xref: /openbmc/linux/mm/memory.c (revision 75020f2d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/memory.c
4  *
5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
6  */
7 
8 /*
9  * demand-loading started 01.12.91 - seems it is high on the list of
10  * things wanted, and it should be easy to implement. - Linus
11  */
12 
13 /*
14  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15  * pages started 02.12.91, seems to work. - Linus.
16  *
17  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18  * would have taken more than the 6M I have free, but it worked well as
19  * far as I could see.
20  *
21  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
22  */
23 
24 /*
25  * Real VM (paging to/from disk) started 18.12.91. Much more work and
26  * thought has to go into this. Oh, well..
27  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
28  *		Found it. Everything seems to work now.
29  * 20.12.91  -  Ok, making the swap-device changeable like the root.
30  */
31 
32 /*
33  * 05.04.94  -  Multi-page memory management added for v1.1.
34  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
35  *
36  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
37  *		(Gerhard.Wichert@pdb.siemens.de)
38  *
39  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
40  */
41 
42 #include <linux/kernel_stat.h>
43 #include <linux/mm.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73 #include <linux/perf_event.h>
74 #include <linux/ptrace.h>
75 #include <linux/vmalloc.h>
76 
77 #include <trace/events/kmem.h>
78 
79 #include <asm/io.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
83 #include <asm/tlb.h>
84 #include <asm/tlbflush.h>
85 
86 #include "pgalloc-track.h"
87 #include "internal.h"
88 
89 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
91 #endif
92 
93 #ifndef CONFIG_NEED_MULTIPLE_NODES
94 /* use the per-pgdat data instead for discontigmem - mbligh */
95 unsigned long max_mapnr;
96 EXPORT_SYMBOL(max_mapnr);
97 
98 struct page *mem_map;
99 EXPORT_SYMBOL(mem_map);
100 #endif
101 
102 /*
103  * A number of key systems in x86 including ioremap() rely on the assumption
104  * that high_memory defines the upper bound on direct map memory, then end
105  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
106  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
107  * and ZONE_HIGHMEM.
108  */
109 void *high_memory;
110 EXPORT_SYMBOL(high_memory);
111 
112 /*
113  * Randomize the address space (stacks, mmaps, brk, etc.).
114  *
115  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
116  *   as ancient (libc5 based) binaries can segfault. )
117  */
118 int randomize_va_space __read_mostly =
119 #ifdef CONFIG_COMPAT_BRK
120 					1;
121 #else
122 					2;
123 #endif
124 
125 #ifndef arch_faults_on_old_pte
126 static inline bool arch_faults_on_old_pte(void)
127 {
128 	/*
129 	 * Those arches which don't have hw access flag feature need to
130 	 * implement their own helper. By default, "true" means pagefault
131 	 * will be hit on old pte.
132 	 */
133 	return true;
134 }
135 #endif
136 
137 #ifndef arch_wants_old_prefaulted_pte
138 static inline bool arch_wants_old_prefaulted_pte(void)
139 {
140 	/*
141 	 * Transitioning a PTE from 'old' to 'young' can be expensive on
142 	 * some architectures, even if it's performed in hardware. By
143 	 * default, "false" means prefaulted entries will be 'young'.
144 	 */
145 	return false;
146 }
147 #endif
148 
149 static int __init disable_randmaps(char *s)
150 {
151 	randomize_va_space = 0;
152 	return 1;
153 }
154 __setup("norandmaps", disable_randmaps);
155 
156 unsigned long zero_pfn __read_mostly;
157 EXPORT_SYMBOL(zero_pfn);
158 
159 unsigned long highest_memmap_pfn __read_mostly;
160 
161 /*
162  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
163  */
164 static int __init init_zero_pfn(void)
165 {
166 	zero_pfn = page_to_pfn(ZERO_PAGE(0));
167 	return 0;
168 }
169 early_initcall(init_zero_pfn);
170 
171 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
172 {
173 	trace_rss_stat(mm, member, count);
174 }
175 
176 #if defined(SPLIT_RSS_COUNTING)
177 
178 void sync_mm_rss(struct mm_struct *mm)
179 {
180 	int i;
181 
182 	for (i = 0; i < NR_MM_COUNTERS; i++) {
183 		if (current->rss_stat.count[i]) {
184 			add_mm_counter(mm, i, current->rss_stat.count[i]);
185 			current->rss_stat.count[i] = 0;
186 		}
187 	}
188 	current->rss_stat.events = 0;
189 }
190 
191 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
192 {
193 	struct task_struct *task = current;
194 
195 	if (likely(task->mm == mm))
196 		task->rss_stat.count[member] += val;
197 	else
198 		add_mm_counter(mm, member, val);
199 }
200 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
201 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
202 
203 /* sync counter once per 64 page faults */
204 #define TASK_RSS_EVENTS_THRESH	(64)
205 static void check_sync_rss_stat(struct task_struct *task)
206 {
207 	if (unlikely(task != current))
208 		return;
209 	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
210 		sync_mm_rss(task->mm);
211 }
212 #else /* SPLIT_RSS_COUNTING */
213 
214 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
215 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
216 
217 static void check_sync_rss_stat(struct task_struct *task)
218 {
219 }
220 
221 #endif /* SPLIT_RSS_COUNTING */
222 
223 /*
224  * Note: this doesn't free the actual pages themselves. That
225  * has been handled earlier when unmapping all the memory regions.
226  */
227 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
228 			   unsigned long addr)
229 {
230 	pgtable_t token = pmd_pgtable(*pmd);
231 	pmd_clear(pmd);
232 	pte_free_tlb(tlb, token, addr);
233 	mm_dec_nr_ptes(tlb->mm);
234 }
235 
236 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
237 				unsigned long addr, unsigned long end,
238 				unsigned long floor, unsigned long ceiling)
239 {
240 	pmd_t *pmd;
241 	unsigned long next;
242 	unsigned long start;
243 
244 	start = addr;
245 	pmd = pmd_offset(pud, addr);
246 	do {
247 		next = pmd_addr_end(addr, end);
248 		if (pmd_none_or_clear_bad(pmd))
249 			continue;
250 		free_pte_range(tlb, pmd, addr);
251 	} while (pmd++, addr = next, addr != end);
252 
253 	start &= PUD_MASK;
254 	if (start < floor)
255 		return;
256 	if (ceiling) {
257 		ceiling &= PUD_MASK;
258 		if (!ceiling)
259 			return;
260 	}
261 	if (end - 1 > ceiling - 1)
262 		return;
263 
264 	pmd = pmd_offset(pud, start);
265 	pud_clear(pud);
266 	pmd_free_tlb(tlb, pmd, start);
267 	mm_dec_nr_pmds(tlb->mm);
268 }
269 
270 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
271 				unsigned long addr, unsigned long end,
272 				unsigned long floor, unsigned long ceiling)
273 {
274 	pud_t *pud;
275 	unsigned long next;
276 	unsigned long start;
277 
278 	start = addr;
279 	pud = pud_offset(p4d, addr);
280 	do {
281 		next = pud_addr_end(addr, end);
282 		if (pud_none_or_clear_bad(pud))
283 			continue;
284 		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
285 	} while (pud++, addr = next, addr != end);
286 
287 	start &= P4D_MASK;
288 	if (start < floor)
289 		return;
290 	if (ceiling) {
291 		ceiling &= P4D_MASK;
292 		if (!ceiling)
293 			return;
294 	}
295 	if (end - 1 > ceiling - 1)
296 		return;
297 
298 	pud = pud_offset(p4d, start);
299 	p4d_clear(p4d);
300 	pud_free_tlb(tlb, pud, start);
301 	mm_dec_nr_puds(tlb->mm);
302 }
303 
304 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
305 				unsigned long addr, unsigned long end,
306 				unsigned long floor, unsigned long ceiling)
307 {
308 	p4d_t *p4d;
309 	unsigned long next;
310 	unsigned long start;
311 
312 	start = addr;
313 	p4d = p4d_offset(pgd, addr);
314 	do {
315 		next = p4d_addr_end(addr, end);
316 		if (p4d_none_or_clear_bad(p4d))
317 			continue;
318 		free_pud_range(tlb, p4d, addr, next, floor, ceiling);
319 	} while (p4d++, addr = next, addr != end);
320 
321 	start &= PGDIR_MASK;
322 	if (start < floor)
323 		return;
324 	if (ceiling) {
325 		ceiling &= PGDIR_MASK;
326 		if (!ceiling)
327 			return;
328 	}
329 	if (end - 1 > ceiling - 1)
330 		return;
331 
332 	p4d = p4d_offset(pgd, start);
333 	pgd_clear(pgd);
334 	p4d_free_tlb(tlb, p4d, start);
335 }
336 
337 /*
338  * This function frees user-level page tables of a process.
339  */
340 void free_pgd_range(struct mmu_gather *tlb,
341 			unsigned long addr, unsigned long end,
342 			unsigned long floor, unsigned long ceiling)
343 {
344 	pgd_t *pgd;
345 	unsigned long next;
346 
347 	/*
348 	 * The next few lines have given us lots of grief...
349 	 *
350 	 * Why are we testing PMD* at this top level?  Because often
351 	 * there will be no work to do at all, and we'd prefer not to
352 	 * go all the way down to the bottom just to discover that.
353 	 *
354 	 * Why all these "- 1"s?  Because 0 represents both the bottom
355 	 * of the address space and the top of it (using -1 for the
356 	 * top wouldn't help much: the masks would do the wrong thing).
357 	 * The rule is that addr 0 and floor 0 refer to the bottom of
358 	 * the address space, but end 0 and ceiling 0 refer to the top
359 	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
360 	 * that end 0 case should be mythical).
361 	 *
362 	 * Wherever addr is brought up or ceiling brought down, we must
363 	 * be careful to reject "the opposite 0" before it confuses the
364 	 * subsequent tests.  But what about where end is brought down
365 	 * by PMD_SIZE below? no, end can't go down to 0 there.
366 	 *
367 	 * Whereas we round start (addr) and ceiling down, by different
368 	 * masks at different levels, in order to test whether a table
369 	 * now has no other vmas using it, so can be freed, we don't
370 	 * bother to round floor or end up - the tests don't need that.
371 	 */
372 
373 	addr &= PMD_MASK;
374 	if (addr < floor) {
375 		addr += PMD_SIZE;
376 		if (!addr)
377 			return;
378 	}
379 	if (ceiling) {
380 		ceiling &= PMD_MASK;
381 		if (!ceiling)
382 			return;
383 	}
384 	if (end - 1 > ceiling - 1)
385 		end -= PMD_SIZE;
386 	if (addr > end - 1)
387 		return;
388 	/*
389 	 * We add page table cache pages with PAGE_SIZE,
390 	 * (see pte_free_tlb()), flush the tlb if we need
391 	 */
392 	tlb_change_page_size(tlb, PAGE_SIZE);
393 	pgd = pgd_offset(tlb->mm, addr);
394 	do {
395 		next = pgd_addr_end(addr, end);
396 		if (pgd_none_or_clear_bad(pgd))
397 			continue;
398 		free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
399 	} while (pgd++, addr = next, addr != end);
400 }
401 
402 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
403 		unsigned long floor, unsigned long ceiling)
404 {
405 	while (vma) {
406 		struct vm_area_struct *next = vma->vm_next;
407 		unsigned long addr = vma->vm_start;
408 
409 		/*
410 		 * Hide vma from rmap and truncate_pagecache before freeing
411 		 * pgtables
412 		 */
413 		unlink_anon_vmas(vma);
414 		unlink_file_vma(vma);
415 
416 		if (is_vm_hugetlb_page(vma)) {
417 			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
418 				floor, next ? next->vm_start : ceiling);
419 		} else {
420 			/*
421 			 * Optimization: gather nearby vmas into one call down
422 			 */
423 			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
424 			       && !is_vm_hugetlb_page(next)) {
425 				vma = next;
426 				next = vma->vm_next;
427 				unlink_anon_vmas(vma);
428 				unlink_file_vma(vma);
429 			}
430 			free_pgd_range(tlb, addr, vma->vm_end,
431 				floor, next ? next->vm_start : ceiling);
432 		}
433 		vma = next;
434 	}
435 }
436 
437 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
438 {
439 	spinlock_t *ptl;
440 	pgtable_t new = pte_alloc_one(mm);
441 	if (!new)
442 		return -ENOMEM;
443 
444 	/*
445 	 * Ensure all pte setup (eg. pte page lock and page clearing) are
446 	 * visible before the pte is made visible to other CPUs by being
447 	 * put into page tables.
448 	 *
449 	 * The other side of the story is the pointer chasing in the page
450 	 * table walking code (when walking the page table without locking;
451 	 * ie. most of the time). Fortunately, these data accesses consist
452 	 * of a chain of data-dependent loads, meaning most CPUs (alpha
453 	 * being the notable exception) will already guarantee loads are
454 	 * seen in-order. See the alpha page table accessors for the
455 	 * smp_rmb() barriers in page table walking code.
456 	 */
457 	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
458 
459 	ptl = pmd_lock(mm, pmd);
460 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
461 		mm_inc_nr_ptes(mm);
462 		pmd_populate(mm, pmd, new);
463 		new = NULL;
464 	}
465 	spin_unlock(ptl);
466 	if (new)
467 		pte_free(mm, new);
468 	return 0;
469 }
470 
471 int __pte_alloc_kernel(pmd_t *pmd)
472 {
473 	pte_t *new = pte_alloc_one_kernel(&init_mm);
474 	if (!new)
475 		return -ENOMEM;
476 
477 	smp_wmb(); /* See comment in __pte_alloc */
478 
479 	spin_lock(&init_mm.page_table_lock);
480 	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
481 		pmd_populate_kernel(&init_mm, pmd, new);
482 		new = NULL;
483 	}
484 	spin_unlock(&init_mm.page_table_lock);
485 	if (new)
486 		pte_free_kernel(&init_mm, new);
487 	return 0;
488 }
489 
490 static inline void init_rss_vec(int *rss)
491 {
492 	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
493 }
494 
495 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
496 {
497 	int i;
498 
499 	if (current->mm == mm)
500 		sync_mm_rss(mm);
501 	for (i = 0; i < NR_MM_COUNTERS; i++)
502 		if (rss[i])
503 			add_mm_counter(mm, i, rss[i]);
504 }
505 
506 /*
507  * This function is called to print an error when a bad pte
508  * is found. For example, we might have a PFN-mapped pte in
509  * a region that doesn't allow it.
510  *
511  * The calling function must still handle the error.
512  */
513 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
514 			  pte_t pte, struct page *page)
515 {
516 	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
517 	p4d_t *p4d = p4d_offset(pgd, addr);
518 	pud_t *pud = pud_offset(p4d, addr);
519 	pmd_t *pmd = pmd_offset(pud, addr);
520 	struct address_space *mapping;
521 	pgoff_t index;
522 	static unsigned long resume;
523 	static unsigned long nr_shown;
524 	static unsigned long nr_unshown;
525 
526 	/*
527 	 * Allow a burst of 60 reports, then keep quiet for that minute;
528 	 * or allow a steady drip of one report per second.
529 	 */
530 	if (nr_shown == 60) {
531 		if (time_before(jiffies, resume)) {
532 			nr_unshown++;
533 			return;
534 		}
535 		if (nr_unshown) {
536 			pr_alert("BUG: Bad page map: %lu messages suppressed\n",
537 				 nr_unshown);
538 			nr_unshown = 0;
539 		}
540 		nr_shown = 0;
541 	}
542 	if (nr_shown++ == 0)
543 		resume = jiffies + 60 * HZ;
544 
545 	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
546 	index = linear_page_index(vma, addr);
547 
548 	pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
549 		 current->comm,
550 		 (long long)pte_val(pte), (long long)pmd_val(*pmd));
551 	if (page)
552 		dump_page(page, "bad pte");
553 	pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
554 		 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
555 	pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
556 		 vma->vm_file,
557 		 vma->vm_ops ? vma->vm_ops->fault : NULL,
558 		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
559 		 mapping ? mapping->a_ops->readpage : NULL);
560 	dump_stack();
561 	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
562 }
563 
564 /*
565  * vm_normal_page -- This function gets the "struct page" associated with a pte.
566  *
567  * "Special" mappings do not wish to be associated with a "struct page" (either
568  * it doesn't exist, or it exists but they don't want to touch it). In this
569  * case, NULL is returned here. "Normal" mappings do have a struct page.
570  *
571  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
572  * pte bit, in which case this function is trivial. Secondly, an architecture
573  * may not have a spare pte bit, which requires a more complicated scheme,
574  * described below.
575  *
576  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
577  * special mapping (even if there are underlying and valid "struct pages").
578  * COWed pages of a VM_PFNMAP are always normal.
579  *
580  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
581  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
582  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
583  * mapping will always honor the rule
584  *
585  *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
586  *
587  * And for normal mappings this is false.
588  *
589  * This restricts such mappings to be a linear translation from virtual address
590  * to pfn. To get around this restriction, we allow arbitrary mappings so long
591  * as the vma is not a COW mapping; in that case, we know that all ptes are
592  * special (because none can have been COWed).
593  *
594  *
595  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
596  *
597  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
598  * page" backing, however the difference is that _all_ pages with a struct
599  * page (that is, those where pfn_valid is true) are refcounted and considered
600  * normal pages by the VM. The disadvantage is that pages are refcounted
601  * (which can be slower and simply not an option for some PFNMAP users). The
602  * advantage is that we don't have to follow the strict linearity rule of
603  * PFNMAP mappings in order to support COWable mappings.
604  *
605  */
606 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
607 			    pte_t pte)
608 {
609 	unsigned long pfn = pte_pfn(pte);
610 
611 	if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
612 		if (likely(!pte_special(pte)))
613 			goto check_pfn;
614 		if (vma->vm_ops && vma->vm_ops->find_special_page)
615 			return vma->vm_ops->find_special_page(vma, addr);
616 		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
617 			return NULL;
618 		if (is_zero_pfn(pfn))
619 			return NULL;
620 		if (pte_devmap(pte))
621 			return NULL;
622 
623 		print_bad_pte(vma, addr, pte, NULL);
624 		return NULL;
625 	}
626 
627 	/* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
628 
629 	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
630 		if (vma->vm_flags & VM_MIXEDMAP) {
631 			if (!pfn_valid(pfn))
632 				return NULL;
633 			goto out;
634 		} else {
635 			unsigned long off;
636 			off = (addr - vma->vm_start) >> PAGE_SHIFT;
637 			if (pfn == vma->vm_pgoff + off)
638 				return NULL;
639 			if (!is_cow_mapping(vma->vm_flags))
640 				return NULL;
641 		}
642 	}
643 
644 	if (is_zero_pfn(pfn))
645 		return NULL;
646 
647 check_pfn:
648 	if (unlikely(pfn > highest_memmap_pfn)) {
649 		print_bad_pte(vma, addr, pte, NULL);
650 		return NULL;
651 	}
652 
653 	/*
654 	 * NOTE! We still have PageReserved() pages in the page tables.
655 	 * eg. VDSO mappings can cause them to exist.
656 	 */
657 out:
658 	return pfn_to_page(pfn);
659 }
660 
661 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
662 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
663 				pmd_t pmd)
664 {
665 	unsigned long pfn = pmd_pfn(pmd);
666 
667 	/*
668 	 * There is no pmd_special() but there may be special pmds, e.g.
669 	 * in a direct-access (dax) mapping, so let's just replicate the
670 	 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
671 	 */
672 	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
673 		if (vma->vm_flags & VM_MIXEDMAP) {
674 			if (!pfn_valid(pfn))
675 				return NULL;
676 			goto out;
677 		} else {
678 			unsigned long off;
679 			off = (addr - vma->vm_start) >> PAGE_SHIFT;
680 			if (pfn == vma->vm_pgoff + off)
681 				return NULL;
682 			if (!is_cow_mapping(vma->vm_flags))
683 				return NULL;
684 		}
685 	}
686 
687 	if (pmd_devmap(pmd))
688 		return NULL;
689 	if (is_huge_zero_pmd(pmd))
690 		return NULL;
691 	if (unlikely(pfn > highest_memmap_pfn))
692 		return NULL;
693 
694 	/*
695 	 * NOTE! We still have PageReserved() pages in the page tables.
696 	 * eg. VDSO mappings can cause them to exist.
697 	 */
698 out:
699 	return pfn_to_page(pfn);
700 }
701 #endif
702 
703 /*
704  * copy one vm_area from one task to the other. Assumes the page tables
705  * already present in the new task to be cleared in the whole range
706  * covered by this vma.
707  */
708 
709 static unsigned long
710 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
711 		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
712 		unsigned long addr, int *rss)
713 {
714 	unsigned long vm_flags = vma->vm_flags;
715 	pte_t pte = *src_pte;
716 	struct page *page;
717 	swp_entry_t entry = pte_to_swp_entry(pte);
718 
719 	if (likely(!non_swap_entry(entry))) {
720 		if (swap_duplicate(entry) < 0)
721 			return entry.val;
722 
723 		/* make sure dst_mm is on swapoff's mmlist. */
724 		if (unlikely(list_empty(&dst_mm->mmlist))) {
725 			spin_lock(&mmlist_lock);
726 			if (list_empty(&dst_mm->mmlist))
727 				list_add(&dst_mm->mmlist,
728 						&src_mm->mmlist);
729 			spin_unlock(&mmlist_lock);
730 		}
731 		rss[MM_SWAPENTS]++;
732 	} else if (is_migration_entry(entry)) {
733 		page = migration_entry_to_page(entry);
734 
735 		rss[mm_counter(page)]++;
736 
737 		if (is_write_migration_entry(entry) &&
738 				is_cow_mapping(vm_flags)) {
739 			/*
740 			 * COW mappings require pages in both
741 			 * parent and child to be set to read.
742 			 */
743 			make_migration_entry_read(&entry);
744 			pte = swp_entry_to_pte(entry);
745 			if (pte_swp_soft_dirty(*src_pte))
746 				pte = pte_swp_mksoft_dirty(pte);
747 			if (pte_swp_uffd_wp(*src_pte))
748 				pte = pte_swp_mkuffd_wp(pte);
749 			set_pte_at(src_mm, addr, src_pte, pte);
750 		}
751 	} else if (is_device_private_entry(entry)) {
752 		page = device_private_entry_to_page(entry);
753 
754 		/*
755 		 * Update rss count even for unaddressable pages, as
756 		 * they should treated just like normal pages in this
757 		 * respect.
758 		 *
759 		 * We will likely want to have some new rss counters
760 		 * for unaddressable pages, at some point. But for now
761 		 * keep things as they are.
762 		 */
763 		get_page(page);
764 		rss[mm_counter(page)]++;
765 		page_dup_rmap(page, false);
766 
767 		/*
768 		 * We do not preserve soft-dirty information, because so
769 		 * far, checkpoint/restore is the only feature that
770 		 * requires that. And checkpoint/restore does not work
771 		 * when a device driver is involved (you cannot easily
772 		 * save and restore device driver state).
773 		 */
774 		if (is_write_device_private_entry(entry) &&
775 		    is_cow_mapping(vm_flags)) {
776 			make_device_private_entry_read(&entry);
777 			pte = swp_entry_to_pte(entry);
778 			if (pte_swp_uffd_wp(*src_pte))
779 				pte = pte_swp_mkuffd_wp(pte);
780 			set_pte_at(src_mm, addr, src_pte, pte);
781 		}
782 	}
783 	set_pte_at(dst_mm, addr, dst_pte, pte);
784 	return 0;
785 }
786 
787 /*
788  * Copy a present and normal page if necessary.
789  *
790  * NOTE! The usual case is that this doesn't need to do
791  * anything, and can just return a positive value. That
792  * will let the caller know that it can just increase
793  * the page refcount and re-use the pte the traditional
794  * way.
795  *
796  * But _if_ we need to copy it because it needs to be
797  * pinned in the parent (and the child should get its own
798  * copy rather than just a reference to the same page),
799  * we'll do that here and return zero to let the caller
800  * know we're done.
801  *
802  * And if we need a pre-allocated page but don't yet have
803  * one, return a negative error to let the preallocation
804  * code know so that it can do so outside the page table
805  * lock.
806  */
807 static inline int
808 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
809 		  pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
810 		  struct page **prealloc, pte_t pte, struct page *page)
811 {
812 	struct page *new_page;
813 
814 	/*
815 	 * What we want to do is to check whether this page may
816 	 * have been pinned by the parent process.  If so,
817 	 * instead of wrprotect the pte on both sides, we copy
818 	 * the page immediately so that we'll always guarantee
819 	 * the pinned page won't be randomly replaced in the
820 	 * future.
821 	 *
822 	 * The page pinning checks are just "has this mm ever
823 	 * seen pinning", along with the (inexact) check of
824 	 * the page count. That might give false positives for
825 	 * for pinning, but it will work correctly.
826 	 */
827 	if (likely(!page_needs_cow_for_dma(src_vma, page)))
828 		return 1;
829 
830 	new_page = *prealloc;
831 	if (!new_page)
832 		return -EAGAIN;
833 
834 	/*
835 	 * We have a prealloc page, all good!  Take it
836 	 * over and copy the page & arm it.
837 	 */
838 	*prealloc = NULL;
839 	copy_user_highpage(new_page, page, addr, src_vma);
840 	__SetPageUptodate(new_page);
841 	page_add_new_anon_rmap(new_page, dst_vma, addr, false);
842 	lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
843 	rss[mm_counter(new_page)]++;
844 
845 	/* All done, just insert the new page copy in the child */
846 	pte = mk_pte(new_page, dst_vma->vm_page_prot);
847 	pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
848 	set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
849 	return 0;
850 }
851 
852 /*
853  * Copy one pte.  Returns 0 if succeeded, or -EAGAIN if one preallocated page
854  * is required to copy this pte.
855  */
856 static inline int
857 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
858 		 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
859 		 struct page **prealloc)
860 {
861 	struct mm_struct *src_mm = src_vma->vm_mm;
862 	unsigned long vm_flags = src_vma->vm_flags;
863 	pte_t pte = *src_pte;
864 	struct page *page;
865 
866 	page = vm_normal_page(src_vma, addr, pte);
867 	if (page) {
868 		int retval;
869 
870 		retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
871 					   addr, rss, prealloc, pte, page);
872 		if (retval <= 0)
873 			return retval;
874 
875 		get_page(page);
876 		page_dup_rmap(page, false);
877 		rss[mm_counter(page)]++;
878 	}
879 
880 	/*
881 	 * If it's a COW mapping, write protect it both
882 	 * in the parent and the child
883 	 */
884 	if (is_cow_mapping(vm_flags) && pte_write(pte)) {
885 		ptep_set_wrprotect(src_mm, addr, src_pte);
886 		pte = pte_wrprotect(pte);
887 	}
888 
889 	/*
890 	 * If it's a shared mapping, mark it clean in
891 	 * the child
892 	 */
893 	if (vm_flags & VM_SHARED)
894 		pte = pte_mkclean(pte);
895 	pte = pte_mkold(pte);
896 
897 	/*
898 	 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
899 	 * does not have the VM_UFFD_WP, which means that the uffd
900 	 * fork event is not enabled.
901 	 */
902 	if (!(vm_flags & VM_UFFD_WP))
903 		pte = pte_clear_uffd_wp(pte);
904 
905 	set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
906 	return 0;
907 }
908 
909 static inline struct page *
910 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
911 		   unsigned long addr)
912 {
913 	struct page *new_page;
914 
915 	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
916 	if (!new_page)
917 		return NULL;
918 
919 	if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
920 		put_page(new_page);
921 		return NULL;
922 	}
923 	cgroup_throttle_swaprate(new_page, GFP_KERNEL);
924 
925 	return new_page;
926 }
927 
928 static int
929 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
930 	       pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
931 	       unsigned long end)
932 {
933 	struct mm_struct *dst_mm = dst_vma->vm_mm;
934 	struct mm_struct *src_mm = src_vma->vm_mm;
935 	pte_t *orig_src_pte, *orig_dst_pte;
936 	pte_t *src_pte, *dst_pte;
937 	spinlock_t *src_ptl, *dst_ptl;
938 	int progress, ret = 0;
939 	int rss[NR_MM_COUNTERS];
940 	swp_entry_t entry = (swp_entry_t){0};
941 	struct page *prealloc = NULL;
942 
943 again:
944 	progress = 0;
945 	init_rss_vec(rss);
946 
947 	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
948 	if (!dst_pte) {
949 		ret = -ENOMEM;
950 		goto out;
951 	}
952 	src_pte = pte_offset_map(src_pmd, addr);
953 	src_ptl = pte_lockptr(src_mm, src_pmd);
954 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
955 	orig_src_pte = src_pte;
956 	orig_dst_pte = dst_pte;
957 	arch_enter_lazy_mmu_mode();
958 
959 	do {
960 		/*
961 		 * We are holding two locks at this point - either of them
962 		 * could generate latencies in another task on another CPU.
963 		 */
964 		if (progress >= 32) {
965 			progress = 0;
966 			if (need_resched() ||
967 			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
968 				break;
969 		}
970 		if (pte_none(*src_pte)) {
971 			progress++;
972 			continue;
973 		}
974 		if (unlikely(!pte_present(*src_pte))) {
975 			entry.val = copy_nonpresent_pte(dst_mm, src_mm,
976 							dst_pte, src_pte,
977 							src_vma, addr, rss);
978 			if (entry.val)
979 				break;
980 			progress += 8;
981 			continue;
982 		}
983 		/* copy_present_pte() will clear `*prealloc' if consumed */
984 		ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
985 				       addr, rss, &prealloc);
986 		/*
987 		 * If we need a pre-allocated page for this pte, drop the
988 		 * locks, allocate, and try again.
989 		 */
990 		if (unlikely(ret == -EAGAIN))
991 			break;
992 		if (unlikely(prealloc)) {
993 			/*
994 			 * pre-alloc page cannot be reused by next time so as
995 			 * to strictly follow mempolicy (e.g., alloc_page_vma()
996 			 * will allocate page according to address).  This
997 			 * could only happen if one pinned pte changed.
998 			 */
999 			put_page(prealloc);
1000 			prealloc = NULL;
1001 		}
1002 		progress += 8;
1003 	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1004 
1005 	arch_leave_lazy_mmu_mode();
1006 	spin_unlock(src_ptl);
1007 	pte_unmap(orig_src_pte);
1008 	add_mm_rss_vec(dst_mm, rss);
1009 	pte_unmap_unlock(orig_dst_pte, dst_ptl);
1010 	cond_resched();
1011 
1012 	if (entry.val) {
1013 		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1014 			ret = -ENOMEM;
1015 			goto out;
1016 		}
1017 		entry.val = 0;
1018 	} else if (ret) {
1019 		WARN_ON_ONCE(ret != -EAGAIN);
1020 		prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1021 		if (!prealloc)
1022 			return -ENOMEM;
1023 		/* We've captured and resolved the error. Reset, try again. */
1024 		ret = 0;
1025 	}
1026 	if (addr != end)
1027 		goto again;
1028 out:
1029 	if (unlikely(prealloc))
1030 		put_page(prealloc);
1031 	return ret;
1032 }
1033 
1034 static inline int
1035 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1036 	       pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1037 	       unsigned long end)
1038 {
1039 	struct mm_struct *dst_mm = dst_vma->vm_mm;
1040 	struct mm_struct *src_mm = src_vma->vm_mm;
1041 	pmd_t *src_pmd, *dst_pmd;
1042 	unsigned long next;
1043 
1044 	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1045 	if (!dst_pmd)
1046 		return -ENOMEM;
1047 	src_pmd = pmd_offset(src_pud, addr);
1048 	do {
1049 		next = pmd_addr_end(addr, end);
1050 		if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1051 			|| pmd_devmap(*src_pmd)) {
1052 			int err;
1053 			VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1054 			err = copy_huge_pmd(dst_mm, src_mm,
1055 					    dst_pmd, src_pmd, addr, src_vma);
1056 			if (err == -ENOMEM)
1057 				return -ENOMEM;
1058 			if (!err)
1059 				continue;
1060 			/* fall through */
1061 		}
1062 		if (pmd_none_or_clear_bad(src_pmd))
1063 			continue;
1064 		if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1065 				   addr, next))
1066 			return -ENOMEM;
1067 	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1068 	return 0;
1069 }
1070 
1071 static inline int
1072 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1073 	       p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1074 	       unsigned long end)
1075 {
1076 	struct mm_struct *dst_mm = dst_vma->vm_mm;
1077 	struct mm_struct *src_mm = src_vma->vm_mm;
1078 	pud_t *src_pud, *dst_pud;
1079 	unsigned long next;
1080 
1081 	dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1082 	if (!dst_pud)
1083 		return -ENOMEM;
1084 	src_pud = pud_offset(src_p4d, addr);
1085 	do {
1086 		next = pud_addr_end(addr, end);
1087 		if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1088 			int err;
1089 
1090 			VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1091 			err = copy_huge_pud(dst_mm, src_mm,
1092 					    dst_pud, src_pud, addr, src_vma);
1093 			if (err == -ENOMEM)
1094 				return -ENOMEM;
1095 			if (!err)
1096 				continue;
1097 			/* fall through */
1098 		}
1099 		if (pud_none_or_clear_bad(src_pud))
1100 			continue;
1101 		if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1102 				   addr, next))
1103 			return -ENOMEM;
1104 	} while (dst_pud++, src_pud++, addr = next, addr != end);
1105 	return 0;
1106 }
1107 
1108 static inline int
1109 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1110 	       pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1111 	       unsigned long end)
1112 {
1113 	struct mm_struct *dst_mm = dst_vma->vm_mm;
1114 	p4d_t *src_p4d, *dst_p4d;
1115 	unsigned long next;
1116 
1117 	dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1118 	if (!dst_p4d)
1119 		return -ENOMEM;
1120 	src_p4d = p4d_offset(src_pgd, addr);
1121 	do {
1122 		next = p4d_addr_end(addr, end);
1123 		if (p4d_none_or_clear_bad(src_p4d))
1124 			continue;
1125 		if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1126 				   addr, next))
1127 			return -ENOMEM;
1128 	} while (dst_p4d++, src_p4d++, addr = next, addr != end);
1129 	return 0;
1130 }
1131 
1132 int
1133 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1134 {
1135 	pgd_t *src_pgd, *dst_pgd;
1136 	unsigned long next;
1137 	unsigned long addr = src_vma->vm_start;
1138 	unsigned long end = src_vma->vm_end;
1139 	struct mm_struct *dst_mm = dst_vma->vm_mm;
1140 	struct mm_struct *src_mm = src_vma->vm_mm;
1141 	struct mmu_notifier_range range;
1142 	bool is_cow;
1143 	int ret;
1144 
1145 	/*
1146 	 * Don't copy ptes where a page fault will fill them correctly.
1147 	 * Fork becomes much lighter when there are big shared or private
1148 	 * readonly mappings. The tradeoff is that copy_page_range is more
1149 	 * efficient than faulting.
1150 	 */
1151 	if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1152 	    !src_vma->anon_vma)
1153 		return 0;
1154 
1155 	if (is_vm_hugetlb_page(src_vma))
1156 		return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1157 
1158 	if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1159 		/*
1160 		 * We do not free on error cases below as remove_vma
1161 		 * gets called on error from higher level routine
1162 		 */
1163 		ret = track_pfn_copy(src_vma);
1164 		if (ret)
1165 			return ret;
1166 	}
1167 
1168 	/*
1169 	 * We need to invalidate the secondary MMU mappings only when
1170 	 * there could be a permission downgrade on the ptes of the
1171 	 * parent mm. And a permission downgrade will only happen if
1172 	 * is_cow_mapping() returns true.
1173 	 */
1174 	is_cow = is_cow_mapping(src_vma->vm_flags);
1175 
1176 	if (is_cow) {
1177 		mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1178 					0, src_vma, src_mm, addr, end);
1179 		mmu_notifier_invalidate_range_start(&range);
1180 		/*
1181 		 * Disabling preemption is not needed for the write side, as
1182 		 * the read side doesn't spin, but goes to the mmap_lock.
1183 		 *
1184 		 * Use the raw variant of the seqcount_t write API to avoid
1185 		 * lockdep complaining about preemptibility.
1186 		 */
1187 		mmap_assert_write_locked(src_mm);
1188 		raw_write_seqcount_begin(&src_mm->write_protect_seq);
1189 	}
1190 
1191 	ret = 0;
1192 	dst_pgd = pgd_offset(dst_mm, addr);
1193 	src_pgd = pgd_offset(src_mm, addr);
1194 	do {
1195 		next = pgd_addr_end(addr, end);
1196 		if (pgd_none_or_clear_bad(src_pgd))
1197 			continue;
1198 		if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1199 					    addr, next))) {
1200 			ret = -ENOMEM;
1201 			break;
1202 		}
1203 	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1204 
1205 	if (is_cow) {
1206 		raw_write_seqcount_end(&src_mm->write_protect_seq);
1207 		mmu_notifier_invalidate_range_end(&range);
1208 	}
1209 	return ret;
1210 }
1211 
1212 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1213 				struct vm_area_struct *vma, pmd_t *pmd,
1214 				unsigned long addr, unsigned long end,
1215 				struct zap_details *details)
1216 {
1217 	struct mm_struct *mm = tlb->mm;
1218 	int force_flush = 0;
1219 	int rss[NR_MM_COUNTERS];
1220 	spinlock_t *ptl;
1221 	pte_t *start_pte;
1222 	pte_t *pte;
1223 	swp_entry_t entry;
1224 
1225 	tlb_change_page_size(tlb, PAGE_SIZE);
1226 again:
1227 	init_rss_vec(rss);
1228 	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1229 	pte = start_pte;
1230 	flush_tlb_batched_pending(mm);
1231 	arch_enter_lazy_mmu_mode();
1232 	do {
1233 		pte_t ptent = *pte;
1234 		if (pte_none(ptent))
1235 			continue;
1236 
1237 		if (need_resched())
1238 			break;
1239 
1240 		if (pte_present(ptent)) {
1241 			struct page *page;
1242 
1243 			page = vm_normal_page(vma, addr, ptent);
1244 			if (unlikely(details) && page) {
1245 				/*
1246 				 * unmap_shared_mapping_pages() wants to
1247 				 * invalidate cache without truncating:
1248 				 * unmap shared but keep private pages.
1249 				 */
1250 				if (details->check_mapping &&
1251 				    details->check_mapping != page_rmapping(page))
1252 					continue;
1253 			}
1254 			ptent = ptep_get_and_clear_full(mm, addr, pte,
1255 							tlb->fullmm);
1256 			tlb_remove_tlb_entry(tlb, pte, addr);
1257 			if (unlikely(!page))
1258 				continue;
1259 
1260 			if (!PageAnon(page)) {
1261 				if (pte_dirty(ptent)) {
1262 					force_flush = 1;
1263 					set_page_dirty(page);
1264 				}
1265 				if (pte_young(ptent) &&
1266 				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1267 					mark_page_accessed(page);
1268 			}
1269 			rss[mm_counter(page)]--;
1270 			page_remove_rmap(page, false);
1271 			if (unlikely(page_mapcount(page) < 0))
1272 				print_bad_pte(vma, addr, ptent, page);
1273 			if (unlikely(__tlb_remove_page(tlb, page))) {
1274 				force_flush = 1;
1275 				addr += PAGE_SIZE;
1276 				break;
1277 			}
1278 			continue;
1279 		}
1280 
1281 		entry = pte_to_swp_entry(ptent);
1282 		if (is_device_private_entry(entry)) {
1283 			struct page *page = device_private_entry_to_page(entry);
1284 
1285 			if (unlikely(details && details->check_mapping)) {
1286 				/*
1287 				 * unmap_shared_mapping_pages() wants to
1288 				 * invalidate cache without truncating:
1289 				 * unmap shared but keep private pages.
1290 				 */
1291 				if (details->check_mapping !=
1292 				    page_rmapping(page))
1293 					continue;
1294 			}
1295 
1296 			pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1297 			rss[mm_counter(page)]--;
1298 			page_remove_rmap(page, false);
1299 			put_page(page);
1300 			continue;
1301 		}
1302 
1303 		/* If details->check_mapping, we leave swap entries. */
1304 		if (unlikely(details))
1305 			continue;
1306 
1307 		if (!non_swap_entry(entry))
1308 			rss[MM_SWAPENTS]--;
1309 		else if (is_migration_entry(entry)) {
1310 			struct page *page;
1311 
1312 			page = migration_entry_to_page(entry);
1313 			rss[mm_counter(page)]--;
1314 		}
1315 		if (unlikely(!free_swap_and_cache(entry)))
1316 			print_bad_pte(vma, addr, ptent, NULL);
1317 		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1318 	} while (pte++, addr += PAGE_SIZE, addr != end);
1319 
1320 	add_mm_rss_vec(mm, rss);
1321 	arch_leave_lazy_mmu_mode();
1322 
1323 	/* Do the actual TLB flush before dropping ptl */
1324 	if (force_flush)
1325 		tlb_flush_mmu_tlbonly(tlb);
1326 	pte_unmap_unlock(start_pte, ptl);
1327 
1328 	/*
1329 	 * If we forced a TLB flush (either due to running out of
1330 	 * batch buffers or because we needed to flush dirty TLB
1331 	 * entries before releasing the ptl), free the batched
1332 	 * memory too. Restart if we didn't do everything.
1333 	 */
1334 	if (force_flush) {
1335 		force_flush = 0;
1336 		tlb_flush_mmu(tlb);
1337 	}
1338 
1339 	if (addr != end) {
1340 		cond_resched();
1341 		goto again;
1342 	}
1343 
1344 	return addr;
1345 }
1346 
1347 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1348 				struct vm_area_struct *vma, pud_t *pud,
1349 				unsigned long addr, unsigned long end,
1350 				struct zap_details *details)
1351 {
1352 	pmd_t *pmd;
1353 	unsigned long next;
1354 
1355 	pmd = pmd_offset(pud, addr);
1356 	do {
1357 		next = pmd_addr_end(addr, end);
1358 		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1359 			if (next - addr != HPAGE_PMD_SIZE)
1360 				__split_huge_pmd(vma, pmd, addr, false, NULL);
1361 			else if (zap_huge_pmd(tlb, vma, pmd, addr))
1362 				goto next;
1363 			/* fall through */
1364 		}
1365 		/*
1366 		 * Here there can be other concurrent MADV_DONTNEED or
1367 		 * trans huge page faults running, and if the pmd is
1368 		 * none or trans huge it can change under us. This is
1369 		 * because MADV_DONTNEED holds the mmap_lock in read
1370 		 * mode.
1371 		 */
1372 		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1373 			goto next;
1374 		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1375 next:
1376 		cond_resched();
1377 	} while (pmd++, addr = next, addr != end);
1378 
1379 	return addr;
1380 }
1381 
1382 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1383 				struct vm_area_struct *vma, p4d_t *p4d,
1384 				unsigned long addr, unsigned long end,
1385 				struct zap_details *details)
1386 {
1387 	pud_t *pud;
1388 	unsigned long next;
1389 
1390 	pud = pud_offset(p4d, addr);
1391 	do {
1392 		next = pud_addr_end(addr, end);
1393 		if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1394 			if (next - addr != HPAGE_PUD_SIZE) {
1395 				mmap_assert_locked(tlb->mm);
1396 				split_huge_pud(vma, pud, addr);
1397 			} else if (zap_huge_pud(tlb, vma, pud, addr))
1398 				goto next;
1399 			/* fall through */
1400 		}
1401 		if (pud_none_or_clear_bad(pud))
1402 			continue;
1403 		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1404 next:
1405 		cond_resched();
1406 	} while (pud++, addr = next, addr != end);
1407 
1408 	return addr;
1409 }
1410 
1411 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1412 				struct vm_area_struct *vma, pgd_t *pgd,
1413 				unsigned long addr, unsigned long end,
1414 				struct zap_details *details)
1415 {
1416 	p4d_t *p4d;
1417 	unsigned long next;
1418 
1419 	p4d = p4d_offset(pgd, addr);
1420 	do {
1421 		next = p4d_addr_end(addr, end);
1422 		if (p4d_none_or_clear_bad(p4d))
1423 			continue;
1424 		next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1425 	} while (p4d++, addr = next, addr != end);
1426 
1427 	return addr;
1428 }
1429 
1430 void unmap_page_range(struct mmu_gather *tlb,
1431 			     struct vm_area_struct *vma,
1432 			     unsigned long addr, unsigned long end,
1433 			     struct zap_details *details)
1434 {
1435 	pgd_t *pgd;
1436 	unsigned long next;
1437 
1438 	BUG_ON(addr >= end);
1439 	tlb_start_vma(tlb, vma);
1440 	pgd = pgd_offset(vma->vm_mm, addr);
1441 	do {
1442 		next = pgd_addr_end(addr, end);
1443 		if (pgd_none_or_clear_bad(pgd))
1444 			continue;
1445 		next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1446 	} while (pgd++, addr = next, addr != end);
1447 	tlb_end_vma(tlb, vma);
1448 }
1449 
1450 
1451 static void unmap_single_vma(struct mmu_gather *tlb,
1452 		struct vm_area_struct *vma, unsigned long start_addr,
1453 		unsigned long end_addr,
1454 		struct zap_details *details)
1455 {
1456 	unsigned long start = max(vma->vm_start, start_addr);
1457 	unsigned long end;
1458 
1459 	if (start >= vma->vm_end)
1460 		return;
1461 	end = min(vma->vm_end, end_addr);
1462 	if (end <= vma->vm_start)
1463 		return;
1464 
1465 	if (vma->vm_file)
1466 		uprobe_munmap(vma, start, end);
1467 
1468 	if (unlikely(vma->vm_flags & VM_PFNMAP))
1469 		untrack_pfn(vma, 0, 0);
1470 
1471 	if (start != end) {
1472 		if (unlikely(is_vm_hugetlb_page(vma))) {
1473 			/*
1474 			 * It is undesirable to test vma->vm_file as it
1475 			 * should be non-null for valid hugetlb area.
1476 			 * However, vm_file will be NULL in the error
1477 			 * cleanup path of mmap_region. When
1478 			 * hugetlbfs ->mmap method fails,
1479 			 * mmap_region() nullifies vma->vm_file
1480 			 * before calling this function to clean up.
1481 			 * Since no pte has actually been setup, it is
1482 			 * safe to do nothing in this case.
1483 			 */
1484 			if (vma->vm_file) {
1485 				i_mmap_lock_write(vma->vm_file->f_mapping);
1486 				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1487 				i_mmap_unlock_write(vma->vm_file->f_mapping);
1488 			}
1489 		} else
1490 			unmap_page_range(tlb, vma, start, end, details);
1491 	}
1492 }
1493 
1494 /**
1495  * unmap_vmas - unmap a range of memory covered by a list of vma's
1496  * @tlb: address of the caller's struct mmu_gather
1497  * @vma: the starting vma
1498  * @start_addr: virtual address at which to start unmapping
1499  * @end_addr: virtual address at which to end unmapping
1500  *
1501  * Unmap all pages in the vma list.
1502  *
1503  * Only addresses between `start' and `end' will be unmapped.
1504  *
1505  * The VMA list must be sorted in ascending virtual address order.
1506  *
1507  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1508  * range after unmap_vmas() returns.  So the only responsibility here is to
1509  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1510  * drops the lock and schedules.
1511  */
1512 void unmap_vmas(struct mmu_gather *tlb,
1513 		struct vm_area_struct *vma, unsigned long start_addr,
1514 		unsigned long end_addr)
1515 {
1516 	struct mmu_notifier_range range;
1517 
1518 	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1519 				start_addr, end_addr);
1520 	mmu_notifier_invalidate_range_start(&range);
1521 	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1522 		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1523 	mmu_notifier_invalidate_range_end(&range);
1524 }
1525 
1526 /**
1527  * zap_page_range - remove user pages in a given range
1528  * @vma: vm_area_struct holding the applicable pages
1529  * @start: starting address of pages to zap
1530  * @size: number of bytes to zap
1531  *
1532  * Caller must protect the VMA list
1533  */
1534 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1535 		unsigned long size)
1536 {
1537 	struct mmu_notifier_range range;
1538 	struct mmu_gather tlb;
1539 
1540 	lru_add_drain();
1541 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1542 				start, start + size);
1543 	tlb_gather_mmu(&tlb, vma->vm_mm);
1544 	update_hiwater_rss(vma->vm_mm);
1545 	mmu_notifier_invalidate_range_start(&range);
1546 	for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1547 		unmap_single_vma(&tlb, vma, start, range.end, NULL);
1548 	mmu_notifier_invalidate_range_end(&range);
1549 	tlb_finish_mmu(&tlb);
1550 }
1551 
1552 /**
1553  * zap_page_range_single - remove user pages in a given range
1554  * @vma: vm_area_struct holding the applicable pages
1555  * @address: starting address of pages to zap
1556  * @size: number of bytes to zap
1557  * @details: details of shared cache invalidation
1558  *
1559  * The range must fit into one VMA.
1560  */
1561 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1562 		unsigned long size, struct zap_details *details)
1563 {
1564 	struct mmu_notifier_range range;
1565 	struct mmu_gather tlb;
1566 
1567 	lru_add_drain();
1568 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1569 				address, address + size);
1570 	tlb_gather_mmu(&tlb, vma->vm_mm);
1571 	update_hiwater_rss(vma->vm_mm);
1572 	mmu_notifier_invalidate_range_start(&range);
1573 	unmap_single_vma(&tlb, vma, address, range.end, details);
1574 	mmu_notifier_invalidate_range_end(&range);
1575 	tlb_finish_mmu(&tlb);
1576 }
1577 
1578 /**
1579  * zap_vma_ptes - remove ptes mapping the vma
1580  * @vma: vm_area_struct holding ptes to be zapped
1581  * @address: starting address of pages to zap
1582  * @size: number of bytes to zap
1583  *
1584  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1585  *
1586  * The entire address range must be fully contained within the vma.
1587  *
1588  */
1589 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1590 		unsigned long size)
1591 {
1592 	if (address < vma->vm_start || address + size > vma->vm_end ||
1593 	    		!(vma->vm_flags & VM_PFNMAP))
1594 		return;
1595 
1596 	zap_page_range_single(vma, address, size, NULL);
1597 }
1598 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1599 
1600 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1601 {
1602 	pgd_t *pgd;
1603 	p4d_t *p4d;
1604 	pud_t *pud;
1605 	pmd_t *pmd;
1606 
1607 	pgd = pgd_offset(mm, addr);
1608 	p4d = p4d_alloc(mm, pgd, addr);
1609 	if (!p4d)
1610 		return NULL;
1611 	pud = pud_alloc(mm, p4d, addr);
1612 	if (!pud)
1613 		return NULL;
1614 	pmd = pmd_alloc(mm, pud, addr);
1615 	if (!pmd)
1616 		return NULL;
1617 
1618 	VM_BUG_ON(pmd_trans_huge(*pmd));
1619 	return pmd;
1620 }
1621 
1622 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1623 			spinlock_t **ptl)
1624 {
1625 	pmd_t *pmd = walk_to_pmd(mm, addr);
1626 
1627 	if (!pmd)
1628 		return NULL;
1629 	return pte_alloc_map_lock(mm, pmd, addr, ptl);
1630 }
1631 
1632 static int validate_page_before_insert(struct page *page)
1633 {
1634 	if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1635 		return -EINVAL;
1636 	flush_dcache_page(page);
1637 	return 0;
1638 }
1639 
1640 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1641 			unsigned long addr, struct page *page, pgprot_t prot)
1642 {
1643 	if (!pte_none(*pte))
1644 		return -EBUSY;
1645 	/* Ok, finally just insert the thing.. */
1646 	get_page(page);
1647 	inc_mm_counter_fast(mm, mm_counter_file(page));
1648 	page_add_file_rmap(page, false);
1649 	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1650 	return 0;
1651 }
1652 
1653 /*
1654  * This is the old fallback for page remapping.
1655  *
1656  * For historical reasons, it only allows reserved pages. Only
1657  * old drivers should use this, and they needed to mark their
1658  * pages reserved for the old functions anyway.
1659  */
1660 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1661 			struct page *page, pgprot_t prot)
1662 {
1663 	struct mm_struct *mm = vma->vm_mm;
1664 	int retval;
1665 	pte_t *pte;
1666 	spinlock_t *ptl;
1667 
1668 	retval = validate_page_before_insert(page);
1669 	if (retval)
1670 		goto out;
1671 	retval = -ENOMEM;
1672 	pte = get_locked_pte(mm, addr, &ptl);
1673 	if (!pte)
1674 		goto out;
1675 	retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1676 	pte_unmap_unlock(pte, ptl);
1677 out:
1678 	return retval;
1679 }
1680 
1681 #ifdef pte_index
1682 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1683 			unsigned long addr, struct page *page, pgprot_t prot)
1684 {
1685 	int err;
1686 
1687 	if (!page_count(page))
1688 		return -EINVAL;
1689 	err = validate_page_before_insert(page);
1690 	if (err)
1691 		return err;
1692 	return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1693 }
1694 
1695 /* insert_pages() amortizes the cost of spinlock operations
1696  * when inserting pages in a loop. Arch *must* define pte_index.
1697  */
1698 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1699 			struct page **pages, unsigned long *num, pgprot_t prot)
1700 {
1701 	pmd_t *pmd = NULL;
1702 	pte_t *start_pte, *pte;
1703 	spinlock_t *pte_lock;
1704 	struct mm_struct *const mm = vma->vm_mm;
1705 	unsigned long curr_page_idx = 0;
1706 	unsigned long remaining_pages_total = *num;
1707 	unsigned long pages_to_write_in_pmd;
1708 	int ret;
1709 more:
1710 	ret = -EFAULT;
1711 	pmd = walk_to_pmd(mm, addr);
1712 	if (!pmd)
1713 		goto out;
1714 
1715 	pages_to_write_in_pmd = min_t(unsigned long,
1716 		remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1717 
1718 	/* Allocate the PTE if necessary; takes PMD lock once only. */
1719 	ret = -ENOMEM;
1720 	if (pte_alloc(mm, pmd))
1721 		goto out;
1722 
1723 	while (pages_to_write_in_pmd) {
1724 		int pte_idx = 0;
1725 		const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1726 
1727 		start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1728 		for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1729 			int err = insert_page_in_batch_locked(mm, pte,
1730 				addr, pages[curr_page_idx], prot);
1731 			if (unlikely(err)) {
1732 				pte_unmap_unlock(start_pte, pte_lock);
1733 				ret = err;
1734 				remaining_pages_total -= pte_idx;
1735 				goto out;
1736 			}
1737 			addr += PAGE_SIZE;
1738 			++curr_page_idx;
1739 		}
1740 		pte_unmap_unlock(start_pte, pte_lock);
1741 		pages_to_write_in_pmd -= batch_size;
1742 		remaining_pages_total -= batch_size;
1743 	}
1744 	if (remaining_pages_total)
1745 		goto more;
1746 	ret = 0;
1747 out:
1748 	*num = remaining_pages_total;
1749 	return ret;
1750 }
1751 #endif  /* ifdef pte_index */
1752 
1753 /**
1754  * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1755  * @vma: user vma to map to
1756  * @addr: target start user address of these pages
1757  * @pages: source kernel pages
1758  * @num: in: number of pages to map. out: number of pages that were *not*
1759  * mapped. (0 means all pages were successfully mapped).
1760  *
1761  * Preferred over vm_insert_page() when inserting multiple pages.
1762  *
1763  * In case of error, we may have mapped a subset of the provided
1764  * pages. It is the caller's responsibility to account for this case.
1765  *
1766  * The same restrictions apply as in vm_insert_page().
1767  */
1768 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1769 			struct page **pages, unsigned long *num)
1770 {
1771 #ifdef pte_index
1772 	const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1773 
1774 	if (addr < vma->vm_start || end_addr >= vma->vm_end)
1775 		return -EFAULT;
1776 	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1777 		BUG_ON(mmap_read_trylock(vma->vm_mm));
1778 		BUG_ON(vma->vm_flags & VM_PFNMAP);
1779 		vma->vm_flags |= VM_MIXEDMAP;
1780 	}
1781 	/* Defer page refcount checking till we're about to map that page. */
1782 	return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1783 #else
1784 	unsigned long idx = 0, pgcount = *num;
1785 	int err = -EINVAL;
1786 
1787 	for (; idx < pgcount; ++idx) {
1788 		err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1789 		if (err)
1790 			break;
1791 	}
1792 	*num = pgcount - idx;
1793 	return err;
1794 #endif  /* ifdef pte_index */
1795 }
1796 EXPORT_SYMBOL(vm_insert_pages);
1797 
1798 /**
1799  * vm_insert_page - insert single page into user vma
1800  * @vma: user vma to map to
1801  * @addr: target user address of this page
1802  * @page: source kernel page
1803  *
1804  * This allows drivers to insert individual pages they've allocated
1805  * into a user vma.
1806  *
1807  * The page has to be a nice clean _individual_ kernel allocation.
1808  * If you allocate a compound page, you need to have marked it as
1809  * such (__GFP_COMP), or manually just split the page up yourself
1810  * (see split_page()).
1811  *
1812  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1813  * took an arbitrary page protection parameter. This doesn't allow
1814  * that. Your vma protection will have to be set up correctly, which
1815  * means that if you want a shared writable mapping, you'd better
1816  * ask for a shared writable mapping!
1817  *
1818  * The page does not need to be reserved.
1819  *
1820  * Usually this function is called from f_op->mmap() handler
1821  * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1822  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1823  * function from other places, for example from page-fault handler.
1824  *
1825  * Return: %0 on success, negative error code otherwise.
1826  */
1827 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1828 			struct page *page)
1829 {
1830 	if (addr < vma->vm_start || addr >= vma->vm_end)
1831 		return -EFAULT;
1832 	if (!page_count(page))
1833 		return -EINVAL;
1834 	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1835 		BUG_ON(mmap_read_trylock(vma->vm_mm));
1836 		BUG_ON(vma->vm_flags & VM_PFNMAP);
1837 		vma->vm_flags |= VM_MIXEDMAP;
1838 	}
1839 	return insert_page(vma, addr, page, vma->vm_page_prot);
1840 }
1841 EXPORT_SYMBOL(vm_insert_page);
1842 
1843 /*
1844  * __vm_map_pages - maps range of kernel pages into user vma
1845  * @vma: user vma to map to
1846  * @pages: pointer to array of source kernel pages
1847  * @num: number of pages in page array
1848  * @offset: user's requested vm_pgoff
1849  *
1850  * This allows drivers to map range of kernel pages into a user vma.
1851  *
1852  * Return: 0 on success and error code otherwise.
1853  */
1854 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1855 				unsigned long num, unsigned long offset)
1856 {
1857 	unsigned long count = vma_pages(vma);
1858 	unsigned long uaddr = vma->vm_start;
1859 	int ret, i;
1860 
1861 	/* Fail if the user requested offset is beyond the end of the object */
1862 	if (offset >= num)
1863 		return -ENXIO;
1864 
1865 	/* Fail if the user requested size exceeds available object size */
1866 	if (count > num - offset)
1867 		return -ENXIO;
1868 
1869 	for (i = 0; i < count; i++) {
1870 		ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1871 		if (ret < 0)
1872 			return ret;
1873 		uaddr += PAGE_SIZE;
1874 	}
1875 
1876 	return 0;
1877 }
1878 
1879 /**
1880  * vm_map_pages - maps range of kernel pages starts with non zero offset
1881  * @vma: user vma to map to
1882  * @pages: pointer to array of source kernel pages
1883  * @num: number of pages in page array
1884  *
1885  * Maps an object consisting of @num pages, catering for the user's
1886  * requested vm_pgoff
1887  *
1888  * If we fail to insert any page into the vma, the function will return
1889  * immediately leaving any previously inserted pages present.  Callers
1890  * from the mmap handler may immediately return the error as their caller
1891  * will destroy the vma, removing any successfully inserted pages. Other
1892  * callers should make their own arrangements for calling unmap_region().
1893  *
1894  * Context: Process context. Called by mmap handlers.
1895  * Return: 0 on success and error code otherwise.
1896  */
1897 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1898 				unsigned long num)
1899 {
1900 	return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1901 }
1902 EXPORT_SYMBOL(vm_map_pages);
1903 
1904 /**
1905  * vm_map_pages_zero - map range of kernel pages starts with zero offset
1906  * @vma: user vma to map to
1907  * @pages: pointer to array of source kernel pages
1908  * @num: number of pages in page array
1909  *
1910  * Similar to vm_map_pages(), except that it explicitly sets the offset
1911  * to 0. This function is intended for the drivers that did not consider
1912  * vm_pgoff.
1913  *
1914  * Context: Process context. Called by mmap handlers.
1915  * Return: 0 on success and error code otherwise.
1916  */
1917 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1918 				unsigned long num)
1919 {
1920 	return __vm_map_pages(vma, pages, num, 0);
1921 }
1922 EXPORT_SYMBOL(vm_map_pages_zero);
1923 
1924 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1925 			pfn_t pfn, pgprot_t prot, bool mkwrite)
1926 {
1927 	struct mm_struct *mm = vma->vm_mm;
1928 	pte_t *pte, entry;
1929 	spinlock_t *ptl;
1930 
1931 	pte = get_locked_pte(mm, addr, &ptl);
1932 	if (!pte)
1933 		return VM_FAULT_OOM;
1934 	if (!pte_none(*pte)) {
1935 		if (mkwrite) {
1936 			/*
1937 			 * For read faults on private mappings the PFN passed
1938 			 * in may not match the PFN we have mapped if the
1939 			 * mapped PFN is a writeable COW page.  In the mkwrite
1940 			 * case we are creating a writable PTE for a shared
1941 			 * mapping and we expect the PFNs to match. If they
1942 			 * don't match, we are likely racing with block
1943 			 * allocation and mapping invalidation so just skip the
1944 			 * update.
1945 			 */
1946 			if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1947 				WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1948 				goto out_unlock;
1949 			}
1950 			entry = pte_mkyoung(*pte);
1951 			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1952 			if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1953 				update_mmu_cache(vma, addr, pte);
1954 		}
1955 		goto out_unlock;
1956 	}
1957 
1958 	/* Ok, finally just insert the thing.. */
1959 	if (pfn_t_devmap(pfn))
1960 		entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1961 	else
1962 		entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1963 
1964 	if (mkwrite) {
1965 		entry = pte_mkyoung(entry);
1966 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1967 	}
1968 
1969 	set_pte_at(mm, addr, pte, entry);
1970 	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1971 
1972 out_unlock:
1973 	pte_unmap_unlock(pte, ptl);
1974 	return VM_FAULT_NOPAGE;
1975 }
1976 
1977 /**
1978  * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1979  * @vma: user vma to map to
1980  * @addr: target user address of this page
1981  * @pfn: source kernel pfn
1982  * @pgprot: pgprot flags for the inserted page
1983  *
1984  * This is exactly like vmf_insert_pfn(), except that it allows drivers
1985  * to override pgprot on a per-page basis.
1986  *
1987  * This only makes sense for IO mappings, and it makes no sense for
1988  * COW mappings.  In general, using multiple vmas is preferable;
1989  * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1990  * impractical.
1991  *
1992  * See vmf_insert_mixed_prot() for a discussion of the implication of using
1993  * a value of @pgprot different from that of @vma->vm_page_prot.
1994  *
1995  * Context: Process context.  May allocate using %GFP_KERNEL.
1996  * Return: vm_fault_t value.
1997  */
1998 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1999 			unsigned long pfn, pgprot_t pgprot)
2000 {
2001 	/*
2002 	 * Technically, architectures with pte_special can avoid all these
2003 	 * restrictions (same for remap_pfn_range).  However we would like
2004 	 * consistency in testing and feature parity among all, so we should
2005 	 * try to keep these invariants in place for everybody.
2006 	 */
2007 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2008 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2009 						(VM_PFNMAP|VM_MIXEDMAP));
2010 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2011 	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2012 
2013 	if (addr < vma->vm_start || addr >= vma->vm_end)
2014 		return VM_FAULT_SIGBUS;
2015 
2016 	if (!pfn_modify_allowed(pfn, pgprot))
2017 		return VM_FAULT_SIGBUS;
2018 
2019 	track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2020 
2021 	return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2022 			false);
2023 }
2024 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2025 
2026 /**
2027  * vmf_insert_pfn - insert single pfn into user vma
2028  * @vma: user vma to map to
2029  * @addr: target user address of this page
2030  * @pfn: source kernel pfn
2031  *
2032  * Similar to vm_insert_page, this allows drivers to insert individual pages
2033  * they've allocated into a user vma. Same comments apply.
2034  *
2035  * This function should only be called from a vm_ops->fault handler, and
2036  * in that case the handler should return the result of this function.
2037  *
2038  * vma cannot be a COW mapping.
2039  *
2040  * As this is called only for pages that do not currently exist, we
2041  * do not need to flush old virtual caches or the TLB.
2042  *
2043  * Context: Process context.  May allocate using %GFP_KERNEL.
2044  * Return: vm_fault_t value.
2045  */
2046 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2047 			unsigned long pfn)
2048 {
2049 	return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2050 }
2051 EXPORT_SYMBOL(vmf_insert_pfn);
2052 
2053 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2054 {
2055 	/* these checks mirror the abort conditions in vm_normal_page */
2056 	if (vma->vm_flags & VM_MIXEDMAP)
2057 		return true;
2058 	if (pfn_t_devmap(pfn))
2059 		return true;
2060 	if (pfn_t_special(pfn))
2061 		return true;
2062 	if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2063 		return true;
2064 	return false;
2065 }
2066 
2067 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2068 		unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2069 		bool mkwrite)
2070 {
2071 	int err;
2072 
2073 	BUG_ON(!vm_mixed_ok(vma, pfn));
2074 
2075 	if (addr < vma->vm_start || addr >= vma->vm_end)
2076 		return VM_FAULT_SIGBUS;
2077 
2078 	track_pfn_insert(vma, &pgprot, pfn);
2079 
2080 	if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2081 		return VM_FAULT_SIGBUS;
2082 
2083 	/*
2084 	 * If we don't have pte special, then we have to use the pfn_valid()
2085 	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2086 	 * refcount the page if pfn_valid is true (hence insert_page rather
2087 	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2088 	 * without pte special, it would there be refcounted as a normal page.
2089 	 */
2090 	if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2091 	    !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2092 		struct page *page;
2093 
2094 		/*
2095 		 * At this point we are committed to insert_page()
2096 		 * regardless of whether the caller specified flags that
2097 		 * result in pfn_t_has_page() == false.
2098 		 */
2099 		page = pfn_to_page(pfn_t_to_pfn(pfn));
2100 		err = insert_page(vma, addr, page, pgprot);
2101 	} else {
2102 		return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2103 	}
2104 
2105 	if (err == -ENOMEM)
2106 		return VM_FAULT_OOM;
2107 	if (err < 0 && err != -EBUSY)
2108 		return VM_FAULT_SIGBUS;
2109 
2110 	return VM_FAULT_NOPAGE;
2111 }
2112 
2113 /**
2114  * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2115  * @vma: user vma to map to
2116  * @addr: target user address of this page
2117  * @pfn: source kernel pfn
2118  * @pgprot: pgprot flags for the inserted page
2119  *
2120  * This is exactly like vmf_insert_mixed(), except that it allows drivers
2121  * to override pgprot on a per-page basis.
2122  *
2123  * Typically this function should be used by drivers to set caching- and
2124  * encryption bits different than those of @vma->vm_page_prot, because
2125  * the caching- or encryption mode may not be known at mmap() time.
2126  * This is ok as long as @vma->vm_page_prot is not used by the core vm
2127  * to set caching and encryption bits for those vmas (except for COW pages).
2128  * This is ensured by core vm only modifying these page table entries using
2129  * functions that don't touch caching- or encryption bits, using pte_modify()
2130  * if needed. (See for example mprotect()).
2131  * Also when new page-table entries are created, this is only done using the
2132  * fault() callback, and never using the value of vma->vm_page_prot,
2133  * except for page-table entries that point to anonymous pages as the result
2134  * of COW.
2135  *
2136  * Context: Process context.  May allocate using %GFP_KERNEL.
2137  * Return: vm_fault_t value.
2138  */
2139 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2140 				 pfn_t pfn, pgprot_t pgprot)
2141 {
2142 	return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2143 }
2144 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2145 
2146 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2147 		pfn_t pfn)
2148 {
2149 	return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2150 }
2151 EXPORT_SYMBOL(vmf_insert_mixed);
2152 
2153 /*
2154  *  If the insertion of PTE failed because someone else already added a
2155  *  different entry in the mean time, we treat that as success as we assume
2156  *  the same entry was actually inserted.
2157  */
2158 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2159 		unsigned long addr, pfn_t pfn)
2160 {
2161 	return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2162 }
2163 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2164 
2165 /*
2166  * maps a range of physical memory into the requested pages. the old
2167  * mappings are removed. any references to nonexistent pages results
2168  * in null mappings (currently treated as "copy-on-access")
2169  */
2170 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2171 			unsigned long addr, unsigned long end,
2172 			unsigned long pfn, pgprot_t prot)
2173 {
2174 	pte_t *pte, *mapped_pte;
2175 	spinlock_t *ptl;
2176 	int err = 0;
2177 
2178 	mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2179 	if (!pte)
2180 		return -ENOMEM;
2181 	arch_enter_lazy_mmu_mode();
2182 	do {
2183 		BUG_ON(!pte_none(*pte));
2184 		if (!pfn_modify_allowed(pfn, prot)) {
2185 			err = -EACCES;
2186 			break;
2187 		}
2188 		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2189 		pfn++;
2190 	} while (pte++, addr += PAGE_SIZE, addr != end);
2191 	arch_leave_lazy_mmu_mode();
2192 	pte_unmap_unlock(mapped_pte, ptl);
2193 	return err;
2194 }
2195 
2196 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2197 			unsigned long addr, unsigned long end,
2198 			unsigned long pfn, pgprot_t prot)
2199 {
2200 	pmd_t *pmd;
2201 	unsigned long next;
2202 	int err;
2203 
2204 	pfn -= addr >> PAGE_SHIFT;
2205 	pmd = pmd_alloc(mm, pud, addr);
2206 	if (!pmd)
2207 		return -ENOMEM;
2208 	VM_BUG_ON(pmd_trans_huge(*pmd));
2209 	do {
2210 		next = pmd_addr_end(addr, end);
2211 		err = remap_pte_range(mm, pmd, addr, next,
2212 				pfn + (addr >> PAGE_SHIFT), prot);
2213 		if (err)
2214 			return err;
2215 	} while (pmd++, addr = next, addr != end);
2216 	return 0;
2217 }
2218 
2219 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2220 			unsigned long addr, unsigned long end,
2221 			unsigned long pfn, pgprot_t prot)
2222 {
2223 	pud_t *pud;
2224 	unsigned long next;
2225 	int err;
2226 
2227 	pfn -= addr >> PAGE_SHIFT;
2228 	pud = pud_alloc(mm, p4d, addr);
2229 	if (!pud)
2230 		return -ENOMEM;
2231 	do {
2232 		next = pud_addr_end(addr, end);
2233 		err = remap_pmd_range(mm, pud, addr, next,
2234 				pfn + (addr >> PAGE_SHIFT), prot);
2235 		if (err)
2236 			return err;
2237 	} while (pud++, addr = next, addr != end);
2238 	return 0;
2239 }
2240 
2241 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2242 			unsigned long addr, unsigned long end,
2243 			unsigned long pfn, pgprot_t prot)
2244 {
2245 	p4d_t *p4d;
2246 	unsigned long next;
2247 	int err;
2248 
2249 	pfn -= addr >> PAGE_SHIFT;
2250 	p4d = p4d_alloc(mm, pgd, addr);
2251 	if (!p4d)
2252 		return -ENOMEM;
2253 	do {
2254 		next = p4d_addr_end(addr, end);
2255 		err = remap_pud_range(mm, p4d, addr, next,
2256 				pfn + (addr >> PAGE_SHIFT), prot);
2257 		if (err)
2258 			return err;
2259 	} while (p4d++, addr = next, addr != end);
2260 	return 0;
2261 }
2262 
2263 /*
2264  * Variant of remap_pfn_range that does not call track_pfn_remap.  The caller
2265  * must have pre-validated the caching bits of the pgprot_t.
2266  */
2267 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2268 		unsigned long pfn, unsigned long size, pgprot_t prot)
2269 {
2270 	pgd_t *pgd;
2271 	unsigned long next;
2272 	unsigned long end = addr + PAGE_ALIGN(size);
2273 	struct mm_struct *mm = vma->vm_mm;
2274 	int err;
2275 
2276 	if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2277 		return -EINVAL;
2278 
2279 	/*
2280 	 * Physically remapped pages are special. Tell the
2281 	 * rest of the world about it:
2282 	 *   VM_IO tells people not to look at these pages
2283 	 *	(accesses can have side effects).
2284 	 *   VM_PFNMAP tells the core MM that the base pages are just
2285 	 *	raw PFN mappings, and do not have a "struct page" associated
2286 	 *	with them.
2287 	 *   VM_DONTEXPAND
2288 	 *      Disable vma merging and expanding with mremap().
2289 	 *   VM_DONTDUMP
2290 	 *      Omit vma from core dump, even when VM_IO turned off.
2291 	 *
2292 	 * There's a horrible special case to handle copy-on-write
2293 	 * behaviour that some programs depend on. We mark the "original"
2294 	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2295 	 * See vm_normal_page() for details.
2296 	 */
2297 	if (is_cow_mapping(vma->vm_flags)) {
2298 		if (addr != vma->vm_start || end != vma->vm_end)
2299 			return -EINVAL;
2300 		vma->vm_pgoff = pfn;
2301 	}
2302 
2303 	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2304 
2305 	BUG_ON(addr >= end);
2306 	pfn -= addr >> PAGE_SHIFT;
2307 	pgd = pgd_offset(mm, addr);
2308 	flush_cache_range(vma, addr, end);
2309 	do {
2310 		next = pgd_addr_end(addr, end);
2311 		err = remap_p4d_range(mm, pgd, addr, next,
2312 				pfn + (addr >> PAGE_SHIFT), prot);
2313 		if (err)
2314 			return err;
2315 	} while (pgd++, addr = next, addr != end);
2316 
2317 	return 0;
2318 }
2319 
2320 /**
2321  * remap_pfn_range - remap kernel memory to userspace
2322  * @vma: user vma to map to
2323  * @addr: target page aligned user address to start at
2324  * @pfn: page frame number of kernel physical memory address
2325  * @size: size of mapping area
2326  * @prot: page protection flags for this mapping
2327  *
2328  * Note: this is only safe if the mm semaphore is held when called.
2329  *
2330  * Return: %0 on success, negative error code otherwise.
2331  */
2332 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2333 		    unsigned long pfn, unsigned long size, pgprot_t prot)
2334 {
2335 	int err;
2336 
2337 	err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2338 	if (err)
2339 		return -EINVAL;
2340 
2341 	err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2342 	if (err)
2343 		untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2344 	return err;
2345 }
2346 EXPORT_SYMBOL(remap_pfn_range);
2347 
2348 /**
2349  * vm_iomap_memory - remap memory to userspace
2350  * @vma: user vma to map to
2351  * @start: start of the physical memory to be mapped
2352  * @len: size of area
2353  *
2354  * This is a simplified io_remap_pfn_range() for common driver use. The
2355  * driver just needs to give us the physical memory range to be mapped,
2356  * we'll figure out the rest from the vma information.
2357  *
2358  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2359  * whatever write-combining details or similar.
2360  *
2361  * Return: %0 on success, negative error code otherwise.
2362  */
2363 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2364 {
2365 	unsigned long vm_len, pfn, pages;
2366 
2367 	/* Check that the physical memory area passed in looks valid */
2368 	if (start + len < start)
2369 		return -EINVAL;
2370 	/*
2371 	 * You *really* shouldn't map things that aren't page-aligned,
2372 	 * but we've historically allowed it because IO memory might
2373 	 * just have smaller alignment.
2374 	 */
2375 	len += start & ~PAGE_MASK;
2376 	pfn = start >> PAGE_SHIFT;
2377 	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2378 	if (pfn + pages < pfn)
2379 		return -EINVAL;
2380 
2381 	/* We start the mapping 'vm_pgoff' pages into the area */
2382 	if (vma->vm_pgoff > pages)
2383 		return -EINVAL;
2384 	pfn += vma->vm_pgoff;
2385 	pages -= vma->vm_pgoff;
2386 
2387 	/* Can we fit all of the mapping? */
2388 	vm_len = vma->vm_end - vma->vm_start;
2389 	if (vm_len >> PAGE_SHIFT > pages)
2390 		return -EINVAL;
2391 
2392 	/* Ok, let it rip */
2393 	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2394 }
2395 EXPORT_SYMBOL(vm_iomap_memory);
2396 
2397 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2398 				     unsigned long addr, unsigned long end,
2399 				     pte_fn_t fn, void *data, bool create,
2400 				     pgtbl_mod_mask *mask)
2401 {
2402 	pte_t *pte, *mapped_pte;
2403 	int err = 0;
2404 	spinlock_t *ptl;
2405 
2406 	if (create) {
2407 		mapped_pte = pte = (mm == &init_mm) ?
2408 			pte_alloc_kernel_track(pmd, addr, mask) :
2409 			pte_alloc_map_lock(mm, pmd, addr, &ptl);
2410 		if (!pte)
2411 			return -ENOMEM;
2412 	} else {
2413 		mapped_pte = pte = (mm == &init_mm) ?
2414 			pte_offset_kernel(pmd, addr) :
2415 			pte_offset_map_lock(mm, pmd, addr, &ptl);
2416 	}
2417 
2418 	BUG_ON(pmd_huge(*pmd));
2419 
2420 	arch_enter_lazy_mmu_mode();
2421 
2422 	if (fn) {
2423 		do {
2424 			if (create || !pte_none(*pte)) {
2425 				err = fn(pte++, addr, data);
2426 				if (err)
2427 					break;
2428 			}
2429 		} while (addr += PAGE_SIZE, addr != end);
2430 	}
2431 	*mask |= PGTBL_PTE_MODIFIED;
2432 
2433 	arch_leave_lazy_mmu_mode();
2434 
2435 	if (mm != &init_mm)
2436 		pte_unmap_unlock(mapped_pte, ptl);
2437 	return err;
2438 }
2439 
2440 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2441 				     unsigned long addr, unsigned long end,
2442 				     pte_fn_t fn, void *data, bool create,
2443 				     pgtbl_mod_mask *mask)
2444 {
2445 	pmd_t *pmd;
2446 	unsigned long next;
2447 	int err = 0;
2448 
2449 	BUG_ON(pud_huge(*pud));
2450 
2451 	if (create) {
2452 		pmd = pmd_alloc_track(mm, pud, addr, mask);
2453 		if (!pmd)
2454 			return -ENOMEM;
2455 	} else {
2456 		pmd = pmd_offset(pud, addr);
2457 	}
2458 	do {
2459 		next = pmd_addr_end(addr, end);
2460 		if (pmd_none(*pmd) && !create)
2461 			continue;
2462 		if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2463 			return -EINVAL;
2464 		if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2465 			if (!create)
2466 				continue;
2467 			pmd_clear_bad(pmd);
2468 		}
2469 		err = apply_to_pte_range(mm, pmd, addr, next,
2470 					 fn, data, create, mask);
2471 		if (err)
2472 			break;
2473 	} while (pmd++, addr = next, addr != end);
2474 
2475 	return err;
2476 }
2477 
2478 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2479 				     unsigned long addr, unsigned long end,
2480 				     pte_fn_t fn, void *data, bool create,
2481 				     pgtbl_mod_mask *mask)
2482 {
2483 	pud_t *pud;
2484 	unsigned long next;
2485 	int err = 0;
2486 
2487 	if (create) {
2488 		pud = pud_alloc_track(mm, p4d, addr, mask);
2489 		if (!pud)
2490 			return -ENOMEM;
2491 	} else {
2492 		pud = pud_offset(p4d, addr);
2493 	}
2494 	do {
2495 		next = pud_addr_end(addr, end);
2496 		if (pud_none(*pud) && !create)
2497 			continue;
2498 		if (WARN_ON_ONCE(pud_leaf(*pud)))
2499 			return -EINVAL;
2500 		if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2501 			if (!create)
2502 				continue;
2503 			pud_clear_bad(pud);
2504 		}
2505 		err = apply_to_pmd_range(mm, pud, addr, next,
2506 					 fn, data, create, mask);
2507 		if (err)
2508 			break;
2509 	} while (pud++, addr = next, addr != end);
2510 
2511 	return err;
2512 }
2513 
2514 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2515 				     unsigned long addr, unsigned long end,
2516 				     pte_fn_t fn, void *data, bool create,
2517 				     pgtbl_mod_mask *mask)
2518 {
2519 	p4d_t *p4d;
2520 	unsigned long next;
2521 	int err = 0;
2522 
2523 	if (create) {
2524 		p4d = p4d_alloc_track(mm, pgd, addr, mask);
2525 		if (!p4d)
2526 			return -ENOMEM;
2527 	} else {
2528 		p4d = p4d_offset(pgd, addr);
2529 	}
2530 	do {
2531 		next = p4d_addr_end(addr, end);
2532 		if (p4d_none(*p4d) && !create)
2533 			continue;
2534 		if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2535 			return -EINVAL;
2536 		if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2537 			if (!create)
2538 				continue;
2539 			p4d_clear_bad(p4d);
2540 		}
2541 		err = apply_to_pud_range(mm, p4d, addr, next,
2542 					 fn, data, create, mask);
2543 		if (err)
2544 			break;
2545 	} while (p4d++, addr = next, addr != end);
2546 
2547 	return err;
2548 }
2549 
2550 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2551 				 unsigned long size, pte_fn_t fn,
2552 				 void *data, bool create)
2553 {
2554 	pgd_t *pgd;
2555 	unsigned long start = addr, next;
2556 	unsigned long end = addr + size;
2557 	pgtbl_mod_mask mask = 0;
2558 	int err = 0;
2559 
2560 	if (WARN_ON(addr >= end))
2561 		return -EINVAL;
2562 
2563 	pgd = pgd_offset(mm, addr);
2564 	do {
2565 		next = pgd_addr_end(addr, end);
2566 		if (pgd_none(*pgd) && !create)
2567 			continue;
2568 		if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2569 			return -EINVAL;
2570 		if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2571 			if (!create)
2572 				continue;
2573 			pgd_clear_bad(pgd);
2574 		}
2575 		err = apply_to_p4d_range(mm, pgd, addr, next,
2576 					 fn, data, create, &mask);
2577 		if (err)
2578 			break;
2579 	} while (pgd++, addr = next, addr != end);
2580 
2581 	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2582 		arch_sync_kernel_mappings(start, start + size);
2583 
2584 	return err;
2585 }
2586 
2587 /*
2588  * Scan a region of virtual memory, filling in page tables as necessary
2589  * and calling a provided function on each leaf page table.
2590  */
2591 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2592 			unsigned long size, pte_fn_t fn, void *data)
2593 {
2594 	return __apply_to_page_range(mm, addr, size, fn, data, true);
2595 }
2596 EXPORT_SYMBOL_GPL(apply_to_page_range);
2597 
2598 /*
2599  * Scan a region of virtual memory, calling a provided function on
2600  * each leaf page table where it exists.
2601  *
2602  * Unlike apply_to_page_range, this does _not_ fill in page tables
2603  * where they are absent.
2604  */
2605 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2606 				 unsigned long size, pte_fn_t fn, void *data)
2607 {
2608 	return __apply_to_page_range(mm, addr, size, fn, data, false);
2609 }
2610 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2611 
2612 /*
2613  * handle_pte_fault chooses page fault handler according to an entry which was
2614  * read non-atomically.  Before making any commitment, on those architectures
2615  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2616  * parts, do_swap_page must check under lock before unmapping the pte and
2617  * proceeding (but do_wp_page is only called after already making such a check;
2618  * and do_anonymous_page can safely check later on).
2619  */
2620 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2621 				pte_t *page_table, pte_t orig_pte)
2622 {
2623 	int same = 1;
2624 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2625 	if (sizeof(pte_t) > sizeof(unsigned long)) {
2626 		spinlock_t *ptl = pte_lockptr(mm, pmd);
2627 		spin_lock(ptl);
2628 		same = pte_same(*page_table, orig_pte);
2629 		spin_unlock(ptl);
2630 	}
2631 #endif
2632 	pte_unmap(page_table);
2633 	return same;
2634 }
2635 
2636 static inline bool cow_user_page(struct page *dst, struct page *src,
2637 				 struct vm_fault *vmf)
2638 {
2639 	bool ret;
2640 	void *kaddr;
2641 	void __user *uaddr;
2642 	bool locked = false;
2643 	struct vm_area_struct *vma = vmf->vma;
2644 	struct mm_struct *mm = vma->vm_mm;
2645 	unsigned long addr = vmf->address;
2646 
2647 	if (likely(src)) {
2648 		copy_user_highpage(dst, src, addr, vma);
2649 		return true;
2650 	}
2651 
2652 	/*
2653 	 * If the source page was a PFN mapping, we don't have
2654 	 * a "struct page" for it. We do a best-effort copy by
2655 	 * just copying from the original user address. If that
2656 	 * fails, we just zero-fill it. Live with it.
2657 	 */
2658 	kaddr = kmap_atomic(dst);
2659 	uaddr = (void __user *)(addr & PAGE_MASK);
2660 
2661 	/*
2662 	 * On architectures with software "accessed" bits, we would
2663 	 * take a double page fault, so mark it accessed here.
2664 	 */
2665 	if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2666 		pte_t entry;
2667 
2668 		vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2669 		locked = true;
2670 		if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2671 			/*
2672 			 * Other thread has already handled the fault
2673 			 * and update local tlb only
2674 			 */
2675 			update_mmu_tlb(vma, addr, vmf->pte);
2676 			ret = false;
2677 			goto pte_unlock;
2678 		}
2679 
2680 		entry = pte_mkyoung(vmf->orig_pte);
2681 		if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2682 			update_mmu_cache(vma, addr, vmf->pte);
2683 	}
2684 
2685 	/*
2686 	 * This really shouldn't fail, because the page is there
2687 	 * in the page tables. But it might just be unreadable,
2688 	 * in which case we just give up and fill the result with
2689 	 * zeroes.
2690 	 */
2691 	if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2692 		if (locked)
2693 			goto warn;
2694 
2695 		/* Re-validate under PTL if the page is still mapped */
2696 		vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2697 		locked = true;
2698 		if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2699 			/* The PTE changed under us, update local tlb */
2700 			update_mmu_tlb(vma, addr, vmf->pte);
2701 			ret = false;
2702 			goto pte_unlock;
2703 		}
2704 
2705 		/*
2706 		 * The same page can be mapped back since last copy attempt.
2707 		 * Try to copy again under PTL.
2708 		 */
2709 		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2710 			/*
2711 			 * Give a warn in case there can be some obscure
2712 			 * use-case
2713 			 */
2714 warn:
2715 			WARN_ON_ONCE(1);
2716 			clear_page(kaddr);
2717 		}
2718 	}
2719 
2720 	ret = true;
2721 
2722 pte_unlock:
2723 	if (locked)
2724 		pte_unmap_unlock(vmf->pte, vmf->ptl);
2725 	kunmap_atomic(kaddr);
2726 	flush_dcache_page(dst);
2727 
2728 	return ret;
2729 }
2730 
2731 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2732 {
2733 	struct file *vm_file = vma->vm_file;
2734 
2735 	if (vm_file)
2736 		return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2737 
2738 	/*
2739 	 * Special mappings (e.g. VDSO) do not have any file so fake
2740 	 * a default GFP_KERNEL for them.
2741 	 */
2742 	return GFP_KERNEL;
2743 }
2744 
2745 /*
2746  * Notify the address space that the page is about to become writable so that
2747  * it can prohibit this or wait for the page to get into an appropriate state.
2748  *
2749  * We do this without the lock held, so that it can sleep if it needs to.
2750  */
2751 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2752 {
2753 	vm_fault_t ret;
2754 	struct page *page = vmf->page;
2755 	unsigned int old_flags = vmf->flags;
2756 
2757 	vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2758 
2759 	if (vmf->vma->vm_file &&
2760 	    IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2761 		return VM_FAULT_SIGBUS;
2762 
2763 	ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2764 	/* Restore original flags so that caller is not surprised */
2765 	vmf->flags = old_flags;
2766 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2767 		return ret;
2768 	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2769 		lock_page(page);
2770 		if (!page->mapping) {
2771 			unlock_page(page);
2772 			return 0; /* retry */
2773 		}
2774 		ret |= VM_FAULT_LOCKED;
2775 	} else
2776 		VM_BUG_ON_PAGE(!PageLocked(page), page);
2777 	return ret;
2778 }
2779 
2780 /*
2781  * Handle dirtying of a page in shared file mapping on a write fault.
2782  *
2783  * The function expects the page to be locked and unlocks it.
2784  */
2785 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2786 {
2787 	struct vm_area_struct *vma = vmf->vma;
2788 	struct address_space *mapping;
2789 	struct page *page = vmf->page;
2790 	bool dirtied;
2791 	bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2792 
2793 	dirtied = set_page_dirty(page);
2794 	VM_BUG_ON_PAGE(PageAnon(page), page);
2795 	/*
2796 	 * Take a local copy of the address_space - page.mapping may be zeroed
2797 	 * by truncate after unlock_page().   The address_space itself remains
2798 	 * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2799 	 * release semantics to prevent the compiler from undoing this copying.
2800 	 */
2801 	mapping = page_rmapping(page);
2802 	unlock_page(page);
2803 
2804 	if (!page_mkwrite)
2805 		file_update_time(vma->vm_file);
2806 
2807 	/*
2808 	 * Throttle page dirtying rate down to writeback speed.
2809 	 *
2810 	 * mapping may be NULL here because some device drivers do not
2811 	 * set page.mapping but still dirty their pages
2812 	 *
2813 	 * Drop the mmap_lock before waiting on IO, if we can. The file
2814 	 * is pinning the mapping, as per above.
2815 	 */
2816 	if ((dirtied || page_mkwrite) && mapping) {
2817 		struct file *fpin;
2818 
2819 		fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2820 		balance_dirty_pages_ratelimited(mapping);
2821 		if (fpin) {
2822 			fput(fpin);
2823 			return VM_FAULT_RETRY;
2824 		}
2825 	}
2826 
2827 	return 0;
2828 }
2829 
2830 /*
2831  * Handle write page faults for pages that can be reused in the current vma
2832  *
2833  * This can happen either due to the mapping being with the VM_SHARED flag,
2834  * or due to us being the last reference standing to the page. In either
2835  * case, all we need to do here is to mark the page as writable and update
2836  * any related book-keeping.
2837  */
2838 static inline void wp_page_reuse(struct vm_fault *vmf)
2839 	__releases(vmf->ptl)
2840 {
2841 	struct vm_area_struct *vma = vmf->vma;
2842 	struct page *page = vmf->page;
2843 	pte_t entry;
2844 	/*
2845 	 * Clear the pages cpupid information as the existing
2846 	 * information potentially belongs to a now completely
2847 	 * unrelated process.
2848 	 */
2849 	if (page)
2850 		page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2851 
2852 	flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2853 	entry = pte_mkyoung(vmf->orig_pte);
2854 	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2855 	if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2856 		update_mmu_cache(vma, vmf->address, vmf->pte);
2857 	pte_unmap_unlock(vmf->pte, vmf->ptl);
2858 	count_vm_event(PGREUSE);
2859 }
2860 
2861 /*
2862  * Handle the case of a page which we actually need to copy to a new page.
2863  *
2864  * Called with mmap_lock locked and the old page referenced, but
2865  * without the ptl held.
2866  *
2867  * High level logic flow:
2868  *
2869  * - Allocate a page, copy the content of the old page to the new one.
2870  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2871  * - Take the PTL. If the pte changed, bail out and release the allocated page
2872  * - If the pte is still the way we remember it, update the page table and all
2873  *   relevant references. This includes dropping the reference the page-table
2874  *   held to the old page, as well as updating the rmap.
2875  * - In any case, unlock the PTL and drop the reference we took to the old page.
2876  */
2877 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2878 {
2879 	struct vm_area_struct *vma = vmf->vma;
2880 	struct mm_struct *mm = vma->vm_mm;
2881 	struct page *old_page = vmf->page;
2882 	struct page *new_page = NULL;
2883 	pte_t entry;
2884 	int page_copied = 0;
2885 	struct mmu_notifier_range range;
2886 
2887 	if (unlikely(anon_vma_prepare(vma)))
2888 		goto oom;
2889 
2890 	if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2891 		new_page = alloc_zeroed_user_highpage_movable(vma,
2892 							      vmf->address);
2893 		if (!new_page)
2894 			goto oom;
2895 	} else {
2896 		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2897 				vmf->address);
2898 		if (!new_page)
2899 			goto oom;
2900 
2901 		if (!cow_user_page(new_page, old_page, vmf)) {
2902 			/*
2903 			 * COW failed, if the fault was solved by other,
2904 			 * it's fine. If not, userspace would re-fault on
2905 			 * the same address and we will handle the fault
2906 			 * from the second attempt.
2907 			 */
2908 			put_page(new_page);
2909 			if (old_page)
2910 				put_page(old_page);
2911 			return 0;
2912 		}
2913 	}
2914 
2915 	if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2916 		goto oom_free_new;
2917 	cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2918 
2919 	__SetPageUptodate(new_page);
2920 
2921 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2922 				vmf->address & PAGE_MASK,
2923 				(vmf->address & PAGE_MASK) + PAGE_SIZE);
2924 	mmu_notifier_invalidate_range_start(&range);
2925 
2926 	/*
2927 	 * Re-check the pte - we dropped the lock
2928 	 */
2929 	vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2930 	if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2931 		if (old_page) {
2932 			if (!PageAnon(old_page)) {
2933 				dec_mm_counter_fast(mm,
2934 						mm_counter_file(old_page));
2935 				inc_mm_counter_fast(mm, MM_ANONPAGES);
2936 			}
2937 		} else {
2938 			inc_mm_counter_fast(mm, MM_ANONPAGES);
2939 		}
2940 		flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2941 		entry = mk_pte(new_page, vma->vm_page_prot);
2942 		entry = pte_sw_mkyoung(entry);
2943 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2944 
2945 		/*
2946 		 * Clear the pte entry and flush it first, before updating the
2947 		 * pte with the new entry, to keep TLBs on different CPUs in
2948 		 * sync. This code used to set the new PTE then flush TLBs, but
2949 		 * that left a window where the new PTE could be loaded into
2950 		 * some TLBs while the old PTE remains in others.
2951 		 */
2952 		ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2953 		page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2954 		lru_cache_add_inactive_or_unevictable(new_page, vma);
2955 		/*
2956 		 * We call the notify macro here because, when using secondary
2957 		 * mmu page tables (such as kvm shadow page tables), we want the
2958 		 * new page to be mapped directly into the secondary page table.
2959 		 */
2960 		set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2961 		update_mmu_cache(vma, vmf->address, vmf->pte);
2962 		if (old_page) {
2963 			/*
2964 			 * Only after switching the pte to the new page may
2965 			 * we remove the mapcount here. Otherwise another
2966 			 * process may come and find the rmap count decremented
2967 			 * before the pte is switched to the new page, and
2968 			 * "reuse" the old page writing into it while our pte
2969 			 * here still points into it and can be read by other
2970 			 * threads.
2971 			 *
2972 			 * The critical issue is to order this
2973 			 * page_remove_rmap with the ptp_clear_flush above.
2974 			 * Those stores are ordered by (if nothing else,)
2975 			 * the barrier present in the atomic_add_negative
2976 			 * in page_remove_rmap.
2977 			 *
2978 			 * Then the TLB flush in ptep_clear_flush ensures that
2979 			 * no process can access the old page before the
2980 			 * decremented mapcount is visible. And the old page
2981 			 * cannot be reused until after the decremented
2982 			 * mapcount is visible. So transitively, TLBs to
2983 			 * old page will be flushed before it can be reused.
2984 			 */
2985 			page_remove_rmap(old_page, false);
2986 		}
2987 
2988 		/* Free the old page.. */
2989 		new_page = old_page;
2990 		page_copied = 1;
2991 	} else {
2992 		update_mmu_tlb(vma, vmf->address, vmf->pte);
2993 	}
2994 
2995 	if (new_page)
2996 		put_page(new_page);
2997 
2998 	pte_unmap_unlock(vmf->pte, vmf->ptl);
2999 	/*
3000 	 * No need to double call mmu_notifier->invalidate_range() callback as
3001 	 * the above ptep_clear_flush_notify() did already call it.
3002 	 */
3003 	mmu_notifier_invalidate_range_only_end(&range);
3004 	if (old_page) {
3005 		/*
3006 		 * Don't let another task, with possibly unlocked vma,
3007 		 * keep the mlocked page.
3008 		 */
3009 		if (page_copied && (vma->vm_flags & VM_LOCKED)) {
3010 			lock_page(old_page);	/* LRU manipulation */
3011 			if (PageMlocked(old_page))
3012 				munlock_vma_page(old_page);
3013 			unlock_page(old_page);
3014 		}
3015 		put_page(old_page);
3016 	}
3017 	return page_copied ? VM_FAULT_WRITE : 0;
3018 oom_free_new:
3019 	put_page(new_page);
3020 oom:
3021 	if (old_page)
3022 		put_page(old_page);
3023 	return VM_FAULT_OOM;
3024 }
3025 
3026 /**
3027  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3028  *			  writeable once the page is prepared
3029  *
3030  * @vmf: structure describing the fault
3031  *
3032  * This function handles all that is needed to finish a write page fault in a
3033  * shared mapping due to PTE being read-only once the mapped page is prepared.
3034  * It handles locking of PTE and modifying it.
3035  *
3036  * The function expects the page to be locked or other protection against
3037  * concurrent faults / writeback (such as DAX radix tree locks).
3038  *
3039  * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
3040  * we acquired PTE lock.
3041  */
3042 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3043 {
3044 	WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3045 	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3046 				       &vmf->ptl);
3047 	/*
3048 	 * We might have raced with another page fault while we released the
3049 	 * pte_offset_map_lock.
3050 	 */
3051 	if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3052 		update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3053 		pte_unmap_unlock(vmf->pte, vmf->ptl);
3054 		return VM_FAULT_NOPAGE;
3055 	}
3056 	wp_page_reuse(vmf);
3057 	return 0;
3058 }
3059 
3060 /*
3061  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3062  * mapping
3063  */
3064 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3065 {
3066 	struct vm_area_struct *vma = vmf->vma;
3067 
3068 	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3069 		vm_fault_t ret;
3070 
3071 		pte_unmap_unlock(vmf->pte, vmf->ptl);
3072 		vmf->flags |= FAULT_FLAG_MKWRITE;
3073 		ret = vma->vm_ops->pfn_mkwrite(vmf);
3074 		if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3075 			return ret;
3076 		return finish_mkwrite_fault(vmf);
3077 	}
3078 	wp_page_reuse(vmf);
3079 	return VM_FAULT_WRITE;
3080 }
3081 
3082 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3083 	__releases(vmf->ptl)
3084 {
3085 	struct vm_area_struct *vma = vmf->vma;
3086 	vm_fault_t ret = VM_FAULT_WRITE;
3087 
3088 	get_page(vmf->page);
3089 
3090 	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3091 		vm_fault_t tmp;
3092 
3093 		pte_unmap_unlock(vmf->pte, vmf->ptl);
3094 		tmp = do_page_mkwrite(vmf);
3095 		if (unlikely(!tmp || (tmp &
3096 				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3097 			put_page(vmf->page);
3098 			return tmp;
3099 		}
3100 		tmp = finish_mkwrite_fault(vmf);
3101 		if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3102 			unlock_page(vmf->page);
3103 			put_page(vmf->page);
3104 			return tmp;
3105 		}
3106 	} else {
3107 		wp_page_reuse(vmf);
3108 		lock_page(vmf->page);
3109 	}
3110 	ret |= fault_dirty_shared_page(vmf);
3111 	put_page(vmf->page);
3112 
3113 	return ret;
3114 }
3115 
3116 /*
3117  * This routine handles present pages, when users try to write
3118  * to a shared page. It is done by copying the page to a new address
3119  * and decrementing the shared-page counter for the old page.
3120  *
3121  * Note that this routine assumes that the protection checks have been
3122  * done by the caller (the low-level page fault routine in most cases).
3123  * Thus we can safely just mark it writable once we've done any necessary
3124  * COW.
3125  *
3126  * We also mark the page dirty at this point even though the page will
3127  * change only once the write actually happens. This avoids a few races,
3128  * and potentially makes it more efficient.
3129  *
3130  * We enter with non-exclusive mmap_lock (to exclude vma changes,
3131  * but allow concurrent faults), with pte both mapped and locked.
3132  * We return with mmap_lock still held, but pte unmapped and unlocked.
3133  */
3134 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3135 	__releases(vmf->ptl)
3136 {
3137 	struct vm_area_struct *vma = vmf->vma;
3138 
3139 	if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3140 		pte_unmap_unlock(vmf->pte, vmf->ptl);
3141 		return handle_userfault(vmf, VM_UFFD_WP);
3142 	}
3143 
3144 	/*
3145 	 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3146 	 * is flushed in this case before copying.
3147 	 */
3148 	if (unlikely(userfaultfd_wp(vmf->vma) &&
3149 		     mm_tlb_flush_pending(vmf->vma->vm_mm)))
3150 		flush_tlb_page(vmf->vma, vmf->address);
3151 
3152 	vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3153 	if (!vmf->page) {
3154 		/*
3155 		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3156 		 * VM_PFNMAP VMA.
3157 		 *
3158 		 * We should not cow pages in a shared writeable mapping.
3159 		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3160 		 */
3161 		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3162 				     (VM_WRITE|VM_SHARED))
3163 			return wp_pfn_shared(vmf);
3164 
3165 		pte_unmap_unlock(vmf->pte, vmf->ptl);
3166 		return wp_page_copy(vmf);
3167 	}
3168 
3169 	/*
3170 	 * Take out anonymous pages first, anonymous shared vmas are
3171 	 * not dirty accountable.
3172 	 */
3173 	if (PageAnon(vmf->page)) {
3174 		struct page *page = vmf->page;
3175 
3176 		/* PageKsm() doesn't necessarily raise the page refcount */
3177 		if (PageKsm(page) || page_count(page) != 1)
3178 			goto copy;
3179 		if (!trylock_page(page))
3180 			goto copy;
3181 		if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3182 			unlock_page(page);
3183 			goto copy;
3184 		}
3185 		/*
3186 		 * Ok, we've got the only map reference, and the only
3187 		 * page count reference, and the page is locked,
3188 		 * it's dark out, and we're wearing sunglasses. Hit it.
3189 		 */
3190 		unlock_page(page);
3191 		wp_page_reuse(vmf);
3192 		return VM_FAULT_WRITE;
3193 	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3194 					(VM_WRITE|VM_SHARED))) {
3195 		return wp_page_shared(vmf);
3196 	}
3197 copy:
3198 	/*
3199 	 * Ok, we need to copy. Oh, well..
3200 	 */
3201 	get_page(vmf->page);
3202 
3203 	pte_unmap_unlock(vmf->pte, vmf->ptl);
3204 	return wp_page_copy(vmf);
3205 }
3206 
3207 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3208 		unsigned long start_addr, unsigned long end_addr,
3209 		struct zap_details *details)
3210 {
3211 	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3212 }
3213 
3214 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3215 					    struct zap_details *details)
3216 {
3217 	struct vm_area_struct *vma;
3218 	pgoff_t vba, vea, zba, zea;
3219 
3220 	vma_interval_tree_foreach(vma, root,
3221 			details->first_index, details->last_index) {
3222 
3223 		vba = vma->vm_pgoff;
3224 		vea = vba + vma_pages(vma) - 1;
3225 		zba = details->first_index;
3226 		if (zba < vba)
3227 			zba = vba;
3228 		zea = details->last_index;
3229 		if (zea > vea)
3230 			zea = vea;
3231 
3232 		unmap_mapping_range_vma(vma,
3233 			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3234 			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3235 				details);
3236 	}
3237 }
3238 
3239 /**
3240  * unmap_mapping_pages() - Unmap pages from processes.
3241  * @mapping: The address space containing pages to be unmapped.
3242  * @start: Index of first page to be unmapped.
3243  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
3244  * @even_cows: Whether to unmap even private COWed pages.
3245  *
3246  * Unmap the pages in this address space from any userspace process which
3247  * has them mmaped.  Generally, you want to remove COWed pages as well when
3248  * a file is being truncated, but not when invalidating pages from the page
3249  * cache.
3250  */
3251 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3252 		pgoff_t nr, bool even_cows)
3253 {
3254 	struct zap_details details = { };
3255 
3256 	details.check_mapping = even_cows ? NULL : mapping;
3257 	details.first_index = start;
3258 	details.last_index = start + nr - 1;
3259 	if (details.last_index < details.first_index)
3260 		details.last_index = ULONG_MAX;
3261 
3262 	i_mmap_lock_write(mapping);
3263 	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3264 		unmap_mapping_range_tree(&mapping->i_mmap, &details);
3265 	i_mmap_unlock_write(mapping);
3266 }
3267 
3268 /**
3269  * unmap_mapping_range - unmap the portion of all mmaps in the specified
3270  * address_space corresponding to the specified byte range in the underlying
3271  * file.
3272  *
3273  * @mapping: the address space containing mmaps to be unmapped.
3274  * @holebegin: byte in first page to unmap, relative to the start of
3275  * the underlying file.  This will be rounded down to a PAGE_SIZE
3276  * boundary.  Note that this is different from truncate_pagecache(), which
3277  * must keep the partial page.  In contrast, we must get rid of
3278  * partial pages.
3279  * @holelen: size of prospective hole in bytes.  This will be rounded
3280  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
3281  * end of the file.
3282  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3283  * but 0 when invalidating pagecache, don't throw away private data.
3284  */
3285 void unmap_mapping_range(struct address_space *mapping,
3286 		loff_t const holebegin, loff_t const holelen, int even_cows)
3287 {
3288 	pgoff_t hba = holebegin >> PAGE_SHIFT;
3289 	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3290 
3291 	/* Check for overflow. */
3292 	if (sizeof(holelen) > sizeof(hlen)) {
3293 		long long holeend =
3294 			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3295 		if (holeend & ~(long long)ULONG_MAX)
3296 			hlen = ULONG_MAX - hba + 1;
3297 	}
3298 
3299 	unmap_mapping_pages(mapping, hba, hlen, even_cows);
3300 }
3301 EXPORT_SYMBOL(unmap_mapping_range);
3302 
3303 /*
3304  * We enter with non-exclusive mmap_lock (to exclude vma changes,
3305  * but allow concurrent faults), and pte mapped but not yet locked.
3306  * We return with pte unmapped and unlocked.
3307  *
3308  * We return with the mmap_lock locked or unlocked in the same cases
3309  * as does filemap_fault().
3310  */
3311 vm_fault_t do_swap_page(struct vm_fault *vmf)
3312 {
3313 	struct vm_area_struct *vma = vmf->vma;
3314 	struct page *page = NULL, *swapcache;
3315 	swp_entry_t entry;
3316 	pte_t pte;
3317 	int locked;
3318 	int exclusive = 0;
3319 	vm_fault_t ret = 0;
3320 	void *shadow = NULL;
3321 
3322 	if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3323 		goto out;
3324 
3325 	entry = pte_to_swp_entry(vmf->orig_pte);
3326 	if (unlikely(non_swap_entry(entry))) {
3327 		if (is_migration_entry(entry)) {
3328 			migration_entry_wait(vma->vm_mm, vmf->pmd,
3329 					     vmf->address);
3330 		} else if (is_device_private_entry(entry)) {
3331 			vmf->page = device_private_entry_to_page(entry);
3332 			ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3333 		} else if (is_hwpoison_entry(entry)) {
3334 			ret = VM_FAULT_HWPOISON;
3335 		} else {
3336 			print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3337 			ret = VM_FAULT_SIGBUS;
3338 		}
3339 		goto out;
3340 	}
3341 
3342 
3343 	delayacct_set_flag(current, DELAYACCT_PF_SWAPIN);
3344 	page = lookup_swap_cache(entry, vma, vmf->address);
3345 	swapcache = page;
3346 
3347 	if (!page) {
3348 		struct swap_info_struct *si = swp_swap_info(entry);
3349 
3350 		if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3351 		    __swap_count(entry) == 1) {
3352 			/* skip swapcache */
3353 			page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3354 							vmf->address);
3355 			if (page) {
3356 				__SetPageLocked(page);
3357 				__SetPageSwapBacked(page);
3358 
3359 				if (mem_cgroup_swapin_charge_page(page,
3360 					vma->vm_mm, GFP_KERNEL, entry)) {
3361 					ret = VM_FAULT_OOM;
3362 					goto out_page;
3363 				}
3364 				mem_cgroup_swapin_uncharge_swap(entry);
3365 
3366 				shadow = get_shadow_from_swap_cache(entry);
3367 				if (shadow)
3368 					workingset_refault(page, shadow);
3369 
3370 				lru_cache_add(page);
3371 
3372 				/* To provide entry to swap_readpage() */
3373 				set_page_private(page, entry.val);
3374 				swap_readpage(page, true);
3375 				set_page_private(page, 0);
3376 			}
3377 		} else {
3378 			page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3379 						vmf);
3380 			swapcache = page;
3381 		}
3382 
3383 		if (!page) {
3384 			/*
3385 			 * Back out if somebody else faulted in this pte
3386 			 * while we released the pte lock.
3387 			 */
3388 			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3389 					vmf->address, &vmf->ptl);
3390 			if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3391 				ret = VM_FAULT_OOM;
3392 			delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3393 			goto unlock;
3394 		}
3395 
3396 		/* Had to read the page from swap area: Major fault */
3397 		ret = VM_FAULT_MAJOR;
3398 		count_vm_event(PGMAJFAULT);
3399 		count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3400 	} else if (PageHWPoison(page)) {
3401 		/*
3402 		 * hwpoisoned dirty swapcache pages are kept for killing
3403 		 * owner processes (which may be unknown at hwpoison time)
3404 		 */
3405 		ret = VM_FAULT_HWPOISON;
3406 		delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3407 		goto out_release;
3408 	}
3409 
3410 	locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3411 
3412 	delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3413 	if (!locked) {
3414 		ret |= VM_FAULT_RETRY;
3415 		goto out_release;
3416 	}
3417 
3418 	/*
3419 	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3420 	 * release the swapcache from under us.  The page pin, and pte_same
3421 	 * test below, are not enough to exclude that.  Even if it is still
3422 	 * swapcache, we need to check that the page's swap has not changed.
3423 	 */
3424 	if (unlikely((!PageSwapCache(page) ||
3425 			page_private(page) != entry.val)) && swapcache)
3426 		goto out_page;
3427 
3428 	page = ksm_might_need_to_copy(page, vma, vmf->address);
3429 	if (unlikely(!page)) {
3430 		ret = VM_FAULT_OOM;
3431 		page = swapcache;
3432 		goto out_page;
3433 	}
3434 
3435 	cgroup_throttle_swaprate(page, GFP_KERNEL);
3436 
3437 	/*
3438 	 * Back out if somebody else already faulted in this pte.
3439 	 */
3440 	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3441 			&vmf->ptl);
3442 	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3443 		goto out_nomap;
3444 
3445 	if (unlikely(!PageUptodate(page))) {
3446 		ret = VM_FAULT_SIGBUS;
3447 		goto out_nomap;
3448 	}
3449 
3450 	/*
3451 	 * The page isn't present yet, go ahead with the fault.
3452 	 *
3453 	 * Be careful about the sequence of operations here.
3454 	 * To get its accounting right, reuse_swap_page() must be called
3455 	 * while the page is counted on swap but not yet in mapcount i.e.
3456 	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3457 	 * must be called after the swap_free(), or it will never succeed.
3458 	 */
3459 
3460 	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3461 	dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3462 	pte = mk_pte(page, vma->vm_page_prot);
3463 	if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3464 		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3465 		vmf->flags &= ~FAULT_FLAG_WRITE;
3466 		ret |= VM_FAULT_WRITE;
3467 		exclusive = RMAP_EXCLUSIVE;
3468 	}
3469 	flush_icache_page(vma, page);
3470 	if (pte_swp_soft_dirty(vmf->orig_pte))
3471 		pte = pte_mksoft_dirty(pte);
3472 	if (pte_swp_uffd_wp(vmf->orig_pte)) {
3473 		pte = pte_mkuffd_wp(pte);
3474 		pte = pte_wrprotect(pte);
3475 	}
3476 	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3477 	arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3478 	vmf->orig_pte = pte;
3479 
3480 	/* ksm created a completely new copy */
3481 	if (unlikely(page != swapcache && swapcache)) {
3482 		page_add_new_anon_rmap(page, vma, vmf->address, false);
3483 		lru_cache_add_inactive_or_unevictable(page, vma);
3484 	} else {
3485 		do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3486 	}
3487 
3488 	swap_free(entry);
3489 	if (mem_cgroup_swap_full(page) ||
3490 	    (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3491 		try_to_free_swap(page);
3492 	unlock_page(page);
3493 	if (page != swapcache && swapcache) {
3494 		/*
3495 		 * Hold the lock to avoid the swap entry to be reused
3496 		 * until we take the PT lock for the pte_same() check
3497 		 * (to avoid false positives from pte_same). For
3498 		 * further safety release the lock after the swap_free
3499 		 * so that the swap count won't change under a
3500 		 * parallel locked swapcache.
3501 		 */
3502 		unlock_page(swapcache);
3503 		put_page(swapcache);
3504 	}
3505 
3506 	if (vmf->flags & FAULT_FLAG_WRITE) {
3507 		ret |= do_wp_page(vmf);
3508 		if (ret & VM_FAULT_ERROR)
3509 			ret &= VM_FAULT_ERROR;
3510 		goto out;
3511 	}
3512 
3513 	/* No need to invalidate - it was non-present before */
3514 	update_mmu_cache(vma, vmf->address, vmf->pte);
3515 unlock:
3516 	pte_unmap_unlock(vmf->pte, vmf->ptl);
3517 out:
3518 	return ret;
3519 out_nomap:
3520 	pte_unmap_unlock(vmf->pte, vmf->ptl);
3521 out_page:
3522 	unlock_page(page);
3523 out_release:
3524 	put_page(page);
3525 	if (page != swapcache && swapcache) {
3526 		unlock_page(swapcache);
3527 		put_page(swapcache);
3528 	}
3529 	return ret;
3530 }
3531 
3532 /*
3533  * We enter with non-exclusive mmap_lock (to exclude vma changes,
3534  * but allow concurrent faults), and pte mapped but not yet locked.
3535  * We return with mmap_lock still held, but pte unmapped and unlocked.
3536  */
3537 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3538 {
3539 	struct vm_area_struct *vma = vmf->vma;
3540 	struct page *page;
3541 	vm_fault_t ret = 0;
3542 	pte_t entry;
3543 
3544 	/* File mapping without ->vm_ops ? */
3545 	if (vma->vm_flags & VM_SHARED)
3546 		return VM_FAULT_SIGBUS;
3547 
3548 	/*
3549 	 * Use pte_alloc() instead of pte_alloc_map().  We can't run
3550 	 * pte_offset_map() on pmds where a huge pmd might be created
3551 	 * from a different thread.
3552 	 *
3553 	 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3554 	 * parallel threads are excluded by other means.
3555 	 *
3556 	 * Here we only have mmap_read_lock(mm).
3557 	 */
3558 	if (pte_alloc(vma->vm_mm, vmf->pmd))
3559 		return VM_FAULT_OOM;
3560 
3561 	/* See comment in handle_pte_fault() */
3562 	if (unlikely(pmd_trans_unstable(vmf->pmd)))
3563 		return 0;
3564 
3565 	/* Use the zero-page for reads */
3566 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3567 			!mm_forbids_zeropage(vma->vm_mm)) {
3568 		entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3569 						vma->vm_page_prot));
3570 		vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3571 				vmf->address, &vmf->ptl);
3572 		if (!pte_none(*vmf->pte)) {
3573 			update_mmu_tlb(vma, vmf->address, vmf->pte);
3574 			goto unlock;
3575 		}
3576 		ret = check_stable_address_space(vma->vm_mm);
3577 		if (ret)
3578 			goto unlock;
3579 		/* Deliver the page fault to userland, check inside PT lock */
3580 		if (userfaultfd_missing(vma)) {
3581 			pte_unmap_unlock(vmf->pte, vmf->ptl);
3582 			return handle_userfault(vmf, VM_UFFD_MISSING);
3583 		}
3584 		goto setpte;
3585 	}
3586 
3587 	/* Allocate our own private page. */
3588 	if (unlikely(anon_vma_prepare(vma)))
3589 		goto oom;
3590 	page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3591 	if (!page)
3592 		goto oom;
3593 
3594 	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3595 		goto oom_free_page;
3596 	cgroup_throttle_swaprate(page, GFP_KERNEL);
3597 
3598 	/*
3599 	 * The memory barrier inside __SetPageUptodate makes sure that
3600 	 * preceding stores to the page contents become visible before
3601 	 * the set_pte_at() write.
3602 	 */
3603 	__SetPageUptodate(page);
3604 
3605 	entry = mk_pte(page, vma->vm_page_prot);
3606 	entry = pte_sw_mkyoung(entry);
3607 	if (vma->vm_flags & VM_WRITE)
3608 		entry = pte_mkwrite(pte_mkdirty(entry));
3609 
3610 	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3611 			&vmf->ptl);
3612 	if (!pte_none(*vmf->pte)) {
3613 		update_mmu_cache(vma, vmf->address, vmf->pte);
3614 		goto release;
3615 	}
3616 
3617 	ret = check_stable_address_space(vma->vm_mm);
3618 	if (ret)
3619 		goto release;
3620 
3621 	/* Deliver the page fault to userland, check inside PT lock */
3622 	if (userfaultfd_missing(vma)) {
3623 		pte_unmap_unlock(vmf->pte, vmf->ptl);
3624 		put_page(page);
3625 		return handle_userfault(vmf, VM_UFFD_MISSING);
3626 	}
3627 
3628 	inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3629 	page_add_new_anon_rmap(page, vma, vmf->address, false);
3630 	lru_cache_add_inactive_or_unevictable(page, vma);
3631 setpte:
3632 	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3633 
3634 	/* No need to invalidate - it was non-present before */
3635 	update_mmu_cache(vma, vmf->address, vmf->pte);
3636 unlock:
3637 	pte_unmap_unlock(vmf->pte, vmf->ptl);
3638 	return ret;
3639 release:
3640 	put_page(page);
3641 	goto unlock;
3642 oom_free_page:
3643 	put_page(page);
3644 oom:
3645 	return VM_FAULT_OOM;
3646 }
3647 
3648 /*
3649  * The mmap_lock must have been held on entry, and may have been
3650  * released depending on flags and vma->vm_ops->fault() return value.
3651  * See filemap_fault() and __lock_page_retry().
3652  */
3653 static vm_fault_t __do_fault(struct vm_fault *vmf)
3654 {
3655 	struct vm_area_struct *vma = vmf->vma;
3656 	vm_fault_t ret;
3657 
3658 	/*
3659 	 * Preallocate pte before we take page_lock because this might lead to
3660 	 * deadlocks for memcg reclaim which waits for pages under writeback:
3661 	 *				lock_page(A)
3662 	 *				SetPageWriteback(A)
3663 	 *				unlock_page(A)
3664 	 * lock_page(B)
3665 	 *				lock_page(B)
3666 	 * pte_alloc_one
3667 	 *   shrink_page_list
3668 	 *     wait_on_page_writeback(A)
3669 	 *				SetPageWriteback(B)
3670 	 *				unlock_page(B)
3671 	 *				# flush A, B to clear the writeback
3672 	 */
3673 	if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3674 		vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3675 		if (!vmf->prealloc_pte)
3676 			return VM_FAULT_OOM;
3677 		smp_wmb(); /* See comment in __pte_alloc() */
3678 	}
3679 
3680 	ret = vma->vm_ops->fault(vmf);
3681 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3682 			    VM_FAULT_DONE_COW)))
3683 		return ret;
3684 
3685 	if (unlikely(PageHWPoison(vmf->page))) {
3686 		if (ret & VM_FAULT_LOCKED)
3687 			unlock_page(vmf->page);
3688 		put_page(vmf->page);
3689 		vmf->page = NULL;
3690 		return VM_FAULT_HWPOISON;
3691 	}
3692 
3693 	if (unlikely(!(ret & VM_FAULT_LOCKED)))
3694 		lock_page(vmf->page);
3695 	else
3696 		VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3697 
3698 	return ret;
3699 }
3700 
3701 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3702 static void deposit_prealloc_pte(struct vm_fault *vmf)
3703 {
3704 	struct vm_area_struct *vma = vmf->vma;
3705 
3706 	pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3707 	/*
3708 	 * We are going to consume the prealloc table,
3709 	 * count that as nr_ptes.
3710 	 */
3711 	mm_inc_nr_ptes(vma->vm_mm);
3712 	vmf->prealloc_pte = NULL;
3713 }
3714 
3715 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3716 {
3717 	struct vm_area_struct *vma = vmf->vma;
3718 	bool write = vmf->flags & FAULT_FLAG_WRITE;
3719 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3720 	pmd_t entry;
3721 	int i;
3722 	vm_fault_t ret = VM_FAULT_FALLBACK;
3723 
3724 	if (!transhuge_vma_suitable(vma, haddr))
3725 		return ret;
3726 
3727 	page = compound_head(page);
3728 	if (compound_order(page) != HPAGE_PMD_ORDER)
3729 		return ret;
3730 
3731 	/*
3732 	 * Archs like ppc64 need additional space to store information
3733 	 * related to pte entry. Use the preallocated table for that.
3734 	 */
3735 	if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3736 		vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3737 		if (!vmf->prealloc_pte)
3738 			return VM_FAULT_OOM;
3739 		smp_wmb(); /* See comment in __pte_alloc() */
3740 	}
3741 
3742 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3743 	if (unlikely(!pmd_none(*vmf->pmd)))
3744 		goto out;
3745 
3746 	for (i = 0; i < HPAGE_PMD_NR; i++)
3747 		flush_icache_page(vma, page + i);
3748 
3749 	entry = mk_huge_pmd(page, vma->vm_page_prot);
3750 	if (write)
3751 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3752 
3753 	add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3754 	page_add_file_rmap(page, true);
3755 	/*
3756 	 * deposit and withdraw with pmd lock held
3757 	 */
3758 	if (arch_needs_pgtable_deposit())
3759 		deposit_prealloc_pte(vmf);
3760 
3761 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3762 
3763 	update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3764 
3765 	/* fault is handled */
3766 	ret = 0;
3767 	count_vm_event(THP_FILE_MAPPED);
3768 out:
3769 	spin_unlock(vmf->ptl);
3770 	return ret;
3771 }
3772 #else
3773 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3774 {
3775 	return VM_FAULT_FALLBACK;
3776 }
3777 #endif
3778 
3779 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3780 {
3781 	struct vm_area_struct *vma = vmf->vma;
3782 	bool write = vmf->flags & FAULT_FLAG_WRITE;
3783 	bool prefault = vmf->address != addr;
3784 	pte_t entry;
3785 
3786 	flush_icache_page(vma, page);
3787 	entry = mk_pte(page, vma->vm_page_prot);
3788 
3789 	if (prefault && arch_wants_old_prefaulted_pte())
3790 		entry = pte_mkold(entry);
3791 	else
3792 		entry = pte_sw_mkyoung(entry);
3793 
3794 	if (write)
3795 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3796 	/* copy-on-write page */
3797 	if (write && !(vma->vm_flags & VM_SHARED)) {
3798 		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3799 		page_add_new_anon_rmap(page, vma, addr, false);
3800 		lru_cache_add_inactive_or_unevictable(page, vma);
3801 	} else {
3802 		inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3803 		page_add_file_rmap(page, false);
3804 	}
3805 	set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
3806 }
3807 
3808 /**
3809  * finish_fault - finish page fault once we have prepared the page to fault
3810  *
3811  * @vmf: structure describing the fault
3812  *
3813  * This function handles all that is needed to finish a page fault once the
3814  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3815  * given page, adds reverse page mapping, handles memcg charges and LRU
3816  * addition.
3817  *
3818  * The function expects the page to be locked and on success it consumes a
3819  * reference of a page being mapped (for the PTE which maps it).
3820  *
3821  * Return: %0 on success, %VM_FAULT_ code in case of error.
3822  */
3823 vm_fault_t finish_fault(struct vm_fault *vmf)
3824 {
3825 	struct vm_area_struct *vma = vmf->vma;
3826 	struct page *page;
3827 	vm_fault_t ret;
3828 
3829 	/* Did we COW the page? */
3830 	if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3831 		page = vmf->cow_page;
3832 	else
3833 		page = vmf->page;
3834 
3835 	/*
3836 	 * check even for read faults because we might have lost our CoWed
3837 	 * page
3838 	 */
3839 	if (!(vma->vm_flags & VM_SHARED)) {
3840 		ret = check_stable_address_space(vma->vm_mm);
3841 		if (ret)
3842 			return ret;
3843 	}
3844 
3845 	if (pmd_none(*vmf->pmd)) {
3846 		if (PageTransCompound(page)) {
3847 			ret = do_set_pmd(vmf, page);
3848 			if (ret != VM_FAULT_FALLBACK)
3849 				return ret;
3850 		}
3851 
3852 		if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
3853 			return VM_FAULT_OOM;
3854 	}
3855 
3856 	/* See comment in handle_pte_fault() */
3857 	if (pmd_devmap_trans_unstable(vmf->pmd))
3858 		return 0;
3859 
3860 	vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3861 				      vmf->address, &vmf->ptl);
3862 	ret = 0;
3863 	/* Re-check under ptl */
3864 	if (likely(pte_none(*vmf->pte)))
3865 		do_set_pte(vmf, page, vmf->address);
3866 	else
3867 		ret = VM_FAULT_NOPAGE;
3868 
3869 	update_mmu_tlb(vma, vmf->address, vmf->pte);
3870 	pte_unmap_unlock(vmf->pte, vmf->ptl);
3871 	return ret;
3872 }
3873 
3874 static unsigned long fault_around_bytes __read_mostly =
3875 	rounddown_pow_of_two(65536);
3876 
3877 #ifdef CONFIG_DEBUG_FS
3878 static int fault_around_bytes_get(void *data, u64 *val)
3879 {
3880 	*val = fault_around_bytes;
3881 	return 0;
3882 }
3883 
3884 /*
3885  * fault_around_bytes must be rounded down to the nearest page order as it's
3886  * what do_fault_around() expects to see.
3887  */
3888 static int fault_around_bytes_set(void *data, u64 val)
3889 {
3890 	if (val / PAGE_SIZE > PTRS_PER_PTE)
3891 		return -EINVAL;
3892 	if (val > PAGE_SIZE)
3893 		fault_around_bytes = rounddown_pow_of_two(val);
3894 	else
3895 		fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3896 	return 0;
3897 }
3898 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3899 		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3900 
3901 static int __init fault_around_debugfs(void)
3902 {
3903 	debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3904 				   &fault_around_bytes_fops);
3905 	return 0;
3906 }
3907 late_initcall(fault_around_debugfs);
3908 #endif
3909 
3910 /*
3911  * do_fault_around() tries to map few pages around the fault address. The hope
3912  * is that the pages will be needed soon and this will lower the number of
3913  * faults to handle.
3914  *
3915  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3916  * not ready to be mapped: not up-to-date, locked, etc.
3917  *
3918  * This function is called with the page table lock taken. In the split ptlock
3919  * case the page table lock only protects only those entries which belong to
3920  * the page table corresponding to the fault address.
3921  *
3922  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3923  * only once.
3924  *
3925  * fault_around_bytes defines how many bytes we'll try to map.
3926  * do_fault_around() expects it to be set to a power of two less than or equal
3927  * to PTRS_PER_PTE.
3928  *
3929  * The virtual address of the area that we map is naturally aligned to
3930  * fault_around_bytes rounded down to the machine page size
3931  * (and therefore to page order).  This way it's easier to guarantee
3932  * that we don't cross page table boundaries.
3933  */
3934 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3935 {
3936 	unsigned long address = vmf->address, nr_pages, mask;
3937 	pgoff_t start_pgoff = vmf->pgoff;
3938 	pgoff_t end_pgoff;
3939 	int off;
3940 
3941 	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3942 	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3943 
3944 	address = max(address & mask, vmf->vma->vm_start);
3945 	off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3946 	start_pgoff -= off;
3947 
3948 	/*
3949 	 *  end_pgoff is either the end of the page table, the end of
3950 	 *  the vma or nr_pages from start_pgoff, depending what is nearest.
3951 	 */
3952 	end_pgoff = start_pgoff -
3953 		((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3954 		PTRS_PER_PTE - 1;
3955 	end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3956 			start_pgoff + nr_pages - 1);
3957 
3958 	if (pmd_none(*vmf->pmd)) {
3959 		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3960 		if (!vmf->prealloc_pte)
3961 			return VM_FAULT_OOM;
3962 		smp_wmb(); /* See comment in __pte_alloc() */
3963 	}
3964 
3965 	return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3966 }
3967 
3968 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3969 {
3970 	struct vm_area_struct *vma = vmf->vma;
3971 	vm_fault_t ret = 0;
3972 
3973 	/*
3974 	 * Let's call ->map_pages() first and use ->fault() as fallback
3975 	 * if page by the offset is not ready to be mapped (cold cache or
3976 	 * something).
3977 	 */
3978 	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3979 		ret = do_fault_around(vmf);
3980 		if (ret)
3981 			return ret;
3982 	}
3983 
3984 	ret = __do_fault(vmf);
3985 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3986 		return ret;
3987 
3988 	ret |= finish_fault(vmf);
3989 	unlock_page(vmf->page);
3990 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3991 		put_page(vmf->page);
3992 	return ret;
3993 }
3994 
3995 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3996 {
3997 	struct vm_area_struct *vma = vmf->vma;
3998 	vm_fault_t ret;
3999 
4000 	if (unlikely(anon_vma_prepare(vma)))
4001 		return VM_FAULT_OOM;
4002 
4003 	vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4004 	if (!vmf->cow_page)
4005 		return VM_FAULT_OOM;
4006 
4007 	if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4008 		put_page(vmf->cow_page);
4009 		return VM_FAULT_OOM;
4010 	}
4011 	cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4012 
4013 	ret = __do_fault(vmf);
4014 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4015 		goto uncharge_out;
4016 	if (ret & VM_FAULT_DONE_COW)
4017 		return ret;
4018 
4019 	copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4020 	__SetPageUptodate(vmf->cow_page);
4021 
4022 	ret |= finish_fault(vmf);
4023 	unlock_page(vmf->page);
4024 	put_page(vmf->page);
4025 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4026 		goto uncharge_out;
4027 	return ret;
4028 uncharge_out:
4029 	put_page(vmf->cow_page);
4030 	return ret;
4031 }
4032 
4033 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4034 {
4035 	struct vm_area_struct *vma = vmf->vma;
4036 	vm_fault_t ret, tmp;
4037 
4038 	ret = __do_fault(vmf);
4039 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4040 		return ret;
4041 
4042 	/*
4043 	 * Check if the backing address space wants to know that the page is
4044 	 * about to become writable
4045 	 */
4046 	if (vma->vm_ops->page_mkwrite) {
4047 		unlock_page(vmf->page);
4048 		tmp = do_page_mkwrite(vmf);
4049 		if (unlikely(!tmp ||
4050 				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4051 			put_page(vmf->page);
4052 			return tmp;
4053 		}
4054 	}
4055 
4056 	ret |= finish_fault(vmf);
4057 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4058 					VM_FAULT_RETRY))) {
4059 		unlock_page(vmf->page);
4060 		put_page(vmf->page);
4061 		return ret;
4062 	}
4063 
4064 	ret |= fault_dirty_shared_page(vmf);
4065 	return ret;
4066 }
4067 
4068 /*
4069  * We enter with non-exclusive mmap_lock (to exclude vma changes,
4070  * but allow concurrent faults).
4071  * The mmap_lock may have been released depending on flags and our
4072  * return value.  See filemap_fault() and __lock_page_or_retry().
4073  * If mmap_lock is released, vma may become invalid (for example
4074  * by other thread calling munmap()).
4075  */
4076 static vm_fault_t do_fault(struct vm_fault *vmf)
4077 {
4078 	struct vm_area_struct *vma = vmf->vma;
4079 	struct mm_struct *vm_mm = vma->vm_mm;
4080 	vm_fault_t ret;
4081 
4082 	/*
4083 	 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4084 	 */
4085 	if (!vma->vm_ops->fault) {
4086 		/*
4087 		 * If we find a migration pmd entry or a none pmd entry, which
4088 		 * should never happen, return SIGBUS
4089 		 */
4090 		if (unlikely(!pmd_present(*vmf->pmd)))
4091 			ret = VM_FAULT_SIGBUS;
4092 		else {
4093 			vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4094 						       vmf->pmd,
4095 						       vmf->address,
4096 						       &vmf->ptl);
4097 			/*
4098 			 * Make sure this is not a temporary clearing of pte
4099 			 * by holding ptl and checking again. A R/M/W update
4100 			 * of pte involves: take ptl, clearing the pte so that
4101 			 * we don't have concurrent modification by hardware
4102 			 * followed by an update.
4103 			 */
4104 			if (unlikely(pte_none(*vmf->pte)))
4105 				ret = VM_FAULT_SIGBUS;
4106 			else
4107 				ret = VM_FAULT_NOPAGE;
4108 
4109 			pte_unmap_unlock(vmf->pte, vmf->ptl);
4110 		}
4111 	} else if (!(vmf->flags & FAULT_FLAG_WRITE))
4112 		ret = do_read_fault(vmf);
4113 	else if (!(vma->vm_flags & VM_SHARED))
4114 		ret = do_cow_fault(vmf);
4115 	else
4116 		ret = do_shared_fault(vmf);
4117 
4118 	/* preallocated pagetable is unused: free it */
4119 	if (vmf->prealloc_pte) {
4120 		pte_free(vm_mm, vmf->prealloc_pte);
4121 		vmf->prealloc_pte = NULL;
4122 	}
4123 	return ret;
4124 }
4125 
4126 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4127 				unsigned long addr, int page_nid,
4128 				int *flags)
4129 {
4130 	get_page(page);
4131 
4132 	count_vm_numa_event(NUMA_HINT_FAULTS);
4133 	if (page_nid == numa_node_id()) {
4134 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4135 		*flags |= TNF_FAULT_LOCAL;
4136 	}
4137 
4138 	return mpol_misplaced(page, vma, addr);
4139 }
4140 
4141 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4142 {
4143 	struct vm_area_struct *vma = vmf->vma;
4144 	struct page *page = NULL;
4145 	int page_nid = NUMA_NO_NODE;
4146 	int last_cpupid;
4147 	int target_nid;
4148 	pte_t pte, old_pte;
4149 	bool was_writable = pte_savedwrite(vmf->orig_pte);
4150 	int flags = 0;
4151 
4152 	/*
4153 	 * The "pte" at this point cannot be used safely without
4154 	 * validation through pte_unmap_same(). It's of NUMA type but
4155 	 * the pfn may be screwed if the read is non atomic.
4156 	 */
4157 	vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4158 	spin_lock(vmf->ptl);
4159 	if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4160 		pte_unmap_unlock(vmf->pte, vmf->ptl);
4161 		goto out;
4162 	}
4163 
4164 	/* Get the normal PTE  */
4165 	old_pte = ptep_get(vmf->pte);
4166 	pte = pte_modify(old_pte, vma->vm_page_prot);
4167 
4168 	page = vm_normal_page(vma, vmf->address, pte);
4169 	if (!page)
4170 		goto out_map;
4171 
4172 	/* TODO: handle PTE-mapped THP */
4173 	if (PageCompound(page))
4174 		goto out_map;
4175 
4176 	/*
4177 	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4178 	 * much anyway since they can be in shared cache state. This misses
4179 	 * the case where a mapping is writable but the process never writes
4180 	 * to it but pte_write gets cleared during protection updates and
4181 	 * pte_dirty has unpredictable behaviour between PTE scan updates,
4182 	 * background writeback, dirty balancing and application behaviour.
4183 	 */
4184 	if (!was_writable)
4185 		flags |= TNF_NO_GROUP;
4186 
4187 	/*
4188 	 * Flag if the page is shared between multiple address spaces. This
4189 	 * is later used when determining whether to group tasks together
4190 	 */
4191 	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4192 		flags |= TNF_SHARED;
4193 
4194 	last_cpupid = page_cpupid_last(page);
4195 	page_nid = page_to_nid(page);
4196 	target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4197 			&flags);
4198 	if (target_nid == NUMA_NO_NODE) {
4199 		put_page(page);
4200 		goto out_map;
4201 	}
4202 	pte_unmap_unlock(vmf->pte, vmf->ptl);
4203 
4204 	/* Migrate to the requested node */
4205 	if (migrate_misplaced_page(page, vma, target_nid)) {
4206 		page_nid = target_nid;
4207 		flags |= TNF_MIGRATED;
4208 	} else {
4209 		flags |= TNF_MIGRATE_FAIL;
4210 		vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4211 		spin_lock(vmf->ptl);
4212 		if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4213 			pte_unmap_unlock(vmf->pte, vmf->ptl);
4214 			goto out;
4215 		}
4216 		goto out_map;
4217 	}
4218 
4219 out:
4220 	if (page_nid != NUMA_NO_NODE)
4221 		task_numa_fault(last_cpupid, page_nid, 1, flags);
4222 	return 0;
4223 out_map:
4224 	/*
4225 	 * Make it present again, depending on how arch implements
4226 	 * non-accessible ptes, some can allow access by kernel mode.
4227 	 */
4228 	old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4229 	pte = pte_modify(old_pte, vma->vm_page_prot);
4230 	pte = pte_mkyoung(pte);
4231 	if (was_writable)
4232 		pte = pte_mkwrite(pte);
4233 	ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4234 	update_mmu_cache(vma, vmf->address, vmf->pte);
4235 	pte_unmap_unlock(vmf->pte, vmf->ptl);
4236 	goto out;
4237 }
4238 
4239 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4240 {
4241 	if (vma_is_anonymous(vmf->vma))
4242 		return do_huge_pmd_anonymous_page(vmf);
4243 	if (vmf->vma->vm_ops->huge_fault)
4244 		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4245 	return VM_FAULT_FALLBACK;
4246 }
4247 
4248 /* `inline' is required to avoid gcc 4.1.2 build error */
4249 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4250 {
4251 	if (vma_is_anonymous(vmf->vma)) {
4252 		if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4253 			return handle_userfault(vmf, VM_UFFD_WP);
4254 		return do_huge_pmd_wp_page(vmf, orig_pmd);
4255 	}
4256 	if (vmf->vma->vm_ops->huge_fault) {
4257 		vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4258 
4259 		if (!(ret & VM_FAULT_FALLBACK))
4260 			return ret;
4261 	}
4262 
4263 	/* COW or write-notify handled on pte level: split pmd. */
4264 	__split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4265 
4266 	return VM_FAULT_FALLBACK;
4267 }
4268 
4269 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4270 {
4271 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) &&			\
4272 	defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4273 	/* No support for anonymous transparent PUD pages yet */
4274 	if (vma_is_anonymous(vmf->vma))
4275 		goto split;
4276 	if (vmf->vma->vm_ops->huge_fault) {
4277 		vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4278 
4279 		if (!(ret & VM_FAULT_FALLBACK))
4280 			return ret;
4281 	}
4282 split:
4283 	/* COW or write-notify not handled on PUD level: split pud.*/
4284 	__split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4285 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4286 	return VM_FAULT_FALLBACK;
4287 }
4288 
4289 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4290 {
4291 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4292 	/* No support for anonymous transparent PUD pages yet */
4293 	if (vma_is_anonymous(vmf->vma))
4294 		return VM_FAULT_FALLBACK;
4295 	if (vmf->vma->vm_ops->huge_fault)
4296 		return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4297 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4298 	return VM_FAULT_FALLBACK;
4299 }
4300 
4301 /*
4302  * These routines also need to handle stuff like marking pages dirty
4303  * and/or accessed for architectures that don't do it in hardware (most
4304  * RISC architectures).  The early dirtying is also good on the i386.
4305  *
4306  * There is also a hook called "update_mmu_cache()" that architectures
4307  * with external mmu caches can use to update those (ie the Sparc or
4308  * PowerPC hashed page tables that act as extended TLBs).
4309  *
4310  * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4311  * concurrent faults).
4312  *
4313  * The mmap_lock may have been released depending on flags and our return value.
4314  * See filemap_fault() and __lock_page_or_retry().
4315  */
4316 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4317 {
4318 	pte_t entry;
4319 
4320 	if (unlikely(pmd_none(*vmf->pmd))) {
4321 		/*
4322 		 * Leave __pte_alloc() until later: because vm_ops->fault may
4323 		 * want to allocate huge page, and if we expose page table
4324 		 * for an instant, it will be difficult to retract from
4325 		 * concurrent faults and from rmap lookups.
4326 		 */
4327 		vmf->pte = NULL;
4328 	} else {
4329 		/*
4330 		 * If a huge pmd materialized under us just retry later.  Use
4331 		 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4332 		 * of pmd_trans_huge() to ensure the pmd didn't become
4333 		 * pmd_trans_huge under us and then back to pmd_none, as a
4334 		 * result of MADV_DONTNEED running immediately after a huge pmd
4335 		 * fault in a different thread of this mm, in turn leading to a
4336 		 * misleading pmd_trans_huge() retval. All we have to ensure is
4337 		 * that it is a regular pmd that we can walk with
4338 		 * pte_offset_map() and we can do that through an atomic read
4339 		 * in C, which is what pmd_trans_unstable() provides.
4340 		 */
4341 		if (pmd_devmap_trans_unstable(vmf->pmd))
4342 			return 0;
4343 		/*
4344 		 * A regular pmd is established and it can't morph into a huge
4345 		 * pmd from under us anymore at this point because we hold the
4346 		 * mmap_lock read mode and khugepaged takes it in write mode.
4347 		 * So now it's safe to run pte_offset_map().
4348 		 */
4349 		vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4350 		vmf->orig_pte = *vmf->pte;
4351 
4352 		/*
4353 		 * some architectures can have larger ptes than wordsize,
4354 		 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4355 		 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4356 		 * accesses.  The code below just needs a consistent view
4357 		 * for the ifs and we later double check anyway with the
4358 		 * ptl lock held. So here a barrier will do.
4359 		 */
4360 		barrier();
4361 		if (pte_none(vmf->orig_pte)) {
4362 			pte_unmap(vmf->pte);
4363 			vmf->pte = NULL;
4364 		}
4365 	}
4366 
4367 	if (!vmf->pte) {
4368 		if (vma_is_anonymous(vmf->vma))
4369 			return do_anonymous_page(vmf);
4370 		else
4371 			return do_fault(vmf);
4372 	}
4373 
4374 	if (!pte_present(vmf->orig_pte))
4375 		return do_swap_page(vmf);
4376 
4377 	if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4378 		return do_numa_page(vmf);
4379 
4380 	vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4381 	spin_lock(vmf->ptl);
4382 	entry = vmf->orig_pte;
4383 	if (unlikely(!pte_same(*vmf->pte, entry))) {
4384 		update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4385 		goto unlock;
4386 	}
4387 	if (vmf->flags & FAULT_FLAG_WRITE) {
4388 		if (!pte_write(entry))
4389 			return do_wp_page(vmf);
4390 		entry = pte_mkdirty(entry);
4391 	}
4392 	entry = pte_mkyoung(entry);
4393 	if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4394 				vmf->flags & FAULT_FLAG_WRITE)) {
4395 		update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4396 	} else {
4397 		/* Skip spurious TLB flush for retried page fault */
4398 		if (vmf->flags & FAULT_FLAG_TRIED)
4399 			goto unlock;
4400 		/*
4401 		 * This is needed only for protection faults but the arch code
4402 		 * is not yet telling us if this is a protection fault or not.
4403 		 * This still avoids useless tlb flushes for .text page faults
4404 		 * with threads.
4405 		 */
4406 		if (vmf->flags & FAULT_FLAG_WRITE)
4407 			flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4408 	}
4409 unlock:
4410 	pte_unmap_unlock(vmf->pte, vmf->ptl);
4411 	return 0;
4412 }
4413 
4414 /*
4415  * By the time we get here, we already hold the mm semaphore
4416  *
4417  * The mmap_lock may have been released depending on flags and our
4418  * return value.  See filemap_fault() and __lock_page_or_retry().
4419  */
4420 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4421 		unsigned long address, unsigned int flags)
4422 {
4423 	struct vm_fault vmf = {
4424 		.vma = vma,
4425 		.address = address & PAGE_MASK,
4426 		.flags = flags,
4427 		.pgoff = linear_page_index(vma, address),
4428 		.gfp_mask = __get_fault_gfp_mask(vma),
4429 	};
4430 	unsigned int dirty = flags & FAULT_FLAG_WRITE;
4431 	struct mm_struct *mm = vma->vm_mm;
4432 	pgd_t *pgd;
4433 	p4d_t *p4d;
4434 	vm_fault_t ret;
4435 
4436 	pgd = pgd_offset(mm, address);
4437 	p4d = p4d_alloc(mm, pgd, address);
4438 	if (!p4d)
4439 		return VM_FAULT_OOM;
4440 
4441 	vmf.pud = pud_alloc(mm, p4d, address);
4442 	if (!vmf.pud)
4443 		return VM_FAULT_OOM;
4444 retry_pud:
4445 	if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4446 		ret = create_huge_pud(&vmf);
4447 		if (!(ret & VM_FAULT_FALLBACK))
4448 			return ret;
4449 	} else {
4450 		pud_t orig_pud = *vmf.pud;
4451 
4452 		barrier();
4453 		if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4454 
4455 			/* NUMA case for anonymous PUDs would go here */
4456 
4457 			if (dirty && !pud_write(orig_pud)) {
4458 				ret = wp_huge_pud(&vmf, orig_pud);
4459 				if (!(ret & VM_FAULT_FALLBACK))
4460 					return ret;
4461 			} else {
4462 				huge_pud_set_accessed(&vmf, orig_pud);
4463 				return 0;
4464 			}
4465 		}
4466 	}
4467 
4468 	vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4469 	if (!vmf.pmd)
4470 		return VM_FAULT_OOM;
4471 
4472 	/* Huge pud page fault raced with pmd_alloc? */
4473 	if (pud_trans_unstable(vmf.pud))
4474 		goto retry_pud;
4475 
4476 	if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4477 		ret = create_huge_pmd(&vmf);
4478 		if (!(ret & VM_FAULT_FALLBACK))
4479 			return ret;
4480 	} else {
4481 		pmd_t orig_pmd = *vmf.pmd;
4482 
4483 		barrier();
4484 		if (unlikely(is_swap_pmd(orig_pmd))) {
4485 			VM_BUG_ON(thp_migration_supported() &&
4486 					  !is_pmd_migration_entry(orig_pmd));
4487 			if (is_pmd_migration_entry(orig_pmd))
4488 				pmd_migration_entry_wait(mm, vmf.pmd);
4489 			return 0;
4490 		}
4491 		if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4492 			if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4493 				return do_huge_pmd_numa_page(&vmf, orig_pmd);
4494 
4495 			if (dirty && !pmd_write(orig_pmd)) {
4496 				ret = wp_huge_pmd(&vmf, orig_pmd);
4497 				if (!(ret & VM_FAULT_FALLBACK))
4498 					return ret;
4499 			} else {
4500 				huge_pmd_set_accessed(&vmf, orig_pmd);
4501 				return 0;
4502 			}
4503 		}
4504 	}
4505 
4506 	return handle_pte_fault(&vmf);
4507 }
4508 
4509 /**
4510  * mm_account_fault - Do page fault accounting
4511  *
4512  * @regs: the pt_regs struct pointer.  When set to NULL, will skip accounting
4513  *        of perf event counters, but we'll still do the per-task accounting to
4514  *        the task who triggered this page fault.
4515  * @address: the faulted address.
4516  * @flags: the fault flags.
4517  * @ret: the fault retcode.
4518  *
4519  * This will take care of most of the page fault accounting.  Meanwhile, it
4520  * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4521  * updates.  However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4522  * still be in per-arch page fault handlers at the entry of page fault.
4523  */
4524 static inline void mm_account_fault(struct pt_regs *regs,
4525 				    unsigned long address, unsigned int flags,
4526 				    vm_fault_t ret)
4527 {
4528 	bool major;
4529 
4530 	/*
4531 	 * We don't do accounting for some specific faults:
4532 	 *
4533 	 * - Unsuccessful faults (e.g. when the address wasn't valid).  That
4534 	 *   includes arch_vma_access_permitted() failing before reaching here.
4535 	 *   So this is not a "this many hardware page faults" counter.  We
4536 	 *   should use the hw profiling for that.
4537 	 *
4538 	 * - Incomplete faults (VM_FAULT_RETRY).  They will only be counted
4539 	 *   once they're completed.
4540 	 */
4541 	if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4542 		return;
4543 
4544 	/*
4545 	 * We define the fault as a major fault when the final successful fault
4546 	 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4547 	 * handle it immediately previously).
4548 	 */
4549 	major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4550 
4551 	if (major)
4552 		current->maj_flt++;
4553 	else
4554 		current->min_flt++;
4555 
4556 	/*
4557 	 * If the fault is done for GUP, regs will be NULL.  We only do the
4558 	 * accounting for the per thread fault counters who triggered the
4559 	 * fault, and we skip the perf event updates.
4560 	 */
4561 	if (!regs)
4562 		return;
4563 
4564 	if (major)
4565 		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4566 	else
4567 		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4568 }
4569 
4570 /*
4571  * By the time we get here, we already hold the mm semaphore
4572  *
4573  * The mmap_lock may have been released depending on flags and our
4574  * return value.  See filemap_fault() and __lock_page_or_retry().
4575  */
4576 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4577 			   unsigned int flags, struct pt_regs *regs)
4578 {
4579 	vm_fault_t ret;
4580 
4581 	__set_current_state(TASK_RUNNING);
4582 
4583 	count_vm_event(PGFAULT);
4584 	count_memcg_event_mm(vma->vm_mm, PGFAULT);
4585 
4586 	/* do counter updates before entering really critical section. */
4587 	check_sync_rss_stat(current);
4588 
4589 	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4590 					    flags & FAULT_FLAG_INSTRUCTION,
4591 					    flags & FAULT_FLAG_REMOTE))
4592 		return VM_FAULT_SIGSEGV;
4593 
4594 	/*
4595 	 * Enable the memcg OOM handling for faults triggered in user
4596 	 * space.  Kernel faults are handled more gracefully.
4597 	 */
4598 	if (flags & FAULT_FLAG_USER)
4599 		mem_cgroup_enter_user_fault();
4600 
4601 	if (unlikely(is_vm_hugetlb_page(vma)))
4602 		ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4603 	else
4604 		ret = __handle_mm_fault(vma, address, flags);
4605 
4606 	if (flags & FAULT_FLAG_USER) {
4607 		mem_cgroup_exit_user_fault();
4608 		/*
4609 		 * The task may have entered a memcg OOM situation but
4610 		 * if the allocation error was handled gracefully (no
4611 		 * VM_FAULT_OOM), there is no need to kill anything.
4612 		 * Just clean up the OOM state peacefully.
4613 		 */
4614 		if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4615 			mem_cgroup_oom_synchronize(false);
4616 	}
4617 
4618 	mm_account_fault(regs, address, flags, ret);
4619 
4620 	return ret;
4621 }
4622 EXPORT_SYMBOL_GPL(handle_mm_fault);
4623 
4624 #ifndef __PAGETABLE_P4D_FOLDED
4625 /*
4626  * Allocate p4d page table.
4627  * We've already handled the fast-path in-line.
4628  */
4629 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4630 {
4631 	p4d_t *new = p4d_alloc_one(mm, address);
4632 	if (!new)
4633 		return -ENOMEM;
4634 
4635 	smp_wmb(); /* See comment in __pte_alloc */
4636 
4637 	spin_lock(&mm->page_table_lock);
4638 	if (pgd_present(*pgd))		/* Another has populated it */
4639 		p4d_free(mm, new);
4640 	else
4641 		pgd_populate(mm, pgd, new);
4642 	spin_unlock(&mm->page_table_lock);
4643 	return 0;
4644 }
4645 #endif /* __PAGETABLE_P4D_FOLDED */
4646 
4647 #ifndef __PAGETABLE_PUD_FOLDED
4648 /*
4649  * Allocate page upper directory.
4650  * We've already handled the fast-path in-line.
4651  */
4652 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4653 {
4654 	pud_t *new = pud_alloc_one(mm, address);
4655 	if (!new)
4656 		return -ENOMEM;
4657 
4658 	smp_wmb(); /* See comment in __pte_alloc */
4659 
4660 	spin_lock(&mm->page_table_lock);
4661 	if (!p4d_present(*p4d)) {
4662 		mm_inc_nr_puds(mm);
4663 		p4d_populate(mm, p4d, new);
4664 	} else	/* Another has populated it */
4665 		pud_free(mm, new);
4666 	spin_unlock(&mm->page_table_lock);
4667 	return 0;
4668 }
4669 #endif /* __PAGETABLE_PUD_FOLDED */
4670 
4671 #ifndef __PAGETABLE_PMD_FOLDED
4672 /*
4673  * Allocate page middle directory.
4674  * We've already handled the fast-path in-line.
4675  */
4676 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4677 {
4678 	spinlock_t *ptl;
4679 	pmd_t *new = pmd_alloc_one(mm, address);
4680 	if (!new)
4681 		return -ENOMEM;
4682 
4683 	smp_wmb(); /* See comment in __pte_alloc */
4684 
4685 	ptl = pud_lock(mm, pud);
4686 	if (!pud_present(*pud)) {
4687 		mm_inc_nr_pmds(mm);
4688 		pud_populate(mm, pud, new);
4689 	} else	/* Another has populated it */
4690 		pmd_free(mm, new);
4691 	spin_unlock(ptl);
4692 	return 0;
4693 }
4694 #endif /* __PAGETABLE_PMD_FOLDED */
4695 
4696 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4697 			  struct mmu_notifier_range *range, pte_t **ptepp,
4698 			  pmd_t **pmdpp, spinlock_t **ptlp)
4699 {
4700 	pgd_t *pgd;
4701 	p4d_t *p4d;
4702 	pud_t *pud;
4703 	pmd_t *pmd;
4704 	pte_t *ptep;
4705 
4706 	pgd = pgd_offset(mm, address);
4707 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4708 		goto out;
4709 
4710 	p4d = p4d_offset(pgd, address);
4711 	if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4712 		goto out;
4713 
4714 	pud = pud_offset(p4d, address);
4715 	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4716 		goto out;
4717 
4718 	pmd = pmd_offset(pud, address);
4719 	VM_BUG_ON(pmd_trans_huge(*pmd));
4720 
4721 	if (pmd_huge(*pmd)) {
4722 		if (!pmdpp)
4723 			goto out;
4724 
4725 		if (range) {
4726 			mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4727 						NULL, mm, address & PMD_MASK,
4728 						(address & PMD_MASK) + PMD_SIZE);
4729 			mmu_notifier_invalidate_range_start(range);
4730 		}
4731 		*ptlp = pmd_lock(mm, pmd);
4732 		if (pmd_huge(*pmd)) {
4733 			*pmdpp = pmd;
4734 			return 0;
4735 		}
4736 		spin_unlock(*ptlp);
4737 		if (range)
4738 			mmu_notifier_invalidate_range_end(range);
4739 	}
4740 
4741 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4742 		goto out;
4743 
4744 	if (range) {
4745 		mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4746 					address & PAGE_MASK,
4747 					(address & PAGE_MASK) + PAGE_SIZE);
4748 		mmu_notifier_invalidate_range_start(range);
4749 	}
4750 	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4751 	if (!pte_present(*ptep))
4752 		goto unlock;
4753 	*ptepp = ptep;
4754 	return 0;
4755 unlock:
4756 	pte_unmap_unlock(ptep, *ptlp);
4757 	if (range)
4758 		mmu_notifier_invalidate_range_end(range);
4759 out:
4760 	return -EINVAL;
4761 }
4762 
4763 /**
4764  * follow_pte - look up PTE at a user virtual address
4765  * @mm: the mm_struct of the target address space
4766  * @address: user virtual address
4767  * @ptepp: location to store found PTE
4768  * @ptlp: location to store the lock for the PTE
4769  *
4770  * On a successful return, the pointer to the PTE is stored in @ptepp;
4771  * the corresponding lock is taken and its location is stored in @ptlp.
4772  * The contents of the PTE are only stable until @ptlp is released;
4773  * any further use, if any, must be protected against invalidation
4774  * with MMU notifiers.
4775  *
4776  * Only IO mappings and raw PFN mappings are allowed.  The mmap semaphore
4777  * should be taken for read.
4778  *
4779  * KVM uses this function.  While it is arguably less bad than ``follow_pfn``,
4780  * it is not a good general-purpose API.
4781  *
4782  * Return: zero on success, -ve otherwise.
4783  */
4784 int follow_pte(struct mm_struct *mm, unsigned long address,
4785 	       pte_t **ptepp, spinlock_t **ptlp)
4786 {
4787 	return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4788 }
4789 EXPORT_SYMBOL_GPL(follow_pte);
4790 
4791 /**
4792  * follow_pfn - look up PFN at a user virtual address
4793  * @vma: memory mapping
4794  * @address: user virtual address
4795  * @pfn: location to store found PFN
4796  *
4797  * Only IO mappings and raw PFN mappings are allowed.
4798  *
4799  * This function does not allow the caller to read the permissions
4800  * of the PTE.  Do not use it.
4801  *
4802  * Return: zero and the pfn at @pfn on success, -ve otherwise.
4803  */
4804 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4805 	unsigned long *pfn)
4806 {
4807 	int ret = -EINVAL;
4808 	spinlock_t *ptl;
4809 	pte_t *ptep;
4810 
4811 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4812 		return ret;
4813 
4814 	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4815 	if (ret)
4816 		return ret;
4817 	*pfn = pte_pfn(*ptep);
4818 	pte_unmap_unlock(ptep, ptl);
4819 	return 0;
4820 }
4821 EXPORT_SYMBOL(follow_pfn);
4822 
4823 #ifdef CONFIG_HAVE_IOREMAP_PROT
4824 int follow_phys(struct vm_area_struct *vma,
4825 		unsigned long address, unsigned int flags,
4826 		unsigned long *prot, resource_size_t *phys)
4827 {
4828 	int ret = -EINVAL;
4829 	pte_t *ptep, pte;
4830 	spinlock_t *ptl;
4831 
4832 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4833 		goto out;
4834 
4835 	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4836 		goto out;
4837 	pte = *ptep;
4838 
4839 	if ((flags & FOLL_WRITE) && !pte_write(pte))
4840 		goto unlock;
4841 
4842 	*prot = pgprot_val(pte_pgprot(pte));
4843 	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4844 
4845 	ret = 0;
4846 unlock:
4847 	pte_unmap_unlock(ptep, ptl);
4848 out:
4849 	return ret;
4850 }
4851 
4852 /**
4853  * generic_access_phys - generic implementation for iomem mmap access
4854  * @vma: the vma to access
4855  * @addr: userspace address, not relative offset within @vma
4856  * @buf: buffer to read/write
4857  * @len: length of transfer
4858  * @write: set to FOLL_WRITE when writing, otherwise reading
4859  *
4860  * This is a generic implementation for &vm_operations_struct.access for an
4861  * iomem mapping. This callback is used by access_process_vm() when the @vma is
4862  * not page based.
4863  */
4864 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4865 			void *buf, int len, int write)
4866 {
4867 	resource_size_t phys_addr;
4868 	unsigned long prot = 0;
4869 	void __iomem *maddr;
4870 	pte_t *ptep, pte;
4871 	spinlock_t *ptl;
4872 	int offset = offset_in_page(addr);
4873 	int ret = -EINVAL;
4874 
4875 	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4876 		return -EINVAL;
4877 
4878 retry:
4879 	if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
4880 		return -EINVAL;
4881 	pte = *ptep;
4882 	pte_unmap_unlock(ptep, ptl);
4883 
4884 	prot = pgprot_val(pte_pgprot(pte));
4885 	phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4886 
4887 	if ((write & FOLL_WRITE) && !pte_write(pte))
4888 		return -EINVAL;
4889 
4890 	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4891 	if (!maddr)
4892 		return -ENOMEM;
4893 
4894 	if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
4895 		goto out_unmap;
4896 
4897 	if (!pte_same(pte, *ptep)) {
4898 		pte_unmap_unlock(ptep, ptl);
4899 		iounmap(maddr);
4900 
4901 		goto retry;
4902 	}
4903 
4904 	if (write)
4905 		memcpy_toio(maddr + offset, buf, len);
4906 	else
4907 		memcpy_fromio(buf, maddr + offset, len);
4908 	ret = len;
4909 	pte_unmap_unlock(ptep, ptl);
4910 out_unmap:
4911 	iounmap(maddr);
4912 
4913 	return ret;
4914 }
4915 EXPORT_SYMBOL_GPL(generic_access_phys);
4916 #endif
4917 
4918 /*
4919  * Access another process' address space as given in mm.
4920  */
4921 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
4922 		       int len, unsigned int gup_flags)
4923 {
4924 	struct vm_area_struct *vma;
4925 	void *old_buf = buf;
4926 	int write = gup_flags & FOLL_WRITE;
4927 
4928 	if (mmap_read_lock_killable(mm))
4929 		return 0;
4930 
4931 	/* ignore errors, just check how much was successfully transferred */
4932 	while (len) {
4933 		int bytes, ret, offset;
4934 		void *maddr;
4935 		struct page *page = NULL;
4936 
4937 		ret = get_user_pages_remote(mm, addr, 1,
4938 				gup_flags, &page, &vma, NULL);
4939 		if (ret <= 0) {
4940 #ifndef CONFIG_HAVE_IOREMAP_PROT
4941 			break;
4942 #else
4943 			/*
4944 			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4945 			 * we can access using slightly different code.
4946 			 */
4947 			vma = find_vma(mm, addr);
4948 			if (!vma || vma->vm_start > addr)
4949 				break;
4950 			if (vma->vm_ops && vma->vm_ops->access)
4951 				ret = vma->vm_ops->access(vma, addr, buf,
4952 							  len, write);
4953 			if (ret <= 0)
4954 				break;
4955 			bytes = ret;
4956 #endif
4957 		} else {
4958 			bytes = len;
4959 			offset = addr & (PAGE_SIZE-1);
4960 			if (bytes > PAGE_SIZE-offset)
4961 				bytes = PAGE_SIZE-offset;
4962 
4963 			maddr = kmap(page);
4964 			if (write) {
4965 				copy_to_user_page(vma, page, addr,
4966 						  maddr + offset, buf, bytes);
4967 				set_page_dirty_lock(page);
4968 			} else {
4969 				copy_from_user_page(vma, page, addr,
4970 						    buf, maddr + offset, bytes);
4971 			}
4972 			kunmap(page);
4973 			put_page(page);
4974 		}
4975 		len -= bytes;
4976 		buf += bytes;
4977 		addr += bytes;
4978 	}
4979 	mmap_read_unlock(mm);
4980 
4981 	return buf - old_buf;
4982 }
4983 
4984 /**
4985  * access_remote_vm - access another process' address space
4986  * @mm:		the mm_struct of the target address space
4987  * @addr:	start address to access
4988  * @buf:	source or destination buffer
4989  * @len:	number of bytes to transfer
4990  * @gup_flags:	flags modifying lookup behaviour
4991  *
4992  * The caller must hold a reference on @mm.
4993  *
4994  * Return: number of bytes copied from source to destination.
4995  */
4996 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4997 		void *buf, int len, unsigned int gup_flags)
4998 {
4999 	return __access_remote_vm(mm, addr, buf, len, gup_flags);
5000 }
5001 
5002 /*
5003  * Access another process' address space.
5004  * Source/target buffer must be kernel space,
5005  * Do not walk the page table directly, use get_user_pages
5006  */
5007 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5008 		void *buf, int len, unsigned int gup_flags)
5009 {
5010 	struct mm_struct *mm;
5011 	int ret;
5012 
5013 	mm = get_task_mm(tsk);
5014 	if (!mm)
5015 		return 0;
5016 
5017 	ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5018 
5019 	mmput(mm);
5020 
5021 	return ret;
5022 }
5023 EXPORT_SYMBOL_GPL(access_process_vm);
5024 
5025 /*
5026  * Print the name of a VMA.
5027  */
5028 void print_vma_addr(char *prefix, unsigned long ip)
5029 {
5030 	struct mm_struct *mm = current->mm;
5031 	struct vm_area_struct *vma;
5032 
5033 	/*
5034 	 * we might be running from an atomic context so we cannot sleep
5035 	 */
5036 	if (!mmap_read_trylock(mm))
5037 		return;
5038 
5039 	vma = find_vma(mm, ip);
5040 	if (vma && vma->vm_file) {
5041 		struct file *f = vma->vm_file;
5042 		char *buf = (char *)__get_free_page(GFP_NOWAIT);
5043 		if (buf) {
5044 			char *p;
5045 
5046 			p = file_path(f, buf, PAGE_SIZE);
5047 			if (IS_ERR(p))
5048 				p = "?";
5049 			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5050 					vma->vm_start,
5051 					vma->vm_end - vma->vm_start);
5052 			free_page((unsigned long)buf);
5053 		}
5054 	}
5055 	mmap_read_unlock(mm);
5056 }
5057 
5058 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5059 void __might_fault(const char *file, int line)
5060 {
5061 	/*
5062 	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5063 	 * holding the mmap_lock, this is safe because kernel memory doesn't
5064 	 * get paged out, therefore we'll never actually fault, and the
5065 	 * below annotations will generate false positives.
5066 	 */
5067 	if (uaccess_kernel())
5068 		return;
5069 	if (pagefault_disabled())
5070 		return;
5071 	__might_sleep(file, line, 0);
5072 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5073 	if (current->mm)
5074 		might_lock_read(&current->mm->mmap_lock);
5075 #endif
5076 }
5077 EXPORT_SYMBOL(__might_fault);
5078 #endif
5079 
5080 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5081 /*
5082  * Process all subpages of the specified huge page with the specified
5083  * operation.  The target subpage will be processed last to keep its
5084  * cache lines hot.
5085  */
5086 static inline void process_huge_page(
5087 	unsigned long addr_hint, unsigned int pages_per_huge_page,
5088 	void (*process_subpage)(unsigned long addr, int idx, void *arg),
5089 	void *arg)
5090 {
5091 	int i, n, base, l;
5092 	unsigned long addr = addr_hint &
5093 		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5094 
5095 	/* Process target subpage last to keep its cache lines hot */
5096 	might_sleep();
5097 	n = (addr_hint - addr) / PAGE_SIZE;
5098 	if (2 * n <= pages_per_huge_page) {
5099 		/* If target subpage in first half of huge page */
5100 		base = 0;
5101 		l = n;
5102 		/* Process subpages at the end of huge page */
5103 		for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5104 			cond_resched();
5105 			process_subpage(addr + i * PAGE_SIZE, i, arg);
5106 		}
5107 	} else {
5108 		/* If target subpage in second half of huge page */
5109 		base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5110 		l = pages_per_huge_page - n;
5111 		/* Process subpages at the begin of huge page */
5112 		for (i = 0; i < base; i++) {
5113 			cond_resched();
5114 			process_subpage(addr + i * PAGE_SIZE, i, arg);
5115 		}
5116 	}
5117 	/*
5118 	 * Process remaining subpages in left-right-left-right pattern
5119 	 * towards the target subpage
5120 	 */
5121 	for (i = 0; i < l; i++) {
5122 		int left_idx = base + i;
5123 		int right_idx = base + 2 * l - 1 - i;
5124 
5125 		cond_resched();
5126 		process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5127 		cond_resched();
5128 		process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5129 	}
5130 }
5131 
5132 static void clear_gigantic_page(struct page *page,
5133 				unsigned long addr,
5134 				unsigned int pages_per_huge_page)
5135 {
5136 	int i;
5137 	struct page *p = page;
5138 
5139 	might_sleep();
5140 	for (i = 0; i < pages_per_huge_page;
5141 	     i++, p = mem_map_next(p, page, i)) {
5142 		cond_resched();
5143 		clear_user_highpage(p, addr + i * PAGE_SIZE);
5144 	}
5145 }
5146 
5147 static void clear_subpage(unsigned long addr, int idx, void *arg)
5148 {
5149 	struct page *page = arg;
5150 
5151 	clear_user_highpage(page + idx, addr);
5152 }
5153 
5154 void clear_huge_page(struct page *page,
5155 		     unsigned long addr_hint, unsigned int pages_per_huge_page)
5156 {
5157 	unsigned long addr = addr_hint &
5158 		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5159 
5160 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5161 		clear_gigantic_page(page, addr, pages_per_huge_page);
5162 		return;
5163 	}
5164 
5165 	process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5166 }
5167 
5168 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5169 				    unsigned long addr,
5170 				    struct vm_area_struct *vma,
5171 				    unsigned int pages_per_huge_page)
5172 {
5173 	int i;
5174 	struct page *dst_base = dst;
5175 	struct page *src_base = src;
5176 
5177 	for (i = 0; i < pages_per_huge_page; ) {
5178 		cond_resched();
5179 		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5180 
5181 		i++;
5182 		dst = mem_map_next(dst, dst_base, i);
5183 		src = mem_map_next(src, src_base, i);
5184 	}
5185 }
5186 
5187 struct copy_subpage_arg {
5188 	struct page *dst;
5189 	struct page *src;
5190 	struct vm_area_struct *vma;
5191 };
5192 
5193 static void copy_subpage(unsigned long addr, int idx, void *arg)
5194 {
5195 	struct copy_subpage_arg *copy_arg = arg;
5196 
5197 	copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5198 			   addr, copy_arg->vma);
5199 }
5200 
5201 void copy_user_huge_page(struct page *dst, struct page *src,
5202 			 unsigned long addr_hint, struct vm_area_struct *vma,
5203 			 unsigned int pages_per_huge_page)
5204 {
5205 	unsigned long addr = addr_hint &
5206 		~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5207 	struct copy_subpage_arg arg = {
5208 		.dst = dst,
5209 		.src = src,
5210 		.vma = vma,
5211 	};
5212 
5213 	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5214 		copy_user_gigantic_page(dst, src, addr, vma,
5215 					pages_per_huge_page);
5216 		return;
5217 	}
5218 
5219 	process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5220 }
5221 
5222 long copy_huge_page_from_user(struct page *dst_page,
5223 				const void __user *usr_src,
5224 				unsigned int pages_per_huge_page,
5225 				bool allow_pagefault)
5226 {
5227 	void *src = (void *)usr_src;
5228 	void *page_kaddr;
5229 	unsigned long i, rc = 0;
5230 	unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5231 	struct page *subpage = dst_page;
5232 
5233 	for (i = 0; i < pages_per_huge_page;
5234 	     i++, subpage = mem_map_next(subpage, dst_page, i)) {
5235 		if (allow_pagefault)
5236 			page_kaddr = kmap(subpage);
5237 		else
5238 			page_kaddr = kmap_atomic(subpage);
5239 		rc = copy_from_user(page_kaddr,
5240 				(const void __user *)(src + i * PAGE_SIZE),
5241 				PAGE_SIZE);
5242 		if (allow_pagefault)
5243 			kunmap(subpage);
5244 		else
5245 			kunmap_atomic(page_kaddr);
5246 
5247 		ret_val -= (PAGE_SIZE - rc);
5248 		if (rc)
5249 			break;
5250 
5251 		cond_resched();
5252 	}
5253 	return ret_val;
5254 }
5255 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5256 
5257 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5258 
5259 static struct kmem_cache *page_ptl_cachep;
5260 
5261 void __init ptlock_cache_init(void)
5262 {
5263 	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5264 			SLAB_PANIC, NULL);
5265 }
5266 
5267 bool ptlock_alloc(struct page *page)
5268 {
5269 	spinlock_t *ptl;
5270 
5271 	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5272 	if (!ptl)
5273 		return false;
5274 	page->ptl = ptl;
5275 	return true;
5276 }
5277 
5278 void ptlock_free(struct page *page)
5279 {
5280 	kmem_cache_free(page_ptl_cachep, page->ptl);
5281 }
5282 #endif
5283