Loading...
1/*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/mm.h>
11#include <linux/sched.h>
12#include <linux/sched/coredump.h>
13#include <linux/sched/numa_balancing.h>
14#include <linux/highmem.h>
15#include <linux/hugetlb.h>
16#include <linux/mmu_notifier.h>
17#include <linux/rmap.h>
18#include <linux/swap.h>
19#include <linux/shrinker.h>
20#include <linux/mm_inline.h>
21#include <linux/swapops.h>
22#include <linux/dax.h>
23#include <linux/khugepaged.h>
24#include <linux/freezer.h>
25#include <linux/pfn_t.h>
26#include <linux/mman.h>
27#include <linux/memremap.h>
28#include <linux/pagemap.h>
29#include <linux/debugfs.h>
30#include <linux/migrate.h>
31#include <linux/hashtable.h>
32#include <linux/userfaultfd_k.h>
33#include <linux/page_idle.h>
34#include <linux/shmem_fs.h>
35#include <linux/oom.h>
36
37#include <asm/tlb.h>
38#include <asm/pgalloc.h>
39#include "internal.h"
40
41/*
42 * By default, transparent hugepage support is disabled in order to avoid
43 * risking an increased memory footprint for applications that are not
44 * guaranteed to benefit from it. When transparent hugepage support is
45 * enabled, it is for all mappings, and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
48 */
49unsigned long transparent_hugepage_flags __read_mostly =
50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
52#endif
53#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
55#endif
56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
59
60static struct shrinker deferred_split_shrinker;
61
62static atomic_t huge_zero_refcount;
63struct page *huge_zero_page __read_mostly;
64
65static struct page *get_huge_zero_page(void)
66{
67 struct page *zero_page;
68retry:
69 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
70 return READ_ONCE(huge_zero_page);
71
72 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
73 HPAGE_PMD_ORDER);
74 if (!zero_page) {
75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
76 return NULL;
77 }
78 count_vm_event(THP_ZERO_PAGE_ALLOC);
79 preempt_disable();
80 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
81 preempt_enable();
82 __free_pages(zero_page, compound_order(zero_page));
83 goto retry;
84 }
85
86 /* We take additional reference here. It will be put back by shrinker */
87 atomic_set(&huge_zero_refcount, 2);
88 preempt_enable();
89 return READ_ONCE(huge_zero_page);
90}
91
92static void put_huge_zero_page(void)
93{
94 /*
95 * Counter should never go to zero here. Only shrinker can put
96 * last reference.
97 */
98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
99}
100
101struct page *mm_get_huge_zero_page(struct mm_struct *mm)
102{
103 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
104 return READ_ONCE(huge_zero_page);
105
106 if (!get_huge_zero_page())
107 return NULL;
108
109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
110 put_huge_zero_page();
111
112 return READ_ONCE(huge_zero_page);
113}
114
115void mm_put_huge_zero_page(struct mm_struct *mm)
116{
117 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
118 put_huge_zero_page();
119}
120
121static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
122 struct shrink_control *sc)
123{
124 /* we can free zero page only if last reference remains */
125 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
126}
127
128static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
129 struct shrink_control *sc)
130{
131 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
132 struct page *zero_page = xchg(&huge_zero_page, NULL);
133 BUG_ON(zero_page == NULL);
134 __free_pages(zero_page, compound_order(zero_page));
135 return HPAGE_PMD_NR;
136 }
137
138 return 0;
139}
140
141static struct shrinker huge_zero_page_shrinker = {
142 .count_objects = shrink_huge_zero_page_count,
143 .scan_objects = shrink_huge_zero_page_scan,
144 .seeks = DEFAULT_SEEKS,
145};
146
147#ifdef CONFIG_SYSFS
148static ssize_t enabled_show(struct kobject *kobj,
149 struct kobj_attribute *attr, char *buf)
150{
151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
152 return sprintf(buf, "[always] madvise never\n");
153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
154 return sprintf(buf, "always [madvise] never\n");
155 else
156 return sprintf(buf, "always madvise [never]\n");
157}
158
159static ssize_t enabled_store(struct kobject *kobj,
160 struct kobj_attribute *attr,
161 const char *buf, size_t count)
162{
163 ssize_t ret = count;
164
165 if (!memcmp("always", buf,
166 min(sizeof("always")-1, count))) {
167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
168 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
169 } else if (!memcmp("madvise", buf,
170 min(sizeof("madvise")-1, count))) {
171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
173 } else if (!memcmp("never", buf,
174 min(sizeof("never")-1, count))) {
175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
177 } else
178 ret = -EINVAL;
179
180 if (ret > 0) {
181 int err = start_stop_khugepaged();
182 if (err)
183 ret = err;
184 }
185 return ret;
186}
187static struct kobj_attribute enabled_attr =
188 __ATTR(enabled, 0644, enabled_show, enabled_store);
189
190ssize_t single_hugepage_flag_show(struct kobject *kobj,
191 struct kobj_attribute *attr, char *buf,
192 enum transparent_hugepage_flag flag)
193{
194 return sprintf(buf, "%d\n",
195 !!test_bit(flag, &transparent_hugepage_flags));
196}
197
198ssize_t single_hugepage_flag_store(struct kobject *kobj,
199 struct kobj_attribute *attr,
200 const char *buf, size_t count,
201 enum transparent_hugepage_flag flag)
202{
203 unsigned long value;
204 int ret;
205
206 ret = kstrtoul(buf, 10, &value);
207 if (ret < 0)
208 return ret;
209 if (value > 1)
210 return -EINVAL;
211
212 if (value)
213 set_bit(flag, &transparent_hugepage_flags);
214 else
215 clear_bit(flag, &transparent_hugepage_flags);
216
217 return count;
218}
219
220static ssize_t defrag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf)
222{
223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
224 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
226 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
228 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
230 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
231 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
232}
233
234static ssize_t defrag_store(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 const char *buf, size_t count)
237{
238 if (!memcmp("always", buf,
239 min(sizeof("always")-1, count))) {
240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
244 } else if (!memcmp("defer+madvise", buf,
245 min(sizeof("defer+madvise")-1, count))) {
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
250 } else if (!memcmp("defer", buf,
251 min(sizeof("defer")-1, count))) {
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
256 } else if (!memcmp("madvise", buf,
257 min(sizeof("madvise")-1, count))) {
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
262 } else if (!memcmp("never", buf,
263 min(sizeof("never")-1, count))) {
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
268 } else
269 return -EINVAL;
270
271 return count;
272}
273static struct kobj_attribute defrag_attr =
274 __ATTR(defrag, 0644, defrag_show, defrag_store);
275
276static ssize_t use_zero_page_show(struct kobject *kobj,
277 struct kobj_attribute *attr, char *buf)
278{
279 return single_hugepage_flag_show(kobj, attr, buf,
280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
281}
282static ssize_t use_zero_page_store(struct kobject *kobj,
283 struct kobj_attribute *attr, const char *buf, size_t count)
284{
285 return single_hugepage_flag_store(kobj, attr, buf, count,
286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
287}
288static struct kobj_attribute use_zero_page_attr =
289 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
290
291static ssize_t hpage_pmd_size_show(struct kobject *kobj,
292 struct kobj_attribute *attr, char *buf)
293{
294 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
295}
296static struct kobj_attribute hpage_pmd_size_attr =
297 __ATTR_RO(hpage_pmd_size);
298
299#ifdef CONFIG_DEBUG_VM
300static ssize_t debug_cow_show(struct kobject *kobj,
301 struct kobj_attribute *attr, char *buf)
302{
303 return single_hugepage_flag_show(kobj, attr, buf,
304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
305}
306static ssize_t debug_cow_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
309{
310 return single_hugepage_flag_store(kobj, attr, buf, count,
311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312}
313static struct kobj_attribute debug_cow_attr =
314 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
315#endif /* CONFIG_DEBUG_VM */
316
317static struct attribute *hugepage_attr[] = {
318 &enabled_attr.attr,
319 &defrag_attr.attr,
320 &use_zero_page_attr.attr,
321 &hpage_pmd_size_attr.attr,
322#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
323 &shmem_enabled_attr.attr,
324#endif
325#ifdef CONFIG_DEBUG_VM
326 &debug_cow_attr.attr,
327#endif
328 NULL,
329};
330
331static const struct attribute_group hugepage_attr_group = {
332 .attrs = hugepage_attr,
333};
334
335static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
336{
337 int err;
338
339 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
340 if (unlikely(!*hugepage_kobj)) {
341 pr_err("failed to create transparent hugepage kobject\n");
342 return -ENOMEM;
343 }
344
345 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
346 if (err) {
347 pr_err("failed to register transparent hugepage group\n");
348 goto delete_obj;
349 }
350
351 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
352 if (err) {
353 pr_err("failed to register transparent hugepage group\n");
354 goto remove_hp_group;
355 }
356
357 return 0;
358
359remove_hp_group:
360 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
361delete_obj:
362 kobject_put(*hugepage_kobj);
363 return err;
364}
365
366static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
367{
368 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
369 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
370 kobject_put(hugepage_kobj);
371}
372#else
373static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
374{
375 return 0;
376}
377
378static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
379{
380}
381#endif /* CONFIG_SYSFS */
382
383static int __init hugepage_init(void)
384{
385 int err;
386 struct kobject *hugepage_kobj;
387
388 if (!has_transparent_hugepage()) {
389 transparent_hugepage_flags = 0;
390 return -EINVAL;
391 }
392
393 /*
394 * hugepages can't be allocated by the buddy allocator
395 */
396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
397 /*
398 * we use page->mapping and page->index in second tail page
399 * as list_head: assuming THP order >= 2
400 */
401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
402
403 err = hugepage_init_sysfs(&hugepage_kobj);
404 if (err)
405 goto err_sysfs;
406
407 err = khugepaged_init();
408 if (err)
409 goto err_slab;
410
411 err = register_shrinker(&huge_zero_page_shrinker);
412 if (err)
413 goto err_hzp_shrinker;
414 err = register_shrinker(&deferred_split_shrinker);
415 if (err)
416 goto err_split_shrinker;
417
418 /*
419 * By default disable transparent hugepages on smaller systems,
420 * where the extra memory used could hurt more than TLB overhead
421 * is likely to save. The admin can still enable it through /sys.
422 */
423 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
424 transparent_hugepage_flags = 0;
425 return 0;
426 }
427
428 err = start_stop_khugepaged();
429 if (err)
430 goto err_khugepaged;
431
432 return 0;
433err_khugepaged:
434 unregister_shrinker(&deferred_split_shrinker);
435err_split_shrinker:
436 unregister_shrinker(&huge_zero_page_shrinker);
437err_hzp_shrinker:
438 khugepaged_destroy();
439err_slab:
440 hugepage_exit_sysfs(hugepage_kobj);
441err_sysfs:
442 return err;
443}
444subsys_initcall(hugepage_init);
445
446static int __init setup_transparent_hugepage(char *str)
447{
448 int ret = 0;
449 if (!str)
450 goto out;
451 if (!strcmp(str, "always")) {
452 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
453 &transparent_hugepage_flags);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455 &transparent_hugepage_flags);
456 ret = 1;
457 } else if (!strcmp(str, "madvise")) {
458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
459 &transparent_hugepage_flags);
460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
461 &transparent_hugepage_flags);
462 ret = 1;
463 } else if (!strcmp(str, "never")) {
464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
465 &transparent_hugepage_flags);
466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
467 &transparent_hugepage_flags);
468 ret = 1;
469 }
470out:
471 if (!ret)
472 pr_warn("transparent_hugepage= cannot parse, ignored\n");
473 return ret;
474}
475__setup("transparent_hugepage=", setup_transparent_hugepage);
476
477pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
478{
479 if (likely(vma->vm_flags & VM_WRITE))
480 pmd = pmd_mkwrite(pmd);
481 return pmd;
482}
483
484static inline struct list_head *page_deferred_list(struct page *page)
485{
486 /*
487 * ->lru in the tail pages is occupied by compound_head.
488 * Let's use ->mapping + ->index in the second tail page as list_head.
489 */
490 return (struct list_head *)&page[2].mapping;
491}
492
493void prep_transhuge_page(struct page *page)
494{
495 /*
496 * we use page->mapping and page->indexlru in second tail page
497 * as list_head: assuming THP order >= 2
498 */
499
500 INIT_LIST_HEAD(page_deferred_list(page));
501 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
502}
503
504unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
505 loff_t off, unsigned long flags, unsigned long size)
506{
507 unsigned long addr;
508 loff_t off_end = off + len;
509 loff_t off_align = round_up(off, size);
510 unsigned long len_pad;
511
512 if (off_end <= off_align || (off_end - off_align) < size)
513 return 0;
514
515 len_pad = len + size;
516 if (len_pad < len || (off + len_pad) < off)
517 return 0;
518
519 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
520 off >> PAGE_SHIFT, flags);
521 if (IS_ERR_VALUE(addr))
522 return 0;
523
524 addr += (off - addr) & (size - 1);
525 return addr;
526}
527
528unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
529 unsigned long len, unsigned long pgoff, unsigned long flags)
530{
531 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
532
533 if (addr)
534 goto out;
535 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
536 goto out;
537
538 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
539 if (addr)
540 return addr;
541
542 out:
543 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
544}
545EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
546
547static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
548 gfp_t gfp)
549{
550 struct vm_area_struct *vma = vmf->vma;
551 struct mem_cgroup *memcg;
552 pgtable_t pgtable;
553 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
554 int ret = 0;
555
556 VM_BUG_ON_PAGE(!PageCompound(page), page);
557
558 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
559 put_page(page);
560 count_vm_event(THP_FAULT_FALLBACK);
561 return VM_FAULT_FALLBACK;
562 }
563
564 pgtable = pte_alloc_one(vma->vm_mm, haddr);
565 if (unlikely(!pgtable)) {
566 ret = VM_FAULT_OOM;
567 goto release;
568 }
569
570 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
571 /*
572 * The memory barrier inside __SetPageUptodate makes sure that
573 * clear_huge_page writes become visible before the set_pmd_at()
574 * write.
575 */
576 __SetPageUptodate(page);
577
578 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
579 if (unlikely(!pmd_none(*vmf->pmd))) {
580 goto unlock_release;
581 } else {
582 pmd_t entry;
583
584 ret = check_stable_address_space(vma->vm_mm);
585 if (ret)
586 goto unlock_release;
587
588 /* Deliver the page fault to userland */
589 if (userfaultfd_missing(vma)) {
590 int ret;
591
592 spin_unlock(vmf->ptl);
593 mem_cgroup_cancel_charge(page, memcg, true);
594 put_page(page);
595 pte_free(vma->vm_mm, pgtable);
596 ret = handle_userfault(vmf, VM_UFFD_MISSING);
597 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
598 return ret;
599 }
600
601 entry = mk_huge_pmd(page, vma->vm_page_prot);
602 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
603 page_add_new_anon_rmap(page, vma, haddr, true);
604 mem_cgroup_commit_charge(page, memcg, false, true);
605 lru_cache_add_active_or_unevictable(page, vma);
606 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
607 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
608 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
609 mm_inc_nr_ptes(vma->vm_mm);
610 spin_unlock(vmf->ptl);
611 count_vm_event(THP_FAULT_ALLOC);
612 }
613
614 return 0;
615unlock_release:
616 spin_unlock(vmf->ptl);
617release:
618 if (pgtable)
619 pte_free(vma->vm_mm, pgtable);
620 mem_cgroup_cancel_charge(page, memcg, true);
621 put_page(page);
622 return ret;
623
624}
625
626/*
627 * always: directly stall for all thp allocations
628 * defer: wake kswapd and fail if not immediately available
629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
630 * fail if not immediately available
631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
632 * available
633 * never: never stall for any thp allocation
634 */
635static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
636{
637 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
638
639 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
640 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
644 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
645 __GFP_KSWAPD_RECLAIM);
646 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
647 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
648 0);
649 return GFP_TRANSHUGE_LIGHT;
650}
651
652/* Caller must hold page table lock. */
653static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
654 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
655 struct page *zero_page)
656{
657 pmd_t entry;
658 if (!pmd_none(*pmd))
659 return false;
660 entry = mk_pmd(zero_page, vma->vm_page_prot);
661 entry = pmd_mkhuge(entry);
662 if (pgtable)
663 pgtable_trans_huge_deposit(mm, pmd, pgtable);
664 set_pmd_at(mm, haddr, pmd, entry);
665 mm_inc_nr_ptes(mm);
666 return true;
667}
668
669int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
670{
671 struct vm_area_struct *vma = vmf->vma;
672 gfp_t gfp;
673 struct page *page;
674 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
675
676 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
677 return VM_FAULT_FALLBACK;
678 if (unlikely(anon_vma_prepare(vma)))
679 return VM_FAULT_OOM;
680 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
681 return VM_FAULT_OOM;
682 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
683 !mm_forbids_zeropage(vma->vm_mm) &&
684 transparent_hugepage_use_zero_page()) {
685 pgtable_t pgtable;
686 struct page *zero_page;
687 bool set;
688 int ret;
689 pgtable = pte_alloc_one(vma->vm_mm, haddr);
690 if (unlikely(!pgtable))
691 return VM_FAULT_OOM;
692 zero_page = mm_get_huge_zero_page(vma->vm_mm);
693 if (unlikely(!zero_page)) {
694 pte_free(vma->vm_mm, pgtable);
695 count_vm_event(THP_FAULT_FALLBACK);
696 return VM_FAULT_FALLBACK;
697 }
698 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
699 ret = 0;
700 set = false;
701 if (pmd_none(*vmf->pmd)) {
702 ret = check_stable_address_space(vma->vm_mm);
703 if (ret) {
704 spin_unlock(vmf->ptl);
705 } else if (userfaultfd_missing(vma)) {
706 spin_unlock(vmf->ptl);
707 ret = handle_userfault(vmf, VM_UFFD_MISSING);
708 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
709 } else {
710 set_huge_zero_page(pgtable, vma->vm_mm, vma,
711 haddr, vmf->pmd, zero_page);
712 spin_unlock(vmf->ptl);
713 set = true;
714 }
715 } else
716 spin_unlock(vmf->ptl);
717 if (!set)
718 pte_free(vma->vm_mm, pgtable);
719 return ret;
720 }
721 gfp = alloc_hugepage_direct_gfpmask(vma);
722 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
723 if (unlikely(!page)) {
724 count_vm_event(THP_FAULT_FALLBACK);
725 return VM_FAULT_FALLBACK;
726 }
727 prep_transhuge_page(page);
728 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
729}
730
731static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
732 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
733 pgtable_t pgtable)
734{
735 struct mm_struct *mm = vma->vm_mm;
736 pmd_t entry;
737 spinlock_t *ptl;
738
739 ptl = pmd_lock(mm, pmd);
740 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
741 if (pfn_t_devmap(pfn))
742 entry = pmd_mkdevmap(entry);
743 if (write) {
744 entry = pmd_mkyoung(pmd_mkdirty(entry));
745 entry = maybe_pmd_mkwrite(entry, vma);
746 }
747
748 if (pgtable) {
749 pgtable_trans_huge_deposit(mm, pmd, pgtable);
750 mm_inc_nr_ptes(mm);
751 }
752
753 set_pmd_at(mm, addr, pmd, entry);
754 update_mmu_cache_pmd(vma, addr, pmd);
755 spin_unlock(ptl);
756}
757
758int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
759 pmd_t *pmd, pfn_t pfn, bool write)
760{
761 pgprot_t pgprot = vma->vm_page_prot;
762 pgtable_t pgtable = NULL;
763 /*
764 * If we had pmd_special, we could avoid all these restrictions,
765 * but we need to be consistent with PTEs and architectures that
766 * can't support a 'special' bit.
767 */
768 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
769 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
770 (VM_PFNMAP|VM_MIXEDMAP));
771 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
772 BUG_ON(!pfn_t_devmap(pfn));
773
774 if (addr < vma->vm_start || addr >= vma->vm_end)
775 return VM_FAULT_SIGBUS;
776
777 if (arch_needs_pgtable_deposit()) {
778 pgtable = pte_alloc_one(vma->vm_mm, addr);
779 if (!pgtable)
780 return VM_FAULT_OOM;
781 }
782
783 track_pfn_insert(vma, &pgprot, pfn);
784
785 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
786 return VM_FAULT_NOPAGE;
787}
788EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
789
790#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
791static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
792{
793 if (likely(vma->vm_flags & VM_WRITE))
794 pud = pud_mkwrite(pud);
795 return pud;
796}
797
798static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
799 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
800{
801 struct mm_struct *mm = vma->vm_mm;
802 pud_t entry;
803 spinlock_t *ptl;
804
805 ptl = pud_lock(mm, pud);
806 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
807 if (pfn_t_devmap(pfn))
808 entry = pud_mkdevmap(entry);
809 if (write) {
810 entry = pud_mkyoung(pud_mkdirty(entry));
811 entry = maybe_pud_mkwrite(entry, vma);
812 }
813 set_pud_at(mm, addr, pud, entry);
814 update_mmu_cache_pud(vma, addr, pud);
815 spin_unlock(ptl);
816}
817
818int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
819 pud_t *pud, pfn_t pfn, bool write)
820{
821 pgprot_t pgprot = vma->vm_page_prot;
822 /*
823 * If we had pud_special, we could avoid all these restrictions,
824 * but we need to be consistent with PTEs and architectures that
825 * can't support a 'special' bit.
826 */
827 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
828 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
829 (VM_PFNMAP|VM_MIXEDMAP));
830 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
831 BUG_ON(!pfn_t_devmap(pfn));
832
833 if (addr < vma->vm_start || addr >= vma->vm_end)
834 return VM_FAULT_SIGBUS;
835
836 track_pfn_insert(vma, &pgprot, pfn);
837
838 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
839 return VM_FAULT_NOPAGE;
840}
841EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
842#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
843
844static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
845 pmd_t *pmd, int flags)
846{
847 pmd_t _pmd;
848
849 _pmd = pmd_mkyoung(*pmd);
850 if (flags & FOLL_WRITE)
851 _pmd = pmd_mkdirty(_pmd);
852 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
853 pmd, _pmd, flags & FOLL_WRITE))
854 update_mmu_cache_pmd(vma, addr, pmd);
855}
856
857struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
858 pmd_t *pmd, int flags)
859{
860 unsigned long pfn = pmd_pfn(*pmd);
861 struct mm_struct *mm = vma->vm_mm;
862 struct dev_pagemap *pgmap;
863 struct page *page;
864
865 assert_spin_locked(pmd_lockptr(mm, pmd));
866
867 /*
868 * When we COW a devmap PMD entry, we split it into PTEs, so we should
869 * not be in this function with `flags & FOLL_COW` set.
870 */
871 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
872
873 if (flags & FOLL_WRITE && !pmd_write(*pmd))
874 return NULL;
875
876 if (pmd_present(*pmd) && pmd_devmap(*pmd))
877 /* pass */;
878 else
879 return NULL;
880
881 if (flags & FOLL_TOUCH)
882 touch_pmd(vma, addr, pmd, flags);
883
884 /*
885 * device mapped pages can only be returned if the
886 * caller will manage the page reference count.
887 */
888 if (!(flags & FOLL_GET))
889 return ERR_PTR(-EEXIST);
890
891 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
892 pgmap = get_dev_pagemap(pfn, NULL);
893 if (!pgmap)
894 return ERR_PTR(-EFAULT);
895 page = pfn_to_page(pfn);
896 get_page(page);
897 put_dev_pagemap(pgmap);
898
899 return page;
900}
901
902int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
903 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
904 struct vm_area_struct *vma)
905{
906 spinlock_t *dst_ptl, *src_ptl;
907 struct page *src_page;
908 pmd_t pmd;
909 pgtable_t pgtable = NULL;
910 int ret = -ENOMEM;
911
912 /* Skip if can be re-fill on fault */
913 if (!vma_is_anonymous(vma))
914 return 0;
915
916 pgtable = pte_alloc_one(dst_mm, addr);
917 if (unlikely(!pgtable))
918 goto out;
919
920 dst_ptl = pmd_lock(dst_mm, dst_pmd);
921 src_ptl = pmd_lockptr(src_mm, src_pmd);
922 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
923
924 ret = -EAGAIN;
925 pmd = *src_pmd;
926
927#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
928 if (unlikely(is_swap_pmd(pmd))) {
929 swp_entry_t entry = pmd_to_swp_entry(pmd);
930
931 VM_BUG_ON(!is_pmd_migration_entry(pmd));
932 if (is_write_migration_entry(entry)) {
933 make_migration_entry_read(&entry);
934 pmd = swp_entry_to_pmd(entry);
935 if (pmd_swp_soft_dirty(*src_pmd))
936 pmd = pmd_swp_mksoft_dirty(pmd);
937 set_pmd_at(src_mm, addr, src_pmd, pmd);
938 }
939 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
940 mm_inc_nr_ptes(dst_mm);
941 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
942 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
943 ret = 0;
944 goto out_unlock;
945 }
946#endif
947
948 if (unlikely(!pmd_trans_huge(pmd))) {
949 pte_free(dst_mm, pgtable);
950 goto out_unlock;
951 }
952 /*
953 * When page table lock is held, the huge zero pmd should not be
954 * under splitting since we don't split the page itself, only pmd to
955 * a page table.
956 */
957 if (is_huge_zero_pmd(pmd)) {
958 struct page *zero_page;
959 /*
960 * get_huge_zero_page() will never allocate a new page here,
961 * since we already have a zero page to copy. It just takes a
962 * reference.
963 */
964 zero_page = mm_get_huge_zero_page(dst_mm);
965 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
966 zero_page);
967 ret = 0;
968 goto out_unlock;
969 }
970
971 src_page = pmd_page(pmd);
972 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
973 get_page(src_page);
974 page_dup_rmap(src_page, true);
975 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
976 mm_inc_nr_ptes(dst_mm);
977 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
978
979 pmdp_set_wrprotect(src_mm, addr, src_pmd);
980 pmd = pmd_mkold(pmd_wrprotect(pmd));
981 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
982
983 ret = 0;
984out_unlock:
985 spin_unlock(src_ptl);
986 spin_unlock(dst_ptl);
987out:
988 return ret;
989}
990
991#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
992static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
993 pud_t *pud, int flags)
994{
995 pud_t _pud;
996
997 _pud = pud_mkyoung(*pud);
998 if (flags & FOLL_WRITE)
999 _pud = pud_mkdirty(_pud);
1000 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1001 pud, _pud, flags & FOLL_WRITE))
1002 update_mmu_cache_pud(vma, addr, pud);
1003}
1004
1005struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1006 pud_t *pud, int flags)
1007{
1008 unsigned long pfn = pud_pfn(*pud);
1009 struct mm_struct *mm = vma->vm_mm;
1010 struct dev_pagemap *pgmap;
1011 struct page *page;
1012
1013 assert_spin_locked(pud_lockptr(mm, pud));
1014
1015 if (flags & FOLL_WRITE && !pud_write(*pud))
1016 return NULL;
1017
1018 if (pud_present(*pud) && pud_devmap(*pud))
1019 /* pass */;
1020 else
1021 return NULL;
1022
1023 if (flags & FOLL_TOUCH)
1024 touch_pud(vma, addr, pud, flags);
1025
1026 /*
1027 * device mapped pages can only be returned if the
1028 * caller will manage the page reference count.
1029 */
1030 if (!(flags & FOLL_GET))
1031 return ERR_PTR(-EEXIST);
1032
1033 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1034 pgmap = get_dev_pagemap(pfn, NULL);
1035 if (!pgmap)
1036 return ERR_PTR(-EFAULT);
1037 page = pfn_to_page(pfn);
1038 get_page(page);
1039 put_dev_pagemap(pgmap);
1040
1041 return page;
1042}
1043
1044int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1046 struct vm_area_struct *vma)
1047{
1048 spinlock_t *dst_ptl, *src_ptl;
1049 pud_t pud;
1050 int ret;
1051
1052 dst_ptl = pud_lock(dst_mm, dst_pud);
1053 src_ptl = pud_lockptr(src_mm, src_pud);
1054 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1055
1056 ret = -EAGAIN;
1057 pud = *src_pud;
1058 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1059 goto out_unlock;
1060
1061 /*
1062 * When page table lock is held, the huge zero pud should not be
1063 * under splitting since we don't split the page itself, only pud to
1064 * a page table.
1065 */
1066 if (is_huge_zero_pud(pud)) {
1067 /* No huge zero pud yet */
1068 }
1069
1070 pudp_set_wrprotect(src_mm, addr, src_pud);
1071 pud = pud_mkold(pud_wrprotect(pud));
1072 set_pud_at(dst_mm, addr, dst_pud, pud);
1073
1074 ret = 0;
1075out_unlock:
1076 spin_unlock(src_ptl);
1077 spin_unlock(dst_ptl);
1078 return ret;
1079}
1080
1081void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1082{
1083 pud_t entry;
1084 unsigned long haddr;
1085 bool write = vmf->flags & FAULT_FLAG_WRITE;
1086
1087 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1088 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1089 goto unlock;
1090
1091 entry = pud_mkyoung(orig_pud);
1092 if (write)
1093 entry = pud_mkdirty(entry);
1094 haddr = vmf->address & HPAGE_PUD_MASK;
1095 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1096 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1097
1098unlock:
1099 spin_unlock(vmf->ptl);
1100}
1101#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1102
1103void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1104{
1105 pmd_t entry;
1106 unsigned long haddr;
1107 bool write = vmf->flags & FAULT_FLAG_WRITE;
1108
1109 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1110 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1111 goto unlock;
1112
1113 entry = pmd_mkyoung(orig_pmd);
1114 if (write)
1115 entry = pmd_mkdirty(entry);
1116 haddr = vmf->address & HPAGE_PMD_MASK;
1117 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1118 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1119
1120unlock:
1121 spin_unlock(vmf->ptl);
1122}
1123
1124static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1125 struct page *page)
1126{
1127 struct vm_area_struct *vma = vmf->vma;
1128 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1129 struct mem_cgroup *memcg;
1130 pgtable_t pgtable;
1131 pmd_t _pmd;
1132 int ret = 0, i;
1133 struct page **pages;
1134 unsigned long mmun_start; /* For mmu_notifiers */
1135 unsigned long mmun_end; /* For mmu_notifiers */
1136
1137 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1138 GFP_KERNEL);
1139 if (unlikely(!pages)) {
1140 ret |= VM_FAULT_OOM;
1141 goto out;
1142 }
1143
1144 for (i = 0; i < HPAGE_PMD_NR; i++) {
1145 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1146 vmf->address, page_to_nid(page));
1147 if (unlikely(!pages[i] ||
1148 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1149 GFP_KERNEL, &memcg, false))) {
1150 if (pages[i])
1151 put_page(pages[i]);
1152 while (--i >= 0) {
1153 memcg = (void *)page_private(pages[i]);
1154 set_page_private(pages[i], 0);
1155 mem_cgroup_cancel_charge(pages[i], memcg,
1156 false);
1157 put_page(pages[i]);
1158 }
1159 kfree(pages);
1160 ret |= VM_FAULT_OOM;
1161 goto out;
1162 }
1163 set_page_private(pages[i], (unsigned long)memcg);
1164 }
1165
1166 for (i = 0; i < HPAGE_PMD_NR; i++) {
1167 copy_user_highpage(pages[i], page + i,
1168 haddr + PAGE_SIZE * i, vma);
1169 __SetPageUptodate(pages[i]);
1170 cond_resched();
1171 }
1172
1173 mmun_start = haddr;
1174 mmun_end = haddr + HPAGE_PMD_SIZE;
1175 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1176
1177 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1178 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1179 goto out_free_pages;
1180 VM_BUG_ON_PAGE(!PageHead(page), page);
1181
1182 /*
1183 * Leave pmd empty until pte is filled note we must notify here as
1184 * concurrent CPU thread might write to new page before the call to
1185 * mmu_notifier_invalidate_range_end() happens which can lead to a
1186 * device seeing memory write in different order than CPU.
1187 *
1188 * See Documentation/vm/mmu_notifier.txt
1189 */
1190 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1191
1192 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1193 pmd_populate(vma->vm_mm, &_pmd, pgtable);
1194
1195 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1196 pte_t entry;
1197 entry = mk_pte(pages[i], vma->vm_page_prot);
1198 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1199 memcg = (void *)page_private(pages[i]);
1200 set_page_private(pages[i], 0);
1201 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1202 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1203 lru_cache_add_active_or_unevictable(pages[i], vma);
1204 vmf->pte = pte_offset_map(&_pmd, haddr);
1205 VM_BUG_ON(!pte_none(*vmf->pte));
1206 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1207 pte_unmap(vmf->pte);
1208 }
1209 kfree(pages);
1210
1211 smp_wmb(); /* make pte visible before pmd */
1212 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1213 page_remove_rmap(page, true);
1214 spin_unlock(vmf->ptl);
1215
1216 /*
1217 * No need to double call mmu_notifier->invalidate_range() callback as
1218 * the above pmdp_huge_clear_flush_notify() did already call it.
1219 */
1220 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1221 mmun_end);
1222
1223 ret |= VM_FAULT_WRITE;
1224 put_page(page);
1225
1226out:
1227 return ret;
1228
1229out_free_pages:
1230 spin_unlock(vmf->ptl);
1231 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1232 for (i = 0; i < HPAGE_PMD_NR; i++) {
1233 memcg = (void *)page_private(pages[i]);
1234 set_page_private(pages[i], 0);
1235 mem_cgroup_cancel_charge(pages[i], memcg, false);
1236 put_page(pages[i]);
1237 }
1238 kfree(pages);
1239 goto out;
1240}
1241
1242int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1243{
1244 struct vm_area_struct *vma = vmf->vma;
1245 struct page *page = NULL, *new_page;
1246 struct mem_cgroup *memcg;
1247 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1248 unsigned long mmun_start; /* For mmu_notifiers */
1249 unsigned long mmun_end; /* For mmu_notifiers */
1250 gfp_t huge_gfp; /* for allocation and charge */
1251 int ret = 0;
1252
1253 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1254 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1255 if (is_huge_zero_pmd(orig_pmd))
1256 goto alloc;
1257 spin_lock(vmf->ptl);
1258 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1259 goto out_unlock;
1260
1261 page = pmd_page(orig_pmd);
1262 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1263 /*
1264 * We can only reuse the page if nobody else maps the huge page or it's
1265 * part.
1266 */
1267 if (!trylock_page(page)) {
1268 get_page(page);
1269 spin_unlock(vmf->ptl);
1270 lock_page(page);
1271 spin_lock(vmf->ptl);
1272 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1273 unlock_page(page);
1274 put_page(page);
1275 goto out_unlock;
1276 }
1277 put_page(page);
1278 }
1279 if (reuse_swap_page(page, NULL)) {
1280 pmd_t entry;
1281 entry = pmd_mkyoung(orig_pmd);
1282 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1283 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1284 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1285 ret |= VM_FAULT_WRITE;
1286 unlock_page(page);
1287 goto out_unlock;
1288 }
1289 unlock_page(page);
1290 get_page(page);
1291 spin_unlock(vmf->ptl);
1292alloc:
1293 if (transparent_hugepage_enabled(vma) &&
1294 !transparent_hugepage_debug_cow()) {
1295 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1296 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1297 } else
1298 new_page = NULL;
1299
1300 if (likely(new_page)) {
1301 prep_transhuge_page(new_page);
1302 } else {
1303 if (!page) {
1304 split_huge_pmd(vma, vmf->pmd, vmf->address);
1305 ret |= VM_FAULT_FALLBACK;
1306 } else {
1307 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1308 if (ret & VM_FAULT_OOM) {
1309 split_huge_pmd(vma, vmf->pmd, vmf->address);
1310 ret |= VM_FAULT_FALLBACK;
1311 }
1312 put_page(page);
1313 }
1314 count_vm_event(THP_FAULT_FALLBACK);
1315 goto out;
1316 }
1317
1318 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1319 huge_gfp, &memcg, true))) {
1320 put_page(new_page);
1321 split_huge_pmd(vma, vmf->pmd, vmf->address);
1322 if (page)
1323 put_page(page);
1324 ret |= VM_FAULT_FALLBACK;
1325 count_vm_event(THP_FAULT_FALLBACK);
1326 goto out;
1327 }
1328
1329 count_vm_event(THP_FAULT_ALLOC);
1330
1331 if (!page)
1332 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1333 else
1334 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1335 __SetPageUptodate(new_page);
1336
1337 mmun_start = haddr;
1338 mmun_end = haddr + HPAGE_PMD_SIZE;
1339 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1340
1341 spin_lock(vmf->ptl);
1342 if (page)
1343 put_page(page);
1344 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345 spin_unlock(vmf->ptl);
1346 mem_cgroup_cancel_charge(new_page, memcg, true);
1347 put_page(new_page);
1348 goto out_mn;
1349 } else {
1350 pmd_t entry;
1351 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1352 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1353 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1354 page_add_new_anon_rmap(new_page, vma, haddr, true);
1355 mem_cgroup_commit_charge(new_page, memcg, false, true);
1356 lru_cache_add_active_or_unevictable(new_page, vma);
1357 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1358 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1359 if (!page) {
1360 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1361 } else {
1362 VM_BUG_ON_PAGE(!PageHead(page), page);
1363 page_remove_rmap(page, true);
1364 put_page(page);
1365 }
1366 ret |= VM_FAULT_WRITE;
1367 }
1368 spin_unlock(vmf->ptl);
1369out_mn:
1370 /*
1371 * No need to double call mmu_notifier->invalidate_range() callback as
1372 * the above pmdp_huge_clear_flush_notify() did already call it.
1373 */
1374 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1375 mmun_end);
1376out:
1377 return ret;
1378out_unlock:
1379 spin_unlock(vmf->ptl);
1380 return ret;
1381}
1382
1383/*
1384 * FOLL_FORCE can write to even unwritable pmd's, but only
1385 * after we've gone through a COW cycle and they are dirty.
1386 */
1387static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1388{
1389 return pmd_write(pmd) ||
1390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1391}
1392
1393struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1394 unsigned long addr,
1395 pmd_t *pmd,
1396 unsigned int flags)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct page *page = NULL;
1400
1401 assert_spin_locked(pmd_lockptr(mm, pmd));
1402
1403 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1404 goto out;
1405
1406 /* Avoid dumping huge zero page */
1407 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1408 return ERR_PTR(-EFAULT);
1409
1410 /* Full NUMA hinting faults to serialise migration in fault paths */
1411 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1412 goto out;
1413
1414 page = pmd_page(*pmd);
1415 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1416 if (flags & FOLL_TOUCH)
1417 touch_pmd(vma, addr, pmd, flags);
1418 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1419 /*
1420 * We don't mlock() pte-mapped THPs. This way we can avoid
1421 * leaking mlocked pages into non-VM_LOCKED VMAs.
1422 *
1423 * For anon THP:
1424 *
1425 * In most cases the pmd is the only mapping of the page as we
1426 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1427 * writable private mappings in populate_vma_page_range().
1428 *
1429 * The only scenario when we have the page shared here is if we
1430 * mlocking read-only mapping shared over fork(). We skip
1431 * mlocking such pages.
1432 *
1433 * For file THP:
1434 *
1435 * We can expect PageDoubleMap() to be stable under page lock:
1436 * for file pages we set it in page_add_file_rmap(), which
1437 * requires page to be locked.
1438 */
1439
1440 if (PageAnon(page) && compound_mapcount(page) != 1)
1441 goto skip_mlock;
1442 if (PageDoubleMap(page) || !page->mapping)
1443 goto skip_mlock;
1444 if (!trylock_page(page))
1445 goto skip_mlock;
1446 lru_add_drain();
1447 if (page->mapping && !PageDoubleMap(page))
1448 mlock_vma_page(page);
1449 unlock_page(page);
1450 }
1451skip_mlock:
1452 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1453 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1454 if (flags & FOLL_GET)
1455 get_page(page);
1456
1457out:
1458 return page;
1459}
1460
1461/* NUMA hinting page fault entry point for trans huge pmds */
1462int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1463{
1464 struct vm_area_struct *vma = vmf->vma;
1465 struct anon_vma *anon_vma = NULL;
1466 struct page *page;
1467 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1468 int page_nid = -1, this_nid = numa_node_id();
1469 int target_nid, last_cpupid = -1;
1470 bool page_locked;
1471 bool migrated = false;
1472 bool was_writable;
1473 int flags = 0;
1474
1475 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1476 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1477 goto out_unlock;
1478
1479 /*
1480 * If there are potential migrations, wait for completion and retry
1481 * without disrupting NUMA hinting information. Do not relock and
1482 * check_same as the page may no longer be mapped.
1483 */
1484 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1485 page = pmd_page(*vmf->pmd);
1486 if (!get_page_unless_zero(page))
1487 goto out_unlock;
1488 spin_unlock(vmf->ptl);
1489 wait_on_page_locked(page);
1490 put_page(page);
1491 goto out;
1492 }
1493
1494 page = pmd_page(pmd);
1495 BUG_ON(is_huge_zero_page(page));
1496 page_nid = page_to_nid(page);
1497 last_cpupid = page_cpupid_last(page);
1498 count_vm_numa_event(NUMA_HINT_FAULTS);
1499 if (page_nid == this_nid) {
1500 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1501 flags |= TNF_FAULT_LOCAL;
1502 }
1503
1504 /* See similar comment in do_numa_page for explanation */
1505 if (!pmd_savedwrite(pmd))
1506 flags |= TNF_NO_GROUP;
1507
1508 /*
1509 * Acquire the page lock to serialise THP migrations but avoid dropping
1510 * page_table_lock if at all possible
1511 */
1512 page_locked = trylock_page(page);
1513 target_nid = mpol_misplaced(page, vma, haddr);
1514 if (target_nid == -1) {
1515 /* If the page was locked, there are no parallel migrations */
1516 if (page_locked)
1517 goto clear_pmdnuma;
1518 }
1519
1520 /* Migration could have started since the pmd_trans_migrating check */
1521 if (!page_locked) {
1522 page_nid = -1;
1523 if (!get_page_unless_zero(page))
1524 goto out_unlock;
1525 spin_unlock(vmf->ptl);
1526 wait_on_page_locked(page);
1527 put_page(page);
1528 goto out;
1529 }
1530
1531 /*
1532 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1533 * to serialises splits
1534 */
1535 get_page(page);
1536 spin_unlock(vmf->ptl);
1537 anon_vma = page_lock_anon_vma_read(page);
1538
1539 /* Confirm the PMD did not change while page_table_lock was released */
1540 spin_lock(vmf->ptl);
1541 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1542 unlock_page(page);
1543 put_page(page);
1544 page_nid = -1;
1545 goto out_unlock;
1546 }
1547
1548 /* Bail if we fail to protect against THP splits for any reason */
1549 if (unlikely(!anon_vma)) {
1550 put_page(page);
1551 page_nid = -1;
1552 goto clear_pmdnuma;
1553 }
1554
1555 /*
1556 * Since we took the NUMA fault, we must have observed the !accessible
1557 * bit. Make sure all other CPUs agree with that, to avoid them
1558 * modifying the page we're about to migrate.
1559 *
1560 * Must be done under PTL such that we'll observe the relevant
1561 * inc_tlb_flush_pending().
1562 *
1563 * We are not sure a pending tlb flush here is for a huge page
1564 * mapping or not. Hence use the tlb range variant
1565 */
1566 if (mm_tlb_flush_pending(vma->vm_mm))
1567 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1568
1569 /*
1570 * Migrate the THP to the requested node, returns with page unlocked
1571 * and access rights restored.
1572 */
1573 spin_unlock(vmf->ptl);
1574
1575 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1576 vmf->pmd, pmd, vmf->address, page, target_nid);
1577 if (migrated) {
1578 flags |= TNF_MIGRATED;
1579 page_nid = target_nid;
1580 } else
1581 flags |= TNF_MIGRATE_FAIL;
1582
1583 goto out;
1584clear_pmdnuma:
1585 BUG_ON(!PageLocked(page));
1586 was_writable = pmd_savedwrite(pmd);
1587 pmd = pmd_modify(pmd, vma->vm_page_prot);
1588 pmd = pmd_mkyoung(pmd);
1589 if (was_writable)
1590 pmd = pmd_mkwrite(pmd);
1591 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1592 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1593 unlock_page(page);
1594out_unlock:
1595 spin_unlock(vmf->ptl);
1596
1597out:
1598 if (anon_vma)
1599 page_unlock_anon_vma_read(anon_vma);
1600
1601 if (page_nid != -1)
1602 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1603 flags);
1604
1605 return 0;
1606}
1607
1608/*
1609 * Return true if we do MADV_FREE successfully on entire pmd page.
1610 * Otherwise, return false.
1611 */
1612bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1613 pmd_t *pmd, unsigned long addr, unsigned long next)
1614{
1615 spinlock_t *ptl;
1616 pmd_t orig_pmd;
1617 struct page *page;
1618 struct mm_struct *mm = tlb->mm;
1619 bool ret = false;
1620
1621 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1622
1623 ptl = pmd_trans_huge_lock(pmd, vma);
1624 if (!ptl)
1625 goto out_unlocked;
1626
1627 orig_pmd = *pmd;
1628 if (is_huge_zero_pmd(orig_pmd))
1629 goto out;
1630
1631 if (unlikely(!pmd_present(orig_pmd))) {
1632 VM_BUG_ON(thp_migration_supported() &&
1633 !is_pmd_migration_entry(orig_pmd));
1634 goto out;
1635 }
1636
1637 page = pmd_page(orig_pmd);
1638 /*
1639 * If other processes are mapping this page, we couldn't discard
1640 * the page unless they all do MADV_FREE so let's skip the page.
1641 */
1642 if (page_mapcount(page) != 1)
1643 goto out;
1644
1645 if (!trylock_page(page))
1646 goto out;
1647
1648 /*
1649 * If user want to discard part-pages of THP, split it so MADV_FREE
1650 * will deactivate only them.
1651 */
1652 if (next - addr != HPAGE_PMD_SIZE) {
1653 get_page(page);
1654 spin_unlock(ptl);
1655 split_huge_page(page);
1656 unlock_page(page);
1657 put_page(page);
1658 goto out_unlocked;
1659 }
1660
1661 if (PageDirty(page))
1662 ClearPageDirty(page);
1663 unlock_page(page);
1664
1665 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1666 pmdp_invalidate(vma, addr, pmd);
1667 orig_pmd = pmd_mkold(orig_pmd);
1668 orig_pmd = pmd_mkclean(orig_pmd);
1669
1670 set_pmd_at(mm, addr, pmd, orig_pmd);
1671 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1672 }
1673
1674 mark_page_lazyfree(page);
1675 ret = true;
1676out:
1677 spin_unlock(ptl);
1678out_unlocked:
1679 return ret;
1680}
1681
1682static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1683{
1684 pgtable_t pgtable;
1685
1686 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1687 pte_free(mm, pgtable);
1688 mm_dec_nr_ptes(mm);
1689}
1690
1691int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1692 pmd_t *pmd, unsigned long addr)
1693{
1694 pmd_t orig_pmd;
1695 spinlock_t *ptl;
1696
1697 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1698
1699 ptl = __pmd_trans_huge_lock(pmd, vma);
1700 if (!ptl)
1701 return 0;
1702 /*
1703 * For architectures like ppc64 we look at deposited pgtable
1704 * when calling pmdp_huge_get_and_clear. So do the
1705 * pgtable_trans_huge_withdraw after finishing pmdp related
1706 * operations.
1707 */
1708 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1709 tlb->fullmm);
1710 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1711 if (vma_is_dax(vma)) {
1712 if (arch_needs_pgtable_deposit())
1713 zap_deposited_table(tlb->mm, pmd);
1714 spin_unlock(ptl);
1715 if (is_huge_zero_pmd(orig_pmd))
1716 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1717 } else if (is_huge_zero_pmd(orig_pmd)) {
1718 zap_deposited_table(tlb->mm, pmd);
1719 spin_unlock(ptl);
1720 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1721 } else {
1722 struct page *page = NULL;
1723 int flush_needed = 1;
1724
1725 if (pmd_present(orig_pmd)) {
1726 page = pmd_page(orig_pmd);
1727 page_remove_rmap(page, true);
1728 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1729 VM_BUG_ON_PAGE(!PageHead(page), page);
1730 } else if (thp_migration_supported()) {
1731 swp_entry_t entry;
1732
1733 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1734 entry = pmd_to_swp_entry(orig_pmd);
1735 page = pfn_to_page(swp_offset(entry));
1736 flush_needed = 0;
1737 } else
1738 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1739
1740 if (PageAnon(page)) {
1741 zap_deposited_table(tlb->mm, pmd);
1742 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1743 } else {
1744 if (arch_needs_pgtable_deposit())
1745 zap_deposited_table(tlb->mm, pmd);
1746 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1747 }
1748
1749 spin_unlock(ptl);
1750 if (flush_needed)
1751 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1752 }
1753 return 1;
1754}
1755
1756#ifndef pmd_move_must_withdraw
1757static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1758 spinlock_t *old_pmd_ptl,
1759 struct vm_area_struct *vma)
1760{
1761 /*
1762 * With split pmd lock we also need to move preallocated
1763 * PTE page table if new_pmd is on different PMD page table.
1764 *
1765 * We also don't deposit and withdraw tables for file pages.
1766 */
1767 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1768}
1769#endif
1770
1771static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1772{
1773#ifdef CONFIG_MEM_SOFT_DIRTY
1774 if (unlikely(is_pmd_migration_entry(pmd)))
1775 pmd = pmd_swp_mksoft_dirty(pmd);
1776 else if (pmd_present(pmd))
1777 pmd = pmd_mksoft_dirty(pmd);
1778#endif
1779 return pmd;
1780}
1781
1782bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1783 unsigned long new_addr, unsigned long old_end,
1784 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1785{
1786 spinlock_t *old_ptl, *new_ptl;
1787 pmd_t pmd;
1788 struct mm_struct *mm = vma->vm_mm;
1789 bool force_flush = false;
1790
1791 if ((old_addr & ~HPAGE_PMD_MASK) ||
1792 (new_addr & ~HPAGE_PMD_MASK) ||
1793 old_end - old_addr < HPAGE_PMD_SIZE)
1794 return false;
1795
1796 /*
1797 * The destination pmd shouldn't be established, free_pgtables()
1798 * should have release it.
1799 */
1800 if (WARN_ON(!pmd_none(*new_pmd))) {
1801 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1802 return false;
1803 }
1804
1805 /*
1806 * We don't have to worry about the ordering of src and dst
1807 * ptlocks because exclusive mmap_sem prevents deadlock.
1808 */
1809 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1810 if (old_ptl) {
1811 new_ptl = pmd_lockptr(mm, new_pmd);
1812 if (new_ptl != old_ptl)
1813 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1814 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1815 if (pmd_present(pmd) && pmd_dirty(pmd))
1816 force_flush = true;
1817 VM_BUG_ON(!pmd_none(*new_pmd));
1818
1819 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1820 pgtable_t pgtable;
1821 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1822 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1823 }
1824 pmd = move_soft_dirty_pmd(pmd);
1825 set_pmd_at(mm, new_addr, new_pmd, pmd);
1826 if (new_ptl != old_ptl)
1827 spin_unlock(new_ptl);
1828 if (force_flush)
1829 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1830 else
1831 *need_flush = true;
1832 spin_unlock(old_ptl);
1833 return true;
1834 }
1835 return false;
1836}
1837
1838/*
1839 * Returns
1840 * - 0 if PMD could not be locked
1841 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1842 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1843 */
1844int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1845 unsigned long addr, pgprot_t newprot, int prot_numa)
1846{
1847 struct mm_struct *mm = vma->vm_mm;
1848 spinlock_t *ptl;
1849 pmd_t entry;
1850 bool preserve_write;
1851 int ret;
1852
1853 ptl = __pmd_trans_huge_lock(pmd, vma);
1854 if (!ptl)
1855 return 0;
1856
1857 preserve_write = prot_numa && pmd_write(*pmd);
1858 ret = 1;
1859
1860#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1861 if (is_swap_pmd(*pmd)) {
1862 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1863
1864 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1865 if (is_write_migration_entry(entry)) {
1866 pmd_t newpmd;
1867 /*
1868 * A protection check is difficult so
1869 * just be safe and disable write
1870 */
1871 make_migration_entry_read(&entry);
1872 newpmd = swp_entry_to_pmd(entry);
1873 if (pmd_swp_soft_dirty(*pmd))
1874 newpmd = pmd_swp_mksoft_dirty(newpmd);
1875 set_pmd_at(mm, addr, pmd, newpmd);
1876 }
1877 goto unlock;
1878 }
1879#endif
1880
1881 /*
1882 * Avoid trapping faults against the zero page. The read-only
1883 * data is likely to be read-cached on the local CPU and
1884 * local/remote hits to the zero page are not interesting.
1885 */
1886 if (prot_numa && is_huge_zero_pmd(*pmd))
1887 goto unlock;
1888
1889 if (prot_numa && pmd_protnone(*pmd))
1890 goto unlock;
1891
1892 /*
1893 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1894 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1895 * which is also under down_read(mmap_sem):
1896 *
1897 * CPU0: CPU1:
1898 * change_huge_pmd(prot_numa=1)
1899 * pmdp_huge_get_and_clear_notify()
1900 * madvise_dontneed()
1901 * zap_pmd_range()
1902 * pmd_trans_huge(*pmd) == 0 (without ptl)
1903 * // skip the pmd
1904 * set_pmd_at();
1905 * // pmd is re-established
1906 *
1907 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1908 * which may break userspace.
1909 *
1910 * pmdp_invalidate() is required to make sure we don't miss
1911 * dirty/young flags set by hardware.
1912 */
1913 entry = pmdp_invalidate(vma, addr, pmd);
1914
1915 entry = pmd_modify(entry, newprot);
1916 if (preserve_write)
1917 entry = pmd_mk_savedwrite(entry);
1918 ret = HPAGE_PMD_NR;
1919 set_pmd_at(mm, addr, pmd, entry);
1920 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1921unlock:
1922 spin_unlock(ptl);
1923 return ret;
1924}
1925
1926/*
1927 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1928 *
1929 * Note that if it returns page table lock pointer, this routine returns without
1930 * unlocking page table lock. So callers must unlock it.
1931 */
1932spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1933{
1934 spinlock_t *ptl;
1935 ptl = pmd_lock(vma->vm_mm, pmd);
1936 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1937 pmd_devmap(*pmd)))
1938 return ptl;
1939 spin_unlock(ptl);
1940 return NULL;
1941}
1942
1943/*
1944 * Returns true if a given pud maps a thp, false otherwise.
1945 *
1946 * Note that if it returns true, this routine returns without unlocking page
1947 * table lock. So callers must unlock it.
1948 */
1949spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1950{
1951 spinlock_t *ptl;
1952
1953 ptl = pud_lock(vma->vm_mm, pud);
1954 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1955 return ptl;
1956 spin_unlock(ptl);
1957 return NULL;
1958}
1959
1960#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1961int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1962 pud_t *pud, unsigned long addr)
1963{
1964 pud_t orig_pud;
1965 spinlock_t *ptl;
1966
1967 ptl = __pud_trans_huge_lock(pud, vma);
1968 if (!ptl)
1969 return 0;
1970 /*
1971 * For architectures like ppc64 we look at deposited pgtable
1972 * when calling pudp_huge_get_and_clear. So do the
1973 * pgtable_trans_huge_withdraw after finishing pudp related
1974 * operations.
1975 */
1976 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1977 tlb->fullmm);
1978 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1979 if (vma_is_dax(vma)) {
1980 spin_unlock(ptl);
1981 /* No zero page support yet */
1982 } else {
1983 /* No support for anonymous PUD pages yet */
1984 BUG();
1985 }
1986 return 1;
1987}
1988
1989static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1990 unsigned long haddr)
1991{
1992 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1993 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1994 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1995 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1996
1997 count_vm_event(THP_SPLIT_PUD);
1998
1999 pudp_huge_clear_flush_notify(vma, haddr, pud);
2000}
2001
2002void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2003 unsigned long address)
2004{
2005 spinlock_t *ptl;
2006 struct mm_struct *mm = vma->vm_mm;
2007 unsigned long haddr = address & HPAGE_PUD_MASK;
2008
2009 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2010 ptl = pud_lock(mm, pud);
2011 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2012 goto out;
2013 __split_huge_pud_locked(vma, pud, haddr);
2014
2015out:
2016 spin_unlock(ptl);
2017 /*
2018 * No need to double call mmu_notifier->invalidate_range() callback as
2019 * the above pudp_huge_clear_flush_notify() did already call it.
2020 */
2021 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2022 HPAGE_PUD_SIZE);
2023}
2024#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2025
2026static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2027 unsigned long haddr, pmd_t *pmd)
2028{
2029 struct mm_struct *mm = vma->vm_mm;
2030 pgtable_t pgtable;
2031 pmd_t _pmd;
2032 int i;
2033
2034 /*
2035 * Leave pmd empty until pte is filled note that it is fine to delay
2036 * notification until mmu_notifier_invalidate_range_end() as we are
2037 * replacing a zero pmd write protected page with a zero pte write
2038 * protected page.
2039 *
2040 * See Documentation/vm/mmu_notifier.txt
2041 */
2042 pmdp_huge_clear_flush(vma, haddr, pmd);
2043
2044 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2045 pmd_populate(mm, &_pmd, pgtable);
2046
2047 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2048 pte_t *pte, entry;
2049 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2050 entry = pte_mkspecial(entry);
2051 pte = pte_offset_map(&_pmd, haddr);
2052 VM_BUG_ON(!pte_none(*pte));
2053 set_pte_at(mm, haddr, pte, entry);
2054 pte_unmap(pte);
2055 }
2056 smp_wmb(); /* make pte visible before pmd */
2057 pmd_populate(mm, pmd, pgtable);
2058}
2059
2060static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2061 unsigned long haddr, bool freeze)
2062{
2063 struct mm_struct *mm = vma->vm_mm;
2064 struct page *page;
2065 pgtable_t pgtable;
2066 pmd_t old_pmd, _pmd;
2067 bool young, write, soft_dirty, pmd_migration = false;
2068 unsigned long addr;
2069 int i;
2070
2071 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2072 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2073 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2074 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2075 && !pmd_devmap(*pmd));
2076
2077 count_vm_event(THP_SPLIT_PMD);
2078
2079 if (!vma_is_anonymous(vma)) {
2080 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2081 /*
2082 * We are going to unmap this huge page. So
2083 * just go ahead and zap it
2084 */
2085 if (arch_needs_pgtable_deposit())
2086 zap_deposited_table(mm, pmd);
2087 if (vma_is_dax(vma))
2088 return;
2089 page = pmd_page(_pmd);
2090 if (!PageReferenced(page) && pmd_young(_pmd))
2091 SetPageReferenced(page);
2092 page_remove_rmap(page, true);
2093 put_page(page);
2094 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2095 return;
2096 } else if (is_huge_zero_pmd(*pmd)) {
2097 /*
2098 * FIXME: Do we want to invalidate secondary mmu by calling
2099 * mmu_notifier_invalidate_range() see comments below inside
2100 * __split_huge_pmd() ?
2101 *
2102 * We are going from a zero huge page write protected to zero
2103 * small page also write protected so it does not seems useful
2104 * to invalidate secondary mmu at this time.
2105 */
2106 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2107 }
2108
2109 /*
2110 * Up to this point the pmd is present and huge and userland has the
2111 * whole access to the hugepage during the split (which happens in
2112 * place). If we overwrite the pmd with the not-huge version pointing
2113 * to the pte here (which of course we could if all CPUs were bug
2114 * free), userland could trigger a small page size TLB miss on the
2115 * small sized TLB while the hugepage TLB entry is still established in
2116 * the huge TLB. Some CPU doesn't like that.
2117 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2118 * 383 on page 93. Intel should be safe but is also warns that it's
2119 * only safe if the permission and cache attributes of the two entries
2120 * loaded in the two TLB is identical (which should be the case here).
2121 * But it is generally safer to never allow small and huge TLB entries
2122 * for the same virtual address to be loaded simultaneously. So instead
2123 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2124 * current pmd notpresent (atomically because here the pmd_trans_huge
2125 * must remain set at all times on the pmd until the split is complete
2126 * for this pmd), then we flush the SMP TLB and finally we write the
2127 * non-huge version of the pmd entry with pmd_populate.
2128 */
2129 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2130
2131#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2132 pmd_migration = is_pmd_migration_entry(old_pmd);
2133 if (pmd_migration) {
2134 swp_entry_t entry;
2135
2136 entry = pmd_to_swp_entry(old_pmd);
2137 page = pfn_to_page(swp_offset(entry));
2138 } else
2139#endif
2140 page = pmd_page(old_pmd);
2141 VM_BUG_ON_PAGE(!page_count(page), page);
2142 page_ref_add(page, HPAGE_PMD_NR - 1);
2143 if (pmd_dirty(old_pmd))
2144 SetPageDirty(page);
2145 write = pmd_write(old_pmd);
2146 young = pmd_young(old_pmd);
2147 soft_dirty = pmd_soft_dirty(old_pmd);
2148
2149 /*
2150 * Withdraw the table only after we mark the pmd entry invalid.
2151 * This's critical for some architectures (Power).
2152 */
2153 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2154 pmd_populate(mm, &_pmd, pgtable);
2155
2156 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2157 pte_t entry, *pte;
2158 /*
2159 * Note that NUMA hinting access restrictions are not
2160 * transferred to avoid any possibility of altering
2161 * permissions across VMAs.
2162 */
2163 if (freeze || pmd_migration) {
2164 swp_entry_t swp_entry;
2165 swp_entry = make_migration_entry(page + i, write);
2166 entry = swp_entry_to_pte(swp_entry);
2167 if (soft_dirty)
2168 entry = pte_swp_mksoft_dirty(entry);
2169 } else {
2170 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2171 entry = maybe_mkwrite(entry, vma);
2172 if (!write)
2173 entry = pte_wrprotect(entry);
2174 if (!young)
2175 entry = pte_mkold(entry);
2176 if (soft_dirty)
2177 entry = pte_mksoft_dirty(entry);
2178 }
2179 pte = pte_offset_map(&_pmd, addr);
2180 BUG_ON(!pte_none(*pte));
2181 set_pte_at(mm, addr, pte, entry);
2182 atomic_inc(&page[i]._mapcount);
2183 pte_unmap(pte);
2184 }
2185
2186 /*
2187 * Set PG_double_map before dropping compound_mapcount to avoid
2188 * false-negative page_mapped().
2189 */
2190 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2191 for (i = 0; i < HPAGE_PMD_NR; i++)
2192 atomic_inc(&page[i]._mapcount);
2193 }
2194
2195 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2196 /* Last compound_mapcount is gone. */
2197 __dec_node_page_state(page, NR_ANON_THPS);
2198 if (TestClearPageDoubleMap(page)) {
2199 /* No need in mapcount reference anymore */
2200 for (i = 0; i < HPAGE_PMD_NR; i++)
2201 atomic_dec(&page[i]._mapcount);
2202 }
2203 }
2204
2205 smp_wmb(); /* make pte visible before pmd */
2206 pmd_populate(mm, pmd, pgtable);
2207
2208 if (freeze) {
2209 for (i = 0; i < HPAGE_PMD_NR; i++) {
2210 page_remove_rmap(page + i, false);
2211 put_page(page + i);
2212 }
2213 }
2214}
2215
2216void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2217 unsigned long address, bool freeze, struct page *page)
2218{
2219 spinlock_t *ptl;
2220 struct mm_struct *mm = vma->vm_mm;
2221 unsigned long haddr = address & HPAGE_PMD_MASK;
2222
2223 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2224 ptl = pmd_lock(mm, pmd);
2225
2226 /*
2227 * If caller asks to setup a migration entries, we need a page to check
2228 * pmd against. Otherwise we can end up replacing wrong page.
2229 */
2230 VM_BUG_ON(freeze && !page);
2231 if (page && page != pmd_page(*pmd))
2232 goto out;
2233
2234 if (pmd_trans_huge(*pmd)) {
2235 page = pmd_page(*pmd);
2236 if (PageMlocked(page))
2237 clear_page_mlock(page);
2238 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2239 goto out;
2240 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2241out:
2242 spin_unlock(ptl);
2243 /*
2244 * No need to double call mmu_notifier->invalidate_range() callback.
2245 * They are 3 cases to consider inside __split_huge_pmd_locked():
2246 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2247 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2248 * fault will trigger a flush_notify before pointing to a new page
2249 * (it is fine if the secondary mmu keeps pointing to the old zero
2250 * page in the meantime)
2251 * 3) Split a huge pmd into pte pointing to the same page. No need
2252 * to invalidate secondary tlb entry they are all still valid.
2253 * any further changes to individual pte will notify. So no need
2254 * to call mmu_notifier->invalidate_range()
2255 */
2256 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2257 HPAGE_PMD_SIZE);
2258}
2259
2260void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2261 bool freeze, struct page *page)
2262{
2263 pgd_t *pgd;
2264 p4d_t *p4d;
2265 pud_t *pud;
2266 pmd_t *pmd;
2267
2268 pgd = pgd_offset(vma->vm_mm, address);
2269 if (!pgd_present(*pgd))
2270 return;
2271
2272 p4d = p4d_offset(pgd, address);
2273 if (!p4d_present(*p4d))
2274 return;
2275
2276 pud = pud_offset(p4d, address);
2277 if (!pud_present(*pud))
2278 return;
2279
2280 pmd = pmd_offset(pud, address);
2281
2282 __split_huge_pmd(vma, pmd, address, freeze, page);
2283}
2284
2285void vma_adjust_trans_huge(struct vm_area_struct *vma,
2286 unsigned long start,
2287 unsigned long end,
2288 long adjust_next)
2289{
2290 /*
2291 * If the new start address isn't hpage aligned and it could
2292 * previously contain an hugepage: check if we need to split
2293 * an huge pmd.
2294 */
2295 if (start & ~HPAGE_PMD_MASK &&
2296 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2297 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2298 split_huge_pmd_address(vma, start, false, NULL);
2299
2300 /*
2301 * If the new end address isn't hpage aligned and it could
2302 * previously contain an hugepage: check if we need to split
2303 * an huge pmd.
2304 */
2305 if (end & ~HPAGE_PMD_MASK &&
2306 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2307 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2308 split_huge_pmd_address(vma, end, false, NULL);
2309
2310 /*
2311 * If we're also updating the vma->vm_next->vm_start, if the new
2312 * vm_next->vm_start isn't page aligned and it could previously
2313 * contain an hugepage: check if we need to split an huge pmd.
2314 */
2315 if (adjust_next > 0) {
2316 struct vm_area_struct *next = vma->vm_next;
2317 unsigned long nstart = next->vm_start;
2318 nstart += adjust_next << PAGE_SHIFT;
2319 if (nstart & ~HPAGE_PMD_MASK &&
2320 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2321 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2322 split_huge_pmd_address(next, nstart, false, NULL);
2323 }
2324}
2325
2326static void freeze_page(struct page *page)
2327{
2328 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2329 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2330 bool unmap_success;
2331
2332 VM_BUG_ON_PAGE(!PageHead(page), page);
2333
2334 if (PageAnon(page))
2335 ttu_flags |= TTU_SPLIT_FREEZE;
2336
2337 unmap_success = try_to_unmap(page, ttu_flags);
2338 VM_BUG_ON_PAGE(!unmap_success, page);
2339}
2340
2341static void unfreeze_page(struct page *page)
2342{
2343 int i;
2344 if (PageTransHuge(page)) {
2345 remove_migration_ptes(page, page, true);
2346 } else {
2347 for (i = 0; i < HPAGE_PMD_NR; i++)
2348 remove_migration_ptes(page + i, page + i, true);
2349 }
2350}
2351
2352static void __split_huge_page_tail(struct page *head, int tail,
2353 struct lruvec *lruvec, struct list_head *list)
2354{
2355 struct page *page_tail = head + tail;
2356
2357 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2358
2359 /*
2360 * Clone page flags before unfreezing refcount.
2361 *
2362 * After successful get_page_unless_zero() might follow flags change,
2363 * for exmaple lock_page() which set PG_waiters.
2364 */
2365 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2366 page_tail->flags |= (head->flags &
2367 ((1L << PG_referenced) |
2368 (1L << PG_swapbacked) |
2369 (1L << PG_swapcache) |
2370 (1L << PG_mlocked) |
2371 (1L << PG_uptodate) |
2372 (1L << PG_active) |
2373 (1L << PG_locked) |
2374 (1L << PG_unevictable) |
2375 (1L << PG_dirty)));
2376
2377 /* Page flags must be visible before we make the page non-compound. */
2378 smp_wmb();
2379
2380 /*
2381 * Clear PageTail before unfreezing page refcount.
2382 *
2383 * After successful get_page_unless_zero() might follow put_page()
2384 * which needs correct compound_head().
2385 */
2386 clear_compound_head(page_tail);
2387
2388 /* Finally unfreeze refcount. Additional reference from page cache. */
2389 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2390 PageSwapCache(head)));
2391
2392 if (page_is_young(head))
2393 set_page_young(page_tail);
2394 if (page_is_idle(head))
2395 set_page_idle(page_tail);
2396
2397 /* ->mapping in first tail page is compound_mapcount */
2398 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2399 page_tail);
2400 page_tail->mapping = head->mapping;
2401
2402 page_tail->index = head->index + tail;
2403 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2404
2405 /*
2406 * always add to the tail because some iterators expect new
2407 * pages to show after the currently processed elements - e.g.
2408 * migrate_pages
2409 */
2410 lru_add_page_tail(head, page_tail, lruvec, list);
2411}
2412
2413static void __split_huge_page(struct page *page, struct list_head *list,
2414 unsigned long flags)
2415{
2416 struct page *head = compound_head(page);
2417 struct zone *zone = page_zone(head);
2418 struct lruvec *lruvec;
2419 pgoff_t end = -1;
2420 int i;
2421
2422 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2423
2424 /* complete memcg works before add pages to LRU */
2425 mem_cgroup_split_huge_fixup(head);
2426
2427 if (!PageAnon(page))
2428 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2429
2430 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2431 __split_huge_page_tail(head, i, lruvec, list);
2432 /* Some pages can be beyond i_size: drop them from page cache */
2433 if (head[i].index >= end) {
2434 ClearPageDirty(head + i);
2435 __delete_from_page_cache(head + i, NULL);
2436 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2437 shmem_uncharge(head->mapping->host, 1);
2438 put_page(head + i);
2439 }
2440 }
2441
2442 ClearPageCompound(head);
2443 /* See comment in __split_huge_page_tail() */
2444 if (PageAnon(head)) {
2445 /* Additional pin to radix tree of swap cache */
2446 if (PageSwapCache(head))
2447 page_ref_add(head, 2);
2448 else
2449 page_ref_inc(head);
2450 } else {
2451 /* Additional pin to radix tree */
2452 page_ref_add(head, 2);
2453 xa_unlock(&head->mapping->i_pages);
2454 }
2455
2456 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2457
2458 unfreeze_page(head);
2459
2460 for (i = 0; i < HPAGE_PMD_NR; i++) {
2461 struct page *subpage = head + i;
2462 if (subpage == page)
2463 continue;
2464 unlock_page(subpage);
2465
2466 /*
2467 * Subpages may be freed if there wasn't any mapping
2468 * like if add_to_swap() is running on a lru page that
2469 * had its mapping zapped. And freeing these pages
2470 * requires taking the lru_lock so we do the put_page
2471 * of the tail pages after the split is complete.
2472 */
2473 put_page(subpage);
2474 }
2475}
2476
2477int total_mapcount(struct page *page)
2478{
2479 int i, compound, ret;
2480
2481 VM_BUG_ON_PAGE(PageTail(page), page);
2482
2483 if (likely(!PageCompound(page)))
2484 return atomic_read(&page->_mapcount) + 1;
2485
2486 compound = compound_mapcount(page);
2487 if (PageHuge(page))
2488 return compound;
2489 ret = compound;
2490 for (i = 0; i < HPAGE_PMD_NR; i++)
2491 ret += atomic_read(&page[i]._mapcount) + 1;
2492 /* File pages has compound_mapcount included in _mapcount */
2493 if (!PageAnon(page))
2494 return ret - compound * HPAGE_PMD_NR;
2495 if (PageDoubleMap(page))
2496 ret -= HPAGE_PMD_NR;
2497 return ret;
2498}
2499
2500/*
2501 * This calculates accurately how many mappings a transparent hugepage
2502 * has (unlike page_mapcount() which isn't fully accurate). This full
2503 * accuracy is primarily needed to know if copy-on-write faults can
2504 * reuse the page and change the mapping to read-write instead of
2505 * copying them. At the same time this returns the total_mapcount too.
2506 *
2507 * The function returns the highest mapcount any one of the subpages
2508 * has. If the return value is one, even if different processes are
2509 * mapping different subpages of the transparent hugepage, they can
2510 * all reuse it, because each process is reusing a different subpage.
2511 *
2512 * The total_mapcount is instead counting all virtual mappings of the
2513 * subpages. If the total_mapcount is equal to "one", it tells the
2514 * caller all mappings belong to the same "mm" and in turn the
2515 * anon_vma of the transparent hugepage can become the vma->anon_vma
2516 * local one as no other process may be mapping any of the subpages.
2517 *
2518 * It would be more accurate to replace page_mapcount() with
2519 * page_trans_huge_mapcount(), however we only use
2520 * page_trans_huge_mapcount() in the copy-on-write faults where we
2521 * need full accuracy to avoid breaking page pinning, because
2522 * page_trans_huge_mapcount() is slower than page_mapcount().
2523 */
2524int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2525{
2526 int i, ret, _total_mapcount, mapcount;
2527
2528 /* hugetlbfs shouldn't call it */
2529 VM_BUG_ON_PAGE(PageHuge(page), page);
2530
2531 if (likely(!PageTransCompound(page))) {
2532 mapcount = atomic_read(&page->_mapcount) + 1;
2533 if (total_mapcount)
2534 *total_mapcount = mapcount;
2535 return mapcount;
2536 }
2537
2538 page = compound_head(page);
2539
2540 _total_mapcount = ret = 0;
2541 for (i = 0; i < HPAGE_PMD_NR; i++) {
2542 mapcount = atomic_read(&page[i]._mapcount) + 1;
2543 ret = max(ret, mapcount);
2544 _total_mapcount += mapcount;
2545 }
2546 if (PageDoubleMap(page)) {
2547 ret -= 1;
2548 _total_mapcount -= HPAGE_PMD_NR;
2549 }
2550 mapcount = compound_mapcount(page);
2551 ret += mapcount;
2552 _total_mapcount += mapcount;
2553 if (total_mapcount)
2554 *total_mapcount = _total_mapcount;
2555 return ret;
2556}
2557
2558/* Racy check whether the huge page can be split */
2559bool can_split_huge_page(struct page *page, int *pextra_pins)
2560{
2561 int extra_pins;
2562
2563 /* Additional pins from radix tree */
2564 if (PageAnon(page))
2565 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2566 else
2567 extra_pins = HPAGE_PMD_NR;
2568 if (pextra_pins)
2569 *pextra_pins = extra_pins;
2570 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2571}
2572
2573/*
2574 * This function splits huge page into normal pages. @page can point to any
2575 * subpage of huge page to split. Split doesn't change the position of @page.
2576 *
2577 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2578 * The huge page must be locked.
2579 *
2580 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2581 *
2582 * Both head page and tail pages will inherit mapping, flags, and so on from
2583 * the hugepage.
2584 *
2585 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2586 * they are not mapped.
2587 *
2588 * Returns 0 if the hugepage is split successfully.
2589 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2590 * us.
2591 */
2592int split_huge_page_to_list(struct page *page, struct list_head *list)
2593{
2594 struct page *head = compound_head(page);
2595 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2596 struct anon_vma *anon_vma = NULL;
2597 struct address_space *mapping = NULL;
2598 int count, mapcount, extra_pins, ret;
2599 bool mlocked;
2600 unsigned long flags;
2601
2602 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2603 VM_BUG_ON_PAGE(!PageLocked(page), page);
2604 VM_BUG_ON_PAGE(!PageCompound(page), page);
2605
2606 if (PageWriteback(page))
2607 return -EBUSY;
2608
2609 if (PageAnon(head)) {
2610 /*
2611 * The caller does not necessarily hold an mmap_sem that would
2612 * prevent the anon_vma disappearing so we first we take a
2613 * reference to it and then lock the anon_vma for write. This
2614 * is similar to page_lock_anon_vma_read except the write lock
2615 * is taken to serialise against parallel split or collapse
2616 * operations.
2617 */
2618 anon_vma = page_get_anon_vma(head);
2619 if (!anon_vma) {
2620 ret = -EBUSY;
2621 goto out;
2622 }
2623 mapping = NULL;
2624 anon_vma_lock_write(anon_vma);
2625 } else {
2626 mapping = head->mapping;
2627
2628 /* Truncated ? */
2629 if (!mapping) {
2630 ret = -EBUSY;
2631 goto out;
2632 }
2633
2634 anon_vma = NULL;
2635 i_mmap_lock_read(mapping);
2636 }
2637
2638 /*
2639 * Racy check if we can split the page, before freeze_page() will
2640 * split PMDs
2641 */
2642 if (!can_split_huge_page(head, &extra_pins)) {
2643 ret = -EBUSY;
2644 goto out_unlock;
2645 }
2646
2647 mlocked = PageMlocked(page);
2648 freeze_page(head);
2649 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2650
2651 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2652 if (mlocked)
2653 lru_add_drain();
2654
2655 /* prevent PageLRU to go away from under us, and freeze lru stats */
2656 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2657
2658 if (mapping) {
2659 void **pslot;
2660
2661 xa_lock(&mapping->i_pages);
2662 pslot = radix_tree_lookup_slot(&mapping->i_pages,
2663 page_index(head));
2664 /*
2665 * Check if the head page is present in radix tree.
2666 * We assume all tail are present too, if head is there.
2667 */
2668 if (radix_tree_deref_slot_protected(pslot,
2669 &mapping->i_pages.xa_lock) != head)
2670 goto fail;
2671 }
2672
2673 /* Prevent deferred_split_scan() touching ->_refcount */
2674 spin_lock(&pgdata->split_queue_lock);
2675 count = page_count(head);
2676 mapcount = total_mapcount(head);
2677 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2678 if (!list_empty(page_deferred_list(head))) {
2679 pgdata->split_queue_len--;
2680 list_del(page_deferred_list(head));
2681 }
2682 if (mapping)
2683 __dec_node_page_state(page, NR_SHMEM_THPS);
2684 spin_unlock(&pgdata->split_queue_lock);
2685 __split_huge_page(page, list, flags);
2686 if (PageSwapCache(head)) {
2687 swp_entry_t entry = { .val = page_private(head) };
2688
2689 ret = split_swap_cluster(entry);
2690 } else
2691 ret = 0;
2692 } else {
2693 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2694 pr_alert("total_mapcount: %u, page_count(): %u\n",
2695 mapcount, count);
2696 if (PageTail(page))
2697 dump_page(head, NULL);
2698 dump_page(page, "total_mapcount(head) > 0");
2699 BUG();
2700 }
2701 spin_unlock(&pgdata->split_queue_lock);
2702fail: if (mapping)
2703 xa_unlock(&mapping->i_pages);
2704 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2705 unfreeze_page(head);
2706 ret = -EBUSY;
2707 }
2708
2709out_unlock:
2710 if (anon_vma) {
2711 anon_vma_unlock_write(anon_vma);
2712 put_anon_vma(anon_vma);
2713 }
2714 if (mapping)
2715 i_mmap_unlock_read(mapping);
2716out:
2717 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2718 return ret;
2719}
2720
2721void free_transhuge_page(struct page *page)
2722{
2723 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2724 unsigned long flags;
2725
2726 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2727 if (!list_empty(page_deferred_list(page))) {
2728 pgdata->split_queue_len--;
2729 list_del(page_deferred_list(page));
2730 }
2731 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2732 free_compound_page(page);
2733}
2734
2735void deferred_split_huge_page(struct page *page)
2736{
2737 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2738 unsigned long flags;
2739
2740 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2741
2742 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2743 if (list_empty(page_deferred_list(page))) {
2744 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2745 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2746 pgdata->split_queue_len++;
2747 }
2748 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2749}
2750
2751static unsigned long deferred_split_count(struct shrinker *shrink,
2752 struct shrink_control *sc)
2753{
2754 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2755 return READ_ONCE(pgdata->split_queue_len);
2756}
2757
2758static unsigned long deferred_split_scan(struct shrinker *shrink,
2759 struct shrink_control *sc)
2760{
2761 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2762 unsigned long flags;
2763 LIST_HEAD(list), *pos, *next;
2764 struct page *page;
2765 int split = 0;
2766
2767 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2768 /* Take pin on all head pages to avoid freeing them under us */
2769 list_for_each_safe(pos, next, &pgdata->split_queue) {
2770 page = list_entry((void *)pos, struct page, mapping);
2771 page = compound_head(page);
2772 if (get_page_unless_zero(page)) {
2773 list_move(page_deferred_list(page), &list);
2774 } else {
2775 /* We lost race with put_compound_page() */
2776 list_del_init(page_deferred_list(page));
2777 pgdata->split_queue_len--;
2778 }
2779 if (!--sc->nr_to_scan)
2780 break;
2781 }
2782 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2783
2784 list_for_each_safe(pos, next, &list) {
2785 page = list_entry((void *)pos, struct page, mapping);
2786 if (!trylock_page(page))
2787 goto next;
2788 /* split_huge_page() removes page from list on success */
2789 if (!split_huge_page(page))
2790 split++;
2791 unlock_page(page);
2792next:
2793 put_page(page);
2794 }
2795
2796 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2797 list_splice_tail(&list, &pgdata->split_queue);
2798 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2799
2800 /*
2801 * Stop shrinker if we didn't split any page, but the queue is empty.
2802 * This can happen if pages were freed under us.
2803 */
2804 if (!split && list_empty(&pgdata->split_queue))
2805 return SHRINK_STOP;
2806 return split;
2807}
2808
2809static struct shrinker deferred_split_shrinker = {
2810 .count_objects = deferred_split_count,
2811 .scan_objects = deferred_split_scan,
2812 .seeks = DEFAULT_SEEKS,
2813 .flags = SHRINKER_NUMA_AWARE,
2814};
2815
2816#ifdef CONFIG_DEBUG_FS
2817static int split_huge_pages_set(void *data, u64 val)
2818{
2819 struct zone *zone;
2820 struct page *page;
2821 unsigned long pfn, max_zone_pfn;
2822 unsigned long total = 0, split = 0;
2823
2824 if (val != 1)
2825 return -EINVAL;
2826
2827 for_each_populated_zone(zone) {
2828 max_zone_pfn = zone_end_pfn(zone);
2829 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2830 if (!pfn_valid(pfn))
2831 continue;
2832
2833 page = pfn_to_page(pfn);
2834 if (!get_page_unless_zero(page))
2835 continue;
2836
2837 if (zone != page_zone(page))
2838 goto next;
2839
2840 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2841 goto next;
2842
2843 total++;
2844 lock_page(page);
2845 if (!split_huge_page(page))
2846 split++;
2847 unlock_page(page);
2848next:
2849 put_page(page);
2850 }
2851 }
2852
2853 pr_info("%lu of %lu THP split\n", split, total);
2854
2855 return 0;
2856}
2857DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2858 "%llu\n");
2859
2860static int __init split_huge_pages_debugfs(void)
2861{
2862 void *ret;
2863
2864 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2865 &split_huge_pages_fops);
2866 if (!ret)
2867 pr_warn("Failed to create split_huge_pages in debugfs");
2868 return 0;
2869}
2870late_initcall(split_huge_pages_debugfs);
2871#endif
2872
2873#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2874void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2875 struct page *page)
2876{
2877 struct vm_area_struct *vma = pvmw->vma;
2878 struct mm_struct *mm = vma->vm_mm;
2879 unsigned long address = pvmw->address;
2880 pmd_t pmdval;
2881 swp_entry_t entry;
2882 pmd_t pmdswp;
2883
2884 if (!(pvmw->pmd && !pvmw->pte))
2885 return;
2886
2887 mmu_notifier_invalidate_range_start(mm, address,
2888 address + HPAGE_PMD_SIZE);
2889
2890 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2891 pmdval = *pvmw->pmd;
2892 pmdp_invalidate(vma, address, pvmw->pmd);
2893 if (pmd_dirty(pmdval))
2894 set_page_dirty(page);
2895 entry = make_migration_entry(page, pmd_write(pmdval));
2896 pmdswp = swp_entry_to_pmd(entry);
2897 if (pmd_soft_dirty(pmdval))
2898 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2899 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2900 page_remove_rmap(page, true);
2901 put_page(page);
2902
2903 mmu_notifier_invalidate_range_end(mm, address,
2904 address + HPAGE_PMD_SIZE);
2905}
2906
2907void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2908{
2909 struct vm_area_struct *vma = pvmw->vma;
2910 struct mm_struct *mm = vma->vm_mm;
2911 unsigned long address = pvmw->address;
2912 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2913 pmd_t pmde;
2914 swp_entry_t entry;
2915
2916 if (!(pvmw->pmd && !pvmw->pte))
2917 return;
2918
2919 entry = pmd_to_swp_entry(*pvmw->pmd);
2920 get_page(new);
2921 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2922 if (pmd_swp_soft_dirty(*pvmw->pmd))
2923 pmde = pmd_mksoft_dirty(pmde);
2924 if (is_write_migration_entry(entry))
2925 pmde = maybe_pmd_mkwrite(pmde, vma);
2926
2927 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2928 if (PageAnon(new))
2929 page_add_anon_rmap(new, vma, mmun_start, true);
2930 else
2931 page_add_file_rmap(new, true);
2932 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2933 if (vma->vm_flags & VM_LOCKED)
2934 mlock_vma_page(new);
2935 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2936}
2937#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2009 Red Hat, Inc.
4 */
5
6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/sched/mm.h>
11#include <linux/sched/coredump.h>
12#include <linux/sched/numa_balancing.h>
13#include <linux/highmem.h>
14#include <linux/hugetlb.h>
15#include <linux/mmu_notifier.h>
16#include <linux/rmap.h>
17#include <linux/swap.h>
18#include <linux/shrinker.h>
19#include <linux/mm_inline.h>
20#include <linux/swapops.h>
21#include <linux/dax.h>
22#include <linux/khugepaged.h>
23#include <linux/freezer.h>
24#include <linux/pfn_t.h>
25#include <linux/mman.h>
26#include <linux/memremap.h>
27#include <linux/pagemap.h>
28#include <linux/debugfs.h>
29#include <linux/migrate.h>
30#include <linux/hashtable.h>
31#include <linux/userfaultfd_k.h>
32#include <linux/page_idle.h>
33#include <linux/shmem_fs.h>
34#include <linux/oom.h>
35#include <linux/numa.h>
36#include <linux/page_owner.h>
37
38#include <asm/tlb.h>
39#include <asm/pgalloc.h>
40#include "internal.h"
41
42/*
43 * By default, transparent hugepage support is disabled in order to avoid
44 * risking an increased memory footprint for applications that are not
45 * guaranteed to benefit from it. When transparent hugepage support is
46 * enabled, it is for all mappings, and khugepaged scans all mappings.
47 * Defrag is invoked by khugepaged hugepage allocations and by page faults
48 * for all hugepage allocations.
49 */
50unsigned long transparent_hugepage_flags __read_mostly =
51#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
52 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
53#endif
54#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
55 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
56#endif
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
59 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
60
61static struct shrinker deferred_split_shrinker;
62
63static atomic_t huge_zero_refcount;
64struct page *huge_zero_page __read_mostly;
65unsigned long huge_zero_pfn __read_mostly = ~0UL;
66
67static inline bool file_thp_enabled(struct vm_area_struct *vma)
68{
69 return transhuge_vma_enabled(vma, vma->vm_flags) && vma->vm_file &&
70 !inode_is_open_for_write(vma->vm_file->f_inode) &&
71 (vma->vm_flags & VM_EXEC);
72}
73
74bool transparent_hugepage_active(struct vm_area_struct *vma)
75{
76 /* The addr is used to check if the vma size fits */
77 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
78
79 if (!transhuge_vma_suitable(vma, addr))
80 return false;
81 if (vma_is_anonymous(vma))
82 return __transparent_hugepage_enabled(vma);
83 if (vma_is_shmem(vma))
84 return shmem_huge_enabled(vma);
85 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
86 return file_thp_enabled(vma);
87
88 return false;
89}
90
91static bool get_huge_zero_page(void)
92{
93 struct page *zero_page;
94retry:
95 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
96 return true;
97
98 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
99 HPAGE_PMD_ORDER);
100 if (!zero_page) {
101 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
102 return false;
103 }
104 count_vm_event(THP_ZERO_PAGE_ALLOC);
105 preempt_disable();
106 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
107 preempt_enable();
108 __free_pages(zero_page, compound_order(zero_page));
109 goto retry;
110 }
111 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
112
113 /* We take additional reference here. It will be put back by shrinker */
114 atomic_set(&huge_zero_refcount, 2);
115 preempt_enable();
116 return true;
117}
118
119static void put_huge_zero_page(void)
120{
121 /*
122 * Counter should never go to zero here. Only shrinker can put
123 * last reference.
124 */
125 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
126}
127
128struct page *mm_get_huge_zero_page(struct mm_struct *mm)
129{
130 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
131 return READ_ONCE(huge_zero_page);
132
133 if (!get_huge_zero_page())
134 return NULL;
135
136 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
137 put_huge_zero_page();
138
139 return READ_ONCE(huge_zero_page);
140}
141
142void mm_put_huge_zero_page(struct mm_struct *mm)
143{
144 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
145 put_huge_zero_page();
146}
147
148static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
149 struct shrink_control *sc)
150{
151 /* we can free zero page only if last reference remains */
152 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
153}
154
155static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
156 struct shrink_control *sc)
157{
158 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
159 struct page *zero_page = xchg(&huge_zero_page, NULL);
160 BUG_ON(zero_page == NULL);
161 WRITE_ONCE(huge_zero_pfn, ~0UL);
162 __free_pages(zero_page, compound_order(zero_page));
163 return HPAGE_PMD_NR;
164 }
165
166 return 0;
167}
168
169static struct shrinker huge_zero_page_shrinker = {
170 .count_objects = shrink_huge_zero_page_count,
171 .scan_objects = shrink_huge_zero_page_scan,
172 .seeks = DEFAULT_SEEKS,
173};
174
175#ifdef CONFIG_SYSFS
176static ssize_t enabled_show(struct kobject *kobj,
177 struct kobj_attribute *attr, char *buf)
178{
179 const char *output;
180
181 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
182 output = "[always] madvise never";
183 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
184 &transparent_hugepage_flags))
185 output = "always [madvise] never";
186 else
187 output = "always madvise [never]";
188
189 return sysfs_emit(buf, "%s\n", output);
190}
191
192static ssize_t enabled_store(struct kobject *kobj,
193 struct kobj_attribute *attr,
194 const char *buf, size_t count)
195{
196 ssize_t ret = count;
197
198 if (sysfs_streq(buf, "always")) {
199 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
200 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
201 } else if (sysfs_streq(buf, "madvise")) {
202 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
203 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
204 } else if (sysfs_streq(buf, "never")) {
205 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
206 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
207 } else
208 ret = -EINVAL;
209
210 if (ret > 0) {
211 int err = start_stop_khugepaged();
212 if (err)
213 ret = err;
214 }
215 return ret;
216}
217static struct kobj_attribute enabled_attr =
218 __ATTR(enabled, 0644, enabled_show, enabled_store);
219
220ssize_t single_hugepage_flag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf,
222 enum transparent_hugepage_flag flag)
223{
224 return sysfs_emit(buf, "%d\n",
225 !!test_bit(flag, &transparent_hugepage_flags));
226}
227
228ssize_t single_hugepage_flag_store(struct kobject *kobj,
229 struct kobj_attribute *attr,
230 const char *buf, size_t count,
231 enum transparent_hugepage_flag flag)
232{
233 unsigned long value;
234 int ret;
235
236 ret = kstrtoul(buf, 10, &value);
237 if (ret < 0)
238 return ret;
239 if (value > 1)
240 return -EINVAL;
241
242 if (value)
243 set_bit(flag, &transparent_hugepage_flags);
244 else
245 clear_bit(flag, &transparent_hugepage_flags);
246
247 return count;
248}
249
250static ssize_t defrag_show(struct kobject *kobj,
251 struct kobj_attribute *attr, char *buf)
252{
253 const char *output;
254
255 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
256 &transparent_hugepage_flags))
257 output = "[always] defer defer+madvise madvise never";
258 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
259 &transparent_hugepage_flags))
260 output = "always [defer] defer+madvise madvise never";
261 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
262 &transparent_hugepage_flags))
263 output = "always defer [defer+madvise] madvise never";
264 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
265 &transparent_hugepage_flags))
266 output = "always defer defer+madvise [madvise] never";
267 else
268 output = "always defer defer+madvise madvise [never]";
269
270 return sysfs_emit(buf, "%s\n", output);
271}
272
273static ssize_t defrag_store(struct kobject *kobj,
274 struct kobj_attribute *attr,
275 const char *buf, size_t count)
276{
277 if (sysfs_streq(buf, "always")) {
278 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
279 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
280 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
281 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
282 } else if (sysfs_streq(buf, "defer+madvise")) {
283 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
284 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
285 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
286 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
287 } else if (sysfs_streq(buf, "defer")) {
288 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
289 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
290 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
291 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
292 } else if (sysfs_streq(buf, "madvise")) {
293 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
294 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
295 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
296 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
297 } else if (sysfs_streq(buf, "never")) {
298 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
299 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
300 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
301 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
302 } else
303 return -EINVAL;
304
305 return count;
306}
307static struct kobj_attribute defrag_attr =
308 __ATTR(defrag, 0644, defrag_show, defrag_store);
309
310static ssize_t use_zero_page_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf)
312{
313 return single_hugepage_flag_show(kobj, attr, buf,
314 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
315}
316static ssize_t use_zero_page_store(struct kobject *kobj,
317 struct kobj_attribute *attr, const char *buf, size_t count)
318{
319 return single_hugepage_flag_store(kobj, attr, buf, count,
320 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
321}
322static struct kobj_attribute use_zero_page_attr =
323 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
324
325static ssize_t hpage_pmd_size_show(struct kobject *kobj,
326 struct kobj_attribute *attr, char *buf)
327{
328 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
329}
330static struct kobj_attribute hpage_pmd_size_attr =
331 __ATTR_RO(hpage_pmd_size);
332
333static struct attribute *hugepage_attr[] = {
334 &enabled_attr.attr,
335 &defrag_attr.attr,
336 &use_zero_page_attr.attr,
337 &hpage_pmd_size_attr.attr,
338#ifdef CONFIG_SHMEM
339 &shmem_enabled_attr.attr,
340#endif
341 NULL,
342};
343
344static const struct attribute_group hugepage_attr_group = {
345 .attrs = hugepage_attr,
346};
347
348static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
349{
350 int err;
351
352 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
353 if (unlikely(!*hugepage_kobj)) {
354 pr_err("failed to create transparent hugepage kobject\n");
355 return -ENOMEM;
356 }
357
358 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
359 if (err) {
360 pr_err("failed to register transparent hugepage group\n");
361 goto delete_obj;
362 }
363
364 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
365 if (err) {
366 pr_err("failed to register transparent hugepage group\n");
367 goto remove_hp_group;
368 }
369
370 return 0;
371
372remove_hp_group:
373 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
374delete_obj:
375 kobject_put(*hugepage_kobj);
376 return err;
377}
378
379static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
380{
381 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
382 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
383 kobject_put(hugepage_kobj);
384}
385#else
386static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
387{
388 return 0;
389}
390
391static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
392{
393}
394#endif /* CONFIG_SYSFS */
395
396static int __init hugepage_init(void)
397{
398 int err;
399 struct kobject *hugepage_kobj;
400
401 if (!has_transparent_hugepage()) {
402 /*
403 * Hardware doesn't support hugepages, hence disable
404 * DAX PMD support.
405 */
406 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
407 return -EINVAL;
408 }
409
410 /*
411 * hugepages can't be allocated by the buddy allocator
412 */
413 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
414 /*
415 * we use page->mapping and page->index in second tail page
416 * as list_head: assuming THP order >= 2
417 */
418 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
419
420 err = hugepage_init_sysfs(&hugepage_kobj);
421 if (err)
422 goto err_sysfs;
423
424 err = khugepaged_init();
425 if (err)
426 goto err_slab;
427
428 err = register_shrinker(&huge_zero_page_shrinker);
429 if (err)
430 goto err_hzp_shrinker;
431 err = register_shrinker(&deferred_split_shrinker);
432 if (err)
433 goto err_split_shrinker;
434
435 /*
436 * By default disable transparent hugepages on smaller systems,
437 * where the extra memory used could hurt more than TLB overhead
438 * is likely to save. The admin can still enable it through /sys.
439 */
440 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
441 transparent_hugepage_flags = 0;
442 return 0;
443 }
444
445 err = start_stop_khugepaged();
446 if (err)
447 goto err_khugepaged;
448
449 return 0;
450err_khugepaged:
451 unregister_shrinker(&deferred_split_shrinker);
452err_split_shrinker:
453 unregister_shrinker(&huge_zero_page_shrinker);
454err_hzp_shrinker:
455 khugepaged_destroy();
456err_slab:
457 hugepage_exit_sysfs(hugepage_kobj);
458err_sysfs:
459 return err;
460}
461subsys_initcall(hugepage_init);
462
463static int __init setup_transparent_hugepage(char *str)
464{
465 int ret = 0;
466 if (!str)
467 goto out;
468 if (!strcmp(str, "always")) {
469 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
470 &transparent_hugepage_flags);
471 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
472 &transparent_hugepage_flags);
473 ret = 1;
474 } else if (!strcmp(str, "madvise")) {
475 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
476 &transparent_hugepage_flags);
477 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
478 &transparent_hugepage_flags);
479 ret = 1;
480 } else if (!strcmp(str, "never")) {
481 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
482 &transparent_hugepage_flags);
483 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
484 &transparent_hugepage_flags);
485 ret = 1;
486 }
487out:
488 if (!ret)
489 pr_warn("transparent_hugepage= cannot parse, ignored\n");
490 return ret;
491}
492__setup("transparent_hugepage=", setup_transparent_hugepage);
493
494pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
495{
496 if (likely(vma->vm_flags & VM_WRITE))
497 pmd = pmd_mkwrite(pmd);
498 return pmd;
499}
500
501#ifdef CONFIG_MEMCG
502static inline struct deferred_split *get_deferred_split_queue(struct page *page)
503{
504 struct mem_cgroup *memcg = page_memcg(compound_head(page));
505 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
506
507 if (memcg)
508 return &memcg->deferred_split_queue;
509 else
510 return &pgdat->deferred_split_queue;
511}
512#else
513static inline struct deferred_split *get_deferred_split_queue(struct page *page)
514{
515 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
516
517 return &pgdat->deferred_split_queue;
518}
519#endif
520
521void prep_transhuge_page(struct page *page)
522{
523 /*
524 * we use page->mapping and page->indexlru in second tail page
525 * as list_head: assuming THP order >= 2
526 */
527
528 INIT_LIST_HEAD(page_deferred_list(page));
529 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
530}
531
532bool is_transparent_hugepage(struct page *page)
533{
534 if (!PageCompound(page))
535 return false;
536
537 page = compound_head(page);
538 return is_huge_zero_page(page) ||
539 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
540}
541EXPORT_SYMBOL_GPL(is_transparent_hugepage);
542
543static unsigned long __thp_get_unmapped_area(struct file *filp,
544 unsigned long addr, unsigned long len,
545 loff_t off, unsigned long flags, unsigned long size)
546{
547 loff_t off_end = off + len;
548 loff_t off_align = round_up(off, size);
549 unsigned long len_pad, ret;
550
551 if (off_end <= off_align || (off_end - off_align) < size)
552 return 0;
553
554 len_pad = len + size;
555 if (len_pad < len || (off + len_pad) < off)
556 return 0;
557
558 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
559 off >> PAGE_SHIFT, flags);
560
561 /*
562 * The failure might be due to length padding. The caller will retry
563 * without the padding.
564 */
565 if (IS_ERR_VALUE(ret))
566 return 0;
567
568 /*
569 * Do not try to align to THP boundary if allocation at the address
570 * hint succeeds.
571 */
572 if (ret == addr)
573 return addr;
574
575 ret += (off - ret) & (size - 1);
576 return ret;
577}
578
579unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
580 unsigned long len, unsigned long pgoff, unsigned long flags)
581{
582 unsigned long ret;
583 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
584
585 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
586 goto out;
587
588 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
589 if (ret)
590 return ret;
591out:
592 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
593}
594EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
595
596static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
597 struct page *page, gfp_t gfp)
598{
599 struct vm_area_struct *vma = vmf->vma;
600 pgtable_t pgtable;
601 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
602 vm_fault_t ret = 0;
603
604 VM_BUG_ON_PAGE(!PageCompound(page), page);
605
606 if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
607 put_page(page);
608 count_vm_event(THP_FAULT_FALLBACK);
609 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
610 return VM_FAULT_FALLBACK;
611 }
612 cgroup_throttle_swaprate(page, gfp);
613
614 pgtable = pte_alloc_one(vma->vm_mm);
615 if (unlikely(!pgtable)) {
616 ret = VM_FAULT_OOM;
617 goto release;
618 }
619
620 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
621 /*
622 * The memory barrier inside __SetPageUptodate makes sure that
623 * clear_huge_page writes become visible before the set_pmd_at()
624 * write.
625 */
626 __SetPageUptodate(page);
627
628 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
629 if (unlikely(!pmd_none(*vmf->pmd))) {
630 goto unlock_release;
631 } else {
632 pmd_t entry;
633
634 ret = check_stable_address_space(vma->vm_mm);
635 if (ret)
636 goto unlock_release;
637
638 /* Deliver the page fault to userland */
639 if (userfaultfd_missing(vma)) {
640 spin_unlock(vmf->ptl);
641 put_page(page);
642 pte_free(vma->vm_mm, pgtable);
643 ret = handle_userfault(vmf, VM_UFFD_MISSING);
644 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
645 return ret;
646 }
647
648 entry = mk_huge_pmd(page, vma->vm_page_prot);
649 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
650 page_add_new_anon_rmap(page, vma, haddr, true);
651 lru_cache_add_inactive_or_unevictable(page, vma);
652 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
653 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
654 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
655 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
656 mm_inc_nr_ptes(vma->vm_mm);
657 spin_unlock(vmf->ptl);
658 count_vm_event(THP_FAULT_ALLOC);
659 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
660 }
661
662 return 0;
663unlock_release:
664 spin_unlock(vmf->ptl);
665release:
666 if (pgtable)
667 pte_free(vma->vm_mm, pgtable);
668 put_page(page);
669 return ret;
670
671}
672
673/*
674 * always: directly stall for all thp allocations
675 * defer: wake kswapd and fail if not immediately available
676 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
677 * fail if not immediately available
678 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
679 * available
680 * never: never stall for any thp allocation
681 */
682gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
683{
684 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
685
686 /* Always do synchronous compaction */
687 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
688 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
689
690 /* Kick kcompactd and fail quickly */
691 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
692 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
693
694 /* Synchronous compaction if madvised, otherwise kick kcompactd */
695 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
696 return GFP_TRANSHUGE_LIGHT |
697 (vma_madvised ? __GFP_DIRECT_RECLAIM :
698 __GFP_KSWAPD_RECLAIM);
699
700 /* Only do synchronous compaction if madvised */
701 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
702 return GFP_TRANSHUGE_LIGHT |
703 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
704
705 return GFP_TRANSHUGE_LIGHT;
706}
707
708/* Caller must hold page table lock. */
709static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
710 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
711 struct page *zero_page)
712{
713 pmd_t entry;
714 if (!pmd_none(*pmd))
715 return;
716 entry = mk_pmd(zero_page, vma->vm_page_prot);
717 entry = pmd_mkhuge(entry);
718 if (pgtable)
719 pgtable_trans_huge_deposit(mm, pmd, pgtable);
720 set_pmd_at(mm, haddr, pmd, entry);
721 mm_inc_nr_ptes(mm);
722}
723
724vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
725{
726 struct vm_area_struct *vma = vmf->vma;
727 gfp_t gfp;
728 struct page *page;
729 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
730
731 if (!transhuge_vma_suitable(vma, haddr))
732 return VM_FAULT_FALLBACK;
733 if (unlikely(anon_vma_prepare(vma)))
734 return VM_FAULT_OOM;
735 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
736 return VM_FAULT_OOM;
737 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
738 !mm_forbids_zeropage(vma->vm_mm) &&
739 transparent_hugepage_use_zero_page()) {
740 pgtable_t pgtable;
741 struct page *zero_page;
742 vm_fault_t ret;
743 pgtable = pte_alloc_one(vma->vm_mm);
744 if (unlikely(!pgtable))
745 return VM_FAULT_OOM;
746 zero_page = mm_get_huge_zero_page(vma->vm_mm);
747 if (unlikely(!zero_page)) {
748 pte_free(vma->vm_mm, pgtable);
749 count_vm_event(THP_FAULT_FALLBACK);
750 return VM_FAULT_FALLBACK;
751 }
752 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
753 ret = 0;
754 if (pmd_none(*vmf->pmd)) {
755 ret = check_stable_address_space(vma->vm_mm);
756 if (ret) {
757 spin_unlock(vmf->ptl);
758 pte_free(vma->vm_mm, pgtable);
759 } else if (userfaultfd_missing(vma)) {
760 spin_unlock(vmf->ptl);
761 pte_free(vma->vm_mm, pgtable);
762 ret = handle_userfault(vmf, VM_UFFD_MISSING);
763 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
764 } else {
765 set_huge_zero_page(pgtable, vma->vm_mm, vma,
766 haddr, vmf->pmd, zero_page);
767 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
768 spin_unlock(vmf->ptl);
769 }
770 } else {
771 spin_unlock(vmf->ptl);
772 pte_free(vma->vm_mm, pgtable);
773 }
774 return ret;
775 }
776 gfp = vma_thp_gfp_mask(vma);
777 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
778 if (unlikely(!page)) {
779 count_vm_event(THP_FAULT_FALLBACK);
780 return VM_FAULT_FALLBACK;
781 }
782 prep_transhuge_page(page);
783 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
784}
785
786static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
787 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
788 pgtable_t pgtable)
789{
790 struct mm_struct *mm = vma->vm_mm;
791 pmd_t entry;
792 spinlock_t *ptl;
793
794 ptl = pmd_lock(mm, pmd);
795 if (!pmd_none(*pmd)) {
796 if (write) {
797 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
798 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
799 goto out_unlock;
800 }
801 entry = pmd_mkyoung(*pmd);
802 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
803 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
804 update_mmu_cache_pmd(vma, addr, pmd);
805 }
806
807 goto out_unlock;
808 }
809
810 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
811 if (pfn_t_devmap(pfn))
812 entry = pmd_mkdevmap(entry);
813 if (write) {
814 entry = pmd_mkyoung(pmd_mkdirty(entry));
815 entry = maybe_pmd_mkwrite(entry, vma);
816 }
817
818 if (pgtable) {
819 pgtable_trans_huge_deposit(mm, pmd, pgtable);
820 mm_inc_nr_ptes(mm);
821 pgtable = NULL;
822 }
823
824 set_pmd_at(mm, addr, pmd, entry);
825 update_mmu_cache_pmd(vma, addr, pmd);
826
827out_unlock:
828 spin_unlock(ptl);
829 if (pgtable)
830 pte_free(mm, pgtable);
831}
832
833/**
834 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
835 * @vmf: Structure describing the fault
836 * @pfn: pfn to insert
837 * @pgprot: page protection to use
838 * @write: whether it's a write fault
839 *
840 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
841 * also consult the vmf_insert_mixed_prot() documentation when
842 * @pgprot != @vmf->vma->vm_page_prot.
843 *
844 * Return: vm_fault_t value.
845 */
846vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
847 pgprot_t pgprot, bool write)
848{
849 unsigned long addr = vmf->address & PMD_MASK;
850 struct vm_area_struct *vma = vmf->vma;
851 pgtable_t pgtable = NULL;
852
853 /*
854 * If we had pmd_special, we could avoid all these restrictions,
855 * but we need to be consistent with PTEs and architectures that
856 * can't support a 'special' bit.
857 */
858 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
859 !pfn_t_devmap(pfn));
860 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
861 (VM_PFNMAP|VM_MIXEDMAP));
862 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
863
864 if (addr < vma->vm_start || addr >= vma->vm_end)
865 return VM_FAULT_SIGBUS;
866
867 if (arch_needs_pgtable_deposit()) {
868 pgtable = pte_alloc_one(vma->vm_mm);
869 if (!pgtable)
870 return VM_FAULT_OOM;
871 }
872
873 track_pfn_insert(vma, &pgprot, pfn);
874
875 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
876 return VM_FAULT_NOPAGE;
877}
878EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
879
880#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
881static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
882{
883 if (likely(vma->vm_flags & VM_WRITE))
884 pud = pud_mkwrite(pud);
885 return pud;
886}
887
888static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
889 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
890{
891 struct mm_struct *mm = vma->vm_mm;
892 pud_t entry;
893 spinlock_t *ptl;
894
895 ptl = pud_lock(mm, pud);
896 if (!pud_none(*pud)) {
897 if (write) {
898 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
899 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
900 goto out_unlock;
901 }
902 entry = pud_mkyoung(*pud);
903 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
904 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
905 update_mmu_cache_pud(vma, addr, pud);
906 }
907 goto out_unlock;
908 }
909
910 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
911 if (pfn_t_devmap(pfn))
912 entry = pud_mkdevmap(entry);
913 if (write) {
914 entry = pud_mkyoung(pud_mkdirty(entry));
915 entry = maybe_pud_mkwrite(entry, vma);
916 }
917 set_pud_at(mm, addr, pud, entry);
918 update_mmu_cache_pud(vma, addr, pud);
919
920out_unlock:
921 spin_unlock(ptl);
922}
923
924/**
925 * vmf_insert_pfn_pud_prot - insert a pud size pfn
926 * @vmf: Structure describing the fault
927 * @pfn: pfn to insert
928 * @pgprot: page protection to use
929 * @write: whether it's a write fault
930 *
931 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
932 * also consult the vmf_insert_mixed_prot() documentation when
933 * @pgprot != @vmf->vma->vm_page_prot.
934 *
935 * Return: vm_fault_t value.
936 */
937vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
938 pgprot_t pgprot, bool write)
939{
940 unsigned long addr = vmf->address & PUD_MASK;
941 struct vm_area_struct *vma = vmf->vma;
942
943 /*
944 * If we had pud_special, we could avoid all these restrictions,
945 * but we need to be consistent with PTEs and architectures that
946 * can't support a 'special' bit.
947 */
948 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
949 !pfn_t_devmap(pfn));
950 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
951 (VM_PFNMAP|VM_MIXEDMAP));
952 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
953
954 if (addr < vma->vm_start || addr >= vma->vm_end)
955 return VM_FAULT_SIGBUS;
956
957 track_pfn_insert(vma, &pgprot, pfn);
958
959 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
960 return VM_FAULT_NOPAGE;
961}
962EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
963#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
964
965static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
966 pmd_t *pmd, int flags)
967{
968 pmd_t _pmd;
969
970 _pmd = pmd_mkyoung(*pmd);
971 if (flags & FOLL_WRITE)
972 _pmd = pmd_mkdirty(_pmd);
973 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
974 pmd, _pmd, flags & FOLL_WRITE))
975 update_mmu_cache_pmd(vma, addr, pmd);
976}
977
978struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
979 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
980{
981 unsigned long pfn = pmd_pfn(*pmd);
982 struct mm_struct *mm = vma->vm_mm;
983 struct page *page;
984
985 assert_spin_locked(pmd_lockptr(mm, pmd));
986
987 /*
988 * When we COW a devmap PMD entry, we split it into PTEs, so we should
989 * not be in this function with `flags & FOLL_COW` set.
990 */
991 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
992
993 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
994 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
995 (FOLL_PIN | FOLL_GET)))
996 return NULL;
997
998 if (flags & FOLL_WRITE && !pmd_write(*pmd))
999 return NULL;
1000
1001 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1002 /* pass */;
1003 else
1004 return NULL;
1005
1006 if (flags & FOLL_TOUCH)
1007 touch_pmd(vma, addr, pmd, flags);
1008
1009 /*
1010 * device mapped pages can only be returned if the
1011 * caller will manage the page reference count.
1012 */
1013 if (!(flags & (FOLL_GET | FOLL_PIN)))
1014 return ERR_PTR(-EEXIST);
1015
1016 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1017 *pgmap = get_dev_pagemap(pfn, *pgmap);
1018 if (!*pgmap)
1019 return ERR_PTR(-EFAULT);
1020 page = pfn_to_page(pfn);
1021 if (!try_grab_page(page, flags))
1022 page = ERR_PTR(-ENOMEM);
1023
1024 return page;
1025}
1026
1027int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1028 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1029 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1030{
1031 spinlock_t *dst_ptl, *src_ptl;
1032 struct page *src_page;
1033 pmd_t pmd;
1034 pgtable_t pgtable = NULL;
1035 int ret = -ENOMEM;
1036
1037 /* Skip if can be re-fill on fault */
1038 if (!vma_is_anonymous(dst_vma))
1039 return 0;
1040
1041 pgtable = pte_alloc_one(dst_mm);
1042 if (unlikely(!pgtable))
1043 goto out;
1044
1045 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1046 src_ptl = pmd_lockptr(src_mm, src_pmd);
1047 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1048
1049 ret = -EAGAIN;
1050 pmd = *src_pmd;
1051
1052#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1053 if (unlikely(is_swap_pmd(pmd))) {
1054 swp_entry_t entry = pmd_to_swp_entry(pmd);
1055
1056 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1057 if (is_writable_migration_entry(entry)) {
1058 entry = make_readable_migration_entry(
1059 swp_offset(entry));
1060 pmd = swp_entry_to_pmd(entry);
1061 if (pmd_swp_soft_dirty(*src_pmd))
1062 pmd = pmd_swp_mksoft_dirty(pmd);
1063 if (pmd_swp_uffd_wp(*src_pmd))
1064 pmd = pmd_swp_mkuffd_wp(pmd);
1065 set_pmd_at(src_mm, addr, src_pmd, pmd);
1066 }
1067 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1068 mm_inc_nr_ptes(dst_mm);
1069 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1070 if (!userfaultfd_wp(dst_vma))
1071 pmd = pmd_swp_clear_uffd_wp(pmd);
1072 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1073 ret = 0;
1074 goto out_unlock;
1075 }
1076#endif
1077
1078 if (unlikely(!pmd_trans_huge(pmd))) {
1079 pte_free(dst_mm, pgtable);
1080 goto out_unlock;
1081 }
1082 /*
1083 * When page table lock is held, the huge zero pmd should not be
1084 * under splitting since we don't split the page itself, only pmd to
1085 * a page table.
1086 */
1087 if (is_huge_zero_pmd(pmd)) {
1088 /*
1089 * get_huge_zero_page() will never allocate a new page here,
1090 * since we already have a zero page to copy. It just takes a
1091 * reference.
1092 */
1093 mm_get_huge_zero_page(dst_mm);
1094 goto out_zero_page;
1095 }
1096
1097 src_page = pmd_page(pmd);
1098 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1099
1100 /*
1101 * If this page is a potentially pinned page, split and retry the fault
1102 * with smaller page size. Normally this should not happen because the
1103 * userspace should use MADV_DONTFORK upon pinned regions. This is a
1104 * best effort that the pinned pages won't be replaced by another
1105 * random page during the coming copy-on-write.
1106 */
1107 if (unlikely(page_needs_cow_for_dma(src_vma, src_page))) {
1108 pte_free(dst_mm, pgtable);
1109 spin_unlock(src_ptl);
1110 spin_unlock(dst_ptl);
1111 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1112 return -EAGAIN;
1113 }
1114
1115 get_page(src_page);
1116 page_dup_rmap(src_page, true);
1117 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118out_zero_page:
1119 mm_inc_nr_ptes(dst_mm);
1120 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1121 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1122 if (!userfaultfd_wp(dst_vma))
1123 pmd = pmd_clear_uffd_wp(pmd);
1124 pmd = pmd_mkold(pmd_wrprotect(pmd));
1125 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1126
1127 ret = 0;
1128out_unlock:
1129 spin_unlock(src_ptl);
1130 spin_unlock(dst_ptl);
1131out:
1132 return ret;
1133}
1134
1135#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1136static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1137 pud_t *pud, int flags)
1138{
1139 pud_t _pud;
1140
1141 _pud = pud_mkyoung(*pud);
1142 if (flags & FOLL_WRITE)
1143 _pud = pud_mkdirty(_pud);
1144 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1145 pud, _pud, flags & FOLL_WRITE))
1146 update_mmu_cache_pud(vma, addr, pud);
1147}
1148
1149struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1150 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1151{
1152 unsigned long pfn = pud_pfn(*pud);
1153 struct mm_struct *mm = vma->vm_mm;
1154 struct page *page;
1155
1156 assert_spin_locked(pud_lockptr(mm, pud));
1157
1158 if (flags & FOLL_WRITE && !pud_write(*pud))
1159 return NULL;
1160
1161 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1162 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1163 (FOLL_PIN | FOLL_GET)))
1164 return NULL;
1165
1166 if (pud_present(*pud) && pud_devmap(*pud))
1167 /* pass */;
1168 else
1169 return NULL;
1170
1171 if (flags & FOLL_TOUCH)
1172 touch_pud(vma, addr, pud, flags);
1173
1174 /*
1175 * device mapped pages can only be returned if the
1176 * caller will manage the page reference count.
1177 *
1178 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1179 */
1180 if (!(flags & (FOLL_GET | FOLL_PIN)))
1181 return ERR_PTR(-EEXIST);
1182
1183 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1184 *pgmap = get_dev_pagemap(pfn, *pgmap);
1185 if (!*pgmap)
1186 return ERR_PTR(-EFAULT);
1187 page = pfn_to_page(pfn);
1188 if (!try_grab_page(page, flags))
1189 page = ERR_PTR(-ENOMEM);
1190
1191 return page;
1192}
1193
1194int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1195 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1196 struct vm_area_struct *vma)
1197{
1198 spinlock_t *dst_ptl, *src_ptl;
1199 pud_t pud;
1200 int ret;
1201
1202 dst_ptl = pud_lock(dst_mm, dst_pud);
1203 src_ptl = pud_lockptr(src_mm, src_pud);
1204 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1205
1206 ret = -EAGAIN;
1207 pud = *src_pud;
1208 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1209 goto out_unlock;
1210
1211 /*
1212 * When page table lock is held, the huge zero pud should not be
1213 * under splitting since we don't split the page itself, only pud to
1214 * a page table.
1215 */
1216 if (is_huge_zero_pud(pud)) {
1217 /* No huge zero pud yet */
1218 }
1219
1220 /* Please refer to comments in copy_huge_pmd() */
1221 if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1222 spin_unlock(src_ptl);
1223 spin_unlock(dst_ptl);
1224 __split_huge_pud(vma, src_pud, addr);
1225 return -EAGAIN;
1226 }
1227
1228 pudp_set_wrprotect(src_mm, addr, src_pud);
1229 pud = pud_mkold(pud_wrprotect(pud));
1230 set_pud_at(dst_mm, addr, dst_pud, pud);
1231
1232 ret = 0;
1233out_unlock:
1234 spin_unlock(src_ptl);
1235 spin_unlock(dst_ptl);
1236 return ret;
1237}
1238
1239void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1240{
1241 pud_t entry;
1242 unsigned long haddr;
1243 bool write = vmf->flags & FAULT_FLAG_WRITE;
1244
1245 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1246 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1247 goto unlock;
1248
1249 entry = pud_mkyoung(orig_pud);
1250 if (write)
1251 entry = pud_mkdirty(entry);
1252 haddr = vmf->address & HPAGE_PUD_MASK;
1253 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1254 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1255
1256unlock:
1257 spin_unlock(vmf->ptl);
1258}
1259#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1260
1261void huge_pmd_set_accessed(struct vm_fault *vmf)
1262{
1263 pmd_t entry;
1264 unsigned long haddr;
1265 bool write = vmf->flags & FAULT_FLAG_WRITE;
1266 pmd_t orig_pmd = vmf->orig_pmd;
1267
1268 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1269 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1270 goto unlock;
1271
1272 entry = pmd_mkyoung(orig_pmd);
1273 if (write)
1274 entry = pmd_mkdirty(entry);
1275 haddr = vmf->address & HPAGE_PMD_MASK;
1276 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1277 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1278
1279unlock:
1280 spin_unlock(vmf->ptl);
1281}
1282
1283vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1284{
1285 struct vm_area_struct *vma = vmf->vma;
1286 struct page *page;
1287 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1288 pmd_t orig_pmd = vmf->orig_pmd;
1289
1290 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1291 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1292
1293 if (is_huge_zero_pmd(orig_pmd))
1294 goto fallback;
1295
1296 spin_lock(vmf->ptl);
1297
1298 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1299 spin_unlock(vmf->ptl);
1300 return 0;
1301 }
1302
1303 page = pmd_page(orig_pmd);
1304 VM_BUG_ON_PAGE(!PageHead(page), page);
1305
1306 /* Lock page for reuse_swap_page() */
1307 if (!trylock_page(page)) {
1308 get_page(page);
1309 spin_unlock(vmf->ptl);
1310 lock_page(page);
1311 spin_lock(vmf->ptl);
1312 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1313 spin_unlock(vmf->ptl);
1314 unlock_page(page);
1315 put_page(page);
1316 return 0;
1317 }
1318 put_page(page);
1319 }
1320
1321 /*
1322 * We can only reuse the page if nobody else maps the huge page or it's
1323 * part.
1324 */
1325 if (reuse_swap_page(page, NULL)) {
1326 pmd_t entry;
1327 entry = pmd_mkyoung(orig_pmd);
1328 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1329 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1330 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1331 unlock_page(page);
1332 spin_unlock(vmf->ptl);
1333 return VM_FAULT_WRITE;
1334 }
1335
1336 unlock_page(page);
1337 spin_unlock(vmf->ptl);
1338fallback:
1339 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1340 return VM_FAULT_FALLBACK;
1341}
1342
1343/*
1344 * FOLL_FORCE can write to even unwritable pmd's, but only
1345 * after we've gone through a COW cycle and they are dirty.
1346 */
1347static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1348{
1349 return pmd_write(pmd) ||
1350 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1351}
1352
1353struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1354 unsigned long addr,
1355 pmd_t *pmd,
1356 unsigned int flags)
1357{
1358 struct mm_struct *mm = vma->vm_mm;
1359 struct page *page = NULL;
1360
1361 assert_spin_locked(pmd_lockptr(mm, pmd));
1362
1363 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1364 goto out;
1365
1366 /* Avoid dumping huge zero page */
1367 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1368 return ERR_PTR(-EFAULT);
1369
1370 /* Full NUMA hinting faults to serialise migration in fault paths */
1371 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1372 goto out;
1373
1374 page = pmd_page(*pmd);
1375 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1376
1377 if (!try_grab_page(page, flags))
1378 return ERR_PTR(-ENOMEM);
1379
1380 if (flags & FOLL_TOUCH)
1381 touch_pmd(vma, addr, pmd, flags);
1382
1383 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1384 /*
1385 * We don't mlock() pte-mapped THPs. This way we can avoid
1386 * leaking mlocked pages into non-VM_LOCKED VMAs.
1387 *
1388 * For anon THP:
1389 *
1390 * In most cases the pmd is the only mapping of the page as we
1391 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1392 * writable private mappings in populate_vma_page_range().
1393 *
1394 * The only scenario when we have the page shared here is if we
1395 * mlocking read-only mapping shared over fork(). We skip
1396 * mlocking such pages.
1397 *
1398 * For file THP:
1399 *
1400 * We can expect PageDoubleMap() to be stable under page lock:
1401 * for file pages we set it in page_add_file_rmap(), which
1402 * requires page to be locked.
1403 */
1404
1405 if (PageAnon(page) && compound_mapcount(page) != 1)
1406 goto skip_mlock;
1407 if (PageDoubleMap(page) || !page->mapping)
1408 goto skip_mlock;
1409 if (!trylock_page(page))
1410 goto skip_mlock;
1411 if (page->mapping && !PageDoubleMap(page))
1412 mlock_vma_page(page);
1413 unlock_page(page);
1414 }
1415skip_mlock:
1416 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1417 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1418
1419out:
1420 return page;
1421}
1422
1423/* NUMA hinting page fault entry point for trans huge pmds */
1424vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1425{
1426 struct vm_area_struct *vma = vmf->vma;
1427 pmd_t oldpmd = vmf->orig_pmd;
1428 pmd_t pmd;
1429 struct page *page;
1430 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1431 int page_nid = NUMA_NO_NODE;
1432 int target_nid, last_cpupid = -1;
1433 bool migrated = false;
1434 bool was_writable = pmd_savedwrite(oldpmd);
1435 int flags = 0;
1436
1437 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1438 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1439 spin_unlock(vmf->ptl);
1440 goto out;
1441 }
1442
1443 /*
1444 * Since we took the NUMA fault, we must have observed the !accessible
1445 * bit. Make sure all other CPUs agree with that, to avoid them
1446 * modifying the page we're about to migrate.
1447 *
1448 * Must be done under PTL such that we'll observe the relevant
1449 * inc_tlb_flush_pending().
1450 *
1451 * We are not sure a pending tlb flush here is for a huge page
1452 * mapping or not. Hence use the tlb range variant
1453 */
1454 if (mm_tlb_flush_pending(vma->vm_mm)) {
1455 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1456 /*
1457 * change_huge_pmd() released the pmd lock before
1458 * invalidating the secondary MMUs sharing the primary
1459 * MMU pagetables (with ->invalidate_range()). The
1460 * mmu_notifier_invalidate_range_end() (which
1461 * internally calls ->invalidate_range()) in
1462 * change_pmd_range() will run after us, so we can't
1463 * rely on it here and we need an explicit invalidate.
1464 */
1465 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1466 haddr + HPAGE_PMD_SIZE);
1467 }
1468
1469 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1470 page = vm_normal_page_pmd(vma, haddr, pmd);
1471 if (!page)
1472 goto out_map;
1473
1474 /* See similar comment in do_numa_page for explanation */
1475 if (!was_writable)
1476 flags |= TNF_NO_GROUP;
1477
1478 page_nid = page_to_nid(page);
1479 last_cpupid = page_cpupid_last(page);
1480 target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1481 &flags);
1482
1483 if (target_nid == NUMA_NO_NODE) {
1484 put_page(page);
1485 goto out_map;
1486 }
1487
1488 spin_unlock(vmf->ptl);
1489
1490 migrated = migrate_misplaced_page(page, vma, target_nid);
1491 if (migrated) {
1492 flags |= TNF_MIGRATED;
1493 page_nid = target_nid;
1494 } else {
1495 flags |= TNF_MIGRATE_FAIL;
1496 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1497 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1498 spin_unlock(vmf->ptl);
1499 goto out;
1500 }
1501 goto out_map;
1502 }
1503
1504out:
1505 if (page_nid != NUMA_NO_NODE)
1506 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1507 flags);
1508
1509 return 0;
1510
1511out_map:
1512 /* Restore the PMD */
1513 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1514 pmd = pmd_mkyoung(pmd);
1515 if (was_writable)
1516 pmd = pmd_mkwrite(pmd);
1517 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1518 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1519 spin_unlock(vmf->ptl);
1520 goto out;
1521}
1522
1523/*
1524 * Return true if we do MADV_FREE successfully on entire pmd page.
1525 * Otherwise, return false.
1526 */
1527bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1528 pmd_t *pmd, unsigned long addr, unsigned long next)
1529{
1530 spinlock_t *ptl;
1531 pmd_t orig_pmd;
1532 struct page *page;
1533 struct mm_struct *mm = tlb->mm;
1534 bool ret = false;
1535
1536 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1537
1538 ptl = pmd_trans_huge_lock(pmd, vma);
1539 if (!ptl)
1540 goto out_unlocked;
1541
1542 orig_pmd = *pmd;
1543 if (is_huge_zero_pmd(orig_pmd))
1544 goto out;
1545
1546 if (unlikely(!pmd_present(orig_pmd))) {
1547 VM_BUG_ON(thp_migration_supported() &&
1548 !is_pmd_migration_entry(orig_pmd));
1549 goto out;
1550 }
1551
1552 page = pmd_page(orig_pmd);
1553 /*
1554 * If other processes are mapping this page, we couldn't discard
1555 * the page unless they all do MADV_FREE so let's skip the page.
1556 */
1557 if (total_mapcount(page) != 1)
1558 goto out;
1559
1560 if (!trylock_page(page))
1561 goto out;
1562
1563 /*
1564 * If user want to discard part-pages of THP, split it so MADV_FREE
1565 * will deactivate only them.
1566 */
1567 if (next - addr != HPAGE_PMD_SIZE) {
1568 get_page(page);
1569 spin_unlock(ptl);
1570 split_huge_page(page);
1571 unlock_page(page);
1572 put_page(page);
1573 goto out_unlocked;
1574 }
1575
1576 if (PageDirty(page))
1577 ClearPageDirty(page);
1578 unlock_page(page);
1579
1580 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1581 pmdp_invalidate(vma, addr, pmd);
1582 orig_pmd = pmd_mkold(orig_pmd);
1583 orig_pmd = pmd_mkclean(orig_pmd);
1584
1585 set_pmd_at(mm, addr, pmd, orig_pmd);
1586 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1587 }
1588
1589 mark_page_lazyfree(page);
1590 ret = true;
1591out:
1592 spin_unlock(ptl);
1593out_unlocked:
1594 return ret;
1595}
1596
1597static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1598{
1599 pgtable_t pgtable;
1600
1601 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1602 pte_free(mm, pgtable);
1603 mm_dec_nr_ptes(mm);
1604}
1605
1606int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1607 pmd_t *pmd, unsigned long addr)
1608{
1609 pmd_t orig_pmd;
1610 spinlock_t *ptl;
1611
1612 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1613
1614 ptl = __pmd_trans_huge_lock(pmd, vma);
1615 if (!ptl)
1616 return 0;
1617 /*
1618 * For architectures like ppc64 we look at deposited pgtable
1619 * when calling pmdp_huge_get_and_clear. So do the
1620 * pgtable_trans_huge_withdraw after finishing pmdp related
1621 * operations.
1622 */
1623 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1624 tlb->fullmm);
1625 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1626 if (vma_is_special_huge(vma)) {
1627 if (arch_needs_pgtable_deposit())
1628 zap_deposited_table(tlb->mm, pmd);
1629 spin_unlock(ptl);
1630 } else if (is_huge_zero_pmd(orig_pmd)) {
1631 zap_deposited_table(tlb->mm, pmd);
1632 spin_unlock(ptl);
1633 } else {
1634 struct page *page = NULL;
1635 int flush_needed = 1;
1636
1637 if (pmd_present(orig_pmd)) {
1638 page = pmd_page(orig_pmd);
1639 page_remove_rmap(page, true);
1640 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1641 VM_BUG_ON_PAGE(!PageHead(page), page);
1642 } else if (thp_migration_supported()) {
1643 swp_entry_t entry;
1644
1645 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1646 entry = pmd_to_swp_entry(orig_pmd);
1647 page = pfn_swap_entry_to_page(entry);
1648 flush_needed = 0;
1649 } else
1650 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1651
1652 if (PageAnon(page)) {
1653 zap_deposited_table(tlb->mm, pmd);
1654 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1655 } else {
1656 if (arch_needs_pgtable_deposit())
1657 zap_deposited_table(tlb->mm, pmd);
1658 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1659 }
1660
1661 spin_unlock(ptl);
1662 if (flush_needed)
1663 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1664 }
1665 return 1;
1666}
1667
1668#ifndef pmd_move_must_withdraw
1669static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1670 spinlock_t *old_pmd_ptl,
1671 struct vm_area_struct *vma)
1672{
1673 /*
1674 * With split pmd lock we also need to move preallocated
1675 * PTE page table if new_pmd is on different PMD page table.
1676 *
1677 * We also don't deposit and withdraw tables for file pages.
1678 */
1679 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1680}
1681#endif
1682
1683static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1684{
1685#ifdef CONFIG_MEM_SOFT_DIRTY
1686 if (unlikely(is_pmd_migration_entry(pmd)))
1687 pmd = pmd_swp_mksoft_dirty(pmd);
1688 else if (pmd_present(pmd))
1689 pmd = pmd_mksoft_dirty(pmd);
1690#endif
1691 return pmd;
1692}
1693
1694bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1695 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1696{
1697 spinlock_t *old_ptl, *new_ptl;
1698 pmd_t pmd;
1699 struct mm_struct *mm = vma->vm_mm;
1700 bool force_flush = false;
1701
1702 /*
1703 * The destination pmd shouldn't be established, free_pgtables()
1704 * should have release it.
1705 */
1706 if (WARN_ON(!pmd_none(*new_pmd))) {
1707 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1708 return false;
1709 }
1710
1711 /*
1712 * We don't have to worry about the ordering of src and dst
1713 * ptlocks because exclusive mmap_lock prevents deadlock.
1714 */
1715 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1716 if (old_ptl) {
1717 new_ptl = pmd_lockptr(mm, new_pmd);
1718 if (new_ptl != old_ptl)
1719 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1720 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1721 if (pmd_present(pmd))
1722 force_flush = true;
1723 VM_BUG_ON(!pmd_none(*new_pmd));
1724
1725 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1726 pgtable_t pgtable;
1727 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1728 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1729 }
1730 pmd = move_soft_dirty_pmd(pmd);
1731 set_pmd_at(mm, new_addr, new_pmd, pmd);
1732 if (force_flush)
1733 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1734 if (new_ptl != old_ptl)
1735 spin_unlock(new_ptl);
1736 spin_unlock(old_ptl);
1737 return true;
1738 }
1739 return false;
1740}
1741
1742/*
1743 * Returns
1744 * - 0 if PMD could not be locked
1745 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1746 * or if prot_numa but THP migration is not supported
1747 * - HPAGE_PMD_NR if protections changed and TLB flush necessary
1748 */
1749int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1750 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1751{
1752 struct mm_struct *mm = vma->vm_mm;
1753 spinlock_t *ptl;
1754 pmd_t entry;
1755 bool preserve_write;
1756 int ret;
1757 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1758 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1759 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1760
1761 if (prot_numa && !thp_migration_supported())
1762 return 1;
1763
1764 ptl = __pmd_trans_huge_lock(pmd, vma);
1765 if (!ptl)
1766 return 0;
1767
1768 preserve_write = prot_numa && pmd_write(*pmd);
1769 ret = 1;
1770
1771#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1772 if (is_swap_pmd(*pmd)) {
1773 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1774
1775 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1776 if (is_writable_migration_entry(entry)) {
1777 pmd_t newpmd;
1778 /*
1779 * A protection check is difficult so
1780 * just be safe and disable write
1781 */
1782 entry = make_readable_migration_entry(
1783 swp_offset(entry));
1784 newpmd = swp_entry_to_pmd(entry);
1785 if (pmd_swp_soft_dirty(*pmd))
1786 newpmd = pmd_swp_mksoft_dirty(newpmd);
1787 if (pmd_swp_uffd_wp(*pmd))
1788 newpmd = pmd_swp_mkuffd_wp(newpmd);
1789 set_pmd_at(mm, addr, pmd, newpmd);
1790 }
1791 goto unlock;
1792 }
1793#endif
1794
1795 /*
1796 * Avoid trapping faults against the zero page. The read-only
1797 * data is likely to be read-cached on the local CPU and
1798 * local/remote hits to the zero page are not interesting.
1799 */
1800 if (prot_numa && is_huge_zero_pmd(*pmd))
1801 goto unlock;
1802
1803 if (prot_numa && pmd_protnone(*pmd))
1804 goto unlock;
1805
1806 /*
1807 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1808 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1809 * which is also under mmap_read_lock(mm):
1810 *
1811 * CPU0: CPU1:
1812 * change_huge_pmd(prot_numa=1)
1813 * pmdp_huge_get_and_clear_notify()
1814 * madvise_dontneed()
1815 * zap_pmd_range()
1816 * pmd_trans_huge(*pmd) == 0 (without ptl)
1817 * // skip the pmd
1818 * set_pmd_at();
1819 * // pmd is re-established
1820 *
1821 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1822 * which may break userspace.
1823 *
1824 * pmdp_invalidate() is required to make sure we don't miss
1825 * dirty/young flags set by hardware.
1826 */
1827 entry = pmdp_invalidate(vma, addr, pmd);
1828
1829 entry = pmd_modify(entry, newprot);
1830 if (preserve_write)
1831 entry = pmd_mk_savedwrite(entry);
1832 if (uffd_wp) {
1833 entry = pmd_wrprotect(entry);
1834 entry = pmd_mkuffd_wp(entry);
1835 } else if (uffd_wp_resolve) {
1836 /*
1837 * Leave the write bit to be handled by PF interrupt
1838 * handler, then things like COW could be properly
1839 * handled.
1840 */
1841 entry = pmd_clear_uffd_wp(entry);
1842 }
1843 ret = HPAGE_PMD_NR;
1844 set_pmd_at(mm, addr, pmd, entry);
1845 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1846unlock:
1847 spin_unlock(ptl);
1848 return ret;
1849}
1850
1851/*
1852 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1853 *
1854 * Note that if it returns page table lock pointer, this routine returns without
1855 * unlocking page table lock. So callers must unlock it.
1856 */
1857spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1858{
1859 spinlock_t *ptl;
1860 ptl = pmd_lock(vma->vm_mm, pmd);
1861 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1862 pmd_devmap(*pmd)))
1863 return ptl;
1864 spin_unlock(ptl);
1865 return NULL;
1866}
1867
1868/*
1869 * Returns true if a given pud maps a thp, false otherwise.
1870 *
1871 * Note that if it returns true, this routine returns without unlocking page
1872 * table lock. So callers must unlock it.
1873 */
1874spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1875{
1876 spinlock_t *ptl;
1877
1878 ptl = pud_lock(vma->vm_mm, pud);
1879 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1880 return ptl;
1881 spin_unlock(ptl);
1882 return NULL;
1883}
1884
1885#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1886int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1887 pud_t *pud, unsigned long addr)
1888{
1889 spinlock_t *ptl;
1890
1891 ptl = __pud_trans_huge_lock(pud, vma);
1892 if (!ptl)
1893 return 0;
1894 /*
1895 * For architectures like ppc64 we look at deposited pgtable
1896 * when calling pudp_huge_get_and_clear. So do the
1897 * pgtable_trans_huge_withdraw after finishing pudp related
1898 * operations.
1899 */
1900 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1901 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1902 if (vma_is_special_huge(vma)) {
1903 spin_unlock(ptl);
1904 /* No zero page support yet */
1905 } else {
1906 /* No support for anonymous PUD pages yet */
1907 BUG();
1908 }
1909 return 1;
1910}
1911
1912static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1913 unsigned long haddr)
1914{
1915 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1916 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1917 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1918 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1919
1920 count_vm_event(THP_SPLIT_PUD);
1921
1922 pudp_huge_clear_flush_notify(vma, haddr, pud);
1923}
1924
1925void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1926 unsigned long address)
1927{
1928 spinlock_t *ptl;
1929 struct mmu_notifier_range range;
1930
1931 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1932 address & HPAGE_PUD_MASK,
1933 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1934 mmu_notifier_invalidate_range_start(&range);
1935 ptl = pud_lock(vma->vm_mm, pud);
1936 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1937 goto out;
1938 __split_huge_pud_locked(vma, pud, range.start);
1939
1940out:
1941 spin_unlock(ptl);
1942 /*
1943 * No need to double call mmu_notifier->invalidate_range() callback as
1944 * the above pudp_huge_clear_flush_notify() did already call it.
1945 */
1946 mmu_notifier_invalidate_range_only_end(&range);
1947}
1948#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1949
1950static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1951 unsigned long haddr, pmd_t *pmd)
1952{
1953 struct mm_struct *mm = vma->vm_mm;
1954 pgtable_t pgtable;
1955 pmd_t _pmd;
1956 int i;
1957
1958 /*
1959 * Leave pmd empty until pte is filled note that it is fine to delay
1960 * notification until mmu_notifier_invalidate_range_end() as we are
1961 * replacing a zero pmd write protected page with a zero pte write
1962 * protected page.
1963 *
1964 * See Documentation/vm/mmu_notifier.rst
1965 */
1966 pmdp_huge_clear_flush(vma, haddr, pmd);
1967
1968 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1969 pmd_populate(mm, &_pmd, pgtable);
1970
1971 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1972 pte_t *pte, entry;
1973 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1974 entry = pte_mkspecial(entry);
1975 pte = pte_offset_map(&_pmd, haddr);
1976 VM_BUG_ON(!pte_none(*pte));
1977 set_pte_at(mm, haddr, pte, entry);
1978 pte_unmap(pte);
1979 }
1980 smp_wmb(); /* make pte visible before pmd */
1981 pmd_populate(mm, pmd, pgtable);
1982}
1983
1984static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1985 unsigned long haddr, bool freeze)
1986{
1987 struct mm_struct *mm = vma->vm_mm;
1988 struct page *page;
1989 pgtable_t pgtable;
1990 pmd_t old_pmd, _pmd;
1991 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
1992 unsigned long addr;
1993 int i;
1994
1995 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1996 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1997 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1998 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
1999 && !pmd_devmap(*pmd));
2000
2001 count_vm_event(THP_SPLIT_PMD);
2002
2003 if (!vma_is_anonymous(vma)) {
2004 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2005 /*
2006 * We are going to unmap this huge page. So
2007 * just go ahead and zap it
2008 */
2009 if (arch_needs_pgtable_deposit())
2010 zap_deposited_table(mm, pmd);
2011 if (vma_is_special_huge(vma))
2012 return;
2013 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2014 swp_entry_t entry;
2015
2016 entry = pmd_to_swp_entry(old_pmd);
2017 page = pfn_swap_entry_to_page(entry);
2018 } else {
2019 page = pmd_page(old_pmd);
2020 if (!PageDirty(page) && pmd_dirty(old_pmd))
2021 set_page_dirty(page);
2022 if (!PageReferenced(page) && pmd_young(old_pmd))
2023 SetPageReferenced(page);
2024 page_remove_rmap(page, true);
2025 put_page(page);
2026 }
2027 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2028 return;
2029 }
2030
2031 if (is_huge_zero_pmd(*pmd)) {
2032 /*
2033 * FIXME: Do we want to invalidate secondary mmu by calling
2034 * mmu_notifier_invalidate_range() see comments below inside
2035 * __split_huge_pmd() ?
2036 *
2037 * We are going from a zero huge page write protected to zero
2038 * small page also write protected so it does not seems useful
2039 * to invalidate secondary mmu at this time.
2040 */
2041 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2042 }
2043
2044 /*
2045 * Up to this point the pmd is present and huge and userland has the
2046 * whole access to the hugepage during the split (which happens in
2047 * place). If we overwrite the pmd with the not-huge version pointing
2048 * to the pte here (which of course we could if all CPUs were bug
2049 * free), userland could trigger a small page size TLB miss on the
2050 * small sized TLB while the hugepage TLB entry is still established in
2051 * the huge TLB. Some CPU doesn't like that.
2052 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2053 * 383 on page 105. Intel should be safe but is also warns that it's
2054 * only safe if the permission and cache attributes of the two entries
2055 * loaded in the two TLB is identical (which should be the case here).
2056 * But it is generally safer to never allow small and huge TLB entries
2057 * for the same virtual address to be loaded simultaneously. So instead
2058 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2059 * current pmd notpresent (atomically because here the pmd_trans_huge
2060 * must remain set at all times on the pmd until the split is complete
2061 * for this pmd), then we flush the SMP TLB and finally we write the
2062 * non-huge version of the pmd entry with pmd_populate.
2063 */
2064 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2065
2066 pmd_migration = is_pmd_migration_entry(old_pmd);
2067 if (unlikely(pmd_migration)) {
2068 swp_entry_t entry;
2069
2070 entry = pmd_to_swp_entry(old_pmd);
2071 page = pfn_swap_entry_to_page(entry);
2072 write = is_writable_migration_entry(entry);
2073 young = false;
2074 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2075 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2076 } else {
2077 page = pmd_page(old_pmd);
2078 if (pmd_dirty(old_pmd))
2079 SetPageDirty(page);
2080 write = pmd_write(old_pmd);
2081 young = pmd_young(old_pmd);
2082 soft_dirty = pmd_soft_dirty(old_pmd);
2083 uffd_wp = pmd_uffd_wp(old_pmd);
2084 }
2085 VM_BUG_ON_PAGE(!page_count(page), page);
2086 page_ref_add(page, HPAGE_PMD_NR - 1);
2087
2088 /*
2089 * Withdraw the table only after we mark the pmd entry invalid.
2090 * This's critical for some architectures (Power).
2091 */
2092 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2093 pmd_populate(mm, &_pmd, pgtable);
2094
2095 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2096 pte_t entry, *pte;
2097 /*
2098 * Note that NUMA hinting access restrictions are not
2099 * transferred to avoid any possibility of altering
2100 * permissions across VMAs.
2101 */
2102 if (freeze || pmd_migration) {
2103 swp_entry_t swp_entry;
2104 if (write)
2105 swp_entry = make_writable_migration_entry(
2106 page_to_pfn(page + i));
2107 else
2108 swp_entry = make_readable_migration_entry(
2109 page_to_pfn(page + i));
2110 entry = swp_entry_to_pte(swp_entry);
2111 if (soft_dirty)
2112 entry = pte_swp_mksoft_dirty(entry);
2113 if (uffd_wp)
2114 entry = pte_swp_mkuffd_wp(entry);
2115 } else {
2116 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2117 entry = maybe_mkwrite(entry, vma);
2118 if (!write)
2119 entry = pte_wrprotect(entry);
2120 if (!young)
2121 entry = pte_mkold(entry);
2122 if (soft_dirty)
2123 entry = pte_mksoft_dirty(entry);
2124 if (uffd_wp)
2125 entry = pte_mkuffd_wp(entry);
2126 }
2127 pte = pte_offset_map(&_pmd, addr);
2128 BUG_ON(!pte_none(*pte));
2129 set_pte_at(mm, addr, pte, entry);
2130 if (!pmd_migration)
2131 atomic_inc(&page[i]._mapcount);
2132 pte_unmap(pte);
2133 }
2134
2135 if (!pmd_migration) {
2136 /*
2137 * Set PG_double_map before dropping compound_mapcount to avoid
2138 * false-negative page_mapped().
2139 */
2140 if (compound_mapcount(page) > 1 &&
2141 !TestSetPageDoubleMap(page)) {
2142 for (i = 0; i < HPAGE_PMD_NR; i++)
2143 atomic_inc(&page[i]._mapcount);
2144 }
2145
2146 lock_page_memcg(page);
2147 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2148 /* Last compound_mapcount is gone. */
2149 __mod_lruvec_page_state(page, NR_ANON_THPS,
2150 -HPAGE_PMD_NR);
2151 if (TestClearPageDoubleMap(page)) {
2152 /* No need in mapcount reference anymore */
2153 for (i = 0; i < HPAGE_PMD_NR; i++)
2154 atomic_dec(&page[i]._mapcount);
2155 }
2156 }
2157 unlock_page_memcg(page);
2158 }
2159
2160 smp_wmb(); /* make pte visible before pmd */
2161 pmd_populate(mm, pmd, pgtable);
2162
2163 if (freeze) {
2164 for (i = 0; i < HPAGE_PMD_NR; i++) {
2165 page_remove_rmap(page + i, false);
2166 put_page(page + i);
2167 }
2168 }
2169}
2170
2171void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2172 unsigned long address, bool freeze, struct page *page)
2173{
2174 spinlock_t *ptl;
2175 struct mmu_notifier_range range;
2176 bool do_unlock_page = false;
2177 pmd_t _pmd;
2178
2179 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2180 address & HPAGE_PMD_MASK,
2181 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2182 mmu_notifier_invalidate_range_start(&range);
2183 ptl = pmd_lock(vma->vm_mm, pmd);
2184
2185 /*
2186 * If caller asks to setup a migration entries, we need a page to check
2187 * pmd against. Otherwise we can end up replacing wrong page.
2188 */
2189 VM_BUG_ON(freeze && !page);
2190 if (page) {
2191 VM_WARN_ON_ONCE(!PageLocked(page));
2192 if (page != pmd_page(*pmd))
2193 goto out;
2194 }
2195
2196repeat:
2197 if (pmd_trans_huge(*pmd)) {
2198 if (!page) {
2199 page = pmd_page(*pmd);
2200 /*
2201 * An anonymous page must be locked, to ensure that a
2202 * concurrent reuse_swap_page() sees stable mapcount;
2203 * but reuse_swap_page() is not used on shmem or file,
2204 * and page lock must not be taken when zap_pmd_range()
2205 * calls __split_huge_pmd() while i_mmap_lock is held.
2206 */
2207 if (PageAnon(page)) {
2208 if (unlikely(!trylock_page(page))) {
2209 get_page(page);
2210 _pmd = *pmd;
2211 spin_unlock(ptl);
2212 lock_page(page);
2213 spin_lock(ptl);
2214 if (unlikely(!pmd_same(*pmd, _pmd))) {
2215 unlock_page(page);
2216 put_page(page);
2217 page = NULL;
2218 goto repeat;
2219 }
2220 put_page(page);
2221 }
2222 do_unlock_page = true;
2223 }
2224 }
2225 if (PageMlocked(page))
2226 clear_page_mlock(page);
2227 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2228 goto out;
2229 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2230out:
2231 spin_unlock(ptl);
2232 if (do_unlock_page)
2233 unlock_page(page);
2234 /*
2235 * No need to double call mmu_notifier->invalidate_range() callback.
2236 * They are 3 cases to consider inside __split_huge_pmd_locked():
2237 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2238 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2239 * fault will trigger a flush_notify before pointing to a new page
2240 * (it is fine if the secondary mmu keeps pointing to the old zero
2241 * page in the meantime)
2242 * 3) Split a huge pmd into pte pointing to the same page. No need
2243 * to invalidate secondary tlb entry they are all still valid.
2244 * any further changes to individual pte will notify. So no need
2245 * to call mmu_notifier->invalidate_range()
2246 */
2247 mmu_notifier_invalidate_range_only_end(&range);
2248}
2249
2250void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2251 bool freeze, struct page *page)
2252{
2253 pgd_t *pgd;
2254 p4d_t *p4d;
2255 pud_t *pud;
2256 pmd_t *pmd;
2257
2258 pgd = pgd_offset(vma->vm_mm, address);
2259 if (!pgd_present(*pgd))
2260 return;
2261
2262 p4d = p4d_offset(pgd, address);
2263 if (!p4d_present(*p4d))
2264 return;
2265
2266 pud = pud_offset(p4d, address);
2267 if (!pud_present(*pud))
2268 return;
2269
2270 pmd = pmd_offset(pud, address);
2271
2272 __split_huge_pmd(vma, pmd, address, freeze, page);
2273}
2274
2275static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2276{
2277 /*
2278 * If the new address isn't hpage aligned and it could previously
2279 * contain an hugepage: check if we need to split an huge pmd.
2280 */
2281 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2282 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2283 ALIGN(address, HPAGE_PMD_SIZE)))
2284 split_huge_pmd_address(vma, address, false, NULL);
2285}
2286
2287void vma_adjust_trans_huge(struct vm_area_struct *vma,
2288 unsigned long start,
2289 unsigned long end,
2290 long adjust_next)
2291{
2292 /* Check if we need to split start first. */
2293 split_huge_pmd_if_needed(vma, start);
2294
2295 /* Check if we need to split end next. */
2296 split_huge_pmd_if_needed(vma, end);
2297
2298 /*
2299 * If we're also updating the vma->vm_next->vm_start,
2300 * check if we need to split it.
2301 */
2302 if (adjust_next > 0) {
2303 struct vm_area_struct *next = vma->vm_next;
2304 unsigned long nstart = next->vm_start;
2305 nstart += adjust_next;
2306 split_huge_pmd_if_needed(next, nstart);
2307 }
2308}
2309
2310static void unmap_page(struct page *page)
2311{
2312 enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2313 TTU_SYNC;
2314
2315 VM_BUG_ON_PAGE(!PageHead(page), page);
2316
2317 /*
2318 * Anon pages need migration entries to preserve them, but file
2319 * pages can simply be left unmapped, then faulted back on demand.
2320 * If that is ever changed (perhaps for mlock), update remap_page().
2321 */
2322 if (PageAnon(page))
2323 try_to_migrate(page, ttu_flags);
2324 else
2325 try_to_unmap(page, ttu_flags | TTU_IGNORE_MLOCK);
2326
2327 VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2328}
2329
2330static void remap_page(struct page *page, unsigned int nr)
2331{
2332 int i;
2333
2334 /* If unmap_page() uses try_to_migrate() on file, remove this check */
2335 if (!PageAnon(page))
2336 return;
2337 if (PageTransHuge(page)) {
2338 remove_migration_ptes(page, page, true);
2339 } else {
2340 for (i = 0; i < nr; i++)
2341 remove_migration_ptes(page + i, page + i, true);
2342 }
2343}
2344
2345static void lru_add_page_tail(struct page *head, struct page *tail,
2346 struct lruvec *lruvec, struct list_head *list)
2347{
2348 VM_BUG_ON_PAGE(!PageHead(head), head);
2349 VM_BUG_ON_PAGE(PageCompound(tail), head);
2350 VM_BUG_ON_PAGE(PageLRU(tail), head);
2351 lockdep_assert_held(&lruvec->lru_lock);
2352
2353 if (list) {
2354 /* page reclaim is reclaiming a huge page */
2355 VM_WARN_ON(PageLRU(head));
2356 get_page(tail);
2357 list_add_tail(&tail->lru, list);
2358 } else {
2359 /* head is still on lru (and we have it frozen) */
2360 VM_WARN_ON(!PageLRU(head));
2361 SetPageLRU(tail);
2362 list_add_tail(&tail->lru, &head->lru);
2363 }
2364}
2365
2366static void __split_huge_page_tail(struct page *head, int tail,
2367 struct lruvec *lruvec, struct list_head *list)
2368{
2369 struct page *page_tail = head + tail;
2370
2371 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2372
2373 /*
2374 * Clone page flags before unfreezing refcount.
2375 *
2376 * After successful get_page_unless_zero() might follow flags change,
2377 * for example lock_page() which set PG_waiters.
2378 */
2379 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2380 page_tail->flags |= (head->flags &
2381 ((1L << PG_referenced) |
2382 (1L << PG_swapbacked) |
2383 (1L << PG_swapcache) |
2384 (1L << PG_mlocked) |
2385 (1L << PG_uptodate) |
2386 (1L << PG_active) |
2387 (1L << PG_workingset) |
2388 (1L << PG_locked) |
2389 (1L << PG_unevictable) |
2390#ifdef CONFIG_64BIT
2391 (1L << PG_arch_2) |
2392#endif
2393 (1L << PG_dirty)));
2394
2395 /* ->mapping in first tail page is compound_mapcount */
2396 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2397 page_tail);
2398 page_tail->mapping = head->mapping;
2399 page_tail->index = head->index + tail;
2400
2401 /* Page flags must be visible before we make the page non-compound. */
2402 smp_wmb();
2403
2404 /*
2405 * Clear PageTail before unfreezing page refcount.
2406 *
2407 * After successful get_page_unless_zero() might follow put_page()
2408 * which needs correct compound_head().
2409 */
2410 clear_compound_head(page_tail);
2411
2412 /* Finally unfreeze refcount. Additional reference from page cache. */
2413 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2414 PageSwapCache(head)));
2415
2416 if (page_is_young(head))
2417 set_page_young(page_tail);
2418 if (page_is_idle(head))
2419 set_page_idle(page_tail);
2420
2421 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2422
2423 /*
2424 * always add to the tail because some iterators expect new
2425 * pages to show after the currently processed elements - e.g.
2426 * migrate_pages
2427 */
2428 lru_add_page_tail(head, page_tail, lruvec, list);
2429}
2430
2431static void __split_huge_page(struct page *page, struct list_head *list,
2432 pgoff_t end)
2433{
2434 struct page *head = compound_head(page);
2435 struct lruvec *lruvec;
2436 struct address_space *swap_cache = NULL;
2437 unsigned long offset = 0;
2438 unsigned int nr = thp_nr_pages(head);
2439 int i;
2440
2441 /* complete memcg works before add pages to LRU */
2442 split_page_memcg(head, nr);
2443
2444 if (PageAnon(head) && PageSwapCache(head)) {
2445 swp_entry_t entry = { .val = page_private(head) };
2446
2447 offset = swp_offset(entry);
2448 swap_cache = swap_address_space(entry);
2449 xa_lock(&swap_cache->i_pages);
2450 }
2451
2452 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2453 lruvec = lock_page_lruvec(head);
2454
2455 for (i = nr - 1; i >= 1; i--) {
2456 __split_huge_page_tail(head, i, lruvec, list);
2457 /* Some pages can be beyond i_size: drop them from page cache */
2458 if (head[i].index >= end) {
2459 ClearPageDirty(head + i);
2460 __delete_from_page_cache(head + i, NULL);
2461 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2462 shmem_uncharge(head->mapping->host, 1);
2463 put_page(head + i);
2464 } else if (!PageAnon(page)) {
2465 __xa_store(&head->mapping->i_pages, head[i].index,
2466 head + i, 0);
2467 } else if (swap_cache) {
2468 __xa_store(&swap_cache->i_pages, offset + i,
2469 head + i, 0);
2470 }
2471 }
2472
2473 ClearPageCompound(head);
2474 unlock_page_lruvec(lruvec);
2475 /* Caller disabled irqs, so they are still disabled here */
2476
2477 split_page_owner(head, nr);
2478
2479 /* See comment in __split_huge_page_tail() */
2480 if (PageAnon(head)) {
2481 /* Additional pin to swap cache */
2482 if (PageSwapCache(head)) {
2483 page_ref_add(head, 2);
2484 xa_unlock(&swap_cache->i_pages);
2485 } else {
2486 page_ref_inc(head);
2487 }
2488 } else {
2489 /* Additional pin to page cache */
2490 page_ref_add(head, 2);
2491 xa_unlock(&head->mapping->i_pages);
2492 }
2493 local_irq_enable();
2494
2495 remap_page(head, nr);
2496
2497 if (PageSwapCache(head)) {
2498 swp_entry_t entry = { .val = page_private(head) };
2499
2500 split_swap_cluster(entry);
2501 }
2502
2503 for (i = 0; i < nr; i++) {
2504 struct page *subpage = head + i;
2505 if (subpage == page)
2506 continue;
2507 unlock_page(subpage);
2508
2509 /*
2510 * Subpages may be freed if there wasn't any mapping
2511 * like if add_to_swap() is running on a lru page that
2512 * had its mapping zapped. And freeing these pages
2513 * requires taking the lru_lock so we do the put_page
2514 * of the tail pages after the split is complete.
2515 */
2516 put_page(subpage);
2517 }
2518}
2519
2520int total_mapcount(struct page *page)
2521{
2522 int i, compound, nr, ret;
2523
2524 VM_BUG_ON_PAGE(PageTail(page), page);
2525
2526 if (likely(!PageCompound(page)))
2527 return atomic_read(&page->_mapcount) + 1;
2528
2529 compound = compound_mapcount(page);
2530 nr = compound_nr(page);
2531 if (PageHuge(page))
2532 return compound;
2533 ret = compound;
2534 for (i = 0; i < nr; i++)
2535 ret += atomic_read(&page[i]._mapcount) + 1;
2536 /* File pages has compound_mapcount included in _mapcount */
2537 if (!PageAnon(page))
2538 return ret - compound * nr;
2539 if (PageDoubleMap(page))
2540 ret -= nr;
2541 return ret;
2542}
2543
2544/*
2545 * This calculates accurately how many mappings a transparent hugepage
2546 * has (unlike page_mapcount() which isn't fully accurate). This full
2547 * accuracy is primarily needed to know if copy-on-write faults can
2548 * reuse the page and change the mapping to read-write instead of
2549 * copying them. At the same time this returns the total_mapcount too.
2550 *
2551 * The function returns the highest mapcount any one of the subpages
2552 * has. If the return value is one, even if different processes are
2553 * mapping different subpages of the transparent hugepage, they can
2554 * all reuse it, because each process is reusing a different subpage.
2555 *
2556 * The total_mapcount is instead counting all virtual mappings of the
2557 * subpages. If the total_mapcount is equal to "one", it tells the
2558 * caller all mappings belong to the same "mm" and in turn the
2559 * anon_vma of the transparent hugepage can become the vma->anon_vma
2560 * local one as no other process may be mapping any of the subpages.
2561 *
2562 * It would be more accurate to replace page_mapcount() with
2563 * page_trans_huge_mapcount(), however we only use
2564 * page_trans_huge_mapcount() in the copy-on-write faults where we
2565 * need full accuracy to avoid breaking page pinning, because
2566 * page_trans_huge_mapcount() is slower than page_mapcount().
2567 */
2568int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2569{
2570 int i, ret, _total_mapcount, mapcount;
2571
2572 /* hugetlbfs shouldn't call it */
2573 VM_BUG_ON_PAGE(PageHuge(page), page);
2574
2575 if (likely(!PageTransCompound(page))) {
2576 mapcount = atomic_read(&page->_mapcount) + 1;
2577 if (total_mapcount)
2578 *total_mapcount = mapcount;
2579 return mapcount;
2580 }
2581
2582 page = compound_head(page);
2583
2584 _total_mapcount = ret = 0;
2585 for (i = 0; i < thp_nr_pages(page); i++) {
2586 mapcount = atomic_read(&page[i]._mapcount) + 1;
2587 ret = max(ret, mapcount);
2588 _total_mapcount += mapcount;
2589 }
2590 if (PageDoubleMap(page)) {
2591 ret -= 1;
2592 _total_mapcount -= thp_nr_pages(page);
2593 }
2594 mapcount = compound_mapcount(page);
2595 ret += mapcount;
2596 _total_mapcount += mapcount;
2597 if (total_mapcount)
2598 *total_mapcount = _total_mapcount;
2599 return ret;
2600}
2601
2602/* Racy check whether the huge page can be split */
2603bool can_split_huge_page(struct page *page, int *pextra_pins)
2604{
2605 int extra_pins;
2606
2607 /* Additional pins from page cache */
2608 if (PageAnon(page))
2609 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2610 else
2611 extra_pins = thp_nr_pages(page);
2612 if (pextra_pins)
2613 *pextra_pins = extra_pins;
2614 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2615}
2616
2617/*
2618 * This function splits huge page into normal pages. @page can point to any
2619 * subpage of huge page to split. Split doesn't change the position of @page.
2620 *
2621 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2622 * The huge page must be locked.
2623 *
2624 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2625 *
2626 * Both head page and tail pages will inherit mapping, flags, and so on from
2627 * the hugepage.
2628 *
2629 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2630 * they are not mapped.
2631 *
2632 * Returns 0 if the hugepage is split successfully.
2633 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2634 * us.
2635 */
2636int split_huge_page_to_list(struct page *page, struct list_head *list)
2637{
2638 struct page *head = compound_head(page);
2639 struct deferred_split *ds_queue = get_deferred_split_queue(head);
2640 struct anon_vma *anon_vma = NULL;
2641 struct address_space *mapping = NULL;
2642 int extra_pins, ret;
2643 pgoff_t end;
2644
2645 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2646 VM_BUG_ON_PAGE(!PageLocked(head), head);
2647 VM_BUG_ON_PAGE(!PageCompound(head), head);
2648
2649 if (PageWriteback(head))
2650 return -EBUSY;
2651
2652 if (PageAnon(head)) {
2653 /*
2654 * The caller does not necessarily hold an mmap_lock that would
2655 * prevent the anon_vma disappearing so we first we take a
2656 * reference to it and then lock the anon_vma for write. This
2657 * is similar to page_lock_anon_vma_read except the write lock
2658 * is taken to serialise against parallel split or collapse
2659 * operations.
2660 */
2661 anon_vma = page_get_anon_vma(head);
2662 if (!anon_vma) {
2663 ret = -EBUSY;
2664 goto out;
2665 }
2666 end = -1;
2667 mapping = NULL;
2668 anon_vma_lock_write(anon_vma);
2669 } else {
2670 mapping = head->mapping;
2671
2672 /* Truncated ? */
2673 if (!mapping) {
2674 ret = -EBUSY;
2675 goto out;
2676 }
2677
2678 anon_vma = NULL;
2679 i_mmap_lock_read(mapping);
2680
2681 /*
2682 *__split_huge_page() may need to trim off pages beyond EOF:
2683 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2684 * which cannot be nested inside the page tree lock. So note
2685 * end now: i_size itself may be changed at any moment, but
2686 * head page lock is good enough to serialize the trimming.
2687 */
2688 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2689 }
2690
2691 /*
2692 * Racy check if we can split the page, before unmap_page() will
2693 * split PMDs
2694 */
2695 if (!can_split_huge_page(head, &extra_pins)) {
2696 ret = -EBUSY;
2697 goto out_unlock;
2698 }
2699
2700 unmap_page(head);
2701
2702 /* block interrupt reentry in xa_lock and spinlock */
2703 local_irq_disable();
2704 if (mapping) {
2705 XA_STATE(xas, &mapping->i_pages, page_index(head));
2706
2707 /*
2708 * Check if the head page is present in page cache.
2709 * We assume all tail are present too, if head is there.
2710 */
2711 xa_lock(&mapping->i_pages);
2712 if (xas_load(&xas) != head)
2713 goto fail;
2714 }
2715
2716 /* Prevent deferred_split_scan() touching ->_refcount */
2717 spin_lock(&ds_queue->split_queue_lock);
2718 if (page_ref_freeze(head, 1 + extra_pins)) {
2719 if (!list_empty(page_deferred_list(head))) {
2720 ds_queue->split_queue_len--;
2721 list_del(page_deferred_list(head));
2722 }
2723 spin_unlock(&ds_queue->split_queue_lock);
2724 if (mapping) {
2725 int nr = thp_nr_pages(head);
2726
2727 if (PageSwapBacked(head)) {
2728 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2729 -nr);
2730 } else {
2731 __mod_lruvec_page_state(head, NR_FILE_THPS,
2732 -nr);
2733 filemap_nr_thps_dec(mapping);
2734 }
2735 }
2736
2737 __split_huge_page(page, list, end);
2738 ret = 0;
2739 } else {
2740 spin_unlock(&ds_queue->split_queue_lock);
2741fail:
2742 if (mapping)
2743 xa_unlock(&mapping->i_pages);
2744 local_irq_enable();
2745 remap_page(head, thp_nr_pages(head));
2746 ret = -EBUSY;
2747 }
2748
2749out_unlock:
2750 if (anon_vma) {
2751 anon_vma_unlock_write(anon_vma);
2752 put_anon_vma(anon_vma);
2753 }
2754 if (mapping)
2755 i_mmap_unlock_read(mapping);
2756out:
2757 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2758 return ret;
2759}
2760
2761void free_transhuge_page(struct page *page)
2762{
2763 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2764 unsigned long flags;
2765
2766 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2767 if (!list_empty(page_deferred_list(page))) {
2768 ds_queue->split_queue_len--;
2769 list_del(page_deferred_list(page));
2770 }
2771 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2772 free_compound_page(page);
2773}
2774
2775void deferred_split_huge_page(struct page *page)
2776{
2777 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2778#ifdef CONFIG_MEMCG
2779 struct mem_cgroup *memcg = page_memcg(compound_head(page));
2780#endif
2781 unsigned long flags;
2782
2783 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2784
2785 /*
2786 * The try_to_unmap() in page reclaim path might reach here too,
2787 * this may cause a race condition to corrupt deferred split queue.
2788 * And, if page reclaim is already handling the same page, it is
2789 * unnecessary to handle it again in shrinker.
2790 *
2791 * Check PageSwapCache to determine if the page is being
2792 * handled by page reclaim since THP swap would add the page into
2793 * swap cache before calling try_to_unmap().
2794 */
2795 if (PageSwapCache(page))
2796 return;
2797
2798 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2799 if (list_empty(page_deferred_list(page))) {
2800 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2801 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2802 ds_queue->split_queue_len++;
2803#ifdef CONFIG_MEMCG
2804 if (memcg)
2805 set_shrinker_bit(memcg, page_to_nid(page),
2806 deferred_split_shrinker.id);
2807#endif
2808 }
2809 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2810}
2811
2812static unsigned long deferred_split_count(struct shrinker *shrink,
2813 struct shrink_control *sc)
2814{
2815 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2816 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2817
2818#ifdef CONFIG_MEMCG
2819 if (sc->memcg)
2820 ds_queue = &sc->memcg->deferred_split_queue;
2821#endif
2822 return READ_ONCE(ds_queue->split_queue_len);
2823}
2824
2825static unsigned long deferred_split_scan(struct shrinker *shrink,
2826 struct shrink_control *sc)
2827{
2828 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2829 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2830 unsigned long flags;
2831 LIST_HEAD(list), *pos, *next;
2832 struct page *page;
2833 int split = 0;
2834
2835#ifdef CONFIG_MEMCG
2836 if (sc->memcg)
2837 ds_queue = &sc->memcg->deferred_split_queue;
2838#endif
2839
2840 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2841 /* Take pin on all head pages to avoid freeing them under us */
2842 list_for_each_safe(pos, next, &ds_queue->split_queue) {
2843 page = list_entry((void *)pos, struct page, deferred_list);
2844 page = compound_head(page);
2845 if (get_page_unless_zero(page)) {
2846 list_move(page_deferred_list(page), &list);
2847 } else {
2848 /* We lost race with put_compound_page() */
2849 list_del_init(page_deferred_list(page));
2850 ds_queue->split_queue_len--;
2851 }
2852 if (!--sc->nr_to_scan)
2853 break;
2854 }
2855 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2856
2857 list_for_each_safe(pos, next, &list) {
2858 page = list_entry((void *)pos, struct page, deferred_list);
2859 if (!trylock_page(page))
2860 goto next;
2861 /* split_huge_page() removes page from list on success */
2862 if (!split_huge_page(page))
2863 split++;
2864 unlock_page(page);
2865next:
2866 put_page(page);
2867 }
2868
2869 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2870 list_splice_tail(&list, &ds_queue->split_queue);
2871 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2872
2873 /*
2874 * Stop shrinker if we didn't split any page, but the queue is empty.
2875 * This can happen if pages were freed under us.
2876 */
2877 if (!split && list_empty(&ds_queue->split_queue))
2878 return SHRINK_STOP;
2879 return split;
2880}
2881
2882static struct shrinker deferred_split_shrinker = {
2883 .count_objects = deferred_split_count,
2884 .scan_objects = deferred_split_scan,
2885 .seeks = DEFAULT_SEEKS,
2886 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2887 SHRINKER_NONSLAB,
2888};
2889
2890#ifdef CONFIG_DEBUG_FS
2891static void split_huge_pages_all(void)
2892{
2893 struct zone *zone;
2894 struct page *page;
2895 unsigned long pfn, max_zone_pfn;
2896 unsigned long total = 0, split = 0;
2897
2898 pr_debug("Split all THPs\n");
2899 for_each_populated_zone(zone) {
2900 max_zone_pfn = zone_end_pfn(zone);
2901 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2902 if (!pfn_valid(pfn))
2903 continue;
2904
2905 page = pfn_to_page(pfn);
2906 if (!get_page_unless_zero(page))
2907 continue;
2908
2909 if (zone != page_zone(page))
2910 goto next;
2911
2912 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2913 goto next;
2914
2915 total++;
2916 lock_page(page);
2917 if (!split_huge_page(page))
2918 split++;
2919 unlock_page(page);
2920next:
2921 put_page(page);
2922 cond_resched();
2923 }
2924 }
2925
2926 pr_debug("%lu of %lu THP split\n", split, total);
2927}
2928
2929static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2930{
2931 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2932 is_vm_hugetlb_page(vma);
2933}
2934
2935static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2936 unsigned long vaddr_end)
2937{
2938 int ret = 0;
2939 struct task_struct *task;
2940 struct mm_struct *mm;
2941 unsigned long total = 0, split = 0;
2942 unsigned long addr;
2943
2944 vaddr_start &= PAGE_MASK;
2945 vaddr_end &= PAGE_MASK;
2946
2947 /* Find the task_struct from pid */
2948 rcu_read_lock();
2949 task = find_task_by_vpid(pid);
2950 if (!task) {
2951 rcu_read_unlock();
2952 ret = -ESRCH;
2953 goto out;
2954 }
2955 get_task_struct(task);
2956 rcu_read_unlock();
2957
2958 /* Find the mm_struct */
2959 mm = get_task_mm(task);
2960 put_task_struct(task);
2961
2962 if (!mm) {
2963 ret = -EINVAL;
2964 goto out;
2965 }
2966
2967 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2968 pid, vaddr_start, vaddr_end);
2969
2970 mmap_read_lock(mm);
2971 /*
2972 * always increase addr by PAGE_SIZE, since we could have a PTE page
2973 * table filled with PTE-mapped THPs, each of which is distinct.
2974 */
2975 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2976 struct vm_area_struct *vma = find_vma(mm, addr);
2977 unsigned int follflags;
2978 struct page *page;
2979
2980 if (!vma || addr < vma->vm_start)
2981 break;
2982
2983 /* skip special VMA and hugetlb VMA */
2984 if (vma_not_suitable_for_thp_split(vma)) {
2985 addr = vma->vm_end;
2986 continue;
2987 }
2988
2989 /* FOLL_DUMP to ignore special (like zero) pages */
2990 follflags = FOLL_GET | FOLL_DUMP;
2991 page = follow_page(vma, addr, follflags);
2992
2993 if (IS_ERR(page))
2994 continue;
2995 if (!page)
2996 continue;
2997
2998 if (!is_transparent_hugepage(page))
2999 goto next;
3000
3001 total++;
3002 if (!can_split_huge_page(compound_head(page), NULL))
3003 goto next;
3004
3005 if (!trylock_page(page))
3006 goto next;
3007
3008 if (!split_huge_page(page))
3009 split++;
3010
3011 unlock_page(page);
3012next:
3013 put_page(page);
3014 cond_resched();
3015 }
3016 mmap_read_unlock(mm);
3017 mmput(mm);
3018
3019 pr_debug("%lu of %lu THP split\n", split, total);
3020
3021out:
3022 return ret;
3023}
3024
3025static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3026 pgoff_t off_end)
3027{
3028 struct filename *file;
3029 struct file *candidate;
3030 struct address_space *mapping;
3031 int ret = -EINVAL;
3032 pgoff_t index;
3033 int nr_pages = 1;
3034 unsigned long total = 0, split = 0;
3035
3036 file = getname_kernel(file_path);
3037 if (IS_ERR(file))
3038 return ret;
3039
3040 candidate = file_open_name(file, O_RDONLY, 0);
3041 if (IS_ERR(candidate))
3042 goto out;
3043
3044 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3045 file_path, off_start, off_end);
3046
3047 mapping = candidate->f_mapping;
3048
3049 for (index = off_start; index < off_end; index += nr_pages) {
3050 struct page *fpage = pagecache_get_page(mapping, index,
3051 FGP_ENTRY | FGP_HEAD, 0);
3052
3053 nr_pages = 1;
3054 if (xa_is_value(fpage) || !fpage)
3055 continue;
3056
3057 if (!is_transparent_hugepage(fpage))
3058 goto next;
3059
3060 total++;
3061 nr_pages = thp_nr_pages(fpage);
3062
3063 if (!trylock_page(fpage))
3064 goto next;
3065
3066 if (!split_huge_page(fpage))
3067 split++;
3068
3069 unlock_page(fpage);
3070next:
3071 put_page(fpage);
3072 cond_resched();
3073 }
3074
3075 filp_close(candidate, NULL);
3076 ret = 0;
3077
3078 pr_debug("%lu of %lu file-backed THP split\n", split, total);
3079out:
3080 putname(file);
3081 return ret;
3082}
3083
3084#define MAX_INPUT_BUF_SZ 255
3085
3086static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3087 size_t count, loff_t *ppops)
3088{
3089 static DEFINE_MUTEX(split_debug_mutex);
3090 ssize_t ret;
3091 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3092 char input_buf[MAX_INPUT_BUF_SZ];
3093 int pid;
3094 unsigned long vaddr_start, vaddr_end;
3095
3096 ret = mutex_lock_interruptible(&split_debug_mutex);
3097 if (ret)
3098 return ret;
3099
3100 ret = -EFAULT;
3101
3102 memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3103 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3104 goto out;
3105
3106 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3107
3108 if (input_buf[0] == '/') {
3109 char *tok;
3110 char *buf = input_buf;
3111 char file_path[MAX_INPUT_BUF_SZ];
3112 pgoff_t off_start = 0, off_end = 0;
3113 size_t input_len = strlen(input_buf);
3114
3115 tok = strsep(&buf, ",");
3116 if (tok) {
3117 strcpy(file_path, tok);
3118 } else {
3119 ret = -EINVAL;
3120 goto out;
3121 }
3122
3123 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3124 if (ret != 2) {
3125 ret = -EINVAL;
3126 goto out;
3127 }
3128 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3129 if (!ret)
3130 ret = input_len;
3131
3132 goto out;
3133 }
3134
3135 ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3136 if (ret == 1 && pid == 1) {
3137 split_huge_pages_all();
3138 ret = strlen(input_buf);
3139 goto out;
3140 } else if (ret != 3) {
3141 ret = -EINVAL;
3142 goto out;
3143 }
3144
3145 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3146 if (!ret)
3147 ret = strlen(input_buf);
3148out:
3149 mutex_unlock(&split_debug_mutex);
3150 return ret;
3151
3152}
3153
3154static const struct file_operations split_huge_pages_fops = {
3155 .owner = THIS_MODULE,
3156 .write = split_huge_pages_write,
3157 .llseek = no_llseek,
3158};
3159
3160static int __init split_huge_pages_debugfs(void)
3161{
3162 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3163 &split_huge_pages_fops);
3164 return 0;
3165}
3166late_initcall(split_huge_pages_debugfs);
3167#endif
3168
3169#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3170void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3171 struct page *page)
3172{
3173 struct vm_area_struct *vma = pvmw->vma;
3174 struct mm_struct *mm = vma->vm_mm;
3175 unsigned long address = pvmw->address;
3176 pmd_t pmdval;
3177 swp_entry_t entry;
3178 pmd_t pmdswp;
3179
3180 if (!(pvmw->pmd && !pvmw->pte))
3181 return;
3182
3183 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3184 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3185 if (pmd_dirty(pmdval))
3186 set_page_dirty(page);
3187 if (pmd_write(pmdval))
3188 entry = make_writable_migration_entry(page_to_pfn(page));
3189 else
3190 entry = make_readable_migration_entry(page_to_pfn(page));
3191 pmdswp = swp_entry_to_pmd(entry);
3192 if (pmd_soft_dirty(pmdval))
3193 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3194 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3195 page_remove_rmap(page, true);
3196 put_page(page);
3197}
3198
3199void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3200{
3201 struct vm_area_struct *vma = pvmw->vma;
3202 struct mm_struct *mm = vma->vm_mm;
3203 unsigned long address = pvmw->address;
3204 unsigned long mmun_start = address & HPAGE_PMD_MASK;
3205 pmd_t pmde;
3206 swp_entry_t entry;
3207
3208 if (!(pvmw->pmd && !pvmw->pte))
3209 return;
3210
3211 entry = pmd_to_swp_entry(*pvmw->pmd);
3212 get_page(new);
3213 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3214 if (pmd_swp_soft_dirty(*pvmw->pmd))
3215 pmde = pmd_mksoft_dirty(pmde);
3216 if (is_writable_migration_entry(entry))
3217 pmde = maybe_pmd_mkwrite(pmde, vma);
3218 if (pmd_swp_uffd_wp(*pvmw->pmd))
3219 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3220
3221 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3222 if (PageAnon(new))
3223 page_add_anon_rmap(new, vma, mmun_start, true);
3224 else
3225 page_add_file_rmap(new, true);
3226 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3227 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3228 mlock_vma_page(new);
3229 update_mmu_cache_pmd(vma, address, pvmw->pmd);
3230}
3231#endif