linux - mm/memory.c

4123 lines · 110.4 KiB

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 ([email protected])
34	*
35	* 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36	* ([email protected])
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/pfn_t.h>
54	#include <linux/writeback.h>
55	#include <linux/memcontrol.h>
56	#include <linux/mmu_notifier.h>
57	#include <linux/kallsyms.h>
58	#include <linux/swapops.h>
59	#include <linux/elf.h>
60	#include <linux/gfp.h>
61	#include <linux/migrate.h>
62	#include <linux/string.h>
63	#include <linux/dma-debug.h>
64	#include <linux/debugfs.h>
65	#include <linux/userfaultfd_k.h>
66	#include <linux/dax.h>
67
68	#include <asm/io.h>
69	#include <asm/mmu_context.h>
70	#include <asm/pgalloc.h>
71	#include <asm/uaccess.h>
72	#include <asm/tlb.h>
73	#include <asm/tlbflush.h>
74	#include <asm/pgtable.h>
75
76	#include "internal.h"
77
78	#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
79	#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
80	#endif
81
82	#ifndef CONFIG_NEED_MULTIPLE_NODES
83	/* use the per-pgdat data instead for discontigmem - mbligh */
84	unsigned long max_mapnr;
85	struct page *mem_map;
86
87	EXPORT_SYMBOL(max_mapnr);
88	EXPORT_SYMBOL(mem_map);
89	#endif
90
91	/*
92	* A number of key systems in x86 including ioremap() rely on the assumption
93	* that high_memory defines the upper bound on direct map memory, then end
94	* of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
95	* highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96	* and ZONE_HIGHMEM.
97	*/
98	void * high_memory;
99
100	EXPORT_SYMBOL(high_memory);
101
102	/*
103	* Randomize the address space (stacks, mmaps, brk, etc.).
104	*
105	* ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
106	* as ancient (libc5 based) binaries can segfault. )
107	*/
108	int randomize_va_space __read_mostly =
109	#ifdef CONFIG_COMPAT_BRK
110	1;
111	#else
112	2;
113	#endif
114
115	static int __init disable_randmaps(char *s)
116	{
117	randomize_va_space = 0;
118	return 1;
119	}
120	__setup("norandmaps", disable_randmaps);
121
122	unsigned long zero_pfn __read_mostly;
123	unsigned long highest_memmap_pfn __read_mostly;
124
125	EXPORT_SYMBOL(zero_pfn);
126
127	/*
128	* CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
129	*/
130	static int __init init_zero_pfn(void)
131	{
132	zero_pfn = page_to_pfn(ZERO_PAGE(0));
133	return 0;
134	}
135	core_initcall(init_zero_pfn);
136
137
138	#if defined(SPLIT_RSS_COUNTING)
139
140	void sync_mm_rss(struct mm_struct *mm)
141	{
142	int i;
143
144	for (i = 0; i < NR_MM_COUNTERS; i++) {
145	if (current->rss_stat.count[i]) {
146	add_mm_counter(mm, i, current->rss_stat.count[i]);
147	current->rss_stat.count[i] = 0;
148	}
149	}
150	current->rss_stat.events = 0;
151	}
152
153	static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
154	{
155	struct task_struct *task = current;
156
157	if (likely(task->mm == mm))
158	task->rss_stat.count[member] += val;
159	else
160	add_mm_counter(mm, member, val);
161	}
162	#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
163	#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
164
165	/* sync counter once per 64 page faults */
166	#define TASK_RSS_EVENTS_THRESH (64)
167	static void check_sync_rss_stat(struct task_struct *task)
168	{
169	if (unlikely(task != current))
170	return;
171	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
172	sync_mm_rss(task->mm);
173	}
174	#else /* SPLIT_RSS_COUNTING */
175
176	#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
177	#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
178
179	static void check_sync_rss_stat(struct task_struct *task)
180	{
181	}
182
183	#endif /* SPLIT_RSS_COUNTING */
184
185	#ifdef HAVE_GENERIC_MMU_GATHER
186
187	static bool tlb_next_batch(struct mmu_gather *tlb)
188	{
189	struct mmu_gather_batch *batch;
190
191	batch = tlb->active;
192	if (batch->next) {
193	tlb->active = batch->next;
194	return true;
195	}
196
197	if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
198	return false;
199
200	batch = (void *)__get_free_pages(GFP_NOWAIT \| __GFP_NOWARN, 0);
201	if (!batch)
202	return false;
203
204	tlb->batch_count++;
205	batch->next = NULL;
206	batch->nr = 0;
207	batch->max = MAX_GATHER_BATCH;
208
209	tlb->active->next = batch;
210	tlb->active = batch;
211
212	return true;
213	}
214
215	/* tlb_gather_mmu
216	* Called to initialize an (on-stack) mmu_gather structure for page-table
217	* tear-down from @mm. The @fullmm argument is used when @mm is without
218	* users and we're going to destroy the full address space (exit/execve).
219	*/
220	void tlb_gather_mmu(struct mmu_gather tlb, struct mm_struct mm, unsigned long start, unsigned long end)
221	{
222	tlb->mm = mm;
223
224	/* Is it from 0 to ~0? */
225	tlb->fullmm = !(start \| (end+1));
226	tlb->need_flush_all = 0;
227	tlb->local.next = NULL;
228	tlb->local.nr = 0;
229	tlb->local.max = ARRAY_SIZE(tlb->__pages);
230	tlb->active = &tlb->local;
231	tlb->batch_count = 0;
232
233	#ifdef CONFIG_HAVE_RCU_TABLE_FREE
234	tlb->batch = NULL;
235	#endif
236	tlb->page_size = 0;
237
238	__tlb_reset_range(tlb);
239	}
240
241	static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
242	{
243	if (!tlb->end)
244	return;
245
246	tlb_flush(tlb);
247	mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248	#ifdef CONFIG_HAVE_RCU_TABLE_FREE
249	tlb_table_flush(tlb);
250	#endif
251	__tlb_reset_range(tlb);
252	}
253
254	static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255	{
256	struct mmu_gather_batch *batch;
257
258	for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259	free_pages_and_swap_cache(batch->pages, batch->nr);
260	batch->nr = 0;
261	}
262	tlb->active = &tlb->local;
263	}
264
265	void tlb_flush_mmu(struct mmu_gather *tlb)
266	{
267	tlb_flush_mmu_tlbonly(tlb);
268	tlb_flush_mmu_free(tlb);
269	}
270
271	/* tlb_finish_mmu
272	* Called at the end of the shootdown operation to free up any resources
273	* that were required.
274	*/
275	void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
276	{
277	struct mmu_gather_batch batch, next;
278
279	tlb_flush_mmu(tlb);
280
281	/* keep the page table cache within bounds */
282	check_pgt_cache();
283
284	for (batch = tlb->local.next; batch; batch = next) {
285	next = batch->next;
286	free_pages((unsigned long)batch, 0);
287	}
288	tlb->local.next = NULL;
289	}
290
291	/* __tlb_remove_page
292	* Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293	* handling the additional races in SMP caused by other CPUs caching valid
294	* mappings in their TLBs. Returns the number of free page slots left.
295	* When out of page slots we must call tlb_flush_mmu().
296	*returns true if the caller should flush.
297	*/
298	bool __tlb_remove_page_size(struct mmu_gather tlb, struct page page, int page_size)
299	{
300	struct mmu_gather_batch *batch;
301
302	VM_BUG_ON(!tlb->end);
303
304	if (!tlb->page_size)
305	tlb->page_size = page_size;
306	else {
307	if (page_size != tlb->page_size)
308	return true;
309	}
310
311	batch = tlb->active;
312	if (batch->nr == batch->max) {
313	if (!tlb_next_batch(tlb))
314	return true;
315	batch = tlb->active;
316	}
317	VM_BUG_ON_PAGE(batch->nr > batch->max, page);
318
319	batch->pages[batch->nr++] = page;
320	return false;
321	}
322
323	#endif /* HAVE_GENERIC_MMU_GATHER */
324
325	#ifdef CONFIG_HAVE_RCU_TABLE_FREE
326
327	/*
328	* See the comment near struct mmu_table_batch.
329	*/
330
331	static void tlb_remove_table_smp_sync(void *arg)
332	{
333	/* Simply deliver the interrupt */
334	}
335
336	static void tlb_remove_table_one(void *table)
337	{
338	/*
339	* This isn't an RCU grace period and hence the page-tables cannot be
340	* assumed to be actually RCU-freed.
341	*
342	* It is however sufficient for software page-table walkers that rely on
343	* IRQ disabling. See the comment near struct mmu_table_batch.
344	*/
345	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
346	__tlb_remove_table(table);
347	}
348
349	static void tlb_remove_table_rcu(struct rcu_head *head)
350	{
351	struct mmu_table_batch *batch;
352	int i;
353
354	batch = container_of(head, struct mmu_table_batch, rcu);
355
356	for (i = 0; i < batch->nr; i++)
357	__tlb_remove_table(batch->tables[i]);
358
359	free_page((unsigned long)batch);
360	}
361
362	void tlb_table_flush(struct mmu_gather *tlb)
363	{
364	struct mmu_table_batch **batch = &tlb->batch;
365
366	if (*batch) {
367	call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
368	*batch = NULL;
369	}
370	}
371
372	void tlb_remove_table(struct mmu_gather tlb, void table)
373	{
374	struct mmu_table_batch **batch = &tlb->batch;
375
376	/*
377	* When there's less then two users of this mm there cannot be a
378	* concurrent page-table walk.
379	*/
380	if (atomic_read(&tlb->mm->mm_users) < 2) {
381	__tlb_remove_table(table);
382	return;
383	}
384
385	if (*batch == NULL) {
386	batch = (struct mmu_table_batch )__get_free_page(GFP_NOWAIT \| __GFP_NOWARN);
387	if (*batch == NULL) {
388	tlb_remove_table_one(table);
389	return;
390	}
391	(*batch)->nr = 0;
392	}
393	(batch)->tables[(batch)->nr++] = table;
394	if ((*batch)->nr == MAX_TABLE_BATCH)
395	tlb_table_flush(tlb);
396	}
397
398	#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
399
400	/*
401	* Note: this doesn't free the actual pages themselves. That
402	* has been handled earlier when unmapping all the memory regions.
403	*/
404	static void free_pte_range(struct mmu_gather tlb, pmd_t pmd,
405	unsigned long addr)
406	{
407	pgtable_t token = pmd_pgtable(*pmd);
408	pmd_clear(pmd);
409	pte_free_tlb(tlb, token, addr);
410	atomic_long_dec(&tlb->mm->nr_ptes);
411	}
412
413	static inline void free_pmd_range(struct mmu_gather tlb, pud_t pud,
414	unsigned long addr, unsigned long end,
415	unsigned long floor, unsigned long ceiling)
416	{
417	pmd_t *pmd;
418	unsigned long next;
419	unsigned long start;
420
421	start = addr;
422	pmd = pmd_offset(pud, addr);
423	do {
424	next = pmd_addr_end(addr, end);
425	if (pmd_none_or_clear_bad(pmd))
426	continue;
427	free_pte_range(tlb, pmd, addr);
428	} while (pmd++, addr = next, addr != end);
429
430	start &= PUD_MASK;
431	if (start < floor)
432	return;
433	if (ceiling) {
434	ceiling &= PUD_MASK;
435	if (!ceiling)
436	return;
437	}
438	if (end - 1 > ceiling - 1)
439	return;
440
441	pmd = pmd_offset(pud, start);
442	pud_clear(pud);
443	pmd_free_tlb(tlb, pmd, start);
444	mm_dec_nr_pmds(tlb->mm);
445	}
446
447	static inline void free_pud_range(struct mmu_gather tlb, pgd_t pgd,
448	unsigned long addr, unsigned long end,
449	unsigned long floor, unsigned long ceiling)
450	{
451	pud_t *pud;
452	unsigned long next;
453	unsigned long start;
454
455	start = addr;
456	pud = pud_offset(pgd, addr);
457	do {
458	next = pud_addr_end(addr, end);
459	if (pud_none_or_clear_bad(pud))
460	continue;
461	free_pmd_range(tlb, pud, addr, next, floor, ceiling);
462	} while (pud++, addr = next, addr != end);
463
464	start &= PGDIR_MASK;
465	if (start < floor)
466	return;
467	if (ceiling) {
468	ceiling &= PGDIR_MASK;
469	if (!ceiling)
470	return;
471	}
472	if (end - 1 > ceiling - 1)
473	return;
474
475	pud = pud_offset(pgd, start);
476	pgd_clear(pgd);
477	pud_free_tlb(tlb, pud, start);
478	}
479
480	/*
481	* This function frees user-level page tables of a process.
482	*/
483	void free_pgd_range(struct mmu_gather *tlb,
484	unsigned long addr, unsigned long end,
485	unsigned long floor, unsigned long ceiling)
486	{
487	pgd_t *pgd;
488	unsigned long next;
489
490	/*
491	* The next few lines have given us lots of grief...
492	*
493	* Why are we testing PMD* at this top level? Because often
494	* there will be no work to do at all, and we'd prefer not to
495	* go all the way down to the bottom just to discover that.
496	*
497	* Why all these "- 1"s? Because 0 represents both the bottom
498	* of the address space and the top of it (using -1 for the
499	* top wouldn't help much: the masks would do the wrong thing).
500	* The rule is that addr 0 and floor 0 refer to the bottom of
501	* the address space, but end 0 and ceiling 0 refer to the top
502	* Comparisons need to use "end - 1" and "ceiling - 1" (though
503	* that end 0 case should be mythical).
504	*
505	* Wherever addr is brought up or ceiling brought down, we must
506	* be careful to reject "the opposite 0" before it confuses the
507	* subsequent tests. But what about where end is brought down
508	* by PMD_SIZE below? no, end can't go down to 0 there.
509	*
510	* Whereas we round start (addr) and ceiling down, by different
511	* masks at different levels, in order to test whether a table
512	* now has no other vmas using it, so can be freed, we don't
513	* bother to round floor or end up - the tests don't need that.
514	*/
515
516	addr &= PMD_MASK;
517	if (addr < floor) {
518	addr += PMD_SIZE;
519	if (!addr)
520	return;
521	}
522	if (ceiling) {
523	ceiling &= PMD_MASK;
524	if (!ceiling)
525	return;
526	}
527	if (end - 1 > ceiling - 1)
528	end -= PMD_SIZE;
529	if (addr > end - 1)
530	return;
531
532	pgd = pgd_offset(tlb->mm, addr);
533	do {
534	next = pgd_addr_end(addr, end);
535	if (pgd_none_or_clear_bad(pgd))
536	continue;
537	free_pud_range(tlb, pgd, addr, next, floor, ceiling);
538	} while (pgd++, addr = next, addr != end);
539	}
540
541	void free_pgtables(struct mmu_gather tlb, struct vm_area_struct vma,
542	unsigned long floor, unsigned long ceiling)
543	{
544	while (vma) {
545	struct vm_area_struct *next = vma->vm_next;
546	unsigned long addr = vma->vm_start;
547
548	/*
549	* Hide vma from rmap and truncate_pagecache before freeing
550	* pgtables
551	*/
552	unlink_anon_vmas(vma);
553	unlink_file_vma(vma);
554
555	if (is_vm_hugetlb_page(vma)) {
556	hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
557	floor, next? next->vm_start: ceiling);
558	} else {
559	/*
560	* Optimization: gather nearby vmas into one call down
561	*/
562	while (next && next->vm_start <= vma->vm_end + PMD_SIZE
563	&& !is_vm_hugetlb_page(next)) {
564	vma = next;
565	next = vma->vm_next;
566	unlink_anon_vmas(vma);
567	unlink_file_vma(vma);
568	}
569	free_pgd_range(tlb, addr, vma->vm_end,
570	floor, next? next->vm_start: ceiling);
571	}
572	vma = next;
573	}
574	}
575
576	int __pte_alloc(struct mm_struct mm, pmd_t pmd, unsigned long address)
577	{
578	spinlock_t *ptl;
579	pgtable_t new = pte_alloc_one(mm, address);
580	if (!new)
581	return -ENOMEM;
582
583	/*
584	* Ensure all pte setup (eg. pte page lock and page clearing) are
585	* visible before the pte is made visible to other CPUs by being
586	* put into page tables.
587	*
588	* The other side of the story is the pointer chasing in the page
589	* table walking code (when walking the page table without locking;
590	* ie. most of the time). Fortunately, these data accesses consist
591	* of a chain of data-dependent loads, meaning most CPUs (alpha
592	* being the notable exception) will already guarantee loads are
593	* seen in-order. See the alpha page table accessors for the
594	* smp_read_barrier_depends() barriers in page table walking code.
595	*/
596	smp_wmb(); /* Could be smp_wmb__xxx(before\|after)_spin_lock */
597
598	ptl = pmd_lock(mm, pmd);
599	if (likely(pmd_none(pmd))) { / Has another populated it ? */
600	atomic_long_inc(&mm->nr_ptes);
601	pmd_populate(mm, pmd, new);
602	new = NULL;
603	}
604	spin_unlock(ptl);
605	if (new)
606	pte_free(mm, new);
607	return 0;
608	}
609
610	int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
611	{
612	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
613	if (!new)
614	return -ENOMEM;
615
616	smp_wmb(); /* See comment in __pte_alloc */
617
618	spin_lock(&init_mm.page_table_lock);
619	if (likely(pmd_none(pmd))) { / Has another populated it ? */
620	pmd_populate_kernel(&init_mm, pmd, new);
621	new = NULL;
622	}
623	spin_unlock(&init_mm.page_table_lock);
624	if (new)
625	pte_free_kernel(&init_mm, new);
626	return 0;
627	}
628
629	static inline void init_rss_vec(int *rss)
630	{
631	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
632	}
633
634	static inline void add_mm_rss_vec(struct mm_struct mm, int rss)
635	{
636	int i;
637
638	if (current->mm == mm)
639	sync_mm_rss(mm);
640	for (i = 0; i < NR_MM_COUNTERS; i++)
641	if (rss[i])
642	add_mm_counter(mm, i, rss[i]);
643	}
644
645	/*
646	* This function is called to print an error when a bad pte
647	* is found. For example, we might have a PFN-mapped pte in
648	* a region that doesn't allow it.
649	*
650	* The calling function must still handle the error.
651	*/
652	static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
653	pte_t pte, struct page *page)
654	{
655	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
656	pud_t *pud = pud_offset(pgd, addr);
657	pmd_t *pmd = pmd_offset(pud, addr);
658	struct address_space *mapping;
659	pgoff_t index;
660	static unsigned long resume;
661	static unsigned long nr_shown;
662	static unsigned long nr_unshown;
663
664	/*
665	* Allow a burst of 60 reports, then keep quiet for that minute;
666	* or allow a steady drip of one report per second.
667	*/
668	if (nr_shown == 60) {
669	if (time_before(jiffies, resume)) {
670	nr_unshown++;
671	return;
672	}
673	if (nr_unshown) {
674	pr_alert("BUG: Bad page map: %lu messages suppressed\n",
675	nr_unshown);
676	nr_unshown = 0;
677	}
678	nr_shown = 0;
679	}
680	if (nr_shown++ == 0)
681	resume = jiffies + 60 * HZ;
682
683	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
684	index = linear_page_index(vma, addr);
685
686	pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
687	current->comm,
688	(long long)pte_val(pte), (long long)pmd_val(*pmd));
689	if (page)
690	dump_page(page, "bad pte");
691	pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
692	(void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
693	/*
694	* Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
695	*/
696	pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
697	vma->vm_file,
698	vma->vm_ops ? vma->vm_ops->fault : NULL,
699	vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
700	mapping ? mapping->a_ops->readpage : NULL);
701	dump_stack();
702	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
703	}
704
705	/*
706	* vm_normal_page -- This function gets the "struct page" associated with a pte.
707	*
708	* "Special" mappings do not wish to be associated with a "struct page" (either
709	* it doesn't exist, or it exists but they don't want to touch it). In this
710	* case, NULL is returned here. "Normal" mappings do have a struct page.
711	*
712	* There are 2 broad cases. Firstly, an architecture may define a pte_special()
713	* pte bit, in which case this function is trivial. Secondly, an architecture
714	* may not have a spare pte bit, which requires a more complicated scheme,
715	* described below.
716	*
717	* A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
718	* special mapping (even if there are underlying and valid "struct pages").
719	* COWed pages of a VM_PFNMAP are always normal.
720	*
721	* The way we recognize COWed pages within VM_PFNMAP mappings is through the
722	* rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
723	* set, and the vm_pgoff will point to the first PFN mapped: thus every special
724	* mapping will always honor the rule
725	*
726	* pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
727	*
728	* And for normal mappings this is false.
729	*
730	* This restricts such mappings to be a linear translation from virtual address
731	* to pfn. To get around this restriction, we allow arbitrary mappings so long
732	* as the vma is not a COW mapping; in that case, we know that all ptes are
733	* special (because none can have been COWed).
734	*
735	*
736	* In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
737	*
738	* VM_MIXEDMAP mappings can likewise contain memory with or without "struct
739	* page" backing, however the difference is that _all_ pages with a struct
740	* page (that is, those where pfn_valid is true) are refcounted and considered
741	* normal pages by the VM. The disadvantage is that pages are refcounted
742	* (which can be slower and simply not an option for some PFNMAP users). The
743	* advantage is that we don't have to follow the strict linearity rule of
744	* PFNMAP mappings in order to support COWable mappings.
745	*
746	*/
747	#ifdef __HAVE_ARCH_PTE_SPECIAL
748	# define HAVE_PTE_SPECIAL 1
749	#else
750	# define HAVE_PTE_SPECIAL 0
751	#endif
752	struct page vm_normal_page(struct vm_area_struct vma, unsigned long addr,
753	pte_t pte)
754	{
755	unsigned long pfn = pte_pfn(pte);
756
757	if (HAVE_PTE_SPECIAL) {
758	if (likely(!pte_special(pte)))
759	goto check_pfn;
760	if (vma->vm_ops && vma->vm_ops->find_special_page)
761	return vma->vm_ops->find_special_page(vma, addr);
762	if (vma->vm_flags & (VM_PFNMAP \| VM_MIXEDMAP))
763	return NULL;
764	if (!is_zero_pfn(pfn))
765	print_bad_pte(vma, addr, pte, NULL);
766	return NULL;
767	}
768
769	/* !HAVE_PTE_SPECIAL case follows: */
770
771	if (unlikely(vma->vm_flags & (VM_PFNMAP\|VM_MIXEDMAP))) {
772	if (vma->vm_flags & VM_MIXEDMAP) {
773	if (!pfn_valid(pfn))
774	return NULL;
775	goto out;
776	} else {
777	unsigned long off;
778	off = (addr - vma->vm_start) >> PAGE_SHIFT;
779	if (pfn == vma->vm_pgoff + off)
780	return NULL;
781	if (!is_cow_mapping(vma->vm_flags))
782	return NULL;
783	}
784	}
785
786	if (is_zero_pfn(pfn))
787	return NULL;
788	check_pfn:
789	if (unlikely(pfn > highest_memmap_pfn)) {
790	print_bad_pte(vma, addr, pte, NULL);
791	return NULL;
792	}
793
794	/*
795	* NOTE! We still have PageReserved() pages in the page tables.
796	* eg. VDSO mappings can cause them to exist.
797	*/
798	out:
799	return pfn_to_page(pfn);
800	}
801
802	#ifdef CONFIG_TRANSPARENT_HUGEPAGE
803	struct page vm_normal_page_pmd(struct vm_area_struct vma, unsigned long addr,
804	pmd_t pmd)
805	{
806	unsigned long pfn = pmd_pfn(pmd);
807
808	/*
809	* There is no pmd_special() but there may be special pmds, e.g.
810	* in a direct-access (dax) mapping, so let's just replicate the
811	* !HAVE_PTE_SPECIAL case from vm_normal_page() here.
812	*/
813	if (unlikely(vma->vm_flags & (VM_PFNMAP\|VM_MIXEDMAP))) {
814	if (vma->vm_flags & VM_MIXEDMAP) {
815	if (!pfn_valid(pfn))
816	return NULL;
817	goto out;
818	} else {
819	unsigned long off;
820	off = (addr - vma->vm_start) >> PAGE_SHIFT;
821	if (pfn == vma->vm_pgoff + off)
822	return NULL;
823	if (!is_cow_mapping(vma->vm_flags))
824	return NULL;
825	}
826	}
827
828	if (is_zero_pfn(pfn))
829	return NULL;
830	if (unlikely(pfn > highest_memmap_pfn))
831	return NULL;
832
833	/*
834	* NOTE! We still have PageReserved() pages in the page tables.
835	* eg. VDSO mappings can cause them to exist.
836	*/
837	out:
838	return pfn_to_page(pfn);
839	}
840	#endif
841
842	/*
843	* copy one vm_area from one task to the other. Assumes the page tables
844	* already present in the new task to be cleared in the whole range
845	* covered by this vma.
846	*/
847
848	static inline unsigned long
849	copy_one_pte(struct mm_struct dst_mm, struct mm_struct src_mm,
850	pte_t dst_pte, pte_t src_pte, struct vm_area_struct *vma,
851	unsigned long addr, int *rss)
852	{
853	unsigned long vm_flags = vma->vm_flags;
854	pte_t pte = *src_pte;
855	struct page *page;
856
857	/* pte contains position in swap or file, so copy. */
858	if (unlikely(!pte_present(pte))) {
859	swp_entry_t entry = pte_to_swp_entry(pte);
860
861	if (likely(!non_swap_entry(entry))) {
862	if (swap_duplicate(entry) < 0)
863	return entry.val;
864
865	/* make sure dst_mm is on swapoff's mmlist. */
866	if (unlikely(list_empty(&dst_mm->mmlist))) {
867	spin_lock(&mmlist_lock);
868	if (list_empty(&dst_mm->mmlist))
869	list_add(&dst_mm->mmlist,
870	&src_mm->mmlist);
871	spin_unlock(&mmlist_lock);
872	}
873	rss[MM_SWAPENTS]++;
874	} else if (is_migration_entry(entry)) {
875	page = migration_entry_to_page(entry);
876
877	rss[mm_counter(page)]++;
878
879	if (is_write_migration_entry(entry) &&
880	is_cow_mapping(vm_flags)) {
881	/*
882	* COW mappings require pages in both
883	* parent and child to be set to read.
884	*/
885	make_migration_entry_read(&entry);
886	pte = swp_entry_to_pte(entry);
887	if (pte_swp_soft_dirty(*src_pte))
888	pte = pte_swp_mksoft_dirty(pte);
889	set_pte_at(src_mm, addr, src_pte, pte);
890	}
891	}
892	goto out_set_pte;
893	}
894
895	/*
896	* If it's a COW mapping, write protect it both
897	* in the parent and the child
898	*/
899	if (is_cow_mapping(vm_flags)) {
900	ptep_set_wrprotect(src_mm, addr, src_pte);
901	pte = pte_wrprotect(pte);
902	}
903
904	/*
905	* If it's a shared mapping, mark it clean in
906	* the child
907	*/
908	if (vm_flags & VM_SHARED)
909	pte = pte_mkclean(pte);
910	pte = pte_mkold(pte);
911
912	page = vm_normal_page(vma, addr, pte);
913	if (page) {
914	get_page(page);
915	page_dup_rmap(page, false);
916	rss[mm_counter(page)]++;
917	}
918
919	out_set_pte:
920	set_pte_at(dst_mm, addr, dst_pte, pte);
921	return 0;
922	}
923
924	static int copy_pte_range(struct mm_struct dst_mm, struct mm_struct src_mm,
925	pmd_t dst_pmd, pmd_t src_pmd, struct vm_area_struct *vma,
926	unsigned long addr, unsigned long end)
927	{
928	pte_t orig_src_pte, orig_dst_pte;
929	pte_t src_pte, dst_pte;
930	spinlock_t src_ptl, dst_ptl;
931	int progress = 0;
932	int rss[NR_MM_COUNTERS];
933	swp_entry_t entry = (swp_entry_t){0};
934
935	again:
936	init_rss_vec(rss);
937
938	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
939	if (!dst_pte)
940	return -ENOMEM;
941	src_pte = pte_offset_map(src_pmd, addr);
942	src_ptl = pte_lockptr(src_mm, src_pmd);
943	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
944	orig_src_pte = src_pte;
945	orig_dst_pte = dst_pte;
946	arch_enter_lazy_mmu_mode();
947
948	do {
949	/*
950	* We are holding two locks at this point - either of them
951	* could generate latencies in another task on another CPU.
952	*/
953	if (progress >= 32) {
954	progress = 0;
955	if (need_resched() \|\|
956	spin_needbreak(src_ptl) \|\| spin_needbreak(dst_ptl))
957	break;
958	}
959	if (pte_none(*src_pte)) {
960	progress++;
961	continue;
962	}
963	entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
964	vma, addr, rss);
965	if (entry.val)
966	break;
967	progress += 8;
968	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
969
970	arch_leave_lazy_mmu_mode();
971	spin_unlock(src_ptl);
972	pte_unmap(orig_src_pte);
973	add_mm_rss_vec(dst_mm, rss);
974	pte_unmap_unlock(orig_dst_pte, dst_ptl);
975	cond_resched();
976
977	if (entry.val) {
978	if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
979	return -ENOMEM;
980	progress = 0;
981	}
982	if (addr != end)
983	goto again;
984	return 0;
985	}
986
987	static inline int copy_pmd_range(struct mm_struct dst_mm, struct mm_struct src_mm,
988	pud_t dst_pud, pud_t src_pud, struct vm_area_struct *vma,
989	unsigned long addr, unsigned long end)
990	{
991	pmd_t src_pmd, dst_pmd;
992	unsigned long next;
993
994	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
995	if (!dst_pmd)
996	return -ENOMEM;
997	src_pmd = pmd_offset(src_pud, addr);
998	do {
999	next = pmd_addr_end(addr, end);
1000	if (pmd_trans_huge(src_pmd) \|\| pmd_devmap(src_pmd)) {
1001	int err;
1002	VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1003	err = copy_huge_pmd(dst_mm, src_mm,
1004	dst_pmd, src_pmd, addr, vma);
1005	if (err == -ENOMEM)
1006	return -ENOMEM;
1007	if (!err)
1008	continue;
1009	/* fall through */
1010	}
1011	if (pmd_none_or_clear_bad(src_pmd))
1012	continue;
1013	if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1014	vma, addr, next))
1015	return -ENOMEM;
1016	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1017	return 0;
1018	}
1019
1020	static inline int copy_pud_range(struct mm_struct dst_mm, struct mm_struct src_mm,
1021	pgd_t dst_pgd, pgd_t src_pgd, struct vm_area_struct *vma,
1022	unsigned long addr, unsigned long end)
1023	{
1024	pud_t src_pud, dst_pud;
1025	unsigned long next;
1026
1027	dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1028	if (!dst_pud)
1029	return -ENOMEM;
1030	src_pud = pud_offset(src_pgd, addr);
1031	do {
1032	next = pud_addr_end(addr, end);
1033	if (pud_none_or_clear_bad(src_pud))
1034	continue;
1035	if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1036	vma, addr, next))
1037	return -ENOMEM;
1038	} while (dst_pud++, src_pud++, addr = next, addr != end);
1039	return 0;
1040	}
1041
1042	int copy_page_range(struct mm_struct dst_mm, struct mm_struct src_mm,
1043	struct vm_area_struct *vma)
1044	{
1045	pgd_t src_pgd, dst_pgd;
1046	unsigned long next;
1047	unsigned long addr = vma->vm_start;
1048	unsigned long end = vma->vm_end;
1049	unsigned long mmun_start; /* For mmu_notifiers */
1050	unsigned long mmun_end; /* For mmu_notifiers */
1051	bool is_cow;
1052	int ret;
1053
1054	/*
1055	* Don't copy ptes where a page fault will fill them correctly.
1056	* Fork becomes much lighter when there are big shared or private
1057	* readonly mappings. The tradeoff is that copy_page_range is more
1058	* efficient than faulting.
1059	*/
1060	if (!(vma->vm_flags & (VM_HUGETLB \| VM_PFNMAP \| VM_MIXEDMAP)) &&
1061	!vma->anon_vma)
1062	return 0;
1063
1064	if (is_vm_hugetlb_page(vma))
1065	return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1066
1067	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1068	/*
1069	* We do not free on error cases below as remove_vma
1070	* gets called on error from higher level routine
1071	*/
1072	ret = track_pfn_copy(vma);
1073	if (ret)
1074	return ret;
1075	}
1076
1077	/*
1078	* We need to invalidate the secondary MMU mappings only when
1079	* there could be a permission downgrade on the ptes of the
1080	* parent mm. And a permission downgrade will only happen if
1081	* is_cow_mapping() returns true.
1082	*/
1083	is_cow = is_cow_mapping(vma->vm_flags);
1084	mmun_start = addr;
1085	mmun_end = end;
1086	if (is_cow)
1087	mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1088	mmun_end);
1089
1090	ret = 0;
1091	dst_pgd = pgd_offset(dst_mm, addr);
1092	src_pgd = pgd_offset(src_mm, addr);
1093	do {
1094	next = pgd_addr_end(addr, end);
1095	if (pgd_none_or_clear_bad(src_pgd))
1096	continue;
1097	if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1098	vma, addr, next))) {
1099	ret = -ENOMEM;
1100	break;
1101	}
1102	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1103
1104	if (is_cow)
1105	mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1106	return ret;
1107	}
1108
1109	static unsigned long zap_pte_range(struct mmu_gather *tlb,
1110	struct vm_area_struct vma, pmd_t pmd,
1111	unsigned long addr, unsigned long end,
1112	struct zap_details *details)
1113	{
1114	struct mm_struct *mm = tlb->mm;
1115	int force_flush = 0;
1116	int rss[NR_MM_COUNTERS];
1117	spinlock_t *ptl;
1118	pte_t *start_pte;
1119	pte_t *pte;
1120	swp_entry_t entry;
1121	struct page *pending_page = NULL;
1122
1123	again:
1124	init_rss_vec(rss);
1125	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1126	pte = start_pte;
1127	arch_enter_lazy_mmu_mode();
1128	do {
1129	pte_t ptent = *pte;
1130	if (pte_none(ptent)) {
1131	continue;
1132	}
1133
1134	if (pte_present(ptent)) {
1135	struct page *page;
1136
1137	page = vm_normal_page(vma, addr, ptent);
1138	if (unlikely(details) && page) {
1139	/*
1140	* unmap_shared_mapping_pages() wants to
1141	* invalidate cache without truncating:
1142	* unmap shared but keep private pages.
1143	*/
1144	if (details->check_mapping &&
1145	details->check_mapping != page_rmapping(page))
1146	continue;
1147	}
1148	ptent = ptep_get_and_clear_full(mm, addr, pte,
1149	tlb->fullmm);
1150	tlb_remove_tlb_entry(tlb, pte, addr);
1151	if (unlikely(!page))
1152	continue;
1153
1154	if (!PageAnon(page)) {
1155	if (pte_dirty(ptent)) {
1156	/*
1157	* oom_reaper cannot tear down dirty
1158	* pages
1159	*/
1160	if (unlikely(details && details->ignore_dirty))
1161	continue;
1162	force_flush = 1;
1163	set_page_dirty(page);
1164	}
1165	if (pte_young(ptent) &&
1166	likely(!(vma->vm_flags & VM_SEQ_READ)))
1167	mark_page_accessed(page);
1168	}
1169	rss[mm_counter(page)]--;
1170	page_remove_rmap(page, false);
1171	if (unlikely(page_mapcount(page) < 0))
1172	print_bad_pte(vma, addr, ptent, page);
1173	if (unlikely(__tlb_remove_page(tlb, page))) {
1174	force_flush = 1;
1175	pending_page = page;
1176	addr += PAGE_SIZE;
1177	break;
1178	}
1179	continue;
1180	}
1181	/* only check swap_entries if explicitly asked for in details */
1182	if (unlikely(details && !details->check_swap_entries))
1183	continue;
1184
1185	entry = pte_to_swp_entry(ptent);
1186	if (!non_swap_entry(entry))
1187	rss[MM_SWAPENTS]--;
1188	else if (is_migration_entry(entry)) {
1189	struct page *page;
1190
1191	page = migration_entry_to_page(entry);
1192	rss[mm_counter(page)]--;
1193	}
1194	if (unlikely(!free_swap_and_cache(entry)))
1195	print_bad_pte(vma, addr, ptent, NULL);
1196	pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1197	} while (pte++, addr += PAGE_SIZE, addr != end);
1198
1199	add_mm_rss_vec(mm, rss);
1200	arch_leave_lazy_mmu_mode();
1201
1202	/* Do the actual TLB flush before dropping ptl */
1203	if (force_flush)
1204	tlb_flush_mmu_tlbonly(tlb);
1205	pte_unmap_unlock(start_pte, ptl);
1206
1207	/*
1208	* If we forced a TLB flush (either due to running out of
1209	* batch buffers or because we needed to flush dirty TLB
1210	* entries before releasing the ptl), free the batched
1211	* memory too. Restart if we didn't do everything.
1212	*/
1213	if (force_flush) {
1214	force_flush = 0;
1215	tlb_flush_mmu_free(tlb);
1216	if (pending_page) {
1217	/* remove the page with new size */
1218	__tlb_remove_pte_page(tlb, pending_page);
1219	pending_page = NULL;
1220	}
1221	if (addr != end)
1222	goto again;
1223	}
1224
1225	return addr;
1226	}
1227
1228	static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1229	struct vm_area_struct vma, pud_t pud,
1230	unsigned long addr, unsigned long end,
1231	struct zap_details *details)
1232	{
1233	pmd_t *pmd;
1234	unsigned long next;
1235
1236	pmd = pmd_offset(pud, addr);
1237	do {
1238	next = pmd_addr_end(addr, end);
1239	if (pmd_trans_huge(pmd) \|\| pmd_devmap(pmd)) {
1240	if (next - addr != HPAGE_PMD_SIZE) {
1241	VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1242	!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1243	split_huge_pmd(vma, pmd, addr);
1244	} else if (zap_huge_pmd(tlb, vma, pmd, addr))
1245	goto next;
1246	/* fall through */
1247	}
1248	/*
1249	* Here there can be other concurrent MADV_DONTNEED or
1250	* trans huge page faults running, and if the pmd is
1251	* none or trans huge it can change under us. This is
1252	* because MADV_DONTNEED holds the mmap_sem in read
1253	* mode.
1254	*/
1255	if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1256	goto next;
1257	next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1258	next:
1259	cond_resched();
1260	} while (pmd++, addr = next, addr != end);
1261
1262	return addr;
1263	}
1264
1265	static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1266	struct vm_area_struct vma, pgd_t pgd,
1267	unsigned long addr, unsigned long end,
1268	struct zap_details *details)
1269	{
1270	pud_t *pud;
1271	unsigned long next;
1272
1273	pud = pud_offset(pgd, addr);
1274	do {
1275	next = pud_addr_end(addr, end);
1276	if (pud_none_or_clear_bad(pud))
1277	continue;
1278	next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1279	} while (pud++, addr = next, addr != end);
1280
1281	return addr;
1282	}
1283
1284	void unmap_page_range(struct mmu_gather *tlb,
1285	struct vm_area_struct *vma,
1286	unsigned long addr, unsigned long end,
1287	struct zap_details *details)
1288	{
1289	pgd_t *pgd;
1290	unsigned long next;
1291
1292	BUG_ON(addr >= end);
1293	tlb_start_vma(tlb, vma);
1294	pgd = pgd_offset(vma->vm_mm, addr);
1295	do {
1296	next = pgd_addr_end(addr, end);
1297	if (pgd_none_or_clear_bad(pgd))
1298	continue;
1299	next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1300	} while (pgd++, addr = next, addr != end);
1301	tlb_end_vma(tlb, vma);
1302	}
1303
1304
1305	static void unmap_single_vma(struct mmu_gather *tlb,
1306	struct vm_area_struct *vma, unsigned long start_addr,
1307	unsigned long end_addr,
1308	struct zap_details *details)
1309	{
1310	unsigned long start = max(vma->vm_start, start_addr);
1311	unsigned long end;
1312
1313	if (start >= vma->vm_end)
1314	return;
1315	end = min(vma->vm_end, end_addr);
1316	if (end <= vma->vm_start)
1317	return;
1318
1319	if (vma->vm_file)
1320	uprobe_munmap(vma, start, end);
1321
1322	if (unlikely(vma->vm_flags & VM_PFNMAP))
1323	untrack_pfn(vma, 0, 0);
1324
1325	if (start != end) {
1326	if (unlikely(is_vm_hugetlb_page(vma))) {
1327	/*
1328	* It is undesirable to test vma->vm_file as it
1329	* should be non-null for valid hugetlb area.
1330	* However, vm_file will be NULL in the error
1331	* cleanup path of mmap_region. When
1332	* hugetlbfs ->mmap method fails,
1333	* mmap_region() nullifies vma->vm_file
1334	* before calling this function to clean up.
1335	* Since no pte has actually been setup, it is
1336	* safe to do nothing in this case.
1337	*/
1338	if (vma->vm_file) {
1339	i_mmap_lock_write(vma->vm_file->f_mapping);
1340	__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1341	i_mmap_unlock_write(vma->vm_file->f_mapping);
1342	}
1343	} else
1344	unmap_page_range(tlb, vma, start, end, details);
1345	}
1346	}
1347
1348	/**
1349	* unmap_vmas - unmap a range of memory covered by a list of vma's
1350	* @tlb: address of the caller's struct mmu_gather
1351	* @vma: the starting vma
1352	* @start_addr: virtual address at which to start unmapping
1353	* @end_addr: virtual address at which to end unmapping
1354	*
1355	* Unmap all pages in the vma list.
1356	*
1357	* Only addresses between `start' and `end' will be unmapped.
1358	*
1359	* The VMA list must be sorted in ascending virtual address order.
1360	*
1361	* unmap_vmas() assumes that the caller will flush the whole unmapped address
1362	* range after unmap_vmas() returns. So the only responsibility here is to
1363	* ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1364	* drops the lock and schedules.
1365	*/
1366	void unmap_vmas(struct mmu_gather *tlb,
1367	struct vm_area_struct *vma, unsigned long start_addr,
1368	unsigned long end_addr)
1369	{
1370	struct mm_struct *mm = vma->vm_mm;
1371
1372	mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1373	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1374	unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1375	mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1376	}
1377
1378	/**
1379	* zap_page_range - remove user pages in a given range
1380	* @vma: vm_area_struct holding the applicable pages
1381	* @start: starting address of pages to zap
1382	* @size: number of bytes to zap
1383	* @details: details of shared cache invalidation
1384	*
1385	* Caller must protect the VMA list
1386	*/
1387	void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1388	unsigned long size, struct zap_details *details)
1389	{
1390	struct mm_struct *mm = vma->vm_mm;
1391	struct mmu_gather tlb;
1392	unsigned long end = start + size;
1393
1394	lru_add_drain();
1395	tlb_gather_mmu(&tlb, mm, start, end);
1396	update_hiwater_rss(mm);
1397	mmu_notifier_invalidate_range_start(mm, start, end);
1398	for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1399	unmap_single_vma(&tlb, vma, start, end, details);
1400	mmu_notifier_invalidate_range_end(mm, start, end);
1401	tlb_finish_mmu(&tlb, start, end);
1402	}
1403
1404	/**
1405	* zap_page_range_single - remove user pages in a given range
1406	* @vma: vm_area_struct holding the applicable pages
1407	* @address: starting address of pages to zap
1408	* @size: number of bytes to zap
1409	* @details: details of shared cache invalidation
1410	*
1411	* The range must fit into one VMA.
1412	*/
1413	static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1414	unsigned long size, struct zap_details *details)
1415	{
1416	struct mm_struct *mm = vma->vm_mm;
1417	struct mmu_gather tlb;
1418	unsigned long end = address + size;
1419
1420	lru_add_drain();
1421	tlb_gather_mmu(&tlb, mm, address, end);
1422	update_hiwater_rss(mm);
1423	mmu_notifier_invalidate_range_start(mm, address, end);
1424	unmap_single_vma(&tlb, vma, address, end, details);
1425	mmu_notifier_invalidate_range_end(mm, address, end);
1426	tlb_finish_mmu(&tlb, address, end);
1427	}
1428
1429	/**
1430	* zap_vma_ptes - remove ptes mapping the vma
1431	* @vma: vm_area_struct holding ptes to be zapped
1432	* @address: starting address of pages to zap
1433	* @size: number of bytes to zap
1434	*
1435	* This function only unmaps ptes assigned to VM_PFNMAP vmas.
1436	*
1437	* The entire address range must be fully contained within the vma.
1438	*
1439	* Returns 0 if successful.
1440	*/
1441	int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1442	unsigned long size)
1443	{
1444	if (address < vma->vm_start \|\| address + size > vma->vm_end \|\|
1445	!(vma->vm_flags & VM_PFNMAP))
1446	return -1;
1447	zap_page_range_single(vma, address, size, NULL);
1448	return 0;
1449	}
1450	EXPORT_SYMBOL_GPL(zap_vma_ptes);
1451
1452	pte_t __get_locked_pte(struct mm_struct mm, unsigned long addr,
1453	spinlock_t **ptl)
1454	{
1455	pgd_t * pgd = pgd_offset(mm, addr);
1456	pud_t * pud = pud_alloc(mm, pgd, addr);
1457	if (pud) {
1458	pmd_t * pmd = pmd_alloc(mm, pud, addr);
1459	if (pmd) {
1460	VM_BUG_ON(pmd_trans_huge(*pmd));
1461	return pte_alloc_map_lock(mm, pmd, addr, ptl);
1462	}
1463	}
1464	return NULL;
1465	}
1466
1467	/*
1468	* This is the old fallback for page remapping.
1469	*
1470	* For historical reasons, it only allows reserved pages. Only
1471	* old drivers should use this, and they needed to mark their
1472	* pages reserved for the old functions anyway.
1473	*/
1474	static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1475	struct page *page, pgprot_t prot)
1476	{
1477	struct mm_struct *mm = vma->vm_mm;
1478	int retval;
1479	pte_t *pte;
1480	spinlock_t *ptl;
1481
1482	retval = -EINVAL;
1483	if (PageAnon(page))
1484	goto out;
1485	retval = -ENOMEM;
1486	flush_dcache_page(page);
1487	pte = get_locked_pte(mm, addr, &ptl);
1488	if (!pte)
1489	goto out;
1490	retval = -EBUSY;
1491	if (!pte_none(*pte))
1492	goto out_unlock;
1493
1494	/* Ok, finally just insert the thing.. */
1495	get_page(page);
1496	inc_mm_counter_fast(mm, mm_counter_file(page));
1497	page_add_file_rmap(page, false);
1498	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1499
1500	retval = 0;
1501	pte_unmap_unlock(pte, ptl);
1502	return retval;
1503	out_unlock:
1504	pte_unmap_unlock(pte, ptl);
1505	out:
1506	return retval;
1507	}
1508
1509	/**
1510	* vm_insert_page - insert single page into user vma
1511	* @vma: user vma to map to
1512	* @addr: target user address of this page
1513	* @page: source kernel page
1514	*
1515	* This allows drivers to insert individual pages they've allocated
1516	* into a user vma.
1517	*
1518	* The page has to be a nice clean _individual_ kernel allocation.
1519	* If you allocate a compound page, you need to have marked it as
1520	* such (__GFP_COMP), or manually just split the page up yourself
1521	* (see split_page()).
1522	*
1523	* NOTE! Traditionally this was done with "remap_pfn_range()" which
1524	* took an arbitrary page protection parameter. This doesn't allow
1525	* that. Your vma protection will have to be set up correctly, which
1526	* means that if you want a shared writable mapping, you'd better
1527	* ask for a shared writable mapping!
1528	*
1529	* The page does not need to be reserved.
1530	*
1531	* Usually this function is called from f_op->mmap() handler
1532	* under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1533	* Caller must set VM_MIXEDMAP on vma if it wants to call this
1534	* function from other places, for example from page-fault handler.
1535	*/
1536	int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1537	struct page *page)
1538	{
1539	if (addr < vma->vm_start \|\| addr >= vma->vm_end)
1540	return -EFAULT;
1541	if (!page_count(page))
1542	return -EINVAL;
1543	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1544	BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1545	BUG_ON(vma->vm_flags & VM_PFNMAP);
1546	vma->vm_flags \|= VM_MIXEDMAP;
1547	}
1548	return insert_page(vma, addr, page, vma->vm_page_prot);
1549	}
1550	EXPORT_SYMBOL(vm_insert_page);
1551
1552	static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1553	pfn_t pfn, pgprot_t prot)
1554	{
1555	struct mm_struct *mm = vma->vm_mm;
1556	int retval;
1557	pte_t *pte, entry;
1558	spinlock_t *ptl;
1559
1560	retval = -ENOMEM;
1561	pte = get_locked_pte(mm, addr, &ptl);
1562	if (!pte)
1563	goto out;
1564	retval = -EBUSY;
1565	if (!pte_none(*pte))
1566	goto out_unlock;
1567
1568	/* Ok, finally just insert the thing.. */
1569	if (pfn_t_devmap(pfn))
1570	entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1571	else
1572	entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1573	set_pte_at(mm, addr, pte, entry);
1574	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1575
1576	retval = 0;
1577	out_unlock:
1578	pte_unmap_unlock(pte, ptl);
1579	out:
1580	return retval;
1581	}
1582
1583	/**
1584	* vm_insert_pfn - insert single pfn into user vma
1585	* @vma: user vma to map to
1586	* @addr: target user address of this page
1587	* @pfn: source kernel pfn
1588	*
1589	* Similar to vm_insert_page, this allows drivers to insert individual pages
1590	* they've allocated into a user vma. Same comments apply.
1591	*
1592	* This function should only be called from a vm_ops->fault handler, and
1593	* in that case the handler should return NULL.
1594	*
1595	* vma cannot be a COW mapping.
1596	*
1597	* As this is called only for pages that do not currently exist, we
1598	* do not need to flush old virtual caches or the TLB.
1599	*/
1600	int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1601	unsigned long pfn)
1602	{
1603	return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1604	}
1605	EXPORT_SYMBOL(vm_insert_pfn);
1606
1607	/**
1608	* vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1609	* @vma: user vma to map to
1610	* @addr: target user address of this page
1611	* @pfn: source kernel pfn
1612	* @pgprot: pgprot flags for the inserted page
1613	*
1614	* This is exactly like vm_insert_pfn, except that it allows drivers to
1615	* to override pgprot on a per-page basis.
1616	*
1617	* This only makes sense for IO mappings, and it makes no sense for
1618	* cow mappings. In general, using multiple vmas is preferable;
1619	* vm_insert_pfn_prot should only be used if using multiple VMAs is
1620	* impractical.
1621	*/
1622	int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1623	unsigned long pfn, pgprot_t pgprot)
1624	{
1625	int ret;
1626	/*
1627	* Technically, architectures with pte_special can avoid all these
1628	* restrictions (same for remap_pfn_range). However we would like
1629	* consistency in testing and feature parity among all, so we should
1630	* try to keep these invariants in place for everybody.
1631	*/
1632	BUG_ON(!(vma->vm_flags & (VM_PFNMAP\|VM_MIXEDMAP)));
1633	BUG_ON((vma->vm_flags & (VM_PFNMAP\|VM_MIXEDMAP)) ==
1634	(VM_PFNMAP\|VM_MIXEDMAP));
1635	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1636	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1637
1638	if (addr < vma->vm_start \|\| addr >= vma->vm_end)
1639	return -EFAULT;
1640	if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
1641	return -EINVAL;
1642
1643	ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1644
1645	return ret;
1646	}
1647	EXPORT_SYMBOL(vm_insert_pfn_prot);
1648
1649	int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1650	pfn_t pfn)
1651	{
1652	pgprot_t pgprot = vma->vm_page_prot;
1653
1654	BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1655
1656	if (addr < vma->vm_start \|\| addr >= vma->vm_end)
1657	return -EFAULT;
1658	if (track_pfn_insert(vma, &pgprot, pfn))
1659	return -EINVAL;
1660
1661	/*
1662	* If we don't have pte special, then we have to use the pfn_valid()
1663	* based VM_MIXEDMAP scheme (see vm_normal_page), and thus we must
1664	* refcount the page if pfn_valid is true (hence insert_page rather
1665	* than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1666	* without pte special, it would there be refcounted as a normal page.
1667	*/
1668	if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1669	struct page *page;
1670
1671	/*
1672	* At this point we are committed to insert_page()
1673	* regardless of whether the caller specified flags that
1674	* result in pfn_t_has_page() == false.
1675	*/
1676	page = pfn_to_page(pfn_t_to_pfn(pfn));
1677	return insert_page(vma, addr, page, pgprot);
1678	}
1679	return insert_pfn(vma, addr, pfn, pgprot);
1680	}
1681	EXPORT_SYMBOL(vm_insert_mixed);
1682
1683	/*
1684	* maps a range of physical memory into the requested pages. the old
1685	* mappings are removed. any references to nonexistent pages results
1686	* in null mappings (currently treated as "copy-on-access")
1687	*/
1688	static int remap_pte_range(struct mm_struct mm, pmd_t pmd,
1689	unsigned long addr, unsigned long end,
1690	unsigned long pfn, pgprot_t prot)
1691	{
1692	pte_t *pte;
1693	spinlock_t *ptl;
1694
1695	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1696	if (!pte)
1697	return -ENOMEM;
1698	arch_enter_lazy_mmu_mode();
1699	do {
1700	BUG_ON(!pte_none(*pte));
1701	set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1702	pfn++;
1703	} while (pte++, addr += PAGE_SIZE, addr != end);
1704	arch_leave_lazy_mmu_mode();
1705	pte_unmap_unlock(pte - 1, ptl);
1706	return 0;
1707	}
1708
1709	static inline int remap_pmd_range(struct mm_struct mm, pud_t pud,
1710	unsigned long addr, unsigned long end,
1711	unsigned long pfn, pgprot_t prot)
1712	{
1713	pmd_t *pmd;
1714	unsigned long next;
1715
1716	pfn -= addr >> PAGE_SHIFT;
1717	pmd = pmd_alloc(mm, pud, addr);
1718	if (!pmd)
1719	return -ENOMEM;
1720	VM_BUG_ON(pmd_trans_huge(*pmd));
1721	do {
1722	next = pmd_addr_end(addr, end);
1723	if (remap_pte_range(mm, pmd, addr, next,
1724	pfn + (addr >> PAGE_SHIFT), prot))
1725	return -ENOMEM;
1726	} while (pmd++, addr = next, addr != end);
1727	return 0;
1728	}
1729
1730	static inline int remap_pud_range(struct mm_struct mm, pgd_t pgd,
1731	unsigned long addr, unsigned long end,
1732	unsigned long pfn, pgprot_t prot)
1733	{
1734	pud_t *pud;
1735	unsigned long next;
1736
1737	pfn -= addr >> PAGE_SHIFT;
1738	pud = pud_alloc(mm, pgd, addr);
1739	if (!pud)
1740	return -ENOMEM;
1741	do {
1742	next = pud_addr_end(addr, end);
1743	if (remap_pmd_range(mm, pud, addr, next,
1744	pfn + (addr >> PAGE_SHIFT), prot))
1745	return -ENOMEM;
1746	} while (pud++, addr = next, addr != end);
1747	return 0;
1748	}
1749
1750	/**
1751	* remap_pfn_range - remap kernel memory to userspace
1752	* @vma: user vma to map to
1753	* @addr: target user address to start at
1754	* @pfn: physical address of kernel memory
1755	* @size: size of map area
1756	* @prot: page protection flags for this mapping
1757	*
1758	* Note: this is only safe if the mm semaphore is held when called.
1759	*/
1760	int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1761	unsigned long pfn, unsigned long size, pgprot_t prot)
1762	{
1763	pgd_t *pgd;
1764	unsigned long next;
1765	unsigned long end = addr + PAGE_ALIGN(size);
1766	struct mm_struct *mm = vma->vm_mm;
1767	unsigned long remap_pfn = pfn;
1768	int err;
1769
1770	/*
1771	* Physically remapped pages are special. Tell the
1772	* rest of the world about it:
1773	* VM_IO tells people not to look at these pages
1774	* (accesses can have side effects).
1775	* VM_PFNMAP tells the core MM that the base pages are just
1776	* raw PFN mappings, and do not have a "struct page" associated
1777	* with them.
1778	* VM_DONTEXPAND
1779	* Disable vma merging and expanding with mremap().
1780	* VM_DONTDUMP
1781	* Omit vma from core dump, even when VM_IO turned off.
1782	*
1783	* There's a horrible special case to handle copy-on-write
1784	* behaviour that some programs depend on. We mark the "original"
1785	* un-COW'ed pages by matching them up with "vma->vm_pgoff".
1786	* See vm_normal_page() for details.
1787	*/
1788	if (is_cow_mapping(vma->vm_flags)) {
1789	if (addr != vma->vm_start \|\| end != vma->vm_end)
1790	return -EINVAL;
1791	vma->vm_pgoff = pfn;
1792	}
1793
1794	err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1795	if (err)
1796	return -EINVAL;
1797
1798	vma->vm_flags \|= VM_IO \| VM_PFNMAP \| VM_DONTEXPAND \| VM_DONTDUMP;
1799
1800	BUG_ON(addr >= end);
1801	pfn -= addr >> PAGE_SHIFT;
1802	pgd = pgd_offset(mm, addr);
1803	flush_cache_range(vma, addr, end);
1804	do {
1805	next = pgd_addr_end(addr, end);
1806	err = remap_pud_range(mm, pgd, addr, next,
1807	pfn + (addr >> PAGE_SHIFT), prot);
1808	if (err)
1809	break;
1810	} while (pgd++, addr = next, addr != end);
1811
1812	if (err)
1813	untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1814
1815	return err;
1816	}
1817	EXPORT_SYMBOL(remap_pfn_range);
1818
1819	/**
1820	* vm_iomap_memory - remap memory to userspace
1821	* @vma: user vma to map to
1822	* @start: start of area
1823	* @len: size of area
1824	*
1825	* This is a simplified io_remap_pfn_range() for common driver use. The
1826	* driver just needs to give us the physical memory range to be mapped,
1827	* we'll figure out the rest from the vma information.
1828	*
1829	* NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1830	* whatever write-combining details or similar.
1831	*/
1832	int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1833	{
1834	unsigned long vm_len, pfn, pages;
1835
1836	/* Check that the physical memory area passed in looks valid */
1837	if (start + len < start)
1838	return -EINVAL;
1839	/*
1840	* You really shouldn't map things that aren't page-aligned,
1841	* but we've historically allowed it because IO memory might
1842	* just have smaller alignment.
1843	*/
1844	len += start & ~PAGE_MASK;
1845	pfn = start >> PAGE_SHIFT;
1846	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1847	if (pfn + pages < pfn)
1848	return -EINVAL;
1849
1850	/* We start the mapping 'vm_pgoff' pages into the area */
1851	if (vma->vm_pgoff > pages)
1852	return -EINVAL;
1853	pfn += vma->vm_pgoff;
1854	pages -= vma->vm_pgoff;
1855
1856	/* Can we fit all of the mapping? */
1857	vm_len = vma->vm_end - vma->vm_start;
1858	if (vm_len >> PAGE_SHIFT > pages)
1859	return -EINVAL;
1860
1861	/* Ok, let it rip */
1862	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1863	}
1864	EXPORT_SYMBOL(vm_iomap_memory);
1865
1866	static int apply_to_pte_range(struct mm_struct mm, pmd_t pmd,
1867	unsigned long addr, unsigned long end,
1868	pte_fn_t fn, void *data)
1869	{
1870	pte_t *pte;
1871	int err;
1872	pgtable_t token;
1873	spinlock_t *uninitialized_var(ptl);
1874
1875	pte = (mm == &init_mm) ?
1876	pte_alloc_kernel(pmd, addr) :
1877	pte_alloc_map_lock(mm, pmd, addr, &ptl);
1878	if (!pte)
1879	return -ENOMEM;
1880
1881	BUG_ON(pmd_huge(*pmd));
1882
1883	arch_enter_lazy_mmu_mode();
1884
1885	token = pmd_pgtable(*pmd);
1886
1887	do {
1888	err = fn(pte++, token, addr, data);
1889	if (err)
1890	break;
1891	} while (addr += PAGE_SIZE, addr != end);
1892
1893	arch_leave_lazy_mmu_mode();
1894
1895	if (mm != &init_mm)
1896	pte_unmap_unlock(pte-1, ptl);
1897	return err;
1898	}
1899
1900	static int apply_to_pmd_range(struct mm_struct mm, pud_t pud,
1901	unsigned long addr, unsigned long end,
1902	pte_fn_t fn, void *data)
1903	{
1904	pmd_t *pmd;
1905	unsigned long next;
1906	int err;
1907
1908	BUG_ON(pud_huge(*pud));
1909
1910	pmd = pmd_alloc(mm, pud, addr);
1911	if (!pmd)
1912	return -ENOMEM;
1913	do {
1914	next = pmd_addr_end(addr, end);
1915	err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1916	if (err)
1917	break;
1918	} while (pmd++, addr = next, addr != end);
1919	return err;
1920	}
1921
1922	static int apply_to_pud_range(struct mm_struct mm, pgd_t pgd,
1923	unsigned long addr, unsigned long end,
1924	pte_fn_t fn, void *data)
1925	{
1926	pud_t *pud;
1927	unsigned long next;
1928	int err;
1929
1930	pud = pud_alloc(mm, pgd, addr);
1931	if (!pud)
1932	return -ENOMEM;
1933	do {
1934	next = pud_addr_end(addr, end);
1935	err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1936	if (err)
1937	break;
1938	} while (pud++, addr = next, addr != end);
1939	return err;
1940	}
1941
1942	/*
1943	* Scan a region of virtual memory, filling in page tables as necessary
1944	* and calling a provided function on each leaf page table.
1945	*/
1946	int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1947	unsigned long size, pte_fn_t fn, void *data)
1948	{
1949	pgd_t *pgd;
1950	unsigned long next;
1951	unsigned long end = addr + size;
1952	int err;
1953
1954	if (WARN_ON(addr >= end))
1955	return -EINVAL;
1956
1957	pgd = pgd_offset(mm, addr);
1958	do {
1959	next = pgd_addr_end(addr, end);
1960	err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1961	if (err)
1962	break;
1963	} while (pgd++, addr = next, addr != end);
1964
1965	return err;
1966	}
1967	EXPORT_SYMBOL_GPL(apply_to_page_range);
1968
1969	/*
1970	* handle_pte_fault chooses page fault handler according to an entry which was
1971	* read non-atomically. Before making any commitment, on those architectures
1972	* or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1973	* parts, do_swap_page must check under lock before unmapping the pte and
1974	* proceeding (but do_wp_page is only called after already making such a check;
1975	* and do_anonymous_page can safely check later on).
1976	*/
1977	static inline int pte_unmap_same(struct mm_struct mm, pmd_t pmd,
1978	pte_t *page_table, pte_t orig_pte)
1979	{
1980	int same = 1;
1981	#if defined(CONFIG_SMP) \|\| defined(CONFIG_PREEMPT)
1982	if (sizeof(pte_t) > sizeof(unsigned long)) {
1983	spinlock_t *ptl = pte_lockptr(mm, pmd);
1984	spin_lock(ptl);
1985	same = pte_same(*page_table, orig_pte);
1986	spin_unlock(ptl);
1987	}
1988	#endif
1989	pte_unmap(page_table);
1990	return same;
1991	}
1992
1993	static inline void cow_user_page(struct page dst, struct page src, unsigned long va, struct vm_area_struct *vma)
1994	{
1995	debug_dma_assert_idle(src);
1996
1997	/*
1998	* If the source page was a PFN mapping, we don't have
1999	* a "struct page" for it. We do a best-effort copy by
2000	* just copying from the original user address. If that
2001	* fails, we just zero-fill it. Live with it.
2002	*/
2003	if (unlikely(!src)) {
2004	void *kaddr = kmap_atomic(dst);
2005	void __user uaddr = (void __user )(va & PAGE_MASK);
2006
2007	/*
2008	* This really shouldn't fail, because the page is there
2009	* in the page tables. But it might just be unreadable,
2010	* in which case we just give up and fill the result with
2011	* zeroes.
2012	*/
2013	if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2014	clear_page(kaddr);
2015	kunmap_atomic(kaddr);
2016	flush_dcache_page(dst);
2017	} else
2018	copy_user_highpage(dst, src, va, vma);
2019	}
2020
2021	static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2022	{
2023	struct file *vm_file = vma->vm_file;
2024
2025	if (vm_file)
2026	return mapping_gfp_mask(vm_file->f_mapping) \| __GFP_FS \| __GFP_IO;
2027
2028	/*
2029	* Special mappings (e.g. VDSO) do not have any file so fake
2030	* a default GFP_KERNEL for them.
2031	*/
2032	return GFP_KERNEL;
2033	}
2034
2035	/*
2036	* Notify the address space that the page is about to become writable so that
2037	* it can prohibit this or wait for the page to get into an appropriate state.
2038	*
2039	* We do this without the lock held, so that it can sleep if it needs to.
2040	*/
2041	static int do_page_mkwrite(struct vm_area_struct vma, struct page page,
2042	unsigned long address)
2043	{
2044	struct vm_fault vmf;
2045	int ret;
2046
2047	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2048	vmf.pgoff = page->index;
2049	vmf.flags = FAULT_FLAG_WRITE\|FAULT_FLAG_MKWRITE;
2050	vmf.gfp_mask = __get_fault_gfp_mask(vma);
2051	vmf.page = page;
2052	vmf.cow_page = NULL;
2053
2054	ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2055	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE)))
2056	return ret;
2057	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2058	lock_page(page);
2059	if (!page->mapping) {
2060	unlock_page(page);
2061	return 0; /* retry */
2062	}
2063	ret \|= VM_FAULT_LOCKED;
2064	} else
2065	VM_BUG_ON_PAGE(!PageLocked(page), page);
2066	return ret;
2067	}
2068
2069	/*
2070	* Handle write page faults for pages that can be reused in the current vma
2071	*
2072	* This can happen either due to the mapping being with the VM_SHARED flag,
2073	* or due to us being the last reference standing to the page. In either
2074	* case, all we need to do here is to mark the page as writable and update
2075	* any related book-keeping.
2076	*/
2077	static inline int wp_page_reuse(struct fault_env *fe, pte_t orig_pte,
2078	struct page *page, int page_mkwrite, int dirty_shared)
2079	__releases(fe->ptl)
2080	{
2081	struct vm_area_struct *vma = fe->vma;
2082	pte_t entry;
2083	/*
2084	* Clear the pages cpupid information as the existing
2085	* information potentially belongs to a now completely
2086	* unrelated process.
2087	*/
2088	if (page)
2089	page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2090
2091	flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
2092	entry = pte_mkyoung(orig_pte);
2093	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2094	if (ptep_set_access_flags(vma, fe->address, fe->pte, entry, 1))
2095	update_mmu_cache(vma, fe->address, fe->pte);
2096	pte_unmap_unlock(fe->pte, fe->ptl);
2097
2098	if (dirty_shared) {
2099	struct address_space *mapping;
2100	int dirtied;
2101
2102	if (!page_mkwrite)
2103	lock_page(page);
2104
2105	dirtied = set_page_dirty(page);
2106	VM_BUG_ON_PAGE(PageAnon(page), page);
2107	mapping = page->mapping;
2108	unlock_page(page);
2109	put_page(page);
2110
2111	if ((dirtied \|\| page_mkwrite) && mapping) {
2112	/*
2113	* Some device drivers do not set page.mapping
2114	* but still dirty their pages
2115	*/
2116	balance_dirty_pages_ratelimited(mapping);
2117	}
2118
2119	if (!page_mkwrite)
2120	file_update_time(vma->vm_file);
2121	}
2122
2123	return VM_FAULT_WRITE;
2124	}
2125
2126	/*
2127	* Handle the case of a page which we actually need to copy to a new page.
2128	*
2129	* Called with mmap_sem locked and the old page referenced, but
2130	* without the ptl held.
2131	*
2132	* High level logic flow:
2133	*
2134	* - Allocate a page, copy the content of the old page to the new one.
2135	* - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2136	* - Take the PTL. If the pte changed, bail out and release the allocated page
2137	* - If the pte is still the way we remember it, update the page table and all
2138	* relevant references. This includes dropping the reference the page-table
2139	* held to the old page, as well as updating the rmap.
2140	* - In any case, unlock the PTL and drop the reference we took to the old page.
2141	*/
2142	static int wp_page_copy(struct fault_env *fe, pte_t orig_pte,
2143	struct page *old_page)
2144	{
2145	struct vm_area_struct *vma = fe->vma;
2146	struct mm_struct *mm = vma->vm_mm;
2147	struct page *new_page = NULL;
2148	pte_t entry;
2149	int page_copied = 0;
2150	const unsigned long mmun_start = fe->address & PAGE_MASK;
2151	const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2152	struct mem_cgroup *memcg;
2153
2154	if (unlikely(anon_vma_prepare(vma)))
2155	goto oom;
2156
2157	if (is_zero_pfn(pte_pfn(orig_pte))) {
2158	new_page = alloc_zeroed_user_highpage_movable(vma, fe->address);
2159	if (!new_page)
2160	goto oom;
2161	} else {
2162	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2163	fe->address);
2164	if (!new_page)
2165	goto oom;
2166	cow_user_page(new_page, old_page, fe->address, vma);
2167	}
2168
2169	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2170	goto oom_free_new;
2171
2172	__SetPageUptodate(new_page);
2173
2174	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2175
2176	/*
2177	* Re-check the pte - we dropped the lock
2178	*/
2179	fe->pte = pte_offset_map_lock(mm, fe->pmd, fe->address, &fe->ptl);
2180	if (likely(pte_same(*fe->pte, orig_pte))) {
2181	if (old_page) {
2182	if (!PageAnon(old_page)) {
2183	dec_mm_counter_fast(mm,
2184	mm_counter_file(old_page));
2185	inc_mm_counter_fast(mm, MM_ANONPAGES);
2186	}
2187	} else {
2188	inc_mm_counter_fast(mm, MM_ANONPAGES);
2189	}
2190	flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
2191	entry = mk_pte(new_page, vma->vm_page_prot);
2192	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2193	/*
2194	* Clear the pte entry and flush it first, before updating the
2195	* pte with the new entry. This will avoid a race condition
2196	* seen in the presence of one thread doing SMC and another
2197	* thread doing COW.
2198	*/
2199	ptep_clear_flush_notify(vma, fe->address, fe->pte);
2200	page_add_new_anon_rmap(new_page, vma, fe->address, false);
2201	mem_cgroup_commit_charge(new_page, memcg, false, false);
2202	lru_cache_add_active_or_unevictable(new_page, vma);
2203	/*
2204	* We call the notify macro here because, when using secondary
2205	* mmu page tables (such as kvm shadow page tables), we want the
2206	* new page to be mapped directly into the secondary page table.
2207	*/
2208	set_pte_at_notify(mm, fe->address, fe->pte, entry);
2209	update_mmu_cache(vma, fe->address, fe->pte);
2210	if (old_page) {
2211	/*
2212	* Only after switching the pte to the new page may
2213	* we remove the mapcount here. Otherwise another
2214	* process may come and find the rmap count decremented
2215	* before the pte is switched to the new page, and
2216	* "reuse" the old page writing into it while our pte
2217	* here still points into it and can be read by other
2218	* threads.
2219	*
2220	* The critical issue is to order this
2221	* page_remove_rmap with the ptp_clear_flush above.
2222	* Those stores are ordered by (if nothing else,)
2223	* the barrier present in the atomic_add_negative
2224	* in page_remove_rmap.
2225	*
2226	* Then the TLB flush in ptep_clear_flush ensures that
2227	* no process can access the old page before the
2228	* decremented mapcount is visible. And the old page
2229	* cannot be reused until after the decremented
2230	* mapcount is visible. So transitively, TLBs to
2231	* old page will be flushed before it can be reused.
2232	*/
2233	page_remove_rmap(old_page, false);
2234	}
2235
2236	/* Free the old page.. */
2237	new_page = old_page;
2238	page_copied = 1;
2239	} else {
2240	mem_cgroup_cancel_charge(new_page, memcg, false);
2241	}
2242
2243	if (new_page)
2244	put_page(new_page);
2245
2246	pte_unmap_unlock(fe->pte, fe->ptl);
2247	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2248	if (old_page) {
2249	/*
2250	* Don't let another task, with possibly unlocked vma,
2251	* keep the mlocked page.
2252	*/
2253	if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2254	lock_page(old_page); /* LRU manipulation */
2255	if (PageMlocked(old_page))
2256	munlock_vma_page(old_page);
2257	unlock_page(old_page);
2258	}
2259	put_page(old_page);
2260	}
2261	return page_copied ? VM_FAULT_WRITE : 0;
2262	oom_free_new:
2263	put_page(new_page);
2264	oom:
2265	if (old_page)
2266	put_page(old_page);
2267	return VM_FAULT_OOM;
2268	}
2269
2270	/*
2271	* Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2272	* mapping
2273	*/
2274	static int wp_pfn_shared(struct fault_env *fe, pte_t orig_pte)
2275	{
2276	struct vm_area_struct *vma = fe->vma;
2277
2278	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2279	struct vm_fault vmf = {
2280	.page = NULL,
2281	.pgoff = linear_page_index(vma, fe->address),
2282	.virtual_address =
2283	(void __user *)(fe->address & PAGE_MASK),
2284	.flags = FAULT_FLAG_WRITE \| FAULT_FLAG_MKWRITE,
2285	};
2286	int ret;
2287
2288	pte_unmap_unlock(fe->pte, fe->ptl);
2289	ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2290	if (ret & VM_FAULT_ERROR)
2291	return ret;
2292	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2293	&fe->ptl);
2294	/*
2295	* We might have raced with another page fault while we
2296	* released the pte_offset_map_lock.
2297	*/
2298	if (!pte_same(*fe->pte, orig_pte)) {
2299	pte_unmap_unlock(fe->pte, fe->ptl);
2300	return 0;
2301	}
2302	}
2303	return wp_page_reuse(fe, orig_pte, NULL, 0, 0);
2304	}
2305
2306	static int wp_page_shared(struct fault_env *fe, pte_t orig_pte,
2307	struct page *old_page)
2308	__releases(fe->ptl)
2309	{
2310	struct vm_area_struct *vma = fe->vma;
2311	int page_mkwrite = 0;
2312
2313	get_page(old_page);
2314
2315	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2316	int tmp;
2317
2318	pte_unmap_unlock(fe->pte, fe->ptl);
2319	tmp = do_page_mkwrite(vma, old_page, fe->address);
2320	if (unlikely(!tmp \|\| (tmp &
2321	(VM_FAULT_ERROR \| VM_FAULT_NOPAGE)))) {
2322	put_page(old_page);
2323	return tmp;
2324	}
2325	/*
2326	* Since we dropped the lock we need to revalidate
2327	* the PTE as someone else may have changed it. If
2328	* they did, we just return, as we can count on the
2329	* MMU to tell us if they didn't also make it writable.
2330	*/
2331	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2332	&fe->ptl);
2333	if (!pte_same(*fe->pte, orig_pte)) {
2334	unlock_page(old_page);
2335	pte_unmap_unlock(fe->pte, fe->ptl);
2336	put_page(old_page);
2337	return 0;
2338	}
2339	page_mkwrite = 1;
2340	}
2341
2342	return wp_page_reuse(fe, orig_pte, old_page, page_mkwrite, 1);
2343	}
2344
2345	/*
2346	* This routine handles present pages, when users try to write
2347	* to a shared page. It is done by copying the page to a new address
2348	* and decrementing the shared-page counter for the old page.
2349	*
2350	* Note that this routine assumes that the protection checks have been
2351	* done by the caller (the low-level page fault routine in most cases).
2352	* Thus we can safely just mark it writable once we've done any necessary
2353	* COW.
2354	*
2355	* We also mark the page dirty at this point even though the page will
2356	* change only once the write actually happens. This avoids a few races,
2357	* and potentially makes it more efficient.
2358	*
2359	* We enter with non-exclusive mmap_sem (to exclude vma changes,
2360	* but allow concurrent faults), with pte both mapped and locked.
2361	* We return with mmap_sem still held, but pte unmapped and unlocked.
2362	*/
2363	static int do_wp_page(struct fault_env *fe, pte_t orig_pte)
2364	__releases(fe->ptl)
2365	{
2366	struct vm_area_struct *vma = fe->vma;
2367	struct page *old_page;
2368
2369	old_page = vm_normal_page(vma, fe->address, orig_pte);
2370	if (!old_page) {
2371	/*
2372	* VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2373	* VM_PFNMAP VMA.
2374	*
2375	* We should not cow pages in a shared writeable mapping.
2376	* Just mark the pages writable and/or call ops->pfn_mkwrite.
2377	*/
2378	if ((vma->vm_flags & (VM_WRITE\|VM_SHARED)) ==
2379	(VM_WRITE\|VM_SHARED))
2380	return wp_pfn_shared(fe, orig_pte);
2381
2382	pte_unmap_unlock(fe->pte, fe->ptl);
2383	return wp_page_copy(fe, orig_pte, old_page);
2384	}
2385
2386	/*
2387	* Take out anonymous pages first, anonymous shared vmas are
2388	* not dirty accountable.
2389	*/
2390	if (PageAnon(old_page) && !PageKsm(old_page)) {
2391	int total_mapcount;
2392	if (!trylock_page(old_page)) {
2393	get_page(old_page);
2394	pte_unmap_unlock(fe->pte, fe->ptl);
2395	lock_page(old_page);
2396	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2397	fe->address, &fe->ptl);
2398	if (!pte_same(*fe->pte, orig_pte)) {
2399	unlock_page(old_page);
2400	pte_unmap_unlock(fe->pte, fe->ptl);
2401	put_page(old_page);
2402	return 0;
2403	}
2404	put_page(old_page);
2405	}
2406	if (reuse_swap_page(old_page, &total_mapcount)) {
2407	if (total_mapcount == 1) {
2408	/*
2409	* The page is all ours. Move it to
2410	* our anon_vma so the rmap code will
2411	* not search our parent or siblings.
2412	* Protected against the rmap code by
2413	* the page lock.
2414	*/
2415	page_move_anon_rmap(old_page, vma);
2416	}
2417	unlock_page(old_page);
2418	return wp_page_reuse(fe, orig_pte, old_page, 0, 0);
2419	}
2420	unlock_page(old_page);
2421	} else if (unlikely((vma->vm_flags & (VM_WRITE\|VM_SHARED)) ==
2422	(VM_WRITE\|VM_SHARED))) {
2423	return wp_page_shared(fe, orig_pte, old_page);
2424	}
2425
2426	/*
2427	* Ok, we need to copy. Oh, well..
2428	*/
2429	get_page(old_page);
2430
2431	pte_unmap_unlock(fe->pte, fe->ptl);
2432	return wp_page_copy(fe, orig_pte, old_page);
2433	}
2434
2435	static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2436	unsigned long start_addr, unsigned long end_addr,
2437	struct zap_details *details)
2438	{
2439	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2440	}
2441
2442	static inline void unmap_mapping_range_tree(struct rb_root *root,
2443	struct zap_details *details)
2444	{
2445	struct vm_area_struct *vma;
2446	pgoff_t vba, vea, zba, zea;
2447
2448	vma_interval_tree_foreach(vma, root,
2449	details->first_index, details->last_index) {
2450
2451	vba = vma->vm_pgoff;
2452	vea = vba + vma_pages(vma) - 1;
2453	zba = details->first_index;
2454	if (zba < vba)
2455	zba = vba;
2456	zea = details->last_index;
2457	if (zea > vea)
2458	zea = vea;
2459
2460	unmap_mapping_range_vma(vma,
2461	((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2462	((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2463	details);
2464	}
2465	}
2466
2467	/**
2468	* unmap_mapping_range - unmap the portion of all mmaps in the specified
2469	* address_space corresponding to the specified page range in the underlying
2470	* file.
2471	*
2472	* @mapping: the address space containing mmaps to be unmapped.
2473	* @holebegin: byte in first page to unmap, relative to the start of
2474	* the underlying file. This will be rounded down to a PAGE_SIZE
2475	* boundary. Note that this is different from truncate_pagecache(), which
2476	* must keep the partial page. In contrast, we must get rid of
2477	* partial pages.
2478	* @holelen: size of prospective hole in bytes. This will be rounded
2479	* up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2480	* end of the file.
2481	* @even_cows: 1 when truncating a file, unmap even private COWed pages;
2482	* but 0 when invalidating pagecache, don't throw away private data.
2483	*/
2484	void unmap_mapping_range(struct address_space *mapping,
2485	loff_t const holebegin, loff_t const holelen, int even_cows)
2486	{
2487	struct zap_details details = { };
2488	pgoff_t hba = holebegin >> PAGE_SHIFT;
2489	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2490
2491	/* Check for overflow. */
2492	if (sizeof(holelen) > sizeof(hlen)) {
2493	long long holeend =
2494	(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2495	if (holeend & ~(long long)ULONG_MAX)
2496	hlen = ULONG_MAX - hba + 1;
2497	}
2498
2499	details.check_mapping = even_cows? NULL: mapping;
2500	details.first_index = hba;
2501	details.last_index = hba + hlen - 1;
2502	if (details.last_index < details.first_index)
2503	details.last_index = ULONG_MAX;
2504
2505	i_mmap_lock_write(mapping);
2506	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2507	unmap_mapping_range_tree(&mapping->i_mmap, &details);
2508	i_mmap_unlock_write(mapping);
2509	}
2510	EXPORT_SYMBOL(unmap_mapping_range);
2511
2512	/*
2513	* We enter with non-exclusive mmap_sem (to exclude vma changes,
2514	* but allow concurrent faults), and pte mapped but not yet locked.
2515	* We return with pte unmapped and unlocked.
2516	*
2517	* We return with the mmap_sem locked or unlocked in the same cases
2518	* as does filemap_fault().
2519	*/
2520	int do_swap_page(struct fault_env *fe, pte_t orig_pte)
2521	{
2522	struct vm_area_struct *vma = fe->vma;
2523	struct page page, swapcache;
2524	struct mem_cgroup *memcg;
2525	swp_entry_t entry;
2526	pte_t pte;
2527	int locked;
2528	int exclusive = 0;
2529	int ret = 0;
2530
2531	if (!pte_unmap_same(vma->vm_mm, fe->pmd, fe->pte, orig_pte))
2532	goto out;
2533
2534	entry = pte_to_swp_entry(orig_pte);
2535	if (unlikely(non_swap_entry(entry))) {
2536	if (is_migration_entry(entry)) {
2537	migration_entry_wait(vma->vm_mm, fe->pmd, fe->address);
2538	} else if (is_hwpoison_entry(entry)) {
2539	ret = VM_FAULT_HWPOISON;
2540	} else {
2541	print_bad_pte(vma, fe->address, orig_pte, NULL);
2542	ret = VM_FAULT_SIGBUS;
2543	}
2544	goto out;
2545	}
2546	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2547	page = lookup_swap_cache(entry);
2548	if (!page) {
2549	page = swapin_readahead(entry,
2550	GFP_HIGHUSER_MOVABLE, vma, fe->address);
2551	if (!page) {
2552	/*
2553	* Back out if somebody else faulted in this pte
2554	* while we released the pte lock.
2555	*/
2556	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2557	fe->address, &fe->ptl);
2558	if (likely(pte_same(*fe->pte, orig_pte)))
2559	ret = VM_FAULT_OOM;
2560	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2561	goto unlock;
2562	}
2563
2564	/* Had to read the page from swap area: Major fault */
2565	ret = VM_FAULT_MAJOR;
2566	count_vm_event(PGMAJFAULT);
2567	mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2568	} else if (PageHWPoison(page)) {
2569	/*
2570	* hwpoisoned dirty swapcache pages are kept for killing
2571	* owner processes (which may be unknown at hwpoison time)
2572	*/
2573	ret = VM_FAULT_HWPOISON;
2574	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2575	swapcache = page;
2576	goto out_release;
2577	}
2578
2579	swapcache = page;
2580	locked = lock_page_or_retry(page, vma->vm_mm, fe->flags);
2581
2582	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2583	if (!locked) {
2584	ret \|= VM_FAULT_RETRY;
2585	goto out_release;
2586	}
2587
2588	/*
2589	* Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2590	* release the swapcache from under us. The page pin, and pte_same
2591	* test below, are not enough to exclude that. Even if it is still
2592	* swapcache, we need to check that the page's swap has not changed.
2593	*/
2594	if (unlikely(!PageSwapCache(page) \|\| page_private(page) != entry.val))
2595	goto out_page;
2596
2597	page = ksm_might_need_to_copy(page, vma, fe->address);
2598	if (unlikely(!page)) {
2599	ret = VM_FAULT_OOM;
2600	page = swapcache;
2601	goto out_page;
2602	}
2603
2604	if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2605	&memcg, false)) {
2606	ret = VM_FAULT_OOM;
2607	goto out_page;
2608	}
2609
2610	/*
2611	* Back out if somebody else already faulted in this pte.
2612	*/
2613	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2614	&fe->ptl);
2615	if (unlikely(!pte_same(*fe->pte, orig_pte)))
2616	goto out_nomap;
2617
2618	if (unlikely(!PageUptodate(page))) {
2619	ret = VM_FAULT_SIGBUS;
2620	goto out_nomap;
2621	}
2622
2623	/*
2624	* The page isn't present yet, go ahead with the fault.
2625	*
2626	* Be careful about the sequence of operations here.
2627	* To get its accounting right, reuse_swap_page() must be called
2628	* while the page is counted on swap but not yet in mapcount i.e.
2629	* before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2630	* must be called after the swap_free(), or it will never succeed.
2631	*/
2632
2633	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2634	dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2635	pte = mk_pte(page, vma->vm_page_prot);
2636	if ((fe->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2637	pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2638	fe->flags &= ~FAULT_FLAG_WRITE;
2639	ret \|= VM_FAULT_WRITE;
2640	exclusive = RMAP_EXCLUSIVE;
2641	}
2642	flush_icache_page(vma, page);
2643	if (pte_swp_soft_dirty(orig_pte))
2644	pte = pte_mksoft_dirty(pte);
2645	set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
2646	if (page == swapcache) {
2647	do_page_add_anon_rmap(page, vma, fe->address, exclusive);
2648	mem_cgroup_commit_charge(page, memcg, true, false);
2649	activate_page(page);
2650	} else { /* ksm created a completely new copy */
2651	page_add_new_anon_rmap(page, vma, fe->address, false);
2652	mem_cgroup_commit_charge(page, memcg, false, false);
2653	lru_cache_add_active_or_unevictable(page, vma);
2654	}
2655
2656	swap_free(entry);
2657	if (mem_cgroup_swap_full(page) \|\|
2658	(vma->vm_flags & VM_LOCKED) \|\| PageMlocked(page))
2659	try_to_free_swap(page);
2660	unlock_page(page);
2661	if (page != swapcache) {
2662	/*
2663	* Hold the lock to avoid the swap entry to be reused
2664	* until we take the PT lock for the pte_same() check
2665	* (to avoid false positives from pte_same). For
2666	* further safety release the lock after the swap_free
2667	* so that the swap count won't change under a
2668	* parallel locked swapcache.
2669	*/
2670	unlock_page(swapcache);
2671	put_page(swapcache);
2672	}
2673
2674	if (fe->flags & FAULT_FLAG_WRITE) {
2675	ret \|= do_wp_page(fe, pte);
2676	if (ret & VM_FAULT_ERROR)
2677	ret &= VM_FAULT_ERROR;
2678	goto out;
2679	}
2680
2681	/* No need to invalidate - it was non-present before */
2682	update_mmu_cache(vma, fe->address, fe->pte);
2683	unlock:
2684	pte_unmap_unlock(fe->pte, fe->ptl);
2685	out:
2686	return ret;
2687	out_nomap:
2688	mem_cgroup_cancel_charge(page, memcg, false);
2689	pte_unmap_unlock(fe->pte, fe->ptl);
2690	out_page:
2691	unlock_page(page);
2692	out_release:
2693	put_page(page);
2694	if (page != swapcache) {
2695	unlock_page(swapcache);
2696	put_page(swapcache);
2697	}
2698	return ret;
2699	}
2700
2701	/*
2702	* This is like a special single-page "expand_{down\|up}wards()",
2703	* except we must first make sure that 'address{-\|+}PAGE_SIZE'
2704	* doesn't hit another vma.
2705	*/
2706	static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2707	{
2708	address &= PAGE_MASK;
2709	if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2710	struct vm_area_struct *prev = vma->vm_prev;
2711
2712	/*
2713	* Is there a mapping abutting this one below?
2714	*
2715	* That's only ok if it's the same stack mapping
2716	* that has gotten split..
2717	*/
2718	if (prev && prev->vm_end == address)
2719	return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2720
2721	return expand_downwards(vma, address - PAGE_SIZE);
2722	}
2723	if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2724	struct vm_area_struct *next = vma->vm_next;
2725
2726	/* As VM_GROWSDOWN but s/below/above/ */
2727	if (next && next->vm_start == address + PAGE_SIZE)
2728	return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2729
2730	return expand_upwards(vma, address + PAGE_SIZE);
2731	}
2732	return 0;
2733	}
2734
2735	/*
2736	* We enter with non-exclusive mmap_sem (to exclude vma changes,
2737	* but allow concurrent faults), and pte mapped but not yet locked.
2738	* We return with mmap_sem still held, but pte unmapped and unlocked.
2739	*/
2740	static int do_anonymous_page(struct fault_env *fe)
2741	{
2742	struct vm_area_struct *vma = fe->vma;
2743	struct mem_cgroup *memcg;
2744	struct page *page;
2745	pte_t entry;
2746
2747	/* File mapping without ->vm_ops ? */
2748	if (vma->vm_flags & VM_SHARED)
2749	return VM_FAULT_SIGBUS;
2750
2751	/* Check if we need to add a guard page to the stack */
2752	if (check_stack_guard_page(vma, fe->address) < 0)
2753	return VM_FAULT_SIGSEGV;
2754
2755	/*
2756	* Use pte_alloc() instead of pte_alloc_map(). We can't run
2757	* pte_offset_map() on pmds where a huge pmd might be created
2758	* from a different thread.
2759	*
2760	* pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2761	* parallel threads are excluded by other means.
2762	*
2763	* Here we only have down_read(mmap_sem).
2764	*/
2765	if (pte_alloc(vma->vm_mm, fe->pmd, fe->address))
2766	return VM_FAULT_OOM;
2767
2768	/* See the comment in pte_alloc_one_map() */
2769	if (unlikely(pmd_trans_unstable(fe->pmd)))
2770	return 0;
2771
2772	/* Use the zero-page for reads */
2773	if (!(fe->flags & FAULT_FLAG_WRITE) &&
2774	!mm_forbids_zeropage(vma->vm_mm)) {
2775	entry = pte_mkspecial(pfn_pte(my_zero_pfn(fe->address),
2776	vma->vm_page_prot));
2777	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2778	&fe->ptl);
2779	if (!pte_none(*fe->pte))
2780	goto unlock;
2781	/* Deliver the page fault to userland, check inside PT lock */
2782	if (userfaultfd_missing(vma)) {
2783	pte_unmap_unlock(fe->pte, fe->ptl);
2784	return handle_userfault(fe, VM_UFFD_MISSING);
2785	}
2786	goto setpte;
2787	}
2788
2789	/* Allocate our own private page. */
2790	if (unlikely(anon_vma_prepare(vma)))
2791	goto oom;
2792	page = alloc_zeroed_user_highpage_movable(vma, fe->address);
2793	if (!page)
2794	goto oom;
2795
2796	if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2797	goto oom_free_page;
2798
2799	/*
2800	* The memory barrier inside __SetPageUptodate makes sure that
2801	* preceeding stores to the page contents become visible before
2802	* the set_pte_at() write.
2803	*/
2804	__SetPageUptodate(page);
2805
2806	entry = mk_pte(page, vma->vm_page_prot);
2807	if (vma->vm_flags & VM_WRITE)
2808	entry = pte_mkwrite(pte_mkdirty(entry));
2809
2810	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2811	&fe->ptl);
2812	if (!pte_none(*fe->pte))
2813	goto release;
2814
2815	/* Deliver the page fault to userland, check inside PT lock */
2816	if (userfaultfd_missing(vma)) {
2817	pte_unmap_unlock(fe->pte, fe->ptl);
2818	mem_cgroup_cancel_charge(page, memcg, false);
2819	put_page(page);
2820	return handle_userfault(fe, VM_UFFD_MISSING);
2821	}
2822
2823	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2824	page_add_new_anon_rmap(page, vma, fe->address, false);
2825	mem_cgroup_commit_charge(page, memcg, false, false);
2826	lru_cache_add_active_or_unevictable(page, vma);
2827	setpte:
2828	set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
2829
2830	/* No need to invalidate - it was non-present before */
2831	update_mmu_cache(vma, fe->address, fe->pte);
2832	unlock:
2833	pte_unmap_unlock(fe->pte, fe->ptl);
2834	return 0;
2835	release:
2836	mem_cgroup_cancel_charge(page, memcg, false);
2837	put_page(page);
2838	goto unlock;
2839	oom_free_page:
2840	put_page(page);
2841	oom:
2842	return VM_FAULT_OOM;
2843	}
2844
2845	/*
2846	* The mmap_sem must have been held on entry, and may have been
2847	* released depending on flags and vma->vm_ops->fault() return value.
2848	* See filemap_fault() and __lock_page_retry().
2849	*/
2850	static int __do_fault(struct fault_env *fe, pgoff_t pgoff,
2851	struct page cow_page, struct page page, void *entry)
2852	{
2853	struct vm_area_struct *vma = fe->vma;
2854	struct vm_fault vmf;
2855	int ret;
2856
2857	vmf.virtual_address = (void __user *)(fe->address & PAGE_MASK);
2858	vmf.pgoff = pgoff;
2859	vmf.flags = fe->flags;
2860	vmf.page = NULL;
2861	vmf.gfp_mask = __get_fault_gfp_mask(vma);
2862	vmf.cow_page = cow_page;
2863
2864	ret = vma->vm_ops->fault(vma, &vmf);
2865	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE \| VM_FAULT_RETRY)))
2866	return ret;
2867	if (ret & VM_FAULT_DAX_LOCKED) {
2868	*entry = vmf.entry;
2869	return ret;
2870	}
2871
2872	if (unlikely(PageHWPoison(vmf.page))) {
2873	if (ret & VM_FAULT_LOCKED)
2874	unlock_page(vmf.page);
2875	put_page(vmf.page);
2876	return VM_FAULT_HWPOISON;
2877	}
2878
2879	if (unlikely(!(ret & VM_FAULT_LOCKED)))
2880	lock_page(vmf.page);
2881	else
2882	VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2883
2884	*page = vmf.page;
2885	return ret;
2886	}
2887
2888	static int pte_alloc_one_map(struct fault_env *fe)
2889	{
2890	struct vm_area_struct *vma = fe->vma;
2891
2892	if (!pmd_none(*fe->pmd))
2893	goto map_pte;
2894	if (fe->prealloc_pte) {
2895	fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2896	if (unlikely(!pmd_none(*fe->pmd))) {
2897	spin_unlock(fe->ptl);
2898	goto map_pte;
2899	}
2900
2901	atomic_long_inc(&vma->vm_mm->nr_ptes);
2902	pmd_populate(vma->vm_mm, fe->pmd, fe->prealloc_pte);
2903	spin_unlock(fe->ptl);
2904	fe->prealloc_pte = 0;
2905	} else if (unlikely(pte_alloc(vma->vm_mm, fe->pmd, fe->address))) {
2906	return VM_FAULT_OOM;
2907	}
2908	map_pte:
2909	/*
2910	* If a huge pmd materialized under us just retry later. Use
2911	* pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2912	* didn't become pmd_trans_huge under us and then back to pmd_none, as
2913	* a result of MADV_DONTNEED running immediately after a huge pmd fault
2914	* in a different thread of this mm, in turn leading to a misleading
2915	* pmd_trans_huge() retval. All we have to ensure is that it is a
2916	* regular pmd that we can walk with pte_offset_map() and we can do that
2917	* through an atomic read in C, which is what pmd_trans_unstable()
2918	* provides.
2919	*/
2920	if (pmd_trans_unstable(fe->pmd) \|\| pmd_devmap(*fe->pmd))
2921	return VM_FAULT_NOPAGE;
2922
2923	fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2924	&fe->ptl);
2925	return 0;
2926	}
2927
2928	#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
2929
2930	#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2931	static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2932	unsigned long haddr)
2933	{
2934	if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2935	(vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2936	return false;
2937	if (haddr < vma->vm_start \|\| haddr + HPAGE_PMD_SIZE > vma->vm_end)
2938	return false;
2939	return true;
2940	}
2941
2942	static int do_set_pmd(struct fault_env fe, struct page page)
2943	{
2944	struct vm_area_struct *vma = fe->vma;
2945	bool write = fe->flags & FAULT_FLAG_WRITE;
2946	unsigned long haddr = fe->address & HPAGE_PMD_MASK;
2947	pmd_t entry;
2948	int i, ret;
2949
2950	if (!transhuge_vma_suitable(vma, haddr))
2951	return VM_FAULT_FALLBACK;
2952
2953	ret = VM_FAULT_FALLBACK;
2954	page = compound_head(page);
2955
2956	fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2957	if (unlikely(!pmd_none(*fe->pmd)))
2958	goto out;
2959
2960	for (i = 0; i < HPAGE_PMD_NR; i++)
2961	flush_icache_page(vma, page + i);
2962
2963	entry = mk_huge_pmd(page, vma->vm_page_prot);
2964	if (write)
2965	entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2966
2967	add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2968	page_add_file_rmap(page, true);
2969
2970	set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
2971
2972	update_mmu_cache_pmd(vma, haddr, fe->pmd);
2973
2974	/* fault is handled */
2975	ret = 0;
2976	count_vm_event(THP_FILE_MAPPED);
2977	out:
2978	spin_unlock(fe->ptl);
2979	return ret;
2980	}
2981	#else
2982	static int do_set_pmd(struct fault_env fe, struct page page)
2983	{
2984	BUILD_BUG();
2985	return 0;
2986	}
2987	#endif
2988
2989	/**
2990	* alloc_set_pte - setup new PTE entry for given page and add reverse page
2991	* mapping. If needed, the fucntion allocates page table or use pre-allocated.
2992	*
2993	* @fe: fault environment
2994	* @memcg: memcg to charge page (only for private mappings)
2995	* @page: page to map
2996	*
2997	* Caller must take care of unlocking fe->ptl, if fe->pte is non-NULL on return.
2998	*
2999	* Target users are page handler itself and implementations of
3000	* vm_ops->map_pages.
3001	*/
3002	int alloc_set_pte(struct fault_env fe, struct mem_cgroup memcg,
3003	struct page *page)
3004	{
3005	struct vm_area_struct *vma = fe->vma;
3006	bool write = fe->flags & FAULT_FLAG_WRITE;
3007	pte_t entry;
3008	int ret;
3009
3010	if (pmd_none(*fe->pmd) && PageTransCompound(page) &&
3011	IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3012	/* THP on COW? */
3013	VM_BUG_ON_PAGE(memcg, page);
3014
3015	ret = do_set_pmd(fe, page);
3016	if (ret != VM_FAULT_FALLBACK)
3017	return ret;
3018	}
3019
3020	if (!fe->pte) {
3021	ret = pte_alloc_one_map(fe);
3022	if (ret)
3023	return ret;
3024	}
3025
3026	/* Re-check under ptl */
3027	if (unlikely(!pte_none(*fe->pte)))
3028	return VM_FAULT_NOPAGE;
3029
3030	flush_icache_page(vma, page);
3031	entry = mk_pte(page, vma->vm_page_prot);
3032	if (write)
3033	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3034	/* copy-on-write page */
3035	if (write && !(vma->vm_flags & VM_SHARED)) {
3036	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3037	page_add_new_anon_rmap(page, vma, fe->address, false);
3038	mem_cgroup_commit_charge(page, memcg, false, false);
3039	lru_cache_add_active_or_unevictable(page, vma);
3040	} else {
3041	inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3042	page_add_file_rmap(page, false);
3043	}
3044	set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
3045
3046	/* no need to invalidate: a not-present page won't be cached */
3047	update_mmu_cache(vma, fe->address, fe->pte);
3048
3049	return 0;
3050	}
3051
3052	static unsigned long fault_around_bytes __read_mostly =
3053	rounddown_pow_of_two(65536);
3054
3055	#ifdef CONFIG_DEBUG_FS
3056	static int fault_around_bytes_get(void data, u64 val)
3057	{
3058	*val = fault_around_bytes;
3059	return 0;
3060	}
3061
3062	/*
3063	* fault_around_pages() and fault_around_mask() expects fault_around_bytes
3064	* rounded down to nearest page order. It's what do_fault_around() expects to
3065	* see.
3066	*/
3067	static int fault_around_bytes_set(void *data, u64 val)
3068	{
3069	if (val / PAGE_SIZE > PTRS_PER_PTE)
3070	return -EINVAL;
3071	if (val > PAGE_SIZE)
3072	fault_around_bytes = rounddown_pow_of_two(val);
3073	else
3074	fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3075	return 0;
3076	}
3077	DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3078	fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3079
3080	static int __init fault_around_debugfs(void)
3081	{
3082	void *ret;
3083
3084	ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3085	&fault_around_bytes_fops);
3086	if (!ret)
3087	pr_warn("Failed to create fault_around_bytes in debugfs");
3088	return 0;
3089	}
3090	late_initcall(fault_around_debugfs);
3091	#endif
3092
3093	/*
3094	* do_fault_around() tries to map few pages around the fault address. The hope
3095	* is that the pages will be needed soon and this will lower the number of
3096	* faults to handle.
3097	*
3098	* It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3099	* not ready to be mapped: not up-to-date, locked, etc.
3100	*
3101	* This function is called with the page table lock taken. In the split ptlock
3102	* case the page table lock only protects only those entries which belong to
3103	* the page table corresponding to the fault address.
3104	*
3105	* This function doesn't cross the VMA boundaries, in order to call map_pages()
3106	* only once.
3107	*
3108	* fault_around_pages() defines how many pages we'll try to map.
3109	* do_fault_around() expects it to return a power of two less than or equal to
3110	* PTRS_PER_PTE.
3111	*
3112	* The virtual address of the area that we map is naturally aligned to the
3113	* fault_around_pages() value (and therefore to page order). This way it's
3114	* easier to guarantee that we don't cross page table boundaries.
3115	*/
3116	static int do_fault_around(struct fault_env *fe, pgoff_t start_pgoff)
3117	{
3118	unsigned long address = fe->address, nr_pages, mask;
3119	pgoff_t end_pgoff;
3120	int off, ret = 0;
3121
3122	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3123	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3124
3125	fe->address = max(address & mask, fe->vma->vm_start);
3126	off = ((address - fe->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3127	start_pgoff -= off;
3128
3129	/*
3130	* end_pgoff is either end of page table or end of vma
3131	* or fault_around_pages() from start_pgoff, depending what is nearest.
3132	*/
3133	end_pgoff = start_pgoff -
3134	((fe->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3135	PTRS_PER_PTE - 1;
3136	end_pgoff = min3(end_pgoff, vma_pages(fe->vma) + fe->vma->vm_pgoff - 1,
3137	start_pgoff + nr_pages - 1);
3138
3139	if (pmd_none(*fe->pmd)) {
3140	fe->prealloc_pte = pte_alloc_one(fe->vma->vm_mm, fe->address);
3141	if (!fe->prealloc_pte)
3142	goto out;
3143	smp_wmb(); /* See comment in __pte_alloc() */
3144	}
3145
3146	fe->vma->vm_ops->map_pages(fe, start_pgoff, end_pgoff);
3147
3148	/* preallocated pagetable is unused: free it */
3149	if (fe->prealloc_pte) {
3150	pte_free(fe->vma->vm_mm, fe->prealloc_pte);
3151	fe->prealloc_pte = 0;
3152	}
3153	/* Huge page is mapped? Page fault is solved */
3154	if (pmd_trans_huge(*fe->pmd)) {
3155	ret = VM_FAULT_NOPAGE;
3156	goto out;
3157	}
3158
3159	/* ->map_pages() haven't done anything useful. Cold page cache? */
3160	if (!fe->pte)
3161	goto out;
3162
3163	/* check if the page fault is solved */
3164	fe->pte -= (fe->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3165	if (!pte_none(*fe->pte))
3166	ret = VM_FAULT_NOPAGE;
3167	pte_unmap_unlock(fe->pte, fe->ptl);
3168	out:
3169	fe->address = address;
3170	fe->pte = NULL;
3171	return ret;
3172	}
3173
3174	static int do_read_fault(struct fault_env *fe, pgoff_t pgoff)
3175	{
3176	struct vm_area_struct *vma = fe->vma;
3177	struct page *fault_page;
3178	int ret = 0;
3179
3180	/*
3181	* Let's call ->map_pages() first and use ->fault() as fallback
3182	* if page by the offset is not ready to be mapped (cold cache or
3183	* something).
3184	*/
3185	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3186	ret = do_fault_around(fe, pgoff);
3187	if (ret)
3188	return ret;
3189	}
3190
3191	ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
3192	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE \| VM_FAULT_RETRY)))
3193	return ret;
3194
3195	ret \|= alloc_set_pte(fe, NULL, fault_page);
3196	if (fe->pte)
3197	pte_unmap_unlock(fe->pte, fe->ptl);
3198	unlock_page(fault_page);
3199	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE \| VM_FAULT_RETRY)))
3200	put_page(fault_page);
3201	return ret;
3202	}
3203
3204	static int do_cow_fault(struct fault_env *fe, pgoff_t pgoff)
3205	{
3206	struct vm_area_struct *vma = fe->vma;
3207	struct page fault_page, new_page;
3208	void *fault_entry;
3209	struct mem_cgroup *memcg;
3210	int ret;
3211
3212	if (unlikely(anon_vma_prepare(vma)))
3213	return VM_FAULT_OOM;
3214
3215	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, fe->address);
3216	if (!new_page)
3217	return VM_FAULT_OOM;
3218
3219	if (mem_cgroup_try_charge(new_page, vma->vm_mm, GFP_KERNEL,
3220	&memcg, false)) {
3221	put_page(new_page);
3222	return VM_FAULT_OOM;
3223	}
3224
3225	ret = __do_fault(fe, pgoff, new_page, &fault_page, &fault_entry);
3226	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE \| VM_FAULT_RETRY)))
3227	goto uncharge_out;
3228
3229	if (!(ret & VM_FAULT_DAX_LOCKED))
3230	copy_user_highpage(new_page, fault_page, fe->address, vma);
3231	__SetPageUptodate(new_page);
3232
3233	ret \|= alloc_set_pte(fe, memcg, new_page);
3234	if (fe->pte)
3235	pte_unmap_unlock(fe->pte, fe->ptl);
3236	if (!(ret & VM_FAULT_DAX_LOCKED)) {
3237	unlock_page(fault_page);
3238	put_page(fault_page);
3239	} else {
3240	dax_unlock_mapping_entry(vma->vm_file->f_mapping, pgoff);
3241	}
3242	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE \| VM_FAULT_RETRY)))
3243	goto uncharge_out;
3244	return ret;
3245	uncharge_out:
3246	mem_cgroup_cancel_charge(new_page, memcg, false);
3247	put_page(new_page);
3248	return ret;
3249	}
3250
3251	static int do_shared_fault(struct fault_env *fe, pgoff_t pgoff)
3252	{
3253	struct vm_area_struct *vma = fe->vma;
3254	struct page *fault_page;
3255	struct address_space *mapping;
3256	int dirtied = 0;
3257	int ret, tmp;
3258
3259	ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
3260	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE \| VM_FAULT_RETRY)))
3261	return ret;
3262
3263	/*
3264	* Check if the backing address space wants to know that the page is
3265	* about to become writable
3266	*/
3267	if (vma->vm_ops->page_mkwrite) {
3268	unlock_page(fault_page);
3269	tmp = do_page_mkwrite(vma, fault_page, fe->address);
3270	if (unlikely(!tmp \|\|
3271	(tmp & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE)))) {
3272	put_page(fault_page);
3273	return tmp;
3274	}
3275	}
3276
3277	ret \|= alloc_set_pte(fe, NULL, fault_page);
3278	if (fe->pte)
3279	pte_unmap_unlock(fe->pte, fe->ptl);
3280	if (unlikely(ret & (VM_FAULT_ERROR \| VM_FAULT_NOPAGE \|
3281	VM_FAULT_RETRY))) {
3282	unlock_page(fault_page);
3283	put_page(fault_page);
3284	return ret;
3285	}
3286
3287	if (set_page_dirty(fault_page))
3288	dirtied = 1;
3289	/*
3290	* Take a local copy of the address_space - page.mapping may be zeroed
3291	* by truncate after unlock_page(). The address_space itself remains
3292	* pinned by vma->vm_file's reference. We rely on unlock_page()'s
3293	* release semantics to prevent the compiler from undoing this copying.
3294	*/
3295	mapping = page_rmapping(fault_page);
3296	unlock_page(fault_page);
3297	if ((dirtied \|\| vma->vm_ops->page_mkwrite) && mapping) {
3298	/*
3299	* Some device drivers do not set page.mapping but still
3300	* dirty their pages
3301	*/
3302	balance_dirty_pages_ratelimited(mapping);
3303	}
3304
3305	if (!vma->vm_ops->page_mkwrite)
3306	file_update_time(vma->vm_file);
3307
3308	return ret;
3309	}
3310
3311	/*
3312	* We enter with non-exclusive mmap_sem (to exclude vma changes,
3313	* but allow concurrent faults).
3314	* The mmap_sem may have been released depending on flags and our
3315	* return value. See filemap_fault() and __lock_page_or_retry().
3316	*/
3317	static int do_fault(struct fault_env *fe)
3318	{
3319	struct vm_area_struct *vma = fe->vma;
3320	pgoff_t pgoff = linear_page_index(vma, fe->address);
3321
3322	/* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3323	if (!vma->vm_ops->fault)
3324	return VM_FAULT_SIGBUS;
3325	if (!(fe->flags & FAULT_FLAG_WRITE))
3326	return do_read_fault(fe, pgoff);
3327	if (!(vma->vm_flags & VM_SHARED))
3328	return do_cow_fault(fe, pgoff);
3329	return do_shared_fault(fe, pgoff);
3330	}
3331
3332	static int numa_migrate_prep(struct page page, struct vm_area_struct vma,
3333	unsigned long addr, int page_nid,
3334	int *flags)
3335	{
3336	get_page(page);
3337
3338	count_vm_numa_event(NUMA_HINT_FAULTS);
3339	if (page_nid == numa_node_id()) {
3340	count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3341	*flags \|= TNF_FAULT_LOCAL;
3342	}
3343
3344	return mpol_misplaced(page, vma, addr);
3345	}
3346
3347	static int do_numa_page(struct fault_env *fe, pte_t pte)
3348	{
3349	struct vm_area_struct *vma = fe->vma;
3350	struct page *page = NULL;
3351	int page_nid = -1;
3352	int last_cpupid;
3353	int target_nid;
3354	bool migrated = false;
3355	bool was_writable = pte_write(pte);
3356	int flags = 0;
3357
3358	/*
3359	* The "pte" at this point cannot be used safely without
3360	* validation through pte_unmap_same(). It's of NUMA type but
3361	* the pfn may be screwed if the read is non atomic.
3362	*
3363	* We can safely just do a "set_pte_at()", because the old
3364	* page table entry is not accessible, so there would be no
3365	* concurrent hardware modifications to the PTE.
3366	*/
3367	fe->ptl = pte_lockptr(vma->vm_mm, fe->pmd);
3368	spin_lock(fe->ptl);
3369	if (unlikely(!pte_same(*fe->pte, pte))) {
3370	pte_unmap_unlock(fe->pte, fe->ptl);
3371	goto out;
3372	}
3373
3374	/* Make it present again */
3375	pte = pte_modify(pte, vma->vm_page_prot);
3376	pte = pte_mkyoung(pte);
3377	if (was_writable)
3378	pte = pte_mkwrite(pte);
3379	set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
3380	update_mmu_cache(vma, fe->address, fe->pte);
3381
3382	page = vm_normal_page(vma, fe->address, pte);
3383	if (!page) {
3384	pte_unmap_unlock(fe->pte, fe->ptl);
3385	return 0;
3386	}
3387
3388	/* TODO: handle PTE-mapped THP */
3389	if (PageCompound(page)) {
3390	pte_unmap_unlock(fe->pte, fe->ptl);
3391	return 0;
3392	}
3393
3394	/*
3395	* Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3396	* much anyway since they can be in shared cache state. This misses
3397	* the case where a mapping is writable but the process never writes
3398	* to it but pte_write gets cleared during protection updates and
3399	* pte_dirty has unpredictable behaviour between PTE scan updates,
3400	* background writeback, dirty balancing and application behaviour.
3401	*/
3402	if (!pte_write(pte))
3403	flags \|= TNF_NO_GROUP;
3404
3405	/*
3406	* Flag if the page is shared between multiple address spaces. This
3407	* is later used when determining whether to group tasks together
3408	*/
3409	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3410	flags \|= TNF_SHARED;
3411
3412	last_cpupid = page_cpupid_last(page);
3413	page_nid = page_to_nid(page);
3414	target_nid = numa_migrate_prep(page, vma, fe->address, page_nid,
3415	&flags);
3416	pte_unmap_unlock(fe->pte, fe->ptl);
3417	if (target_nid == -1) {
3418	put_page(page);
3419	goto out;
3420	}
3421
3422	/* Migrate to the requested node */
3423	migrated = migrate_misplaced_page(page, vma, target_nid);
3424	if (migrated) {
3425	page_nid = target_nid;
3426	flags \|= TNF_MIGRATED;
3427	} else
3428	flags \|= TNF_MIGRATE_FAIL;
3429
3430	out:
3431	if (page_nid != -1)
3432	task_numa_fault(last_cpupid, page_nid, 1, flags);
3433	return 0;
3434	}
3435
3436	static int create_huge_pmd(struct fault_env *fe)
3437	{
3438	struct vm_area_struct *vma = fe->vma;
3439	if (vma_is_anonymous(vma))
3440	return do_huge_pmd_anonymous_page(fe);
3441	if (vma->vm_ops->pmd_fault)
3442	return vma->vm_ops->pmd_fault(vma, fe->address, fe->pmd,
3443	fe->flags);
3444	return VM_FAULT_FALLBACK;
3445	}
3446
3447	static int wp_huge_pmd(struct fault_env *fe, pmd_t orig_pmd)
3448	{
3449	if (vma_is_anonymous(fe->vma))
3450	return do_huge_pmd_wp_page(fe, orig_pmd);
3451	if (fe->vma->vm_ops->pmd_fault)
3452	return fe->vma->vm_ops->pmd_fault(fe->vma, fe->address, fe->pmd,
3453	fe->flags);
3454
3455	/* COW handled on pte level: split pmd */
3456	VM_BUG_ON_VMA(fe->vma->vm_flags & VM_SHARED, fe->vma);
3457	split_huge_pmd(fe->vma, fe->pmd, fe->address);
3458
3459	return VM_FAULT_FALLBACK;
3460	}
3461
3462	static inline bool vma_is_accessible(struct vm_area_struct *vma)
3463	{
3464	return vma->vm_flags & (VM_READ \| VM_EXEC \| VM_WRITE);
3465	}
3466
3467	/*
3468	* These routines also need to handle stuff like marking pages dirty
3469	* and/or accessed for architectures that don't do it in hardware (most
3470	* RISC architectures). The early dirtying is also good on the i386.
3471	*
3472	* There is also a hook called "update_mmu_cache()" that architectures
3473	* with external mmu caches can use to update those (ie the Sparc or
3474	* PowerPC hashed page tables that act as extended TLBs).
3475	*
3476	* We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3477	* concurrent faults).
3478	*
3479	* The mmap_sem may have been released depending on flags and our return value.
3480	* See filemap_fault() and __lock_page_or_retry().
3481	*/
3482	static int handle_pte_fault(struct fault_env *fe)
3483	{
3484	pte_t entry;
3485
3486	if (unlikely(pmd_none(*fe->pmd))) {
3487	/*
3488	* Leave __pte_alloc() until later: because vm_ops->fault may
3489	* want to allocate huge page, and if we expose page table
3490	* for an instant, it will be difficult to retract from
3491	* concurrent faults and from rmap lookups.
3492	*/
3493	fe->pte = NULL;
3494	} else {
3495	/* See comment in pte_alloc_one_map() */
3496	if (pmd_trans_unstable(fe->pmd) \|\| pmd_devmap(*fe->pmd))
3497	return 0;
3498	/*
3499	* A regular pmd is established and it can't morph into a huge
3500	* pmd from under us anymore at this point because we hold the
3501	* mmap_sem read mode and khugepaged takes it in write mode.
3502	* So now it's safe to run pte_offset_map().
3503	*/
3504	fe->pte = pte_offset_map(fe->pmd, fe->address);
3505
3506	entry = *fe->pte;
3507
3508	/*
3509	* some architectures can have larger ptes than wordsize,
3510	* e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3511	* CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3512	* atomic accesses. The code below just needs a consistent
3513	* view for the ifs and we later double check anyway with the
3514	* ptl lock held. So here a barrier will do.
3515	*/
3516	barrier();
3517	if (pte_none(entry)) {
3518	pte_unmap(fe->pte);
3519	fe->pte = NULL;
3520	}
3521	}
3522
3523	if (!fe->pte) {
3524	if (vma_is_anonymous(fe->vma))
3525	return do_anonymous_page(fe);
3526	else
3527	return do_fault(fe);
3528	}
3529
3530	if (!pte_present(entry))
3531	return do_swap_page(fe, entry);
3532
3533	if (pte_protnone(entry) && vma_is_accessible(fe->vma))
3534	return do_numa_page(fe, entry);
3535
3536	fe->ptl = pte_lockptr(fe->vma->vm_mm, fe->pmd);
3537	spin_lock(fe->ptl);
3538	if (unlikely(!pte_same(*fe->pte, entry)))
3539	goto unlock;
3540	if (fe->flags & FAULT_FLAG_WRITE) {
3541	if (!pte_write(entry))
3542	return do_wp_page(fe, entry);
3543	entry = pte_mkdirty(entry);
3544	}
3545	entry = pte_mkyoung(entry);
3546	if (ptep_set_access_flags(fe->vma, fe->address, fe->pte, entry,
3547	fe->flags & FAULT_FLAG_WRITE)) {
3548	update_mmu_cache(fe->vma, fe->address, fe->pte);
3549	} else {
3550	/*
3551	* This is needed only for protection faults but the arch code
3552	* is not yet telling us if this is a protection fault or not.
3553	* This still avoids useless tlb flushes for .text page faults
3554	* with threads.
3555	*/
3556	if (fe->flags & FAULT_FLAG_WRITE)
3557	flush_tlb_fix_spurious_fault(fe->vma, fe->address);
3558	}
3559	unlock:
3560	pte_unmap_unlock(fe->pte, fe->ptl);
3561	return 0;
3562	}
3563
3564	/*
3565	* By the time we get here, we already hold the mm semaphore
3566	*
3567	* The mmap_sem may have been released depending on flags and our
3568	* return value. See filemap_fault() and __lock_page_or_retry().
3569	*/
3570	static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3571	unsigned int flags)
3572	{
3573	struct fault_env fe = {
3574	.vma = vma,
3575	.address = address,
3576	.flags = flags,
3577	};
3578	struct mm_struct *mm = vma->vm_mm;
3579	pgd_t *pgd;
3580	pud_t *pud;
3581
3582	pgd = pgd_offset(mm, address);
3583	pud = pud_alloc(mm, pgd, address);
3584	if (!pud)
3585	return VM_FAULT_OOM;
3586	fe.pmd = pmd_alloc(mm, pud, address);
3587	if (!fe.pmd)
3588	return VM_FAULT_OOM;
3589	if (pmd_none(*fe.pmd) && transparent_hugepage_enabled(vma)) {
3590	int ret = create_huge_pmd(&fe);
3591	if (!(ret & VM_FAULT_FALLBACK))
3592	return ret;
3593	} else {
3594	pmd_t orig_pmd = *fe.pmd;
3595	int ret;
3596
3597	barrier();
3598	if (pmd_trans_huge(orig_pmd) \|\| pmd_devmap(orig_pmd)) {
3599	if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3600	return do_huge_pmd_numa_page(&fe, orig_pmd);
3601
3602	if ((fe.flags & FAULT_FLAG_WRITE) &&
3603	!pmd_write(orig_pmd)) {
3604	ret = wp_huge_pmd(&fe, orig_pmd);
3605	if (!(ret & VM_FAULT_FALLBACK))
3606	return ret;
3607	} else {
3608	huge_pmd_set_accessed(&fe, orig_pmd);
3609	return 0;
3610	}
3611	}
3612	}
3613
3614	return handle_pte_fault(&fe);
3615	}
3616
3617	/*
3618	* By the time we get here, we already hold the mm semaphore
3619	*
3620	* The mmap_sem may have been released depending on flags and our
3621	* return value. See filemap_fault() and __lock_page_or_retry().
3622	*/
3623	int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3624	unsigned int flags)
3625	{
3626	int ret;
3627
3628	__set_current_state(TASK_RUNNING);
3629
3630	count_vm_event(PGFAULT);
3631	mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
3632
3633	/* do counter updates before entering really critical section. */
3634	check_sync_rss_stat(current);
3635
3636	/*
3637	* Enable the memcg OOM handling for faults triggered in user
3638	* space. Kernel faults are handled more gracefully.
3639	*/
3640	if (flags & FAULT_FLAG_USER)
3641	mem_cgroup_oom_enable();
3642
3643	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3644	flags & FAULT_FLAG_INSTRUCTION,
3645	flags & FAULT_FLAG_REMOTE))
3646	return VM_FAULT_SIGSEGV;
3647
3648	if (unlikely(is_vm_hugetlb_page(vma)))
3649	ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3650	else
3651	ret = __handle_mm_fault(vma, address, flags);
3652
3653	if (flags & FAULT_FLAG_USER) {
3654	mem_cgroup_oom_disable();
3655	/*
3656	* The task may have entered a memcg OOM situation but
3657	* if the allocation error was handled gracefully (no
3658	* VM_FAULT_OOM), there is no need to kill anything.
3659	* Just clean up the OOM state peacefully.
3660	*/
3661	if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3662	mem_cgroup_oom_synchronize(false);
3663	}
3664
3665	/*
3666	* This mm has been already reaped by the oom reaper and so the
3667	* refault cannot be trusted in general. Anonymous refaults would
3668	* lose data and give a zero page instead e.g. This is especially
3669	* problem for use_mm() because regular tasks will just die and
3670	* the corrupted data will not be visible anywhere while kthread
3671	* will outlive the oom victim and potentially propagate the date
3672	* further.
3673	*/
3674	if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3675	&& test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3676	ret = VM_FAULT_SIGBUS;
3677
3678	return ret;
3679	}
3680	EXPORT_SYMBOL_GPL(handle_mm_fault);
3681
3682	#ifndef __PAGETABLE_PUD_FOLDED
3683	/*
3684	* Allocate page upper directory.
3685	* We've already handled the fast-path in-line.
3686	*/
3687	int __pud_alloc(struct mm_struct mm, pgd_t pgd, unsigned long address)
3688	{
3689	pud_t *new = pud_alloc_one(mm, address);
3690	if (!new)
3691	return -ENOMEM;
3692
3693	smp_wmb(); /* See comment in __pte_alloc */
3694
3695	spin_lock(&mm->page_table_lock);
3696	if (pgd_present(pgd)) / Another has populated it */
3697	pud_free(mm, new);
3698	else
3699	pgd_populate(mm, pgd, new);
3700	spin_unlock(&mm->page_table_lock);
3701	return 0;
3702	}
3703	#endif /* __PAGETABLE_PUD_FOLDED */
3704
3705	#ifndef __PAGETABLE_PMD_FOLDED
3706	/*
3707	* Allocate page middle directory.
3708	* We've already handled the fast-path in-line.
3709	*/
3710	int __pmd_alloc(struct mm_struct mm, pud_t pud, unsigned long address)
3711	{
3712	pmd_t *new = pmd_alloc_one(mm, address);
3713	if (!new)
3714	return -ENOMEM;
3715
3716	smp_wmb(); /* See comment in __pte_alloc */
3717
3718	spin_lock(&mm->page_table_lock);
3719	#ifndef __ARCH_HAS_4LEVEL_HACK
3720	if (!pud_present(*pud)) {
3721	mm_inc_nr_pmds(mm);
3722	pud_populate(mm, pud, new);
3723	} else /* Another has populated it */
3724	pmd_free(mm, new);
3725	#else
3726	if (!pgd_present(*pud)) {
3727	mm_inc_nr_pmds(mm);
3728	pgd_populate(mm, pud, new);
3729	} else /* Another has populated it */
3730	pmd_free(mm, new);
3731	#endif /* __ARCH_HAS_4LEVEL_HACK */
3732	spin_unlock(&mm->page_table_lock);
3733	return 0;
3734	}
3735	#endif /* __PAGETABLE_PMD_FOLDED */
3736
3737	static int __follow_pte(struct mm_struct *mm, unsigned long address,
3738	pte_t ptepp, spinlock_t ptlp)
3739	{
3740	pgd_t *pgd;
3741	pud_t *pud;
3742	pmd_t *pmd;
3743	pte_t *ptep;
3744
3745	pgd = pgd_offset(mm, address);
3746	if (pgd_none(pgd) \|\| unlikely(pgd_bad(pgd)))
3747	goto out;
3748
3749	pud = pud_offset(pgd, address);
3750	if (pud_none(pud) \|\| unlikely(pud_bad(pud)))
3751	goto out;
3752
3753	pmd = pmd_offset(pud, address);
3754	VM_BUG_ON(pmd_trans_huge(*pmd));
3755	if (pmd_none(pmd) \|\| unlikely(pmd_bad(pmd)))
3756	goto out;
3757
3758	/* We cannot handle huge page PFN maps. Luckily they don't exist. */
3759	if (pmd_huge(*pmd))
3760	goto out;
3761
3762	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3763	if (!ptep)
3764	goto out;
3765	if (!pte_present(*ptep))
3766	goto unlock;
3767	*ptepp = ptep;
3768	return 0;
3769	unlock:
3770	pte_unmap_unlock(ptep, *ptlp);
3771	out:
3772	return -EINVAL;
3773	}
3774
3775	static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3776	pte_t ptepp, spinlock_t ptlp)
3777	{
3778	int res;
3779
3780	/* (void) is needed to make gcc happy */
3781	(void) __cond_lock(*ptlp,
3782	!(res = __follow_pte(mm, address, ptepp, ptlp)));
3783	return res;
3784	}
3785
3786	/**
3787	* follow_pfn - look up PFN at a user virtual address
3788	* @vma: memory mapping
3789	* @address: user virtual address
3790	* @pfn: location to store found PFN
3791	*
3792	* Only IO mappings and raw PFN mappings are allowed.
3793	*
3794	* Returns zero and the pfn at @pfn on success, -ve otherwise.
3795	*/
3796	int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3797	unsigned long *pfn)
3798	{
3799	int ret = -EINVAL;
3800	spinlock_t *ptl;
3801	pte_t *ptep;
3802
3803	if (!(vma->vm_flags & (VM_IO \| VM_PFNMAP)))
3804	return ret;
3805
3806	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3807	if (ret)
3808	return ret;
3809	pfn = pte_pfn(ptep);
3810	pte_unmap_unlock(ptep, ptl);
3811	return 0;
3812	}
3813	EXPORT_SYMBOL(follow_pfn);
3814
3815	#ifdef CONFIG_HAVE_IOREMAP_PROT
3816	int follow_phys(struct vm_area_struct *vma,
3817	unsigned long address, unsigned int flags,
3818	unsigned long prot, resource_size_t phys)
3819	{
3820	int ret = -EINVAL;
3821	pte_t *ptep, pte;
3822	spinlock_t *ptl;
3823
3824	if (!(vma->vm_flags & (VM_IO \| VM_PFNMAP)))
3825	goto out;
3826
3827	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3828	goto out;
3829	pte = *ptep;
3830
3831	if ((flags & FOLL_WRITE) && !pte_write(pte))
3832	goto unlock;
3833
3834	*prot = pgprot_val(pte_pgprot(pte));
3835	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3836
3837	ret = 0;
3838	unlock:
3839	pte_unmap_unlock(ptep, ptl);
3840	out:
3841	return ret;
3842	}
3843
3844	int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3845	void *buf, int len, int write)
3846	{
3847	resource_size_t phys_addr;
3848	unsigned long prot = 0;
3849	void __iomem *maddr;
3850	int offset = addr & (PAGE_SIZE-1);
3851
3852	if (follow_phys(vma, addr, write, &prot, &phys_addr))
3853	return -EINVAL;
3854
3855	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3856	if (write)
3857	memcpy_toio(maddr + offset, buf, len);
3858	else
3859	memcpy_fromio(buf, maddr + offset, len);
3860	iounmap(maddr);
3861
3862	return len;
3863	}
3864	EXPORT_SYMBOL_GPL(generic_access_phys);
3865	#endif
3866
3867	/*
3868	* Access another process' address space as given in mm. If non-NULL, use the
3869	* given task for page fault accounting.
3870	*/
3871	static int __access_remote_vm(struct task_struct tsk, struct mm_struct mm,
3872	unsigned long addr, void *buf, int len, unsigned int gup_flags)
3873	{
3874	struct vm_area_struct *vma;
3875	void *old_buf = buf;
3876	int write = gup_flags & FOLL_WRITE;
3877
3878	down_read(&mm->mmap_sem);
3879	/* ignore errors, just check how much was successfully transferred */
3880	while (len) {
3881	int bytes, ret, offset;
3882	void *maddr;
3883	struct page *page = NULL;
3884
3885	ret = get_user_pages_remote(tsk, mm, addr, 1,
3886	gup_flags, &page, &vma);
3887	if (ret <= 0) {
3888	#ifndef CONFIG_HAVE_IOREMAP_PROT
3889	break;
3890	#else
3891	/*
3892	* Check if this is a VM_IO \| VM_PFNMAP VMA, which
3893	* we can access using slightly different code.
3894	*/
3895	vma = find_vma(mm, addr);
3896	if (!vma \|\| vma->vm_start > addr)
3897	break;
3898	if (vma->vm_ops && vma->vm_ops->access)
3899	ret = vma->vm_ops->access(vma, addr, buf,
3900	len, write);
3901	if (ret <= 0)
3902	break;
3903	bytes = ret;
3904	#endif
3905	} else {
3906	bytes = len;
3907	offset = addr & (PAGE_SIZE-1);
3908	if (bytes > PAGE_SIZE-offset)
3909	bytes = PAGE_SIZE-offset;
3910
3911	maddr = kmap(page);
3912	if (write) {
3913	copy_to_user_page(vma, page, addr,
3914	maddr + offset, buf, bytes);
3915	set_page_dirty_lock(page);
3916	} else {
3917	copy_from_user_page(vma, page, addr,
3918	buf, maddr + offset, bytes);
3919	}
3920	kunmap(page);
3921	put_page(page);
3922	}
3923	len -= bytes;
3924	buf += bytes;
3925	addr += bytes;
3926	}
3927	up_read(&mm->mmap_sem);
3928
3929	return buf - old_buf;
3930	}
3931
3932	/**
3933	* access_remote_vm - access another process' address space
3934	* @mm: the mm_struct of the target address space
3935	* @addr: start address to access
3936	* @buf: source or destination buffer
3937	* @len: number of bytes to transfer
3938	* @gup_flags: flags modifying lookup behaviour
3939	*
3940	* The caller must hold a reference on @mm.
3941	*/
3942	int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3943	void *buf, int len, unsigned int gup_flags)
3944	{
3945	return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
3946	}
3947
3948	/*
3949	* Access another process' address space.
3950	* Source/target buffer must be kernel space,
3951	* Do not walk the page table directly, use get_user_pages
3952	*/
3953	int access_process_vm(struct task_struct *tsk, unsigned long addr,
3954	void *buf, int len, unsigned int gup_flags)
3955	{
3956	struct mm_struct *mm;
3957	int ret;
3958
3959	mm = get_task_mm(tsk);
3960	if (!mm)
3961	return 0;
3962
3963	ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
3964
3965	mmput(mm);
3966
3967	return ret;
3968	}
3969
3970	/*
3971	* Print the name of a VMA.
3972	*/
3973	void print_vma_addr(char *prefix, unsigned long ip)
3974	{
3975	struct mm_struct *mm = current->mm;
3976	struct vm_area_struct *vma;
3977
3978	/*
3979	* Do not print if we are in atomic
3980	* contexts (in exception stacks, etc.):
3981	*/
3982	if (preempt_count())
3983	return;
3984
3985	down_read(&mm->mmap_sem);
3986	vma = find_vma(mm, ip);
3987	if (vma && vma->vm_file) {
3988	struct file *f = vma->vm_file;
3989	char buf = (char )__get_free_page(GFP_KERNEL);
3990	if (buf) {
3991	char *p;
3992
3993	p = file_path(f, buf, PAGE_SIZE);
3994	if (IS_ERR(p))
3995	p = "?";
3996	printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3997	vma->vm_start,
3998	vma->vm_end - vma->vm_start);
3999	free_page((unsigned long)buf);
4000	}
4001	}
4002	up_read(&mm->mmap_sem);
4003	}
4004
4005	#if defined(CONFIG_PROVE_LOCKING) \|\| defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4006	void __might_fault(const char *file, int line)
4007	{
4008	/*
4009	* Some code (nfs/sunrpc) uses socket ops on kernel memory while
4010	* holding the mmap_sem, this is safe because kernel memory doesn't
4011	* get paged out, therefore we'll never actually fault, and the
4012	* below annotations will generate false positives.
4013	*/
4014	if (segment_eq(get_fs(), KERNEL_DS))
4015	return;
4016	if (pagefault_disabled())
4017	return;
4018	__might_sleep(file, line, 0);
4019	#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4020	if (current->mm)
4021	might_lock_read(&current->mm->mmap_sem);
4022	#endif
4023	}
4024	EXPORT_SYMBOL(__might_fault);
4025	#endif
4026
4027	#if defined(CONFIG_TRANSPARENT_HUGEPAGE) \|\| defined(CONFIG_HUGETLBFS)
4028	static void clear_gigantic_page(struct page *page,
4029	unsigned long addr,
4030	unsigned int pages_per_huge_page)
4031	{
4032	int i;
4033	struct page *p = page;
4034
4035	might_sleep();
4036	for (i = 0; i < pages_per_huge_page;
4037	i++, p = mem_map_next(p, page, i)) {
4038	cond_resched();
4039	clear_user_highpage(p, addr + i * PAGE_SIZE);
4040	}
4041	}
4042	void clear_huge_page(struct page *page,
4043	unsigned long addr, unsigned int pages_per_huge_page)
4044	{
4045	int i;
4046
4047	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4048	clear_gigantic_page(page, addr, pages_per_huge_page);
4049	return;
4050	}
4051
4052	might_sleep();
4053	for (i = 0; i < pages_per_huge_page; i++) {
4054	cond_resched();
4055	clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4056	}
4057	}
4058
4059	static void copy_user_gigantic_page(struct page dst, struct page src,
4060	unsigned long addr,
4061	struct vm_area_struct *vma,
4062	unsigned int pages_per_huge_page)
4063	{
4064	int i;
4065	struct page *dst_base = dst;
4066	struct page *src_base = src;
4067
4068	for (i = 0; i < pages_per_huge_page; ) {
4069	cond_resched();
4070	copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4071
4072	i++;
4073	dst = mem_map_next(dst, dst_base, i);
4074	src = mem_map_next(src, src_base, i);
4075	}
4076	}
4077
4078	void copy_user_huge_page(struct page dst, struct page src,
4079	unsigned long addr, struct vm_area_struct *vma,
4080	unsigned int pages_per_huge_page)
4081	{
4082	int i;
4083
4084	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4085	copy_user_gigantic_page(dst, src, addr, vma,
4086	pages_per_huge_page);
4087	return;
4088	}
4089
4090	might_sleep();
4091	for (i = 0; i < pages_per_huge_page; i++) {
4092	cond_resched();
4093	copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4094	}
4095	}
4096	#endif /* CONFIG_TRANSPARENT_HUGEPAGE \|\| CONFIG_HUGETLBFS */
4097
4098	#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4099
4100	static struct kmem_cache *page_ptl_cachep;
4101
4102	void __init ptlock_cache_init(void)
4103	{
4104	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4105	SLAB_PANIC, NULL);
4106	}
4107
4108	bool ptlock_alloc(struct page *page)
4109	{
4110	spinlock_t *ptl;
4111
4112	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4113	if (!ptl)
4114	return false;
4115	page->ptl = ptl;
4116	return true;
4117	}
4118
4119	void ptlock_free(struct page *page)
4120	{
4121	kmem_cache_free(page_ptl_cachep, page->ptl);
4122	}
4123	#endif