1015 	bool is_cow;
1016 	int ret;
1017 
1018 	/*
1019 	 * Don't copy ptes where a page fault will fill them correctly.
1020 	 * Fork becomes much lighter when there are big shared or private
1021 	 * readonly mappings. The tradeoff is that copy_page_range is more
1022 	 * efficient than faulting.
1023 	 */
1024 	if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1025 			!vma->anon_vma)
1026 		return 0;
1027 
1028 	if (is_vm_hugetlb_page(vma))
1029 		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1030 
1031 	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1032 		/*
1033 		 * We do not free on error cases below as remove_vma
1034 		 * gets called on error from higher level routine
1035 		 */
1036 		ret = track_pfn_copy(vma);
1037 		if (ret)
1038 			return ret;
1039 	}
1040 
1041 	/*
1042 	 * We need to invalidate the secondary MMU mappings only when
1043 	 * there could be a permission downgrade on the ptes of the
1044 	 * parent mm. And a permission downgrade will only happen if
1045 	 * is_cow_mapping() returns true.
1046 	 */
1047 	is_cow = is_cow_mapping(vma->vm_flags);
1048 	mmun_start = addr;
1049 	mmun_end   = end;
1050 	if (is_cow)
1051 		mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1052 						    mmun_end);
1053 
1054 	ret = 0;
1055 	dst_pgd = pgd_offset(dst_mm, addr);
1056 	src_pgd = pgd_offset(src_mm, addr);
1057 	do {
1058 		next = pgd_addr_end(addr, end);
1059 		if (pgd_none_or_clear_bad(src_pgd))
1060 			continue;
1061 		if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1062 					    vma, addr, next))) {
1063 			ret = -ENOMEM;
1064 			break;
1065 		}
1066 	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1067 
1068 	if (is_cow)
1069 		mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1070 	return ret;
1071 }
1072 
1073 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1074 				struct vm_area_struct *vma, pmd_t *pmd,
1075 				unsigned long addr, unsigned long end,
1076 				struct zap_details *details)
1077 {
1078 	struct mm_struct *mm = tlb->mm;
1079 	int force_flush = 0;
1080 	int rss[NR_MM_COUNTERS];
1081 	spinlock_t *ptl;
1082 	pte_t *start_pte;
1083 	pte_t *pte;
1084 	swp_entry_t entry;
1085 
1086 again:
1087 	init_rss_vec(rss);
1088 	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1089 	pte = start_pte;
1090 	arch_enter_lazy_mmu_mode();
1091 	do {
1092 		pte_t ptent = *pte;
1093 		if (pte_none(ptent)) {
1094 			continue;
1095 		}
1096 
1097 		if (pte_present(ptent)) {
1098 			struct page *page;
1099 
1100 			page = vm_normal_page(vma, addr, ptent);
1101 			if (unlikely(details) && page) {
1102 				/*
1103 				 * unmap_shared_mapping_pages() wants to
1104 				 * invalidate cache without truncating:
1105 				 * unmap shared but keep private pages.
1106 				 */
1107 				if (details->check_mapping &&
1108 				    details->check_mapping != page->mapping)
1109 					continue;
1110 			}
1111 			ptent = ptep_get_and_clear_full(mm, addr, pte,
1112 							tlb->fullmm);
1113 			tlb_remove_tlb_entry(tlb, pte, addr);
1114 			if (unlikely(!page))
1115 				continue;
1116 			if (PageAnon(page))
1117 				rss[MM_ANONPAGES]--;
1118 			else {
1119 				if (pte_dirty(ptent)) {
1120 					force_flush = 1;
1121 					set_page_dirty(page);
1122 				}
1123 				if (pte_young(ptent) &&
1124 				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1125 					mark_page_accessed(page);
1126 				rss[MM_FILEPAGES]--;
1127 			}
1128 			page_remove_rmap(page);
1129 			if (unlikely(page_mapcount(page) < 0))
1130 				print_bad_pte(vma, addr, ptent, page);
1131 			if (unlikely(!__tlb_remove_page(tlb, page))) {
1132 				force_flush = 1;
1133 				addr += PAGE_SIZE;
1134 				break;
1135 			}
1136 			continue;
1137 		}
1138 		/* If details->check_mapping, we leave swap entries. */
1139 		if (unlikely(details))
1140 			continue;
1141 
1142 		entry = pte_to_swp_entry(ptent);
1143 		if (!non_swap_entry(entry))
1144 			rss[MM_SWAPENTS]--;
1145 		else if (is_migration_entry(entry)) {
1146 			struct page *page;
1147 
1148 			page = migration_entry_to_page(entry);
1149 
1150 			if (PageAnon(page))
1151 				rss[MM_ANONPAGES]--;
1152 			else
1153 				rss[MM_FILEPAGES]--;
1154 		}
1155 		if (unlikely(!free_swap_and_cache(entry)))
1156 			print_bad_pte(vma, addr, ptent, NULL);
1157 		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1158 	} while (pte++, addr += PAGE_SIZE, addr != end);
1159 
1160 	add_mm_rss_vec(mm, rss);
1161 	arch_leave_lazy_mmu_mode();
1162 
1163 	/* Do the actual TLB flush before dropping ptl */
1164 	if (force_flush)
1165 		tlb_flush_mmu_tlbonly(tlb);
1166 	pte_unmap_unlock(start_pte, ptl);
1167 
1168 	/*
1169 	 * If we forced a TLB flush (either due to running out of
1170 	 * batch buffers or because we needed to flush dirty TLB
1171 	 * entries before releasing the ptl), free the batched
1172 	 * memory too. Restart if we didn't do everything.
1173 	 */
1174 	if (force_flush) {
1175 		force_flush = 0;
1176 		tlb_flush_mmu_free(tlb);
1177 
1178 		if (addr != end)
1179 			goto again;
1180 	}
1181 
1182 	return addr;
1183 }
1184 
1185 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1186 				struct vm_area_struct *vma, pud_t *pud,
1187 				unsigned long addr, unsigned long end,
1188 				struct zap_details *details)
1189 {
1190 	pmd_t *pmd;
1191 	unsigned long next;
1192 
1193 	pmd = pmd_offset(pud, addr);
1194 	do {
1195 		next = pmd_addr_end(addr, end);
1196 		if (pmd_trans_huge(*pmd)) {
1197 			if (next - addr != HPAGE_PMD_SIZE) {
1198 #ifdef CONFIG_DEBUG_VM
1199 				if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1200 					pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1201 						__func__, addr, end,
1202 						vma->vm_start,
1203 						vma->vm_end);
1204 					BUG();
1205 				}
1206 #endif
1207 				split_huge_page_pmd(vma, addr, pmd);
1208 			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
1209 				goto next;
1210 			/* fall through */
1211 		}
1212 		/*
1213 		 * Here there can be other concurrent MADV_DONTNEED or
1214 		 * trans huge page faults running, and if the pmd is
1215 		 * none or trans huge it can change under us. This is
1216 		 * because MADV_DONTNEED holds the mmap_sem in read
1217 		 * mode.
1218 		 */
1219 		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1220 			goto next;
1221 		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1222 next:
1223 		cond_resched();
1224 	} while (pmd++, addr = next, addr != end);
1225 
1226 	return addr;
1227 }
1228 
1229 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1230 				struct vm_area_struct *vma, pgd_t *pgd,
1231 				unsigned long addr, unsigned long end,
1232 				struct zap_details *details)
1233 {
1234 	pud_t *pud;
1235 	unsigned long next;
1236 
1237 	pud = pud_offset(pgd, addr);
1238 	do {
1239 		next = pud_addr_end(addr, end);
1240 		if (pud_none_or_clear_bad(pud))
1241 			continue;
1242 		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1243 	} while (pud++, addr = next, addr != end);
1244 
1245 	return addr;
1246 }
1247 
1248 static void unmap_page_range(struct mmu_gather *tlb,
1249 			     struct vm_area_struct *vma,
1250 			     unsigned long addr, unsigned long end,
1251 			     struct zap_details *details)
1252 {
1253 	pgd_t *pgd;
1254 	unsigned long next;
1255 
1256 	if (details && !details->check_mapping)
1257 		details = NULL;
1258 
1259 	BUG_ON(addr >= end);
1260 	tlb_start_vma(tlb, vma);
1261 	pgd = pgd_offset(vma->vm_mm, addr);
1262 	do {
1263 		next = pgd_addr_end(addr, end);
1264 		if (pgd_none_or_clear_bad(pgd))
1265 			continue;
1266 		next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1267 	} while (pgd++, addr = next, addr != end);
1268 	tlb_end_vma(tlb, vma);
1269 }
1270 
1271 
1272 static void unmap_single_vma(struct mmu_gather *tlb,
1273 		struct vm_area_struct *vma, unsigned long start_addr,
1274 		unsigned long end_addr,
1275 		struct zap_details *details)
1276 {
1277 	unsigned long start = max(vma->vm_start, start_addr);
1278 	unsigned long end;
1279 
1280 	if (start >= vma->vm_end)
1281 		return;
1282 	end = min(vma->vm_end, end_addr);
1283 	if (end <= vma->vm_start)
1284 		return;
1285 
1286 	if (vma->vm_file)
1287 		uprobe_munmap(vma, start, end);
1288 
1289 	if (unlikely(vma->vm_flags & VM_PFNMAP))
1290 		untrack_pfn(vma, 0, 0);
1291 
1292 	if (start != end) {
1293 		if (unlikely(is_vm_hugetlb_page(vma))) {
1294 			/*
1295 			 * It is undesirable to test vma->vm_file as it
1296 			 * should be non-null for valid hugetlb area.
1297 			 * However, vm_file will be NULL in the error
1298 			 * cleanup path of mmap_region. When
1299 			 * hugetlbfs ->mmap method fails,
1300 			 * mmap_region() nullifies vma->vm_file
1301 			 * before calling this function to clean up.
1302 			 * Since no pte has actually been setup, it is
1303 			 * safe to do nothing in this case.
1304 			 */
1305 			if (vma->vm_file) {
1306 				i_mmap_lock_write(vma->vm_file->f_mapping);
1307 				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1308 				i_mmap_unlock_write(vma->vm_file->f_mapping);
1309 			}
1310 		} else
1311 			unmap_page_range(tlb, vma, start, end, details);
1312 	}
1313 }
1314 
1315 /**
1316  * unmap_vmas - unmap a range of memory covered by a list of vma's
1317  * @tlb: address of the caller's struct mmu_gather
1318  * @vma: the starting vma
1319  * @start_addr: virtual address at which to start unmapping
1320  * @end_addr: virtual address at which to end unmapping
1321  *
1322  * Unmap all pages in the vma list.
1323  *
1324  * Only addresses between `start' and `end' will be unmapped.
1325  *
1326  * The VMA list must be sorted in ascending virtual address order.
1327  *
1328  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1329  * range after unmap_vmas() returns.  So the only responsibility here is to
1330  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1331  * drops the lock and schedules.
1332  */
1333 void unmap_vmas(struct mmu_gather *tlb,
1334 		struct vm_area_struct *vma, unsigned long start_addr,
1335 		unsigned long end_addr)
1336 {
1337 	struct mm_struct *mm = vma->vm_mm;
1338 
1339 	mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1340 	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1341 		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1342 	mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1343 }
1344 
1345 /**
1346  * zap_page_range - remove user pages in a given range
1347  * @vma: vm_area_struct holding the applicable pages
1348  * @start: starting address of pages to zap
1349  * @size: number of bytes to zap
1350  * @details: details of shared cache invalidation
1351  *
1352  * Caller must protect the VMA list
1353  */
1354 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1355 		unsigned long size, struct zap_details *details)
1356 {
1357 	struct mm_struct *mm = vma->vm_mm;
1358 	struct mmu_gather tlb;
1359 	unsigned long end = start + size;
1360 
1361 	lru_add_drain();
1362 	tlb_gather_mmu(&tlb, mm, start, end);
1363 	update_hiwater_rss(mm);
1364 	mmu_notifier_invalidate_range_start(mm, start, end);
1365 	for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1366 		unmap_single_vma(&tlb, vma, start, end, details);
1367 	mmu_notifier_invalidate_range_end(mm, start, end);
1368 	tlb_finish_mmu(&tlb, start, end);
1369 }
1370 
1371 /**
1372  * zap_page_range_single - remove user pages in a given range
1373  * @vma: vm_area_struct holding the applicable pages
1374  * @address: starting address of pages to zap
1375  * @size: number of bytes to zap
1376  * @details: details of shared cache invalidation
1377  *
1378  * The range must fit into one VMA.
1379  */
1380 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1381 		unsigned long size, struct zap_details *details)
1382 {
1383 	struct mm_struct *mm = vma->vm_mm;
1384 	struct mmu_gather tlb;
1385 	unsigned long end = address + size;
1386 
1387 	lru_add_drain();
1388 	tlb_gather_mmu(&tlb, mm, address, end);
1389 	update_hiwater_rss(mm);
1390 	mmu_notifier_invalidate_range_start(mm, address, end);
1391 	unmap_single_vma(&tlb, vma, address, end, details);
1392 	mmu_notifier_invalidate_range_end(mm, address, end);
1393 	tlb_finish_mmu(&tlb, address, end);
1394 }
1395 
1396 /**
1397  * zap_vma_ptes - remove ptes mapping the vma
1398  * @vma: vm_area_struct holding ptes to be zapped
1399  * @address: starting address of pages to zap
1400  * @size: number of bytes to zap
1401  *
1402  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1403  *
1404  * The entire address range must be fully contained within the vma.
1405  *
1406  * Returns 0 if successful.
1407  */
1408 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1409 		unsigned long size)
1410 {
1411 	if (address < vma->vm_start || address + size > vma->vm_end ||
1412 	    		!(vma->vm_flags & VM_PFNMAP))
1413 		return -1;
1414 	zap_page_range_single(vma, address, size, NULL);
1415 	return 0;
1416 }
1417 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1418 
1419 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1420 			spinlock_t **ptl)
1421 {
1422 	pgd_t * pgd = pgd_offset(mm, addr);
1423 	pud_t * pud = pud_alloc(mm, pgd, addr);
1424 	if (pud) {
1425 		pmd_t * pmd = pmd_alloc(mm, pud, addr);
1426 		if (pmd) {
1427 			VM_BUG_ON(pmd_trans_huge(*pmd));
1428 			return pte_alloc_map_lock(mm, pmd, addr, ptl);
1429 		}
1430 	}
1431 	return NULL;
1432 }
1433 
1434 /*
1435  * This is the old fallback for page remapping.
1436  *
1437  * For historical reasons, it only allows reserved pages. Only
1438  * old drivers should use this, and they needed to mark their
1439  * pages reserved for the old functions anyway.
1440  */
1441 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1442 			struct page *page, pgprot_t prot)
1443 {
1444 	struct mm_struct *mm = vma->vm_mm;
1445 	int retval;
1446 	pte_t *pte;
1447 	spinlock_t *ptl;
1448 
1449 	retval = -EINVAL;
1450 	if (PageAnon(page))
1451 		goto out;
1452 	retval = -ENOMEM;
1453 	flush_dcache_page(page);
1454 	pte = get_locked_pte(mm, addr, &ptl);
1455 	if (!pte)
1456 		goto out;
1457 	retval = -EBUSY;
1458 	if (!pte_none(*pte))
1459 		goto out_unlock;
1460 
1461 	/* Ok, finally just insert the thing.. */
1462 	get_page(page);
1463 	inc_mm_counter_fast(mm, MM_FILEPAGES);
1464 	page_add_file_rmap(page);
1465 	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1466 
1467 	retval = 0;
1468 	pte_unmap_unlock(pte, ptl);
1469 	return retval;
1470 out_unlock:
1471 	pte_unmap_unlock(pte, ptl);
1472 out:
1473 	return retval;
1474 }
1475 
1476 /**
1477  * vm_insert_page - insert single page into user vma
1478  * @vma: user vma to map to
1479  * @addr: target user address of this page
1480  * @page: source kernel page
1481  *
1482  * This allows drivers to insert individual pages they've allocated
1483  * into a user vma.
1484  *
1485  * The page has to be a nice clean _individual_ kernel allocation.
1486  * If you allocate a compound page, you need to have marked it as
1487  * such (__GFP_COMP), or manually just split the page up yourself
1488  * (see split_page()).
1489  *
1490  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1491  * took an arbitrary page protection parameter. This doesn't allow
1492  * that. Your vma protection will have to be set up correctly, which
1493  * means that if you want a shared writable mapping, you'd better
1494  * ask for a shared writable mapping!
1495  *
1496  * The page does not need to be reserved.
1497  *
1498  * Usually this function is called from f_op->mmap() handler
1499  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1500  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1501  * function from other places, for example from page-fault handler.
1502  */
1503 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1504 			struct page *page)
1505 {
1506 	if (addr < vma->vm_start || addr >= vma->vm_end)
1507 		return -EFAULT;
1508 	if (!page_count(page))
1509 		return -EINVAL;
1510 	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1511 		BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1512 		BUG_ON(vma->vm_flags & VM_PFNMAP);
1513 		vma->vm_flags |= VM_MIXEDMAP;
1514 	}
1515 	return insert_page(vma, addr, page, vma->vm_page_prot);
1516 }
1517 EXPORT_SYMBOL(vm_insert_page);
1518 
1519 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1520 			unsigned long pfn, pgprot_t prot)
1521 {
1522 	struct mm_struct *mm = vma->vm_mm;
1523 	int retval;
1524 	pte_t *pte, entry;
1525 	spinlock_t *ptl;
1526 
1527 	retval = -ENOMEM;
1528 	pte = get_locked_pte(mm, addr, &ptl);
1529 	if (!pte)
1530 		goto out;
1531 	retval = -EBUSY;
1532 	if (!pte_none(*pte))
1533 		goto out_unlock;
1534 
1535 	/* Ok, finally just insert the thing.. */
1536 	entry = pte_mkspecial(pfn_pte(pfn, prot));
1537 	set_pte_at(mm, addr, pte, entry);
1538 	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1539 
1540 	retval = 0;
1541 out_unlock:
1542 	pte_unmap_unlock(pte, ptl);
1543 out:
1544 	return retval;
1545 }
1546 
1547 /**
1548  * vm_insert_pfn - insert single pfn into user vma
1549  * @vma: user vma to map to
1550  * @addr: target user address of this page
1551  * @pfn: source kernel pfn
1552  *
1553  * Similar to vm_insert_page, this allows drivers to insert individual pages
1554  * they've allocated into a user vma. Same comments apply.
1555  *
1556  * This function should only be called from a vm_ops->fault handler, and
1557  * in that case the handler should return NULL.
1558  *
1559  * vma cannot be a COW mapping.
1560  *
1561  * As this is called only for pages that do not currently exist, we
1562  * do not need to flush old virtual caches or the TLB.
1563  */
1564 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1565 			unsigned long pfn)
1566 {
1567 	int ret;
1568 	pgprot_t pgprot = vma->vm_page_prot;
1569 	/*
1570 	 * Technically, architectures with pte_special can avoid all these
1571 	 * restrictions (same for remap_pfn_range).  However we would like
1572 	 * consistency in testing and feature parity among all, so we should
1573 	 * try to keep these invariants in place for everybody.
1574 	 */
1575 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1576 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1577 						(VM_PFNMAP|VM_MIXEDMAP));
1578 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1579 	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1580 
1581 	if (addr < vma->vm_start || addr >= vma->vm_end)
1582 		return -EFAULT;
1583 	if (track_pfn_insert(vma, &pgprot, pfn))
1584 		return -EINVAL;
1585 
1586 	ret = insert_pfn(vma, addr, pfn, pgprot);
1587 
1588 	return ret;
1589 }
1590 EXPORT_SYMBOL(vm_insert_pfn);
1591 
1592 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1593 			unsigned long pfn)
1594 {
1595 	BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1596 
1597 	if (addr < vma->vm_start || addr >= vma->vm_end)
1598 		return -EFAULT;
1599 
1600 	/*
1601 	 * If we don't have pte special, then we have to use the pfn_valid()
1602 	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1603 	 * refcount the page if pfn_valid is true (hence insert_page rather
1604 	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1605 	 * without pte special, it would there be refcounted as a normal page.
1606 	 */
1607 	if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1608 		struct page *page;
1609 
1610 		page = pfn_to_page(pfn);
1611 		return insert_page(vma, addr, page, vma->vm_page_prot);
1612 	}
1613 	return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1614 }
1615 EXPORT_SYMBOL(vm_insert_mixed);
1616 
1617 /*
1618  * maps a range of physical memory into the requested pages. the old
1619  * mappings are removed. any references to nonexistent pages results
1620  * in null mappings (currently treated as "copy-on-access")
1621  */
1622 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1623 			unsigned long addr, unsigned long end,
1624 			unsigned long pfn, pgprot_t prot)
1625 {
1626 	pte_t *pte;
1627 	spinlock_t *ptl;
1628 
1629 	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1630 	if (!pte)
1631 		return -ENOMEM;
1632 	arch_enter_lazy_mmu_mode();
1633 	do {
1634 		BUG_ON(!pte_none(*pte));
1635 		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1636 		pfn++;
1637 	} while (pte++, addr += PAGE_SIZE, addr != end);
1638 	arch_leave_lazy_mmu_mode();
1639 	pte_unmap_unlock(pte - 1, ptl);
1640 	return 0;
1641 }
1642 
1643 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1644 			unsigned long addr, unsigned long end,
1645 			unsigned long pfn, pgprot_t prot)
1646 {
1647 	pmd_t *pmd;
1648 	unsigned long next;
1649 
1650 	pfn -= addr >> PAGE_SHIFT;
1651 	pmd = pmd_alloc(mm, pud, addr);
1652 	if (!pmd)
1653 		return -ENOMEM;
1654 	VM_BUG_ON(pmd_trans_huge(*pmd));
1655 	do {
1656 		next = pmd_addr_end(addr, end);
1657 		if (remap_pte_range(mm, pmd, addr, next,
1658 				pfn + (addr >> PAGE_SHIFT), prot))
1659 			return -ENOMEM;
1660 	} while (pmd++, addr = next, addr != end);
1661 	return 0;
1662 }
1663 
1664 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1665 			unsigned long addr, unsigned long end,
1666 			unsigned long pfn, pgprot_t prot)
1667 {
1668 	pud_t *pud;
1669 	unsigned long next;
1670 
1671 	pfn -= addr >> PAGE_SHIFT;
1672 	pud = pud_alloc(mm, pgd, addr);
1673 	if (!pud)
1674 		return -ENOMEM;
1675 	do {
1676 		next = pud_addr_end(addr, end);
1677 		if (remap_pmd_range(mm, pud, addr, next,
1678 				pfn + (addr >> PAGE_SHIFT), prot))
1679 			return -ENOMEM;
1680 	} while (pud++, addr = next, addr != end);
1681 	return 0;
1682 }
1683 
1684 /**
1685  * remap_pfn_range - remap kernel memory to userspace
1686  * @vma: user vma to map to
1687  * @addr: target user address to start at
1688  * @pfn: physical address of kernel memory
1689  * @size: size of map area
1690  * @prot: page protection flags for this mapping
1691  *
1692  *  Note: this is only safe if the mm semaphore is held when called.
1693  */
1694 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1695 		    unsigned long pfn, unsigned long size, pgprot_t prot)
1696 {
1697 	pgd_t *pgd;
1698 	unsigned long next;
1699 	unsigned long end = addr + PAGE_ALIGN(size);
1700 	struct mm_struct *mm = vma->vm_mm;
1701 	int err;
1702 
1703 	/*
1704 	 * Physically remapped pages are special. Tell the
1705 	 * rest of the world about it:
1706 	 *   VM_IO tells people not to look at these pages
1707 	 *	(accesses can have side effects).
1708 	 *   VM_PFNMAP tells the core MM that the base pages are just
1709 	 *	raw PFN mappings, and do not have a "struct page" associated
1710 	 *	with them.
1711 	 *   VM_DONTEXPAND
1712 	 *      Disable vma merging and expanding with mremap().
1713 	 *   VM_DONTDUMP
1714 	 *      Omit vma from core dump, even when VM_IO turned off.
1715 	 *
1716 	 * There's a horrible special case to handle copy-on-write
1717 	 * behaviour that some programs depend on. We mark the "original"
1718 	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1719 	 * See vm_normal_page() for details.
1720 	 */
1721 	if (is_cow_mapping(vma->vm_flags)) {
1722 		if (addr != vma->vm_start || end != vma->vm_end)
1723 			return -EINVAL;
1724 		vma->vm_pgoff = pfn;
1725 	}
1726 
1727 	err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1728 	if (err)
1729 		return -EINVAL;
1730 
1731 	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1732 
1733 	BUG_ON(addr >= end);
1734 	pfn -= addr >> PAGE_SHIFT;
1735 	pgd = pgd_offset(mm, addr);
1736 	flush_cache_range(vma, addr, end);
1737 	do {
1738 		next = pgd_addr_end(addr, end);
1739 		err = remap_pud_range(mm, pgd, addr, next,
1740 				pfn + (addr >> PAGE_SHIFT), prot);
1741 		if (err)
1742 			break;
1743 	} while (pgd++, addr = next, addr != end);
1744 
1745 	if (err)
1746 		untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1747 
1748 	return err;
1749 }
1750 EXPORT_SYMBOL(remap_pfn_range);
1751 
1752 /**
1753  * vm_iomap_memory - remap memory to userspace
1754  * @vma: user vma to map to
1755  * @start: start of area
1756  * @len: size of area
1757  *
1758  * This is a simplified io_remap_pfn_range() for common driver use. The
1759  * driver just needs to give us the physical memory range to be mapped,
1760  * we'll figure out the rest from the vma information.
1761  *
1762  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1763  * whatever write-combining details or similar.
1764  */
1765 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1766 {
1767 	unsigned long vm_len, pfn, pages;
1768 
1769 	/* Check that the physical memory area passed in looks valid */
1770 	if (start + len < start)
1771 		return -EINVAL;
1772 	/*
1773 	 * You *really* shouldn't map things that aren't page-aligned,
1774 	 * but we've historically allowed it because IO memory might
1775 	 * just have smaller alignment.
1776 	 */
1777 	len += start & ~PAGE_MASK;
1778 	pfn = start >> PAGE_SHIFT;
1779 	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1780 	if (pfn + pages < pfn)
1781 		return -EINVAL;
1782 
1783 	/* We start the mapping 'vm_pgoff' pages into the area */
1784 	if (vma->vm_pgoff > pages)
1785 		return -EINVAL;
1786 	pfn += vma->vm_pgoff;
1787 	pages -= vma->vm_pgoff;
1788 
1789 	/* Can we fit all of the mapping? */
1790 	vm_len = vma->vm_end - vma->vm_start;
1791 	if (vm_len >> PAGE_SHIFT > pages)
1792 		return -EINVAL;
1793 
1794 	/* Ok, let it rip */
1795 	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1796 }
1797 EXPORT_SYMBOL(vm_iomap_memory);
1798 
1799 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1800 				     unsigned long addr, unsigned long end,
1801 				     pte_fn_t fn, void *data)
1802 {
1803 	pte_t *pte;
1804 	int err;
1805 	pgtable_t token;
1806 	spinlock_t *uninitialized_var(ptl);
1807 
1808 	pte = (mm == &init_mm) ?
1809 		pte_alloc_kernel(pmd, addr) :
1810 		pte_alloc_map_lock(mm, pmd, addr, &ptl);
1811 	if (!pte)
1812 		return -ENOMEM;
1813 
1814 	BUG_ON(pmd_huge(*pmd));
1815 
1816 	arch_enter_lazy_mmu_mode();
1817 
1818 	token = pmd_pgtable(*pmd);
1819 
1820 	do {
1821 		err = fn(pte++, token, addr, data);
1822 		if (err)
1823 			break;
1824 	} while (addr += PAGE_SIZE, addr != end);
1825 
1826 	arch_leave_lazy_mmu_mode();
1827 
1828 	if (mm != &init_mm)
1829 		pte_unmap_unlock(pte-1, ptl);
1830 	return err;
1831 }
1832 
1833 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1834 				     unsigned long addr, unsigned long end,
1835 				     pte_fn_t fn, void *data)
1836 {
1837 	pmd_t *pmd;
1838 	unsigned long next;
1839 	int err;
1840 
1841 	BUG_ON(pud_huge(*pud));
1842 
1843 	pmd = pmd_alloc(mm, pud, addr);
1844 	if (!pmd)
1845 		return -ENOMEM;
1846 	do {
1847 		next = pmd_addr_end(addr, end);
1848 		err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1849 		if (err)
1850 			break;
1851 	} while (pmd++, addr = next, addr != end);
1852 	return err;
1853 }
1854 
1855 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1856 				     unsigned long addr, unsigned long end,
1857 				     pte_fn_t fn, void *data)
1858 {
1859 	pud_t *pud;
1860 	unsigned long next;
1861 	int err;
1862 
1863 	pud = pud_alloc(mm, pgd, addr);
1864 	if (!pud)
1865 		return -ENOMEM;
1866 	do {
1867 		next = pud_addr_end(addr, end);
1868 		err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1869 		if (err)
1870 			break;
1871 	} while (pud++, addr = next, addr != end);
1872 	return err;
1873 }
1874 
1875 /*
1876  * Scan a region of virtual memory, filling in page tables as necessary
1877  * and calling a provided function on each leaf page table.
1878  */
1879 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1880 			unsigned long size, pte_fn_t fn, void *data)
1881 {
1882 	pgd_t *pgd;
1883 	unsigned long next;
1884 	unsigned long end = addr + size;
1885 	int err;
1886 
1887 	BUG_ON(addr >= end);
1888 	pgd = pgd_offset(mm, addr);
1889 	do {
1890 		next = pgd_addr_end(addr, end);
1891 		err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1892 		if (err)
1893 			break;
1894 	} while (pgd++, addr = next, addr != end);
1895 
1896 	return err;
1897 }
1898 EXPORT_SYMBOL_GPL(apply_to_page_range);
1899 
1900 /*
1901  * handle_pte_fault chooses page fault handler according to an entry which was
1902  * read non-atomically.  Before making any commitment, on those architectures
1903  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1904  * parts, do_swap_page must check under lock before unmapping the pte and
1905  * proceeding (but do_wp_page is only called after already making such a check;
1906  * and do_anonymous_page can safely check later on).
1907  */
1908 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1909 				pte_t *page_table, pte_t orig_pte)
1910 {
1911 	int same = 1;
1912 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1913 	if (sizeof(pte_t) > sizeof(unsigned long)) {
1914 		spinlock_t *ptl = pte_lockptr(mm, pmd);
1915 		spin_lock(ptl);
1916 		same = pte_same(*page_table, orig_pte);
1917 		spin_unlock(ptl);
1918 	}
1919 #endif
1920 	pte_unmap(page_table);
1921 	return same;
1922 }
1923 
1924 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1925 {
1926 	debug_dma_assert_idle(src);
1927 
1928 	/*
1929 	 * If the source page was a PFN mapping, we don't have
1930 	 * a "struct page" for it. We do a best-effort copy by
1931 	 * just copying from the original user address. If that
1932 	 * fails, we just zero-fill it. Live with it.
1933 	 */
1934 	if (unlikely(!src)) {
1935 		void *kaddr = kmap_atomic(dst);
1936 		void __user *uaddr = (void __user *)(va & PAGE_MASK);
1937 
1938 		/*
1939 		 * This really shouldn't fail, because the page is there
1940 		 * in the page tables. But it might just be unreadable,
1941 		 * in which case we just give up and fill the result with
1942 		 * zeroes.
1943 		 */
1944 		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1945 			clear_page(kaddr);
1946 		kunmap_atomic(kaddr);
1947 		flush_dcache_page(dst);
1948 	} else
1949 		copy_user_highpage(dst, src, va, vma);
1950 }
1951 
1952 /*
1953  * Notify the address space that the page is about to become writable so that
1954  * it can prohibit this or wait for the page to get into an appropriate state.
1955  *
1956  * We do this without the lock held, so that it can sleep if it needs to.
1957  */
1958 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1959 	       unsigned long address)
1960 {
1961 	struct vm_fault vmf;
1962 	int ret;
1963 
1964 	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1965 	vmf.pgoff = page->index;
1966 	vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1967 	vmf.page = page;
1968 	vmf.cow_page = NULL;
1969 
1970 	ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1971 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1972 		return ret;
1973 	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
1974 		lock_page(page);
1975 		if (!page->mapping) {
1976 			unlock_page(page);
1977 			return 0; /* retry */
1978 		}
1979 		ret |= VM_FAULT_LOCKED;
1980 	} else
1981 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1982 	return ret;
1983 }
1984 
1985 /*
1986  * Handle write page faults for pages that can be reused in the current vma
1987  *
1988  * This can happen either due to the mapping being with the VM_SHARED flag,
1989  * or due to us being the last reference standing to the page. In either
1990  * case, all we need to do here is to mark the page as writable and update
1991  * any related book-keeping.
1992  */
1993 static inline int wp_page_reuse(struct mm_struct *mm,
1994 			struct vm_area_struct *vma, unsigned long address,
1995 			pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
1996 			struct page *page, int page_mkwrite,
1997 			int dirty_shared)
1998 	__releases(ptl)
1999 {
2000 	pte_t entry;
2001 	/*
2002 	 * Clear the pages cpupid information as the existing
2003 	 * information potentially belongs to a now completely
2004 	 * unrelated process.
2005 	 */
2006 	if (page)
2007 		page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2008 
2009 	flush_cache_page(vma, address, pte_pfn(orig_pte));
2010 	entry = pte_mkyoung(orig_pte);
2011 	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2012 	if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2013 		update_mmu_cache(vma, address, page_table);
2014 	pte_unmap_unlock(page_table, ptl);
2015 
2016 	if (dirty_shared) {
2017 		struct address_space *mapping;
2018 		int dirtied;
2019 
2020 		if (!page_mkwrite)
2021 			lock_page(page);
2022 
2023 		dirtied = set_page_dirty(page);
2024 		VM_BUG_ON_PAGE(PageAnon(page), page);
2025 		mapping = page->mapping;
2026 		unlock_page(page);
2027 		page_cache_release(page);
2028 
2029 		if ((dirtied || page_mkwrite) && mapping) {
2030 			/*
2031 			 * Some device drivers do not set page.mapping
2032 			 * but still dirty their pages
2033 			 */
2034 			balance_dirty_pages_ratelimited(mapping);
2035 		}
2036 
2037 		if (!page_mkwrite)
2038 			file_update_time(vma->vm_file);
2039 	}
2040 
2041 	return VM_FAULT_WRITE;
2042 }
2043 
2044 /*
2045  * Handle the case of a page which we actually need to copy to a new page.
2046  *
2047  * Called with mmap_sem locked and the old page referenced, but
2048  * without the ptl held.
2049  *
2050  * High level logic flow:
2051  *
2052  * - Allocate a page, copy the content of the old page to the new one.
2053  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2054  * - Take the PTL. If the pte changed, bail out and release the allocated page
2055  * - If the pte is still the way we remember it, update the page table and all
2056  *   relevant references. This includes dropping the reference the page-table
2057  *   held to the old page, as well as updating the rmap.
2058  * - In any case, unlock the PTL and drop the reference we took to the old page.
2059  */
2060 static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2061 			unsigned long address, pte_t *page_table, pmd_t *pmd,
2062 			pte_t orig_pte, struct page *old_page)
2063 {
2064 	struct page *new_page = NULL;
2065 	spinlock_t *ptl = NULL;
2066 	pte_t entry;
2067 	int page_copied = 0;
2068 	const unsigned long mmun_start = address & PAGE_MASK;	/* For mmu_notifiers */
2069 	const unsigned long mmun_end = mmun_start + PAGE_SIZE;	/* For mmu_notifiers */
2070 	struct mem_cgroup *memcg;
2071 
2072 	if (unlikely(anon_vma_prepare(vma)))
2073 		goto oom;
2074 
2075 	if (is_zero_pfn(pte_pfn(orig_pte))) {
2076 		new_page = alloc_zeroed_user_highpage_movable(vma, address);
2077 		if (!new_page)
2078 			goto oom;
2079 	} else {
2080 		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2081 		if (!new_page)
2082 			goto oom;
2083 		cow_user_page(new_page, old_page, address, vma);
2084 	}
2085 
2086 	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
2087 		goto oom_free_new;
2088 
2089 	__SetPageUptodate(new_page);
2090 
2091 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2092 
2093 	/*
2094 	 * Re-check the pte - we dropped the lock
2095 	 */
2096 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2097 	if (likely(pte_same(*page_table, orig_pte))) {
2098 		if (old_page) {
2099 			if (!PageAnon(old_page)) {
2100 				dec_mm_counter_fast(mm, MM_FILEPAGES);
2101 				inc_mm_counter_fast(mm, MM_ANONPAGES);
2102 			}
2103 		} else {
2104 			inc_mm_counter_fast(mm, MM_ANONPAGES);
2105 		}
2106 		flush_cache_page(vma, address, pte_pfn(orig_pte));
2107 		entry = mk_pte(new_page, vma->vm_page_prot);
2108 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2109 		/*
2110 		 * Clear the pte entry and flush it first, before updating the
2111 		 * pte with the new entry. This will avoid a race condition
2112 		 * seen in the presence of one thread doing SMC and another
2113 		 * thread doing COW.
2114 		 */
2115 		ptep_clear_flush_notify(vma, address, page_table);
2116 		page_add_new_anon_rmap(new_page, vma, address);
2117 		mem_cgroup_commit_charge(new_page, memcg, false);
2118 		lru_cache_add_active_or_unevictable(new_page, vma);
2119 		/*
2120 		 * We call the notify macro here because, when using secondary
2121 		 * mmu page tables (such as kvm shadow page tables), we want the
2122 		 * new page to be mapped directly into the secondary page table.
2123 		 */
2124 		set_pte_at_notify(mm, address, page_table, entry);
2125 		update_mmu_cache(vma, address, page_table);
2126 		if (old_page) {
2127 			/*
2128 			 * Only after switching the pte to the new page may
2129 			 * we remove the mapcount here. Otherwise another
2130 			 * process may come and find the rmap count decremented
2131 			 * before the pte is switched to the new page, and
2132 			 * "reuse" the old page writing into it while our pte
2133 			 * here still points into it and can be read by other
2134 			 * threads.
2135 			 *
2136 			 * The critical issue is to order this
2137 			 * page_remove_rmap with the ptp_clear_flush above.
2138 			 * Those stores are ordered by (if nothing else,)
2139 			 * the barrier present in the atomic_add_negative
2140 			 * in page_remove_rmap.
2141 			 *
2142 			 * Then the TLB flush in ptep_clear_flush ensures that
2143 			 * no process can access the old page before the
2144 			 * decremented mapcount is visible. And the old page
2145 			 * cannot be reused until after the decremented
2146 			 * mapcount is visible. So transitively, TLBs to
2147 			 * old page will be flushed before it can be reused.
2148 			 */
2149 			page_remove_rmap(old_page);
2150 		}
2151 
2152 		/* Free the old page.. */
2153 		new_page = old_page;
2154 		page_copied = 1;
2155 	} else {
2156 		mem_cgroup_cancel_charge(new_page, memcg);
2157 	}
2158 
2159 	if (new_page)
2160 		page_cache_release(new_page);
2161 
2162 	pte_unmap_unlock(page_table, ptl);
2163 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2164 	if (old_page) {
2165 		/*
2166 		 * Don't let another task, with possibly unlocked vma,
2167 		 * keep the mlocked page.
2168 		 */
2169 		if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2170 			lock_page(old_page);	/* LRU manipulation */
2171 			munlock_vma_page(old_page);
2172 			unlock_page(old_page);
2173 		}
2174 		page_cache_release(old_page);
2175 	}
2176 	return page_copied ? VM_FAULT_WRITE : 0;
2177 oom_free_new:
2178 	page_cache_release(new_page);
2179 oom:
2180 	if (old_page)
2181 		page_cache_release(old_page);
2182 	return VM_FAULT_OOM;
2183 }
2184 
2185 /*
2186  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2187  * mapping
2188  */
2189 static int wp_pfn_shared(struct mm_struct *mm,
2190 			struct vm_area_struct *vma, unsigned long address,
2191 			pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2192 			pmd_t *pmd)
2193 {
2194 	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2195 		struct vm_fault vmf = {
2196 			.page = NULL,
2197 			.pgoff = linear_page_index(vma, address),
2198 			.virtual_address = (void __user *)(address & PAGE_MASK),
2199 			.flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2200 		};
2201 		int ret;
2202 
2203 		pte_unmap_unlock(page_table, ptl);
2204 		ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2205 		if (ret & VM_FAULT_ERROR)
2206 			return ret;
2207 		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2208 		/*
2209 		 * We might have raced with another page fault while we
2210 		 * released the pte_offset_map_lock.
2211 		 */
2212 		if (!pte_same(*page_table, orig_pte)) {
2213 			pte_unmap_unlock(page_table, ptl);
2214 			return 0;
2215 		}
2216 	}
2217 	return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2218 			     NULL, 0, 0);
2219 }
2220 
2221 static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2222 			  unsigned long address, pte_t *page_table,
2223 			  pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2224 			  struct page *old_page)
2225 	__releases(ptl)
2226 {
2227 	int page_mkwrite = 0;
2228 
2229 	page_cache_get(old_page);
2230 
2231 	/*
2232 	 * Only catch write-faults on shared writable pages,
2233 	 * read-only shared pages can get COWed by
2234 	 * get_user_pages(.write=1, .force=1).
2235 	 */
2236 	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2237 		int tmp;
2238 
2239 		pte_unmap_unlock(page_table, ptl);
2240 		tmp = do_page_mkwrite(vma, old_page, address);
2241 		if (unlikely(!tmp || (tmp &
2242 				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2243 			page_cache_release(old_page);
2244 			return tmp;
2245 		}
2246 		/*
2247 		 * Since we dropped the lock we need to revalidate
2248 		 * the PTE as someone else may have changed it.  If
2249 		 * they did, we just return, as we can count on the
2250 		 * MMU to tell us if they didn't also make it writable.
2251 		 */
2252 		page_table = pte_offset_map_lock(mm, pmd, address,
2253 						 &ptl);
2254 		if (!pte_same(*page_table, orig_pte)) {
2255 			unlock_page(old_page);
2256 			pte_unmap_unlock(page_table, ptl);
2257 			page_cache_release(old_page);
2258 			return 0;
2259 		}
2260 		page_mkwrite = 1;
2261 	}
2262 
2263 	return wp_page_reuse(mm, vma, address, page_table, ptl,
2264 			     orig_pte, old_page, page_mkwrite, 1);
2265 }
2266 
2267 /*
2268  * This routine handles present pages, when users try to write
2269  * to a shared page. It is done by copying the page to a new address
2270  * and decrementing the shared-page counter for the old page.
2271  *
2272  * Note that this routine assumes that the protection checks have been
2273  * done by the caller (the low-level page fault routine in most cases).
2274  * Thus we can safely just mark it writable once we've done any necessary
2275  * COW.
2276  *
2277  * We also mark the page dirty at this point even though the page will
2278  * change only once the write actually happens. This avoids a few races,
2279  * and potentially makes it more efficient.
2280  *
2281  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2282  * but allow concurrent faults), with pte both mapped and locked.
2283  * We return with mmap_sem still held, but pte unmapped and unlocked.
2284  */
2285 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2286 		unsigned long address, pte_t *page_table, pmd_t *pmd,
2287 		spinlock_t *ptl, pte_t orig_pte)
2288 	__releases(ptl)
2289 {
2290 	struct page *old_page;
2291 
2292 	old_page = vm_normal_page(vma, address, orig_pte);
2293 	if (!old_page) {
2294 		/*
2295 		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2296 		 * VM_PFNMAP VMA.
2297 		 *
2298 		 * We should not cow pages in a shared writeable mapping.
2299 		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2300 		 */
2301 		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2302 				     (VM_WRITE|VM_SHARED))
2303 			return wp_pfn_shared(mm, vma, address, page_table, ptl,
2304 					     orig_pte, pmd);
2305 
2306 		pte_unmap_unlock(page_table, ptl);
2307 		return wp_page_copy(mm, vma, address, page_table, pmd,
2308 				    orig_pte, old_page);
2309 	}
2310 
2311 	/*
2312 	 * Take out anonymous pages first, anonymous shared vmas are
2313 	 * not dirty accountable.
2314 	 */
2315 	if (PageAnon(old_page) && !PageKsm(old_page)) {
2316 		if (!trylock_page(old_page)) {
2317 			page_cache_get(old_page);
2318 			pte_unmap_unlock(page_table, ptl);
2319 			lock_page(old_page);
2320 			page_table = pte_offset_map_lock(mm, pmd, address,
2321 							 &ptl);
2322 			if (!pte_same(*page_table, orig_pte)) {
2323 				unlock_page(old_page);
2324 				pte_unmap_unlock(page_table, ptl);
2325 				page_cache_release(old_page);
2326 				return 0;
2327 			}
2328 			page_cache_release(old_page);
2329 		}
2330 		if (reuse_swap_page(old_page)) {
2331 			/*
2332 			 * The page is all ours.  Move it to our anon_vma so
2333 			 * the rmap code will not search our parent or siblings.
2334 			 * Protected against the rmap code by the page lock.
2335 			 */
2336 			page_move_anon_rmap(old_page, vma, address);
2337 			unlock_page(old_page);
2338 			return wp_page_reuse(mm, vma, address, page_table, ptl,
2339 					     orig_pte, old_page, 0, 0);
2340 		}
2341 		unlock_page(old_page);
2342 	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2343 					(VM_WRITE|VM_SHARED))) {
2344 		return wp_page_shared(mm, vma, address, page_table, pmd,
2345 				      ptl, orig_pte, old_page);
2346 	}
2347 
2348 	/*
2349 	 * Ok, we need to copy. Oh, well..
2350 	 */
2351 	page_cache_get(old_page);
2352 
2353 	pte_unmap_unlock(page_table, ptl);
2354 	return wp_page_copy(mm, vma, address, page_table, pmd,
2355 			    orig_pte, old_page);
2356 }
2357 
2358 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2359 		unsigned long start_addr, unsigned long end_addr,
2360 		struct zap_details *details)
2361 {
2362 	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2363 }
2364 
2365 static inline void unmap_mapping_range_tree(struct rb_root *root,
2366 					    struct zap_details *details)
2367 {
2368 	struct vm_area_struct *vma;
2369 	pgoff_t vba, vea, zba, zea;
2370 
2371 	vma_interval_tree_foreach(vma, root,
2372 			details->first_index, details->last_index) {
2373 
2374 		vba = vma->vm_pgoff;
2375 		vea = vba + vma_pages(vma) - 1;
2376 		/* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2377 		zba = details->first_index;
2378 		if (zba < vba)
2379 			zba = vba;
2380 		zea = details->last_index;
2381 		if (zea > vea)
2382 			zea = vea;
2383 
2384 		unmap_mapping_range_vma(vma,
2385 			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2386 			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2387 				details);
2388 	}
2389 }
2390 
2391 /**
2392  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2393  * address_space corresponding to the specified page range in the underlying
2394  * file.
2395  *
2396  * @mapping: the address space containing mmaps to be unmapped.
2397  * @holebegin: byte in first page to unmap, relative to the start of
2398  * the underlying file.  This will be rounded down to a PAGE_SIZE
2399  * boundary.  Note that this is different from truncate_pagecache(), which
2400  * must keep the partial page.  In contrast, we must get rid of
2401  * partial pages.
2402  * @holelen: size of prospective hole in bytes.  This will be rounded
2403  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2404  * end of the file.
2405  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2406  * but 0 when invalidating pagecache, don't throw away private data.
2407  */
2408 void unmap_mapping_range(struct address_space *mapping,
2409 		loff_t const holebegin, loff_t const holelen, int even_cows)
2410 {
2411 	struct zap_details details;
2412 	pgoff_t hba = holebegin >> PAGE_SHIFT;
2413 	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2414 
2415 	/* Check for overflow. */
2416 	if (sizeof(holelen) > sizeof(hlen)) {
2417 		long long holeend =
2418 			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2419 		if (holeend & ~(long long)ULONG_MAX)
2420 			hlen = ULONG_MAX - hba + 1;
2421 	}
2422 
2423 	details.check_mapping = even_cows? NULL: mapping;
2424 	details.first_index = hba;
2425 	details.last_index = hba + hlen - 1;
2426 	if (details.last_index < details.first_index)
2427 		details.last_index = ULONG_MAX;
2428 
2429 
2430 	/* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2431 	i_mmap_lock_write(mapping);
2432 	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2433 		unmap_mapping_range_tree(&mapping->i_mmap, &details);
2434 	i_mmap_unlock_write(mapping);
2435 }
2436 EXPORT_SYMBOL(unmap_mapping_range);
2437 
2438 /*
2439  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2440  * but allow concurrent faults), and pte mapped but not yet locked.
2441  * We return with pte unmapped and unlocked.
2442  *
2443  * We return with the mmap_sem locked or unlocked in the same cases
2444  * as does filemap_fault().
2445  */
2446 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2447 		unsigned long address, pte_t *page_table, pmd_t *pmd,
2448 		unsigned int flags, pte_t orig_pte)
2449 {
2450 	spinlock_t *ptl;
2451 	struct page *page, *swapcache;
2452 	struct mem_cgroup *memcg;
2453 	swp_entry_t entry;
2454 	pte_t pte;
2455 	int locked;
2456 	int exclusive = 0;
2457 	int ret = 0;
2458 
2459 	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2460 		goto out;
2461 
2462 	entry = pte_to_swp_entry(orig_pte);
2463 	if (unlikely(non_swap_entry(entry))) {
2464 		if (is_migration_entry(entry)) {
2465 			migration_entry_wait(mm, pmd, address);
2466 		} else if (is_hwpoison_entry(entry)) {
2467 			ret = VM_FAULT_HWPOISON;
2468 		} else {
2469 			print_bad_pte(vma, address, orig_pte, NULL);
2470 			ret = VM_FAULT_SIGBUS;
2471 		}
2472 		goto out;
2473 	}
2474 	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2475 	page = lookup_swap_cache(entry);
2476 	if (!page) {
2477 		page = swapin_readahead(entry,
2478 					GFP_HIGHUSER_MOVABLE, vma, address);
2479 		if (!page) {
2480 			/*
2481 			 * Back out if somebody else faulted in this pte
2482 			 * while we released the pte lock.
2483 			 */
2484 			page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2485 			if (likely(pte_same(*page_table, orig_pte)))
2486 				ret = VM_FAULT_OOM;
2487 			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2488 			goto unlock;
2489 		}
2490 
2491 		/* Had to read the page from swap area: Major fault */
2492 		ret = VM_FAULT_MAJOR;
2493 		count_vm_event(PGMAJFAULT);
2494 		mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2495 	} else if (PageHWPoison(page)) {
2496 		/*
2497 		 * hwpoisoned dirty swapcache pages are kept for killing
2498 		 * owner processes (which may be unknown at hwpoison time)
2499 		 */
2500 		ret = VM_FAULT_HWPOISON;
2501 		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2502 		swapcache = page;
2503 		goto out_release;
2504 	}
2505 
2506 	swapcache = page;
2507 	locked = lock_page_or_retry(page, mm, flags);
2508 
2509 	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2510 	if (!locked) {
2511 		ret |= VM_FAULT_RETRY;
2512 		goto out_release;
2513 	}
2514 
2515 	/*
2516 	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2517 	 * release the swapcache from under us.  The page pin, and pte_same
2518 	 * test below, are not enough to exclude that.  Even if it is still
2519 	 * swapcache, we need to check that the page's swap has not changed.
2520 	 */
2521 	if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2522 		goto out_page;
2523 
2524 	page = ksm_might_need_to_copy(page, vma, address);
2525 	if (unlikely(!page)) {
2526 		ret = VM_FAULT_OOM;
2527 		page = swapcache;
2528 		goto out_page;
2529 	}
2530 
2531 	if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
2532 		ret = VM_FAULT_OOM;
2533 		goto out_page;
2534 	}
2535 
2536 	/*
2537 	 * Back out if somebody else already faulted in this pte.
2538 	 */
2539 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2540 	if (unlikely(!pte_same(*page_table, orig_pte)))
2541 		goto out_nomap;
2542 
2543 	if (unlikely(!PageUptodate(page))) {
2544 		ret = VM_FAULT_SIGBUS;
2545 		goto out_nomap;
2546 	}
2547 
2548 	/*
2549 	 * The page isn't present yet, go ahead with the fault.
2550 	 *
2551 	 * Be careful about the sequence of operations here.
2552 	 * To get its accounting right, reuse_swap_page() must be called
2553 	 * while the page is counted on swap but not yet in mapcount i.e.
2554 	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2555 	 * must be called after the swap_free(), or it will never succeed.
2556 	 */
2557 
2558 	inc_mm_counter_fast(mm, MM_ANONPAGES);
2559 	dec_mm_counter_fast(mm, MM_SWAPENTS);
2560 	pte = mk_pte(page, vma->vm_page_prot);
2561 	if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2562 		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2563 		flags &= ~FAULT_FLAG_WRITE;
2564 		ret |= VM_FAULT_WRITE;
2565 		exclusive = 1;
2566 	}
2567 	flush_icache_page(vma, page);
2568 	if (pte_swp_soft_dirty(orig_pte))
2569 		pte = pte_mksoft_dirty(pte);
2570 	set_pte_at(mm, address, page_table, pte);
2571 	if (page == swapcache) {
2572 		do_page_add_anon_rmap(page, vma, address, exclusive);
2573 		mem_cgroup_commit_charge(page, memcg, true);
2574 	} else { /* ksm created a completely new copy */
2575 		page_add_new_anon_rmap(page, vma, address);
2576 		mem_cgroup_commit_charge(page, memcg, false);
2577 		lru_cache_add_active_or_unevictable(page, vma);
2578 	}
2579 
2580 	swap_free(entry);
2581 	if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2582 		try_to_free_swap(page);
2583 	unlock_page(page);
2584 	if (page != swapcache) {
2585 		/*
2586 		 * Hold the lock to avoid the swap entry to be reused
2587 		 * until we take the PT lock for the pte_same() check
2588 		 * (to avoid false positives from pte_same). For
2589 		 * further safety release the lock after the swap_free
2590 		 * so that the swap count won't change under a
2591 		 * parallel locked swapcache.
2592 		 */
2593 		unlock_page(swapcache);
2594 		page_cache_release(swapcache);
2595 	}
2596 
2597 	if (flags & FAULT_FLAG_WRITE) {
2598 		ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2599 		if (ret & VM_FAULT_ERROR)
2600 			ret &= VM_FAULT_ERROR;
2601 		goto out;
2602 	}
2603 
2604 	/* No need to invalidate - it was non-present before */
2605 	update_mmu_cache(vma, address, page_table);
2606 unlock:
2607 	pte_unmap_unlock(page_table, ptl);
2608 out:
2609 	return ret;
2610 out_nomap:
2611 	mem_cgroup_cancel_charge(page, memcg);
2612 	pte_unmap_unlock(page_table, ptl);
2613 out_page:
2614 	unlock_page(page);
2615 out_release:
2616 	page_cache_release(page);
2617 	if (page != swapcache) {
2618 		unlock_page(swapcache);
2619 		page_cache_release(swapcache);
2620 	}
2621 	return ret;
2622 }
2623 
2624 /*
2625  * This is like a special single-page "expand_{down|up}wards()",
2626  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2627  * doesn't hit another vma.
2628  */
2629 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2630 {
2631 	address &= PAGE_MASK;
2632 	if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2633 		struct vm_area_struct *prev = vma->vm_prev;
2634 
2635 		/*
2636 		 * Is there a mapping abutting this one below?
2637 		 *
2638 		 * That's only ok if it's the same stack mapping
2639 		 * that has gotten split..
2640 		 */
2641 		if (prev && prev->vm_end == address)
2642 			return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2643 
2644 		return expand_downwards(vma, address - PAGE_SIZE);
2645 	}
2646 	if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2647 		struct vm_area_struct *next = vma->vm_next;
2648 
2649 		/* As VM_GROWSDOWN but s/below/above/ */
2650 		if (next && next->vm_start == address + PAGE_SIZE)
2651 			return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2652 
2653 		return expand_upwards(vma, address + PAGE_SIZE);
2654 	}
2655 	return 0;
2656 }
2657 
2658 /*
2659  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2660  * but allow concurrent faults), and pte mapped but not yet locked.
2661  * We return with mmap_sem still held, but pte unmapped and unlocked.
2662  */
2663 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2664 		unsigned long address, pte_t *page_table, pmd_t *pmd,
2665 		unsigned int flags)
2666 {
2667 	struct mem_cgroup *memcg;
2668 	struct page *page;
2669 	spinlock_t *ptl;
2670 	pte_t entry;
2671 
2672 	pte_unmap(page_table);
2673 
2674 	/* File mapping without ->vm_ops ? */
2675 	if (vma->vm_flags & VM_SHARED)
2676 		return VM_FAULT_SIGBUS;
2677 
2678 	/* Check if we need to add a guard page to the stack */
2679 	if (check_stack_guard_page(vma, address) < 0)
2680 		return VM_FAULT_SIGSEGV;
2681 
2682 	/* Use the zero-page for reads */
2683 	if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2684 		entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2685 						vma->vm_page_prot));
2686 		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2687 		if (!pte_none(*page_table))
2688 			goto unlock;
2689 		/* Deliver the page fault to userland, check inside PT lock */
2690 		if (userfaultfd_missing(vma)) {
2691 			pte_unmap_unlock(page_table, ptl);
2692 			return handle_userfault(vma, address, flags,
2693 						VM_UFFD_MISSING);
2694 		}
2695 		goto setpte;
2696 	}
2697 
2698 	/* Allocate our own private page. */
2699 	if (unlikely(anon_vma_prepare(vma)))
2700 		goto oom;
2701 	page = alloc_zeroed_user_highpage_movable(vma, address);
2702 	if (!page)
2703 		goto oom;
2704 
2705 	if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
2706 		goto oom_free_page;
2707 
2708 	/*
2709 	 * The memory barrier inside __SetPageUptodate makes sure that
2710 	 * preceeding stores to the page contents become visible before
2711 	 * the set_pte_at() write.
2712 	 */
2713 	__SetPageUptodate(page);
2714 
2715 	entry = mk_pte(page, vma->vm_page_prot);
2716 	if (vma->vm_flags & VM_WRITE)
2717 		entry = pte_mkwrite(pte_mkdirty(entry));
2718 
2719 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2720 	if (!pte_none(*page_table))
2721 		goto release;
2722 
2723 	/* Deliver the page fault to userland, check inside PT lock */
2724 	if (userfaultfd_missing(vma)) {
2725 		pte_unmap_unlock(page_table, ptl);
2726 		mem_cgroup_cancel_charge(page, memcg);
2727 		page_cache_release(page);
2728 		return handle_userfault(vma, address, flags,
2729 					VM_UFFD_MISSING);
2730 	}
2731 
2732 	inc_mm_counter_fast(mm, MM_ANONPAGES);
2733 	page_add_new_anon_rmap(page, vma, address);
2734 	mem_cgroup_commit_charge(page, memcg, false);
2735 	lru_cache_add_active_or_unevictable(page, vma);
2736 setpte:
2737 	set_pte_at(mm, address, page_table, entry);
2738 
2739 	/* No need to invalidate - it was non-present before */
2740 	update_mmu_cache(vma, address, page_table);
2741 unlock:
2742 	pte_unmap_unlock(page_table, ptl);
2743 	return 0;
2744 release:
2745 	mem_cgroup_cancel_charge(page, memcg);
2746 	page_cache_release(page);
2747 	goto unlock;
2748 oom_free_page:
2749 	page_cache_release(page);
2750 oom:
2751 	return VM_FAULT_OOM;
2752 }
2753 
2754 /*
2755  * The mmap_sem must have been held on entry, and may have been
2756  * released depending on flags and vma->vm_ops->fault() return value.
2757  * See filemap_fault() and __lock_page_retry().
2758  */
2759 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2760 			pgoff_t pgoff, unsigned int flags,
2761 			struct page *cow_page, struct page **page)
2762 {
2763 	struct vm_fault vmf;
2764 	int ret;
2765 
2766 	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2767 	vmf.pgoff = pgoff;
2768 	vmf.flags = flags;
2769 	vmf.page = NULL;
2770 	vmf.cow_page = cow_page;
2771 
2772 	ret = vma->vm_ops->fault(vma, &vmf);
2773 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2774 		return ret;
2775 	if (!vmf.page)
2776 		goto out;
2777 
2778 	if (unlikely(PageHWPoison(vmf.page))) {
2779 		if (ret & VM_FAULT_LOCKED)
2780 			unlock_page(vmf.page);
2781 		page_cache_release(vmf.page);
2782 		return VM_FAULT_HWPOISON;
2783 	}
2784 
2785 	if (unlikely(!(ret & VM_FAULT_LOCKED)))
2786 		lock_page(vmf.page);
2787 	else
2788 		VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2789 
2790  out:
2791 	*page = vmf.page;
2792 	return ret;
2793 }
2794 
2795 /**
2796  * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2797  *
2798  * @vma: virtual memory area
2799  * @address: user virtual address
2800  * @page: page to map
2801  * @pte: pointer to target page table entry
2802  * @write: true, if new entry is writable
2803  * @anon: true, if it's anonymous page
2804  *
2805  * Caller must hold page table lock relevant for @pte.
2806  *
2807  * Target users are page handler itself and implementations of
2808  * vm_ops->map_pages.
2809  */
2810 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2811 		struct page *page, pte_t *pte, bool write, bool anon)
2812 {
2813 	pte_t entry;
2814 
2815 	flush_icache_page(vma, page);
2816 	entry = mk_pte(page, vma->vm_page_prot);
2817 	if (write)
2818 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2819 	if (anon) {
2820 		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2821 		page_add_new_anon_rmap(page, vma, address);
2822 	} else {
2823 		inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2824 		page_add_file_rmap(page);
2825 	}
2826 	set_pte_at(vma->vm_mm, address, pte, entry);
2827 
2828 	/* no need to invalidate: a not-present page won't be cached */
2829 	update_mmu_cache(vma, address, pte);
2830 }
2831 
2832 static unsigned long fault_around_bytes __read_mostly =
2833 	rounddown_pow_of_two(65536);
2834 
2835 #ifdef CONFIG_DEBUG_FS
2836 static int fault_around_bytes_get(void *data, u64 *val)
2837 {
2838 	*val = fault_around_bytes;
2839 	return 0;
2840 }
2841 
2842 /*
2843  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2844  * rounded down to nearest page order. It's what do_fault_around() expects to
2845  * see.
2846  */
2847 static int fault_around_bytes_set(void *data, u64 val)
2848 {
2849 	if (val / PAGE_SIZE > PTRS_PER_PTE)
2850 		return -EINVAL;
2851 	if (val > PAGE_SIZE)
2852 		fault_around_bytes = rounddown_pow_of_two(val);
2853 	else
2854 		fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2855 	return 0;
2856 }
2857 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2858 		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2859 
2860 static int __init fault_around_debugfs(void)
2861 {
2862 	void *ret;
2863 
2864 	ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2865 			&fault_around_bytes_fops);
2866 	if (!ret)
2867 		pr_warn("Failed to create fault_around_bytes in debugfs");
2868 	return 0;
2869 }
2870 late_initcall(fault_around_debugfs);
2871 #endif
2872 
2873 /*
2874  * do_fault_around() tries to map few pages around the fault address. The hope
2875  * is that the pages will be needed soon and this will lower the number of
2876  * faults to handle.
2877  *
2878  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2879  * not ready to be mapped: not up-to-date, locked, etc.
2880  *
2881  * This function is called with the page table lock taken. In the split ptlock
2882  * case the page table lock only protects only those entries which belong to
2883  * the page table corresponding to the fault address.
2884  *
2885  * This function doesn't cross the VMA boundaries, in order to call map_pages()
2886  * only once.
2887  *
2888  * fault_around_pages() defines how many pages we'll try to map.
2889  * do_fault_around() expects it to return a power of two less than or equal to
2890  * PTRS_PER_PTE.
2891  *
2892  * The virtual address of the area that we map is naturally aligned to the
2893  * fault_around_pages() value (and therefore to page order).  This way it's
2894  * easier to guarantee that we don't cross page table boundaries.
2895  */
2896 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2897 		pte_t *pte, pgoff_t pgoff, unsigned int flags)
2898 {
2899 	unsigned long start_addr, nr_pages, mask;
2900 	pgoff_t max_pgoff;
2901 	struct vm_fault vmf;
2902 	int off;
2903 
2904 	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2905 	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2906 
2907 	start_addr = max(address & mask, vma->vm_start);
2908 	off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2909 	pte -= off;
2910 	pgoff -= off;
2911 
2912 	/*
2913 	 *  max_pgoff is either end of page table or end of vma
2914 	 *  or fault_around_pages() from pgoff, depending what is nearest.
2915 	 */
2916 	max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2917 		PTRS_PER_PTE - 1;
2918 	max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2919 			pgoff + nr_pages - 1);
2920 
2921 	/* Check if it makes any sense to call ->map_pages */
2922 	while (!pte_none(*pte)) {
2923 		if (++pgoff > max_pgoff)
2924 			return;
2925 		start_addr += PAGE_SIZE;
2926 		if (start_addr >= vma->vm_end)
2927 			return;
2928 		pte++;
2929 	}
2930 
2931 	vmf.virtual_address = (void __user *) start_addr;
2932 	vmf.pte = pte;
2933 	vmf.pgoff = pgoff;
2934 	vmf.max_pgoff = max_pgoff;
2935 	vmf.flags = flags;
2936 	vma->vm_ops->map_pages(vma, &vmf);
2937 }
2938 
2939 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2940 		unsigned long address, pmd_t *pmd,
2941 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2942 {
2943 	struct page *fault_page;
2944 	spinlock_t *ptl;
2945 	pte_t *pte;
2946 	int ret = 0;
2947 
2948 	/*
2949 	 * Let's call ->map_pages() first and use ->fault() as fallback
2950 	 * if page by the offset is not ready to be mapped (cold cache or
2951 	 * something).
2952 	 */
2953 	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
2954 		pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2955 		do_fault_around(vma, address, pte, pgoff, flags);
2956 		if (!pte_same(*pte, orig_pte))
2957 			goto unlock_out;
2958 		pte_unmap_unlock(pte, ptl);
2959 	}
2960 
2961 	ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
2962 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2963 		return ret;
2964 
2965 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2966 	if (unlikely(!pte_same(*pte, orig_pte))) {
2967 		pte_unmap_unlock(pte, ptl);
2968 		unlock_page(fault_page);
2969 		page_cache_release(fault_page);
2970 		return ret;
2971 	}
2972 	do_set_pte(vma, address, fault_page, pte, false, false);
2973 	unlock_page(fault_page);
2974 unlock_out:
2975 	pte_unmap_unlock(pte, ptl);
2976 	return ret;
2977 }
2978 
2979 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2980 		unsigned long address, pmd_t *pmd,
2981 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2982 {
2983 	struct page *fault_page, *new_page;
2984 	struct mem_cgroup *memcg;
2985 	spinlock_t *ptl;
2986 	pte_t *pte;
2987 	int ret;
2988 
2989 	if (unlikely(anon_vma_prepare(vma)))
2990 		return VM_FAULT_OOM;
2991 
2992 	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2993 	if (!new_page)
2994 		return VM_FAULT_OOM;
2995 
2996 	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
2997 		page_cache_release(new_page);
2998 		return VM_FAULT_OOM;
2999 	}
3000 
3001 	ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
3002 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3003 		goto uncharge_out;
3004 
3005 	if (fault_page)
3006 		copy_user_highpage(new_page, fault_page, address, vma);
3007 	__SetPageUptodate(new_page);
3008 
3009 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3010 	if (unlikely(!pte_same(*pte, orig_pte))) {
3011 		pte_unmap_unlock(pte, ptl);
3012 		if (fault_page) {
3013 			unlock_page(fault_page);
3014 			page_cache_release(fault_page);
3015 		} else {
3016 			/*
3017 			 * The fault handler has no page to lock, so it holds
3018 			 * i_mmap_lock for read to protect against truncate.
3019 			 */
3020 			i_mmap_unlock_read(vma->vm_file->f_mapping);
3021 		}
3022 		goto uncharge_out;
3023 	}
3024 	do_set_pte(vma, address, new_page, pte, true, true);
3025 	mem_cgroup_commit_charge(new_page, memcg, false);
3026 	lru_cache_add_active_or_unevictable(new_page, vma);
3027 	pte_unmap_unlock(pte, ptl);
3028 	if (fault_page) {
3029 		unlock_page(fault_page);
3030 		page_cache_release(fault_page);
3031 	} else {
3032 		/*
3033 		 * The fault handler has no page to lock, so it holds
3034 		 * i_mmap_lock for read to protect against truncate.
3035 		 */
3036 		i_mmap_unlock_read(vma->vm_file->f_mapping);
3037 	}
3038 	return ret;
3039 uncharge_out:
3040 	mem_cgroup_cancel_charge(new_page, memcg);
3041 	page_cache_release(new_page);
3042 	return ret;
3043 }
3044 
3045 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3046 		unsigned long address, pmd_t *pmd,
3047 		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3048 {
3049 	struct page *fault_page;
3050 	struct address_space *mapping;
3051 	spinlock_t *ptl;
3052 	pte_t *pte;
3053 	int dirtied = 0;
3054 	int ret, tmp;
3055 
3056 	ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3057 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3058 		return ret;
3059 
3060 	/*
3061 	 * Check if the backing address space wants to know that the page is
3062 	 * about to become writable
3063 	 */
3064 	if (vma->vm_ops->page_mkwrite) {
3065 		unlock_page(fault_page);
3066 		tmp = do_page_mkwrite(vma, fault_page, address);
3067 		if (unlikely(!tmp ||
3068 				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3069 			page_cache_release(fault_page);
3070 			return tmp;
3071 		}
3072 	}
3073 
3074 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3075 	if (unlikely(!pte_same(*pte, orig_pte))) {
3076 		pte_unmap_unlock(pte, ptl);
3077 		unlock_page(fault_page);
3078 		page_cache_release(fault_page);
3079 		return ret;
3080 	}
3081 	do_set_pte(vma, address, fault_page, pte, true, false);
3082 	pte_unmap_unlock(pte, ptl);
3083 
3084 	if (set_page_dirty(fault_page))
3085 		dirtied = 1;
3086 	/*
3087 	 * Take a local copy of the address_space - page.mapping may be zeroed
3088 	 * by truncate after unlock_page().   The address_space itself remains
3089 	 * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
3090 	 * release semantics to prevent the compiler from undoing this copying.
3091 	 */
3092 	mapping = fault_page->mapping;
3093 	unlock_page(fault_page);
3094 	if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3095 		/*
3096 		 * Some device drivers do not set page.mapping but still
3097 		 * dirty their pages
3098 		 */
3099 		balance_dirty_pages_ratelimited(mapping);
3100 	}
3101 
3102 	if (!vma->vm_ops->page_mkwrite)
3103 		file_update_time(vma->vm_file);
3104 
3105 	return ret;
3106 }
3107 
3108 /*
3109  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3110  * but allow concurrent faults).
3111  * The mmap_sem may have been released depending on flags and our
3112  * return value.  See filemap_fault() and __lock_page_or_retry().
3113  */
3114 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3115 		unsigned long address, pte_t *page_table, pmd_t *pmd,
3116 		unsigned int flags, pte_t orig_pte)
3117 {
3118 	pgoff_t pgoff = (((address & PAGE_MASK)
3119 			- vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3120 
3121 	pte_unmap(page_table);
3122 	/* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3123 	if (!vma->vm_ops->fault)
3124 		return VM_FAULT_SIGBUS;
3125 	if (!(flags & FAULT_FLAG_WRITE))
3126 		return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3127 				orig_pte);
3128 	if (!(vma->vm_flags & VM_SHARED))
3129 		return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3130 				orig_pte);
3131 	return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3132 }
3133 
3134 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3135 				unsigned long addr, int page_nid,
3136 				int *flags)
3137 {
3138 	get_page(page);
3139 
3140 	count_vm_numa_event(NUMA_HINT_FAULTS);
3141 	if (page_nid == numa_node_id()) {
3142 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3143 		*flags |= TNF_FAULT_LOCAL;
3144 	}
3145 
3146 	return mpol_misplaced(page, vma, addr);
3147 }
3148 
3149 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3150 		   unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3151 {
3152 	struct page *page = NULL;
3153 	spinlock_t *ptl;
3154 	int page_nid = -1;
3155 	int last_cpupid;
3156 	int target_nid;
3157 	bool migrated = false;
3158 	bool was_writable = pte_write(pte);
3159 	int flags = 0;
3160 
3161 	/* A PROT_NONE fault should not end up here */
3162 	BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3163 
3164 	/*
3165 	* The "pte" at this point cannot be used safely without
3166 	* validation through pte_unmap_same(). It's of NUMA type but
3167 	* the pfn may be screwed if the read is non atomic.
3168 	*
3169 	* We can safely just do a "set_pte_at()", because the old
3170 	* page table entry is not accessible, so there would be no
3171 	* concurrent hardware modifications to the PTE.
3172 	*/
3173 	ptl = pte_lockptr(mm, pmd);
3174 	spin_lock(ptl);
3175 	if (unlikely(!pte_same(*ptep, pte))) {
3176 		pte_unmap_unlock(ptep, ptl);
3177 		goto out;
3178 	}
3179 
3180 	/* Make it present again */
3181 	pte = pte_modify(pte, vma->vm_page_prot);
3182 	pte = pte_mkyoung(pte);
3183 	if (was_writable)
3184 		pte = pte_mkwrite(pte);
3185 	set_pte_at(mm, addr, ptep, pte);
3186 	update_mmu_cache(vma, addr, ptep);
3187 
3188 	page = vm_normal_page(vma, addr, pte);
3189 	if (!page) {
3190 		pte_unmap_unlock(ptep, ptl);
3191 		return 0;
3192 	}
3193 
3194 	/*
3195 	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3196 	 * much anyway since they can be in shared cache state. This misses
3197 	 * the case where a mapping is writable but the process never writes
3198 	 * to it but pte_write gets cleared during protection updates and
3199 	 * pte_dirty has unpredictable behaviour between PTE scan updates,
3200 	 * background writeback, dirty balancing and application behaviour.
3201 	 */
3202 	if (!(vma->vm_flags & VM_WRITE))
3203 		flags |= TNF_NO_GROUP;
3204 
3205 	/*
3206 	 * Flag if the page is shared between multiple address spaces. This
3207 	 * is later used when determining whether to group tasks together
3208 	 */
3209 	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3210 		flags |= TNF_SHARED;
3211 
3212 	last_cpupid = page_cpupid_last(page);
3213 	page_nid = page_to_nid(page);
3214 	target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3215 	pte_unmap_unlock(ptep, ptl);
3216 	if (target_nid == -1) {
3217 		put_page(page);
3218 		goto out;
3219 	}
3220 
3221 	/* Migrate to the requested node */
3222 	migrated = migrate_misplaced_page(page, vma, target_nid);
3223 	if (migrated) {
3224 		page_nid = target_nid;
3225 		flags |= TNF_MIGRATED;
3226 	} else
3227 		flags |= TNF_MIGRATE_FAIL;
3228 
3229 out:
3230 	if (page_nid != -1)
3231 		task_numa_fault(last_cpupid, page_nid, 1, flags);
3232 	return 0;
3233 }
3234 
3235 static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3236 			unsigned long address, pmd_t *pmd, unsigned int flags)
3237 {
3238 	if (vma_is_anonymous(vma))
3239 		return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3240 	if (vma->vm_ops->pmd_fault)
3241 		return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3242 	return VM_FAULT_FALLBACK;
3243 }
3244 
3245 static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3246 			unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3247 			unsigned int flags)
3248 {
3249 	if (vma_is_anonymous(vma))
3250 		return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3251 	if (vma->vm_ops->pmd_fault)
3252 		return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3253 	return VM_FAULT_FALLBACK;
3254 }
3255 
3256 /*
3257  * These routines also need to handle stuff like marking pages dirty
3258  * and/or accessed for architectures that don't do it in hardware (most
3259  * RISC architectures).  The early dirtying is also good on the i386.
3260  *
3261  * There is also a hook called "update_mmu_cache()" that architectures
3262  * with external mmu caches can use to update those (ie the Sparc or
3263  * PowerPC hashed page tables that act as extended TLBs).
3264  *
3265  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3266  * but allow concurrent faults), and pte mapped but not yet locked.
3267  * We return with pte unmapped and unlocked.
3268  *
3269  * The mmap_sem may have been released depending on flags and our
3270  * return value.  See filemap_fault() and __lock_page_or_retry().
3271  */
3272 static int handle_pte_fault(struct mm_struct *mm,
3273 		     struct vm_area_struct *vma, unsigned long address,
3274 		     pte_t *pte, pmd_t *pmd, unsigned int flags)
3275 {
3276 	pte_t entry;
3277 	spinlock_t *ptl;
3278 
3279 	/*
3280 	 * some architectures can have larger ptes than wordsize,
3281 	 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3282 	 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3283 	 * The code below just needs a consistent view for the ifs and
3284 	 * we later double check anyway with the ptl lock held. So here
3285 	 * a barrier will do.
3286 	 */
3287 	entry = *pte;
3288 	barrier();
3289 	if (!pte_present(entry)) {
3290 		if (pte_none(entry)) {
3291 			if (vma_is_anonymous(vma))
3292 				return do_anonymous_page(mm, vma, address,
3293 							 pte, pmd, flags);
3294 			else
3295 				return do_fault(mm, vma, address, pte, pmd,
3296 						flags, entry);
3297 		}
3298 		return do_swap_page(mm, vma, address,
3299 					pte, pmd, flags, entry);
3300 	}
3301 
3302 	if (pte_protnone(entry))
3303 		return do_numa_page(mm, vma, address, entry, pte, pmd);
3304 
3305 	ptl = pte_lockptr(mm, pmd);
3306 	spin_lock(ptl);
3307 	if (unlikely(!pte_same(*pte, entry)))
3308 		goto unlock;
3309 	if (flags & FAULT_FLAG_WRITE) {
3310 		if (!pte_write(entry))
3311 			return do_wp_page(mm, vma, address,
3312 					pte, pmd, ptl, entry);
3313 		entry = pte_mkdirty(entry);
3314 	}
3315 	entry = pte_mkyoung(entry);
3316 	if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3317 		update_mmu_cache(vma, address, pte);
3318 	} else {
3319 		/*
3320 		 * This is needed only for protection faults but the arch code
3321 		 * is not yet telling us if this is a protection fault or not.
3322 		 * This still avoids useless tlb flushes for .text page faults
3323 		 * with threads.
3324 		 */
3325 		if (flags & FAULT_FLAG_WRITE)
3326 			flush_tlb_fix_spurious_fault(vma, address);
3327 	}
3328 unlock:
3329 	pte_unmap_unlock(pte, ptl);
3330 	return 0;
3331 }
3332 
3333 /*
3334  * By the time we get here, we already hold the mm semaphore
3335  *
3336  * The mmap_sem may have been released depending on flags and our
3337  * return value.  See filemap_fault() and __lock_page_or_retry().
3338  */
3339 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3340 			     unsigned long address, unsigned int flags)
3341 {
3342 	pgd_t *pgd;
3343 	pud_t *pud;
3344 	pmd_t *pmd;
3345 	pte_t *pte;
3346 
3347 	if (unlikely(is_vm_hugetlb_page(vma)))
3348 		return hugetlb_fault(mm, vma, address, flags);
3349 
3350 	pgd = pgd_offset(mm, address);
3351 	pud = pud_alloc(mm, pgd, address);
3352 	if (!pud)
3353 		return VM_FAULT_OOM;
3354 	pmd = pmd_alloc(mm, pud, address);
3355 	if (!pmd)
3356 		return VM_FAULT_OOM;
3357 	if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3358 		int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3359 		if (!(ret & VM_FAULT_FALLBACK))
3360 			return ret;
3361 	} else {
3362 		pmd_t orig_pmd = *pmd;
3363 		int ret;
3364 
3365 		barrier();
3366 		if (pmd_trans_huge(orig_pmd)) {
3367 			unsigned int dirty = flags & FAULT_FLAG_WRITE;
3368 
3369 			/*
3370 			 * If the pmd is splitting, return and retry the
3371 			 * the fault.  Alternative: wait until the split
3372 			 * is done, and goto retry.
3373 			 */
3374 			if (pmd_trans_splitting(orig_pmd))
3375 				return 0;
3376 
3377 			if (pmd_protnone(orig_pmd))
3378 				return do_huge_pmd_numa_page(mm, vma, address,
3379 							     orig_pmd, pmd);
3380 
3381 			if (dirty && !pmd_write(orig_pmd)) {
3382 				ret = wp_huge_pmd(mm, vma, address, pmd,
3383 							orig_pmd, flags);
3384 				if (!(ret & VM_FAULT_FALLBACK))
3385 					return ret;
3386 			} else {
3387 				huge_pmd_set_accessed(mm, vma, address, pmd,
3388 						      orig_pmd, dirty);
3389 				return 0;
3390 			}
3391 		}
3392 	}
3393 
3394 	/*
3395 	 * Use __pte_alloc instead of pte_alloc_map, because we can't
3396 	 * run pte_offset_map on the pmd, if an huge pmd could
3397 	 * materialize from under us from a different thread.
3398 	 */
3399 	if (unlikely(pmd_none(*pmd)) &&
3400 	    unlikely(__pte_alloc(mm, vma, pmd, address)))
3401 		return VM_FAULT_OOM;
3402 	/* if an huge pmd materialized from under us just retry later */
3403 	if (unlikely(pmd_trans_huge(*pmd)))
3404 		return 0;
3405 	/*
3406 	 * A regular pmd is established and it can't morph into a huge pmd
3407 	 * from under us anymore at this point because we hold the mmap_sem
3408 	 * read mode and khugepaged takes it in write mode. So now it's
3409 	 * safe to run pte_offset_map().
3410 	 */
3411 	pte = pte_offset_map(pmd, address);
3412 
3413 	return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3414 }
3415 
3416 /*
3417  * By the time we get here, we already hold the mm semaphore
3418  *
3419  * The mmap_sem may have been released depending on flags and our
3420  * return value.  See filemap_fault() and __lock_page_or_retry().
3421  */
3422 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3423 		    unsigned long address, unsigned int flags)
3424 {
3425 	int ret;
3426 
3427 	__set_current_state(TASK_RUNNING);
3428 
3429 	count_vm_event(PGFAULT);
3430 	mem_cgroup_count_vm_event(mm, PGFAULT);
3431 
3432 	/* do counter updates before entering really critical section. */
3433 	check_sync_rss_stat(current);
3434 
3435 	/*
3436 	 * Enable the memcg OOM handling for faults triggered in user
3437 	 * space.  Kernel faults are handled more gracefully.
3438 	 */
3439 	if (flags & FAULT_FLAG_USER)
3440 		mem_cgroup_oom_enable();
3441 
3442 	ret = __handle_mm_fault(mm, vma, address, flags);
3443 
3444 	if (flags & FAULT_FLAG_USER) {
3445 		mem_cgroup_oom_disable();
3446                 /*
3447                  * The task may have entered a memcg OOM situation but
3448                  * if the allocation error was handled gracefully (no
3449                  * VM_FAULT_OOM), there is no need to kill anything.
3450                  * Just clean up the OOM state peacefully.
3451                  */
3452                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3453                         mem_cgroup_oom_synchronize(false);
3454 	}
3455 
3456 	return ret;
3457 }
3458 EXPORT_SYMBOL_GPL(handle_mm_fault);
3459 
3460 #ifndef __PAGETABLE_PUD_FOLDED
3461 /*
3462  * Allocate page upper directory.
3463  * We've already handled the fast-path in-line.
3464  */
3465 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3466 {
3467 	pud_t *new = pud_alloc_one(mm, address);
3468 	if (!new)
3469 		return -ENOMEM;
3470 
3471 	smp_wmb(); /* See comment in __pte_alloc */
3472 
3473 	spin_lock(&mm->page_table_lock);
3474 	if (pgd_present(*pgd))		/* Another has populated it */
3475 		pud_free(mm, new);
3476 	else
3477 		pgd_populate(mm, pgd, new);
3478 	spin_unlock(&mm->page_table_lock);
3479 	return 0;
3480 }
3481 #endif /* __PAGETABLE_PUD_FOLDED */
3482 
3483 #ifndef __PAGETABLE_PMD_FOLDED
3484 /*
3485  * Allocate page middle directory.
3486  * We've already handled the fast-path in-line.
3487  */
3488 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3489 {
3490 	pmd_t *new = pmd_alloc_one(mm, address);
3491 	if (!new)
3492 		return -ENOMEM;
3493 
3494 	smp_wmb(); /* See comment in __pte_alloc */
3495 
3496 	spin_lock(&mm->page_table_lock);
3497 #ifndef __ARCH_HAS_4LEVEL_HACK
3498 	if (!pud_present(*pud)) {
3499 		mm_inc_nr_pmds(mm);
3500 		pud_populate(mm, pud, new);
3501 	} else	/* Another has populated it */
3502 		pmd_free(mm, new);
3503 #else
3504 	if (!pgd_present(*pud)) {
3505 		mm_inc_nr_pmds(mm);
3506 		pgd_populate(mm, pud, new);
3507 	} else /* Another has populated it */
3508 		pmd_free(mm, new);
3509 #endif /* __ARCH_HAS_4LEVEL_HACK */
3510 	spin_unlock(&mm->page_table_lock);
3511 	return 0;
3512 }
3513 #endif /* __PAGETABLE_PMD_FOLDED */
3514 
3515 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3516 		pte_t **ptepp, spinlock_t **ptlp)
3517 {
3518 	pgd_t *pgd;
3519 	pud_t *pud;
3520 	pmd_t *pmd;
3521 	pte_t *ptep;
3522 
3523 	pgd = pgd_offset(mm, address);
3524 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3525 		goto out;
3526 
3527 	pud = pud_offset(pgd, address);
3528 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3529 		goto out;
3530 
3531 	pmd = pmd_offset(pud, address);
3532 	VM_BUG_ON(pmd_trans_huge(*pmd));
3533 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3534 		goto out;
3535 
3536 	/* We cannot handle huge page PFN maps. Luckily they don't exist. */
3537 	if (pmd_huge(*pmd))
3538 		goto out;
3539 
3540 	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3541 	if (!ptep)
3542 		goto out;
3543 	if (!pte_present(*ptep))
3544 		goto unlock;
3545 	*ptepp = ptep;
3546 	return 0;
3547 unlock:
3548 	pte_unmap_unlock(ptep, *ptlp);
3549 out:
3550 	return -EINVAL;
3551 }
3552 
3553 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3554 			     pte_t **ptepp, spinlock_t **ptlp)
3555 {
3556 	int res;
3557 
3558 	/* (void) is needed to make gcc happy */
3559 	(void) __cond_lock(*ptlp,
3560 			   !(res = __follow_pte(mm, address, ptepp, ptlp)));
3561 	return res;
3562 }
3563 
3564 /**
3565  * follow_pfn - look up PFN at a user virtual address
3566  * @vma: memory mapping
3567  * @address: user virtual address
3568  * @pfn: location to store found PFN
3569  *
3570  * Only IO mappings and raw PFN mappings are allowed.
3571  *
3572  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3573  */
3574 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3575 	unsigned long *pfn)
3576 {
3577 	int ret = -EINVAL;
3578 	spinlock_t *ptl;
3579 	pte_t *ptep;
3580 
3581 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3582 		return ret;
3583 
3584 	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3585 	if (ret)
3586 		return ret;
3587 	*pfn = pte_pfn(*ptep);
3588 	pte_unmap_unlock(ptep, ptl);
3589 	return 0;
3590 }
3591 EXPORT_SYMBOL(follow_pfn);
3592 
3593 #ifdef CONFIG_HAVE_IOREMAP_PROT
3594 int follow_phys(struct vm_area_struct *vma,
3595 		unsigned long address, unsigned int flags,
3596 		unsigned long *prot, resource_size_t *phys)
3597 {
3598 	int ret = -EINVAL;
3599 	pte_t *ptep, pte;
3600 	spinlock_t *ptl;
3601 
3602 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3603 		goto out;
3604 
3605 	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3606 		goto out;
3607 	pte = *ptep;
3608 
3609 	if ((flags & FOLL_WRITE) && !pte_write(pte))
3610 		goto unlock;
3611 
3612 	*prot = pgprot_val(pte_pgprot(pte));
3613 	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3614 
3615 	ret = 0;
3616 unlock:
3617 	pte_unmap_unlock(ptep, ptl);
3618 out:
3619 	return ret;
3620 }
3621 
3622 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3623 			void *buf, int len, int write)
3624 {
3625 	resource_size_t phys_addr;
3626 	unsigned long prot = 0;
3627 	void __iomem *maddr;
3628 	int offset = addr & (PAGE_SIZE-1);
3629 
3630 	if (follow_phys(vma, addr, write, &prot, &phys_addr))
3631 		return -EINVAL;
3632 
3633 	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3634 	if (write)
3635 		memcpy_toio(maddr + offset, buf, len);
3636 	else
3637 		memcpy_fromio(buf, maddr + offset, len);
3638 	iounmap(maddr);
3639 
3640 	return len;
3641 }
3642 EXPORT_SYMBOL_GPL(generic_access_phys);
3643 #endif
3644 
3645 /*
3646  * Access another process' address space as given in mm.  If non-NULL, use the
3647  * given task for page fault accounting.
3648  */
3649 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3650 		unsigned long addr, void *buf, int len, int write)
3651 {
3652 	struct vm_area_struct *vma;
3653 	void *old_buf = buf;
3654 
3655 	down_read(&mm->mmap_sem);
3656 	/* ignore errors, just check how much was successfully transferred */
3657 	while (len) {
3658 		int bytes, ret, offset;
3659 		void *maddr;
3660 		struct page *page = NULL;
3661 
3662 		ret = get_user_pages(tsk, mm, addr, 1,
3663 				write, 1, &page, &vma);
3664 		if (ret <= 0) {
3665 #ifndef CONFIG_HAVE_IOREMAP_PROT
3666 			break;
3667 #else
3668 			/*
3669 			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3670 			 * we can access using slightly different code.
3671 			 */
3672 			vma = find_vma(mm, addr);
3673 			if (!vma || vma->vm_start > addr)
3674 				break;
3675 			if (vma->vm_ops && vma->vm_ops->access)
3676 				ret = vma->vm_ops->access(vma, addr, buf,
3677 							  len, write);
3678 			if (ret <= 0)
3679 				break;
3680 			bytes = ret;
3681 #endif
3682 		} else {
3683 			bytes = len;
3684 			offset = addr & (PAGE_SIZE-1);
3685 			if (bytes > PAGE_SIZE-offset)
3686 				bytes = PAGE_SIZE-offset;
3687 
3688 			maddr = kmap(page);
3689 			if (write) {
3690 				copy_to_user_page(vma, page, addr,
3691 						  maddr + offset, buf, bytes);
3692 				set_page_dirty_lock(page);
3693 			} else {
3694 				copy_from_user_page(vma, page, addr,
3695 						    buf, maddr + offset, bytes);
3696 			}
3697 			kunmap(page);
3698 			page_cache_release(page);
3699 		}
3700 		len -= bytes;
3701 		buf += bytes;
3702 		addr += bytes;
3703 	}
3704 	up_read(&mm->mmap_sem);
3705 
3706 	return buf - old_buf;
3707 }
3708 
3709 /**
3710  * access_remote_vm - access another process' address space
3711  * @mm:		the mm_struct of the target address space
3712  * @addr:	start address to access
3713  * @buf:	source or destination buffer
3714  * @len:	number of bytes to transfer
3715  * @write:	whether the access is a write
3716  *
3717  * The caller must hold a reference on @mm.
3718  */
3719 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3720 		void *buf, int len, int write)
3721 {
3722 	return __access_remote_vm(NULL, mm, addr, buf, len, write);
3723 }
3724 
3725 /*
3726  * Access another process' address space.
3727  * Source/target buffer must be kernel space,
3728  * Do not walk the page table directly, use get_user_pages
3729  */
3730 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3731 		void *buf, int len, int write)
3732 {
3733 	struct mm_struct *mm;
3734 	int ret;
3735 
3736 	mm = get_task_mm(tsk);
3737 	if (!mm)
3738 		return 0;
3739 
3740 	ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3741 	mmput(mm);
3742 
3743 	return ret;
3744 }
3745 
3746 /*
3747  * Print the name of a VMA.
3748  */
3749 void print_vma_addr(char *prefix, unsigned long ip)
3750 {
3751 	struct mm_struct *mm = current->mm;
3752 	struct vm_area_struct *vma;
3753 
3754 	/*
3755 	 * Do not print if we are in atomic
3756 	 * contexts (in exception stacks, etc.):
3757 	 */
3758 	if (preempt_count())
3759 		return;
3760 
3761 	down_read(&mm->mmap_sem);
3762 	vma = find_vma(mm, ip);
3763 	if (vma && vma->vm_file) {
3764 		struct file *f = vma->vm_file;
3765 		char *buf = (char *)__get_free_page(GFP_KERNEL);
3766 		if (buf) {
3767 			char *p;
3768 
3769 			p = file_path(f, buf, PAGE_SIZE);
3770 			if (IS_ERR(p))
3771 				p = "?";
3772 			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3773 					vma->vm_start,
3774 					vma->vm_end - vma->vm_start);
3775 			free_page((unsigned long)buf);
3776 		}
3777 	}
3778 	up_read(&mm->mmap_sem);
3779 }
3780 
3781 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3782 void __might_fault(const char *file, int line)
3783 {
3784 	/*
3785 	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3786 	 * holding the mmap_sem, this is safe because kernel memory doesn't
3787 	 * get paged out, therefore we'll never actually fault, and the
3788 	 * below annotations will generate false positives.
3789 	 */
3790 	if (segment_eq(get_fs(), KERNEL_DS))
3791 		return;
3792 	if (pagefault_disabled())
3793 		return;
3794 	__might_sleep(file, line, 0);
3795 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3796 	if (current->mm)
3797 		might_lock_read(&current->mm->mmap_sem);
3798 #endif
3799 }
3800 EXPORT_SYMBOL(__might_fault);
3801 #endif
3802 
3803 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3804 static void clear_gigantic_page(struct page *page,
3805 				unsigned long addr,
3806 				unsigned int pages_per_huge_page)
3807 {
3808 	int i;
3809 	struct page *p = page;
3810 
3811 	might_sleep();
3812 	for (i = 0; i < pages_per_huge_page;
3813 	     i++, p = mem_map_next(p, page, i)) {
3814 		cond_resched();
3815 		clear_user_highpage(p, addr + i * PAGE_SIZE);
3816 	}
3817 }
3818 void clear_huge_page(struct page *page,
3819 		     unsigned long addr, unsigned int pages_per_huge_page)
3820 {
3821 	int i;
3822 
3823 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3824 		clear_gigantic_page(page, addr, pages_per_huge_page);
3825 		return;
3826 	}
3827 
3828 	might_sleep();
3829 	for (i = 0; i < pages_per_huge_page; i++) {
3830 		cond_resched();
3831 		clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3832 	}
3833 }
3834 
3835 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3836 				    unsigned long addr,
3837 				    struct vm_area_struct *vma,
3838 				    unsigned int pages_per_huge_page)
3839 {
3840 	int i;
3841 	struct page *dst_base = dst;
3842 	struct page *src_base = src;
3843 
3844 	for (i = 0; i < pages_per_huge_page; ) {
3845 		cond_resched();
3846 		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3847 
3848 		i++;
3849 		dst = mem_map_next(dst, dst_base, i);
3850 		src = mem_map_next(src, src_base, i);
3851 	}
3852 }
3853 
3854 void copy_user_huge_page(struct page *dst, struct page *src,
3855 			 unsigned long addr, struct vm_area_struct *vma,
3856 			 unsigned int pages_per_huge_page)
3857 {
3858 	int i;
3859 
3860 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3861 		copy_user_gigantic_page(dst, src, addr, vma,
3862 					pages_per_huge_page);
3863 		return;
3864 	}
3865 
3866 	might_sleep();
3867 	for (i = 0; i < pages_per_huge_page; i++) {
3868 		cond_resched();
3869 		copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3870 	}
3871 }
3872 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3873 
3874 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3875 
3876 static struct kmem_cache *page_ptl_cachep;
3877 
3878 void __init ptlock_cache_init(void)
3879 {
3880 	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3881 			SLAB_PANIC, NULL);
3882 }
3883 
3884 bool ptlock_alloc(struct page *page)
3885 {
3886 	spinlock_t *ptl;
3887 
3888 	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3889 	if (!ptl)
3890 		return false;
3891 	page->ptl = ptl;
3892 	return true;
3893 }
3894 
3895 void ptlock_free(struct page *page)
3896 {
3897 	kmem_cache_free(page_ptl_cachep, page->ptl);
3898 }
3899 #endif
3900