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#include <linux/numa.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;
65
66bool transparent_hugepage_enabled(struct vm_area_struct *vma)
67{
68 if (vma_is_anonymous(vma))
69 return __transparent_hugepage_enabled(vma);
70 if (vma_is_shmem(vma) && shmem_huge_enabled(vma))
71 return __transparent_hugepage_enabled(vma);
72
73 return false;
74}
75
76static struct page *get_huge_zero_page(void)
77{
78 struct page *zero_page;
79retry:
80 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
81 return READ_ONCE(huge_zero_page);
82
83 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
84 HPAGE_PMD_ORDER);
85 if (!zero_page) {
86 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
87 return NULL;
88 }
89 count_vm_event(THP_ZERO_PAGE_ALLOC);
90 preempt_disable();
91 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
92 preempt_enable();
93 __free_pages(zero_page, compound_order(zero_page));
94 goto retry;
95 }
96
97 /* We take additional reference here. It will be put back by shrinker */
98 atomic_set(&huge_zero_refcount, 2);
99 preempt_enable();
100 return READ_ONCE(huge_zero_page);
101}
102
103static void put_huge_zero_page(void)
104{
105 /*
106 * Counter should never go to zero here. Only shrinker can put
107 * last reference.
108 */
109 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
110}
111
112struct page *mm_get_huge_zero_page(struct mm_struct *mm)
113{
114 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
115 return READ_ONCE(huge_zero_page);
116
117 if (!get_huge_zero_page())
118 return NULL;
119
120 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
121 put_huge_zero_page();
122
123 return READ_ONCE(huge_zero_page);
124}
125
126void mm_put_huge_zero_page(struct mm_struct *mm)
127{
128 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
129 put_huge_zero_page();
130}
131
132static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
133 struct shrink_control *sc)
134{
135 /* we can free zero page only if last reference remains */
136 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
137}
138
139static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
140 struct shrink_control *sc)
141{
142 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
143 struct page *zero_page = xchg(&huge_zero_page, NULL);
144 BUG_ON(zero_page == NULL);
145 __free_pages(zero_page, compound_order(zero_page));
146 return HPAGE_PMD_NR;
147 }
148
149 return 0;
150}
151
152static struct shrinker huge_zero_page_shrinker = {
153 .count_objects = shrink_huge_zero_page_count,
154 .scan_objects = shrink_huge_zero_page_scan,
155 .seeks = DEFAULT_SEEKS,
156};
157
158#ifdef CONFIG_SYSFS
159static ssize_t enabled_show(struct kobject *kobj,
160 struct kobj_attribute *attr, char *buf)
161{
162 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
163 return sprintf(buf, "[always] madvise never\n");
164 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
165 return sprintf(buf, "always [madvise] never\n");
166 else
167 return sprintf(buf, "always madvise [never]\n");
168}
169
170static ssize_t enabled_store(struct kobject *kobj,
171 struct kobj_attribute *attr,
172 const char *buf, size_t count)
173{
174 ssize_t ret = count;
175
176 if (!memcmp("always", buf,
177 min(sizeof("always")-1, count))) {
178 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
179 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
180 } else if (!memcmp("madvise", buf,
181 min(sizeof("madvise")-1, count))) {
182 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
183 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
184 } else if (!memcmp("never", buf,
185 min(sizeof("never")-1, count))) {
186 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
187 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
188 } else
189 ret = -EINVAL;
190
191 if (ret > 0) {
192 int err = start_stop_khugepaged();
193 if (err)
194 ret = err;
195 }
196 return ret;
197}
198static struct kobj_attribute enabled_attr =
199 __ATTR(enabled, 0644, enabled_show, enabled_store);
200
201ssize_t single_hugepage_flag_show(struct kobject *kobj,
202 struct kobj_attribute *attr, char *buf,
203 enum transparent_hugepage_flag flag)
204{
205 return sprintf(buf, "%d\n",
206 !!test_bit(flag, &transparent_hugepage_flags));
207}
208
209ssize_t single_hugepage_flag_store(struct kobject *kobj,
210 struct kobj_attribute *attr,
211 const char *buf, size_t count,
212 enum transparent_hugepage_flag flag)
213{
214 unsigned long value;
215 int ret;
216
217 ret = kstrtoul(buf, 10, &value);
218 if (ret < 0)
219 return ret;
220 if (value > 1)
221 return -EINVAL;
222
223 if (value)
224 set_bit(flag, &transparent_hugepage_flags);
225 else
226 clear_bit(flag, &transparent_hugepage_flags);
227
228 return count;
229}
230
231static ssize_t defrag_show(struct kobject *kobj,
232 struct kobj_attribute *attr, char *buf)
233{
234 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
235 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
236 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
237 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
238 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
239 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
240 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
241 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
242 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
243}
244
245static ssize_t defrag_store(struct kobject *kobj,
246 struct kobj_attribute *attr,
247 const char *buf, size_t count)
248{
249 if (!memcmp("always", buf,
250 min(sizeof("always")-1, count))) {
251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
255 } else if (!memcmp("defer+madvise", buf,
256 min(sizeof("defer+madvise")-1, count))) {
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
260 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261 } else if (!memcmp("defer", buf,
262 min(sizeof("defer")-1, count))) {
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
266 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
267 } else if (!memcmp("madvise", buf,
268 min(sizeof("madvise")-1, count))) {
269 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
270 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
271 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
272 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
273 } else if (!memcmp("never", buf,
274 min(sizeof("never")-1, count))) {
275 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
276 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
277 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
278 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
279 } else
280 return -EINVAL;
281
282 return count;
283}
284static struct kobj_attribute defrag_attr =
285 __ATTR(defrag, 0644, defrag_show, defrag_store);
286
287static ssize_t use_zero_page_show(struct kobject *kobj,
288 struct kobj_attribute *attr, char *buf)
289{
290 return single_hugepage_flag_show(kobj, attr, buf,
291 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
292}
293static ssize_t use_zero_page_store(struct kobject *kobj,
294 struct kobj_attribute *attr, const char *buf, size_t count)
295{
296 return single_hugepage_flag_store(kobj, attr, buf, count,
297 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
298}
299static struct kobj_attribute use_zero_page_attr =
300 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
301
302static ssize_t hpage_pmd_size_show(struct kobject *kobj,
303 struct kobj_attribute *attr, char *buf)
304{
305 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
306}
307static struct kobj_attribute hpage_pmd_size_attr =
308 __ATTR_RO(hpage_pmd_size);
309
310#ifdef CONFIG_DEBUG_VM
311static ssize_t debug_cow_show(struct kobject *kobj,
312 struct kobj_attribute *attr, char *buf)
313{
314 return single_hugepage_flag_show(kobj, attr, buf,
315 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
316}
317static ssize_t debug_cow_store(struct kobject *kobj,
318 struct kobj_attribute *attr,
319 const char *buf, size_t count)
320{
321 return single_hugepage_flag_store(kobj, attr, buf, count,
322 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
323}
324static struct kobj_attribute debug_cow_attr =
325 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
326#endif /* CONFIG_DEBUG_VM */
327
328static struct attribute *hugepage_attr[] = {
329 &enabled_attr.attr,
330 &defrag_attr.attr,
331 &use_zero_page_attr.attr,
332 &hpage_pmd_size_attr.attr,
333#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
334 &shmem_enabled_attr.attr,
335#endif
336#ifdef CONFIG_DEBUG_VM
337 &debug_cow_attr.attr,
338#endif
339 NULL,
340};
341
342static const struct attribute_group hugepage_attr_group = {
343 .attrs = hugepage_attr,
344};
345
346static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
347{
348 int err;
349
350 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
351 if (unlikely(!*hugepage_kobj)) {
352 pr_err("failed to create transparent hugepage kobject\n");
353 return -ENOMEM;
354 }
355
356 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
357 if (err) {
358 pr_err("failed to register transparent hugepage group\n");
359 goto delete_obj;
360 }
361
362 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
363 if (err) {
364 pr_err("failed to register transparent hugepage group\n");
365 goto remove_hp_group;
366 }
367
368 return 0;
369
370remove_hp_group:
371 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
372delete_obj:
373 kobject_put(*hugepage_kobj);
374 return err;
375}
376
377static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
378{
379 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
380 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
381 kobject_put(hugepage_kobj);
382}
383#else
384static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
385{
386 return 0;
387}
388
389static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
390{
391}
392#endif /* CONFIG_SYSFS */
393
394static int __init hugepage_init(void)
395{
396 int err;
397 struct kobject *hugepage_kobj;
398
399 if (!has_transparent_hugepage()) {
400 transparent_hugepage_flags = 0;
401 return -EINVAL;
402 }
403
404 /*
405 * hugepages can't be allocated by the buddy allocator
406 */
407 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
408 /*
409 * we use page->mapping and page->index in second tail page
410 * as list_head: assuming THP order >= 2
411 */
412 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
413
414 err = hugepage_init_sysfs(&hugepage_kobj);
415 if (err)
416 goto err_sysfs;
417
418 err = khugepaged_init();
419 if (err)
420 goto err_slab;
421
422 err = register_shrinker(&huge_zero_page_shrinker);
423 if (err)
424 goto err_hzp_shrinker;
425 err = register_shrinker(&deferred_split_shrinker);
426 if (err)
427 goto err_split_shrinker;
428
429 /*
430 * By default disable transparent hugepages on smaller systems,
431 * where the extra memory used could hurt more than TLB overhead
432 * is likely to save. The admin can still enable it through /sys.
433 */
434 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
435 transparent_hugepage_flags = 0;
436 return 0;
437 }
438
439 err = start_stop_khugepaged();
440 if (err)
441 goto err_khugepaged;
442
443 return 0;
444err_khugepaged:
445 unregister_shrinker(&deferred_split_shrinker);
446err_split_shrinker:
447 unregister_shrinker(&huge_zero_page_shrinker);
448err_hzp_shrinker:
449 khugepaged_destroy();
450err_slab:
451 hugepage_exit_sysfs(hugepage_kobj);
452err_sysfs:
453 return err;
454}
455subsys_initcall(hugepage_init);
456
457static int __init setup_transparent_hugepage(char *str)
458{
459 int ret = 0;
460 if (!str)
461 goto out;
462 if (!strcmp(str, "always")) {
463 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
464 &transparent_hugepage_flags);
465 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
466 &transparent_hugepage_flags);
467 ret = 1;
468 } else if (!strcmp(str, "madvise")) {
469 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
470 &transparent_hugepage_flags);
471 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
472 &transparent_hugepage_flags);
473 ret = 1;
474 } else if (!strcmp(str, "never")) {
475 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
476 &transparent_hugepage_flags);
477 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
478 &transparent_hugepage_flags);
479 ret = 1;
480 }
481out:
482 if (!ret)
483 pr_warn("transparent_hugepage= cannot parse, ignored\n");
484 return ret;
485}
486__setup("transparent_hugepage=", setup_transparent_hugepage);
487
488pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
489{
490 if (likely(vma->vm_flags & VM_WRITE))
491 pmd = pmd_mkwrite(pmd);
492 return pmd;
493}
494
495static inline struct list_head *page_deferred_list(struct page *page)
496{
497 /* ->lru in the tail pages is occupied by compound_head. */
498 return &page[2].deferred_list;
499}
500
501void prep_transhuge_page(struct page *page)
502{
503 /*
504 * we use page->mapping and page->indexlru in second tail page
505 * as list_head: assuming THP order >= 2
506 */
507
508 INIT_LIST_HEAD(page_deferred_list(page));
509 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
510}
511
512unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
513 loff_t off, unsigned long flags, unsigned long size)
514{
515 unsigned long addr;
516 loff_t off_end = off + len;
517 loff_t off_align = round_up(off, size);
518 unsigned long len_pad;
519
520 if (off_end <= off_align || (off_end - off_align) < size)
521 return 0;
522
523 len_pad = len + size;
524 if (len_pad < len || (off + len_pad) < off)
525 return 0;
526
527 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
528 off >> PAGE_SHIFT, flags);
529 if (IS_ERR_VALUE(addr))
530 return 0;
531
532 addr += (off - addr) & (size - 1);
533 return addr;
534}
535
536unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
537 unsigned long len, unsigned long pgoff, unsigned long flags)
538{
539 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
540
541 if (addr)
542 goto out;
543 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
544 goto out;
545
546 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
547 if (addr)
548 return addr;
549
550 out:
551 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
552}
553EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
554
555static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
556 struct page *page, gfp_t gfp)
557{
558 struct vm_area_struct *vma = vmf->vma;
559 struct mem_cgroup *memcg;
560 pgtable_t pgtable;
561 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
562 vm_fault_t ret = 0;
563
564 VM_BUG_ON_PAGE(!PageCompound(page), page);
565
566 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
567 put_page(page);
568 count_vm_event(THP_FAULT_FALLBACK);
569 return VM_FAULT_FALLBACK;
570 }
571
572 pgtable = pte_alloc_one(vma->vm_mm);
573 if (unlikely(!pgtable)) {
574 ret = VM_FAULT_OOM;
575 goto release;
576 }
577
578 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
579 /*
580 * The memory barrier inside __SetPageUptodate makes sure that
581 * clear_huge_page writes become visible before the set_pmd_at()
582 * write.
583 */
584 __SetPageUptodate(page);
585
586 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
587 if (unlikely(!pmd_none(*vmf->pmd))) {
588 goto unlock_release;
589 } else {
590 pmd_t entry;
591
592 ret = check_stable_address_space(vma->vm_mm);
593 if (ret)
594 goto unlock_release;
595
596 /* Deliver the page fault to userland */
597 if (userfaultfd_missing(vma)) {
598 vm_fault_t ret2;
599
600 spin_unlock(vmf->ptl);
601 mem_cgroup_cancel_charge(page, memcg, true);
602 put_page(page);
603 pte_free(vma->vm_mm, pgtable);
604 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
605 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
606 return ret2;
607 }
608
609 entry = mk_huge_pmd(page, vma->vm_page_prot);
610 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
611 page_add_new_anon_rmap(page, vma, haddr, true);
612 mem_cgroup_commit_charge(page, memcg, false, true);
613 lru_cache_add_active_or_unevictable(page, vma);
614 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
615 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
616 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
617 mm_inc_nr_ptes(vma->vm_mm);
618 spin_unlock(vmf->ptl);
619 count_vm_event(THP_FAULT_ALLOC);
620 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
621 }
622
623 return 0;
624unlock_release:
625 spin_unlock(vmf->ptl);
626release:
627 if (pgtable)
628 pte_free(vma->vm_mm, pgtable);
629 mem_cgroup_cancel_charge(page, memcg, true);
630 put_page(page);
631 return ret;
632
633}
634
635/*
636 * always: directly stall for all thp allocations
637 * defer: wake kswapd and fail if not immediately available
638 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
639 * fail if not immediately available
640 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
641 * available
642 * never: never stall for any thp allocation
643 */
644static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
645{
646 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
647
648 /* Always do synchronous compaction */
649 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
650 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
651
652 /* Kick kcompactd and fail quickly */
653 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
654 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
655
656 /* Synchronous compaction if madvised, otherwise kick kcompactd */
657 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
658 return GFP_TRANSHUGE_LIGHT |
659 (vma_madvised ? __GFP_DIRECT_RECLAIM :
660 __GFP_KSWAPD_RECLAIM);
661
662 /* Only do synchronous compaction if madvised */
663 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
664 return GFP_TRANSHUGE_LIGHT |
665 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
666
667 return GFP_TRANSHUGE_LIGHT;
668}
669
670/* Caller must hold page table lock. */
671static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
672 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
673 struct page *zero_page)
674{
675 pmd_t entry;
676 if (!pmd_none(*pmd))
677 return false;
678 entry = mk_pmd(zero_page, vma->vm_page_prot);
679 entry = pmd_mkhuge(entry);
680 if (pgtable)
681 pgtable_trans_huge_deposit(mm, pmd, pgtable);
682 set_pmd_at(mm, haddr, pmd, entry);
683 mm_inc_nr_ptes(mm);
684 return true;
685}
686
687vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
688{
689 struct vm_area_struct *vma = vmf->vma;
690 gfp_t gfp;
691 struct page *page;
692 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
693
694 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
695 return VM_FAULT_FALLBACK;
696 if (unlikely(anon_vma_prepare(vma)))
697 return VM_FAULT_OOM;
698 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
699 return VM_FAULT_OOM;
700 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
701 !mm_forbids_zeropage(vma->vm_mm) &&
702 transparent_hugepage_use_zero_page()) {
703 pgtable_t pgtable;
704 struct page *zero_page;
705 bool set;
706 vm_fault_t ret;
707 pgtable = pte_alloc_one(vma->vm_mm);
708 if (unlikely(!pgtable))
709 return VM_FAULT_OOM;
710 zero_page = mm_get_huge_zero_page(vma->vm_mm);
711 if (unlikely(!zero_page)) {
712 pte_free(vma->vm_mm, pgtable);
713 count_vm_event(THP_FAULT_FALLBACK);
714 return VM_FAULT_FALLBACK;
715 }
716 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
717 ret = 0;
718 set = false;
719 if (pmd_none(*vmf->pmd)) {
720 ret = check_stable_address_space(vma->vm_mm);
721 if (ret) {
722 spin_unlock(vmf->ptl);
723 } else if (userfaultfd_missing(vma)) {
724 spin_unlock(vmf->ptl);
725 ret = handle_userfault(vmf, VM_UFFD_MISSING);
726 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
727 } else {
728 set_huge_zero_page(pgtable, vma->vm_mm, vma,
729 haddr, vmf->pmd, zero_page);
730 spin_unlock(vmf->ptl);
731 set = true;
732 }
733 } else
734 spin_unlock(vmf->ptl);
735 if (!set)
736 pte_free(vma->vm_mm, pgtable);
737 return ret;
738 }
739 gfp = alloc_hugepage_direct_gfpmask(vma);
740 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
741 if (unlikely(!page)) {
742 count_vm_event(THP_FAULT_FALLBACK);
743 return VM_FAULT_FALLBACK;
744 }
745 prep_transhuge_page(page);
746 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
747}
748
749static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
750 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
751 pgtable_t pgtable)
752{
753 struct mm_struct *mm = vma->vm_mm;
754 pmd_t entry;
755 spinlock_t *ptl;
756
757 ptl = pmd_lock(mm, pmd);
758 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
759 if (pfn_t_devmap(pfn))
760 entry = pmd_mkdevmap(entry);
761 if (write) {
762 entry = pmd_mkyoung(pmd_mkdirty(entry));
763 entry = maybe_pmd_mkwrite(entry, vma);
764 }
765
766 if (pgtable) {
767 pgtable_trans_huge_deposit(mm, pmd, pgtable);
768 mm_inc_nr_ptes(mm);
769 }
770
771 set_pmd_at(mm, addr, pmd, entry);
772 update_mmu_cache_pmd(vma, addr, pmd);
773 spin_unlock(ptl);
774}
775
776vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
777 pmd_t *pmd, pfn_t pfn, bool write)
778{
779 pgprot_t pgprot = vma->vm_page_prot;
780 pgtable_t pgtable = NULL;
781 /*
782 * If we had pmd_special, we could avoid all these restrictions,
783 * but we need to be consistent with PTEs and architectures that
784 * can't support a 'special' bit.
785 */
786 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
787 !pfn_t_devmap(pfn));
788 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
789 (VM_PFNMAP|VM_MIXEDMAP));
790 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
791
792 if (addr < vma->vm_start || addr >= vma->vm_end)
793 return VM_FAULT_SIGBUS;
794
795 if (arch_needs_pgtable_deposit()) {
796 pgtable = pte_alloc_one(vma->vm_mm);
797 if (!pgtable)
798 return VM_FAULT_OOM;
799 }
800
801 track_pfn_insert(vma, &pgprot, pfn);
802
803 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
804 return VM_FAULT_NOPAGE;
805}
806EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
807
808#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
809static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
810{
811 if (likely(vma->vm_flags & VM_WRITE))
812 pud = pud_mkwrite(pud);
813 return pud;
814}
815
816static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
817 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
818{
819 struct mm_struct *mm = vma->vm_mm;
820 pud_t entry;
821 spinlock_t *ptl;
822
823 ptl = pud_lock(mm, pud);
824 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
825 if (pfn_t_devmap(pfn))
826 entry = pud_mkdevmap(entry);
827 if (write) {
828 entry = pud_mkyoung(pud_mkdirty(entry));
829 entry = maybe_pud_mkwrite(entry, vma);
830 }
831 set_pud_at(mm, addr, pud, entry);
832 update_mmu_cache_pud(vma, addr, pud);
833 spin_unlock(ptl);
834}
835
836vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
837 pud_t *pud, pfn_t pfn, bool write)
838{
839 pgprot_t pgprot = vma->vm_page_prot;
840 /*
841 * If we had pud_special, we could avoid all these restrictions,
842 * but we need to be consistent with PTEs and architectures that
843 * can't support a 'special' bit.
844 */
845 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
846 !pfn_t_devmap(pfn));
847 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
848 (VM_PFNMAP|VM_MIXEDMAP));
849 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
850
851 if (addr < vma->vm_start || addr >= vma->vm_end)
852 return VM_FAULT_SIGBUS;
853
854 track_pfn_insert(vma, &pgprot, pfn);
855
856 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
857 return VM_FAULT_NOPAGE;
858}
859EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
860#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
861
862static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
863 pmd_t *pmd, int flags)
864{
865 pmd_t _pmd;
866
867 _pmd = pmd_mkyoung(*pmd);
868 if (flags & FOLL_WRITE)
869 _pmd = pmd_mkdirty(_pmd);
870 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
871 pmd, _pmd, flags & FOLL_WRITE))
872 update_mmu_cache_pmd(vma, addr, pmd);
873}
874
875struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
876 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
877{
878 unsigned long pfn = pmd_pfn(*pmd);
879 struct mm_struct *mm = vma->vm_mm;
880 struct page *page;
881
882 assert_spin_locked(pmd_lockptr(mm, pmd));
883
884 /*
885 * When we COW a devmap PMD entry, we split it into PTEs, so we should
886 * not be in this function with `flags & FOLL_COW` set.
887 */
888 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
889
890 if (flags & FOLL_WRITE && !pmd_write(*pmd))
891 return NULL;
892
893 if (pmd_present(*pmd) && pmd_devmap(*pmd))
894 /* pass */;
895 else
896 return NULL;
897
898 if (flags & FOLL_TOUCH)
899 touch_pmd(vma, addr, pmd, flags);
900
901 /*
902 * device mapped pages can only be returned if the
903 * caller will manage the page reference count.
904 */
905 if (!(flags & FOLL_GET))
906 return ERR_PTR(-EEXIST);
907
908 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
909 *pgmap = get_dev_pagemap(pfn, *pgmap);
910 if (!*pgmap)
911 return ERR_PTR(-EFAULT);
912 page = pfn_to_page(pfn);
913 get_page(page);
914
915 return page;
916}
917
918int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
919 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
920 struct vm_area_struct *vma)
921{
922 spinlock_t *dst_ptl, *src_ptl;
923 struct page *src_page;
924 pmd_t pmd;
925 pgtable_t pgtable = NULL;
926 int ret = -ENOMEM;
927
928 /* Skip if can be re-fill on fault */
929 if (!vma_is_anonymous(vma))
930 return 0;
931
932 pgtable = pte_alloc_one(dst_mm);
933 if (unlikely(!pgtable))
934 goto out;
935
936 dst_ptl = pmd_lock(dst_mm, dst_pmd);
937 src_ptl = pmd_lockptr(src_mm, src_pmd);
938 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
939
940 ret = -EAGAIN;
941 pmd = *src_pmd;
942
943#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
944 if (unlikely(is_swap_pmd(pmd))) {
945 swp_entry_t entry = pmd_to_swp_entry(pmd);
946
947 VM_BUG_ON(!is_pmd_migration_entry(pmd));
948 if (is_write_migration_entry(entry)) {
949 make_migration_entry_read(&entry);
950 pmd = swp_entry_to_pmd(entry);
951 if (pmd_swp_soft_dirty(*src_pmd))
952 pmd = pmd_swp_mksoft_dirty(pmd);
953 set_pmd_at(src_mm, addr, src_pmd, pmd);
954 }
955 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
956 mm_inc_nr_ptes(dst_mm);
957 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
958 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
959 ret = 0;
960 goto out_unlock;
961 }
962#endif
963
964 if (unlikely(!pmd_trans_huge(pmd))) {
965 pte_free(dst_mm, pgtable);
966 goto out_unlock;
967 }
968 /*
969 * When page table lock is held, the huge zero pmd should not be
970 * under splitting since we don't split the page itself, only pmd to
971 * a page table.
972 */
973 if (is_huge_zero_pmd(pmd)) {
974 struct page *zero_page;
975 /*
976 * get_huge_zero_page() will never allocate a new page here,
977 * since we already have a zero page to copy. It just takes a
978 * reference.
979 */
980 zero_page = mm_get_huge_zero_page(dst_mm);
981 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
982 zero_page);
983 ret = 0;
984 goto out_unlock;
985 }
986
987 src_page = pmd_page(pmd);
988 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
989 get_page(src_page);
990 page_dup_rmap(src_page, true);
991 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
992 mm_inc_nr_ptes(dst_mm);
993 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
994
995 pmdp_set_wrprotect(src_mm, addr, src_pmd);
996 pmd = pmd_mkold(pmd_wrprotect(pmd));
997 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
998
999 ret = 0;
1000out_unlock:
1001 spin_unlock(src_ptl);
1002 spin_unlock(dst_ptl);
1003out:
1004 return ret;
1005}
1006
1007#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1008static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1009 pud_t *pud, int flags)
1010{
1011 pud_t _pud;
1012
1013 _pud = pud_mkyoung(*pud);
1014 if (flags & FOLL_WRITE)
1015 _pud = pud_mkdirty(_pud);
1016 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1017 pud, _pud, flags & FOLL_WRITE))
1018 update_mmu_cache_pud(vma, addr, pud);
1019}
1020
1021struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1022 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1023{
1024 unsigned long pfn = pud_pfn(*pud);
1025 struct mm_struct *mm = vma->vm_mm;
1026 struct page *page;
1027
1028 assert_spin_locked(pud_lockptr(mm, pud));
1029
1030 if (flags & FOLL_WRITE && !pud_write(*pud))
1031 return NULL;
1032
1033 if (pud_present(*pud) && pud_devmap(*pud))
1034 /* pass */;
1035 else
1036 return NULL;
1037
1038 if (flags & FOLL_TOUCH)
1039 touch_pud(vma, addr, pud, flags);
1040
1041 /*
1042 * device mapped pages can only be returned if the
1043 * caller will manage the page reference count.
1044 */
1045 if (!(flags & FOLL_GET))
1046 return ERR_PTR(-EEXIST);
1047
1048 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1049 *pgmap = get_dev_pagemap(pfn, *pgmap);
1050 if (!*pgmap)
1051 return ERR_PTR(-EFAULT);
1052 page = pfn_to_page(pfn);
1053 get_page(page);
1054
1055 return page;
1056}
1057
1058int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1059 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1060 struct vm_area_struct *vma)
1061{
1062 spinlock_t *dst_ptl, *src_ptl;
1063 pud_t pud;
1064 int ret;
1065
1066 dst_ptl = pud_lock(dst_mm, dst_pud);
1067 src_ptl = pud_lockptr(src_mm, src_pud);
1068 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1069
1070 ret = -EAGAIN;
1071 pud = *src_pud;
1072 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1073 goto out_unlock;
1074
1075 /*
1076 * When page table lock is held, the huge zero pud should not be
1077 * under splitting since we don't split the page itself, only pud to
1078 * a page table.
1079 */
1080 if (is_huge_zero_pud(pud)) {
1081 /* No huge zero pud yet */
1082 }
1083
1084 pudp_set_wrprotect(src_mm, addr, src_pud);
1085 pud = pud_mkold(pud_wrprotect(pud));
1086 set_pud_at(dst_mm, addr, dst_pud, pud);
1087
1088 ret = 0;
1089out_unlock:
1090 spin_unlock(src_ptl);
1091 spin_unlock(dst_ptl);
1092 return ret;
1093}
1094
1095void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1096{
1097 pud_t entry;
1098 unsigned long haddr;
1099 bool write = vmf->flags & FAULT_FLAG_WRITE;
1100
1101 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1102 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1103 goto unlock;
1104
1105 entry = pud_mkyoung(orig_pud);
1106 if (write)
1107 entry = pud_mkdirty(entry);
1108 haddr = vmf->address & HPAGE_PUD_MASK;
1109 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1110 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1111
1112unlock:
1113 spin_unlock(vmf->ptl);
1114}
1115#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1116
1117void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1118{
1119 pmd_t entry;
1120 unsigned long haddr;
1121 bool write = vmf->flags & FAULT_FLAG_WRITE;
1122
1123 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1124 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1125 goto unlock;
1126
1127 entry = pmd_mkyoung(orig_pmd);
1128 if (write)
1129 entry = pmd_mkdirty(entry);
1130 haddr = vmf->address & HPAGE_PMD_MASK;
1131 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1132 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1133
1134unlock:
1135 spin_unlock(vmf->ptl);
1136}
1137
1138static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1139 pmd_t orig_pmd, struct page *page)
1140{
1141 struct vm_area_struct *vma = vmf->vma;
1142 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1143 struct mem_cgroup *memcg;
1144 pgtable_t pgtable;
1145 pmd_t _pmd;
1146 int i;
1147 vm_fault_t ret = 0;
1148 struct page **pages;
1149 struct mmu_notifier_range range;
1150
1151 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1152 GFP_KERNEL);
1153 if (unlikely(!pages)) {
1154 ret |= VM_FAULT_OOM;
1155 goto out;
1156 }
1157
1158 for (i = 0; i < HPAGE_PMD_NR; i++) {
1159 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1160 vmf->address, page_to_nid(page));
1161 if (unlikely(!pages[i] ||
1162 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1163 GFP_KERNEL, &memcg, false))) {
1164 if (pages[i])
1165 put_page(pages[i]);
1166 while (--i >= 0) {
1167 memcg = (void *)page_private(pages[i]);
1168 set_page_private(pages[i], 0);
1169 mem_cgroup_cancel_charge(pages[i], memcg,
1170 false);
1171 put_page(pages[i]);
1172 }
1173 kfree(pages);
1174 ret |= VM_FAULT_OOM;
1175 goto out;
1176 }
1177 set_page_private(pages[i], (unsigned long)memcg);
1178 }
1179
1180 for (i = 0; i < HPAGE_PMD_NR; i++) {
1181 copy_user_highpage(pages[i], page + i,
1182 haddr + PAGE_SIZE * i, vma);
1183 __SetPageUptodate(pages[i]);
1184 cond_resched();
1185 }
1186
1187 mmu_notifier_range_init(&range, vma->vm_mm, haddr,
1188 haddr + HPAGE_PMD_SIZE);
1189 mmu_notifier_invalidate_range_start(&range);
1190
1191 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1192 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1193 goto out_free_pages;
1194 VM_BUG_ON_PAGE(!PageHead(page), page);
1195
1196 /*
1197 * Leave pmd empty until pte is filled note we must notify here as
1198 * concurrent CPU thread might write to new page before the call to
1199 * mmu_notifier_invalidate_range_end() happens which can lead to a
1200 * device seeing memory write in different order than CPU.
1201 *
1202 * See Documentation/vm/mmu_notifier.rst
1203 */
1204 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1205
1206 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1207 pmd_populate(vma->vm_mm, &_pmd, pgtable);
1208
1209 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1210 pte_t entry;
1211 entry = mk_pte(pages[i], vma->vm_page_prot);
1212 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1213 memcg = (void *)page_private(pages[i]);
1214 set_page_private(pages[i], 0);
1215 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1216 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1217 lru_cache_add_active_or_unevictable(pages[i], vma);
1218 vmf->pte = pte_offset_map(&_pmd, haddr);
1219 VM_BUG_ON(!pte_none(*vmf->pte));
1220 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1221 pte_unmap(vmf->pte);
1222 }
1223 kfree(pages);
1224
1225 smp_wmb(); /* make pte visible before pmd */
1226 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1227 page_remove_rmap(page, true);
1228 spin_unlock(vmf->ptl);
1229
1230 /*
1231 * No need to double call mmu_notifier->invalidate_range() callback as
1232 * the above pmdp_huge_clear_flush_notify() did already call it.
1233 */
1234 mmu_notifier_invalidate_range_only_end(&range);
1235
1236 ret |= VM_FAULT_WRITE;
1237 put_page(page);
1238
1239out:
1240 return ret;
1241
1242out_free_pages:
1243 spin_unlock(vmf->ptl);
1244 mmu_notifier_invalidate_range_end(&range);
1245 for (i = 0; i < HPAGE_PMD_NR; i++) {
1246 memcg = (void *)page_private(pages[i]);
1247 set_page_private(pages[i], 0);
1248 mem_cgroup_cancel_charge(pages[i], memcg, false);
1249 put_page(pages[i]);
1250 }
1251 kfree(pages);
1252 goto out;
1253}
1254
1255vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1256{
1257 struct vm_area_struct *vma = vmf->vma;
1258 struct page *page = NULL, *new_page;
1259 struct mem_cgroup *memcg;
1260 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1261 struct mmu_notifier_range range;
1262 gfp_t huge_gfp; /* for allocation and charge */
1263 vm_fault_t ret = 0;
1264
1265 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1266 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1267 if (is_huge_zero_pmd(orig_pmd))
1268 goto alloc;
1269 spin_lock(vmf->ptl);
1270 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1271 goto out_unlock;
1272
1273 page = pmd_page(orig_pmd);
1274 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1275 /*
1276 * We can only reuse the page if nobody else maps the huge page or it's
1277 * part.
1278 */
1279 if (!trylock_page(page)) {
1280 get_page(page);
1281 spin_unlock(vmf->ptl);
1282 lock_page(page);
1283 spin_lock(vmf->ptl);
1284 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1285 unlock_page(page);
1286 put_page(page);
1287 goto out_unlock;
1288 }
1289 put_page(page);
1290 }
1291 if (reuse_swap_page(page, NULL)) {
1292 pmd_t entry;
1293 entry = pmd_mkyoung(orig_pmd);
1294 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1295 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1296 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1297 ret |= VM_FAULT_WRITE;
1298 unlock_page(page);
1299 goto out_unlock;
1300 }
1301 unlock_page(page);
1302 get_page(page);
1303 spin_unlock(vmf->ptl);
1304alloc:
1305 if (__transparent_hugepage_enabled(vma) &&
1306 !transparent_hugepage_debug_cow()) {
1307 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1308 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1309 } else
1310 new_page = NULL;
1311
1312 if (likely(new_page)) {
1313 prep_transhuge_page(new_page);
1314 } else {
1315 if (!page) {
1316 split_huge_pmd(vma, vmf->pmd, vmf->address);
1317 ret |= VM_FAULT_FALLBACK;
1318 } else {
1319 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1320 if (ret & VM_FAULT_OOM) {
1321 split_huge_pmd(vma, vmf->pmd, vmf->address);
1322 ret |= VM_FAULT_FALLBACK;
1323 }
1324 put_page(page);
1325 }
1326 count_vm_event(THP_FAULT_FALLBACK);
1327 goto out;
1328 }
1329
1330 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1331 huge_gfp, &memcg, true))) {
1332 put_page(new_page);
1333 split_huge_pmd(vma, vmf->pmd, vmf->address);
1334 if (page)
1335 put_page(page);
1336 ret |= VM_FAULT_FALLBACK;
1337 count_vm_event(THP_FAULT_FALLBACK);
1338 goto out;
1339 }
1340
1341 count_vm_event(THP_FAULT_ALLOC);
1342 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1343
1344 if (!page)
1345 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1346 else
1347 copy_user_huge_page(new_page, page, vmf->address,
1348 vma, HPAGE_PMD_NR);
1349 __SetPageUptodate(new_page);
1350
1351 mmu_notifier_range_init(&range, vma->vm_mm, haddr,
1352 haddr + HPAGE_PMD_SIZE);
1353 mmu_notifier_invalidate_range_start(&range);
1354
1355 spin_lock(vmf->ptl);
1356 if (page)
1357 put_page(page);
1358 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1359 spin_unlock(vmf->ptl);
1360 mem_cgroup_cancel_charge(new_page, memcg, true);
1361 put_page(new_page);
1362 goto out_mn;
1363 } else {
1364 pmd_t entry;
1365 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1366 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1367 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1368 page_add_new_anon_rmap(new_page, vma, haddr, true);
1369 mem_cgroup_commit_charge(new_page, memcg, false, true);
1370 lru_cache_add_active_or_unevictable(new_page, vma);
1371 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1372 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1373 if (!page) {
1374 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1375 } else {
1376 VM_BUG_ON_PAGE(!PageHead(page), page);
1377 page_remove_rmap(page, true);
1378 put_page(page);
1379 }
1380 ret |= VM_FAULT_WRITE;
1381 }
1382 spin_unlock(vmf->ptl);
1383out_mn:
1384 /*
1385 * No need to double call mmu_notifier->invalidate_range() callback as
1386 * the above pmdp_huge_clear_flush_notify() did already call it.
1387 */
1388 mmu_notifier_invalidate_range_only_end(&range);
1389out:
1390 return ret;
1391out_unlock:
1392 spin_unlock(vmf->ptl);
1393 return ret;
1394}
1395
1396/*
1397 * FOLL_FORCE can write to even unwritable pmd's, but only
1398 * after we've gone through a COW cycle and they are dirty.
1399 */
1400static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1401{
1402 return pmd_write(pmd) ||
1403 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1404}
1405
1406struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1407 unsigned long addr,
1408 pmd_t *pmd,
1409 unsigned int flags)
1410{
1411 struct mm_struct *mm = vma->vm_mm;
1412 struct page *page = NULL;
1413
1414 assert_spin_locked(pmd_lockptr(mm, pmd));
1415
1416 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1417 goto out;
1418
1419 /* Avoid dumping huge zero page */
1420 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1421 return ERR_PTR(-EFAULT);
1422
1423 /* Full NUMA hinting faults to serialise migration in fault paths */
1424 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1425 goto out;
1426
1427 page = pmd_page(*pmd);
1428 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1429 if (flags & FOLL_TOUCH)
1430 touch_pmd(vma, addr, pmd, flags);
1431 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1432 /*
1433 * We don't mlock() pte-mapped THPs. This way we can avoid
1434 * leaking mlocked pages into non-VM_LOCKED VMAs.
1435 *
1436 * For anon THP:
1437 *
1438 * In most cases the pmd is the only mapping of the page as we
1439 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1440 * writable private mappings in populate_vma_page_range().
1441 *
1442 * The only scenario when we have the page shared here is if we
1443 * mlocking read-only mapping shared over fork(). We skip
1444 * mlocking such pages.
1445 *
1446 * For file THP:
1447 *
1448 * We can expect PageDoubleMap() to be stable under page lock:
1449 * for file pages we set it in page_add_file_rmap(), which
1450 * requires page to be locked.
1451 */
1452
1453 if (PageAnon(page) && compound_mapcount(page) != 1)
1454 goto skip_mlock;
1455 if (PageDoubleMap(page) || !page->mapping)
1456 goto skip_mlock;
1457 if (!trylock_page(page))
1458 goto skip_mlock;
1459 lru_add_drain();
1460 if (page->mapping && !PageDoubleMap(page))
1461 mlock_vma_page(page);
1462 unlock_page(page);
1463 }
1464skip_mlock:
1465 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1466 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1467 if (flags & FOLL_GET)
1468 get_page(page);
1469
1470out:
1471 return page;
1472}
1473
1474/* NUMA hinting page fault entry point for trans huge pmds */
1475vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1476{
1477 struct vm_area_struct *vma = vmf->vma;
1478 struct anon_vma *anon_vma = NULL;
1479 struct page *page;
1480 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1481 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1482 int target_nid, last_cpupid = -1;
1483 bool page_locked;
1484 bool migrated = false;
1485 bool was_writable;
1486 int flags = 0;
1487
1488 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1489 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1490 goto out_unlock;
1491
1492 /*
1493 * If there are potential migrations, wait for completion and retry
1494 * without disrupting NUMA hinting information. Do not relock and
1495 * check_same as the page may no longer be mapped.
1496 */
1497 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1498 page = pmd_page(*vmf->pmd);
1499 if (!get_page_unless_zero(page))
1500 goto out_unlock;
1501 spin_unlock(vmf->ptl);
1502 put_and_wait_on_page_locked(page);
1503 goto out;
1504 }
1505
1506 page = pmd_page(pmd);
1507 BUG_ON(is_huge_zero_page(page));
1508 page_nid = page_to_nid(page);
1509 last_cpupid = page_cpupid_last(page);
1510 count_vm_numa_event(NUMA_HINT_FAULTS);
1511 if (page_nid == this_nid) {
1512 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1513 flags |= TNF_FAULT_LOCAL;
1514 }
1515
1516 /* See similar comment in do_numa_page for explanation */
1517 if (!pmd_savedwrite(pmd))
1518 flags |= TNF_NO_GROUP;
1519
1520 /*
1521 * Acquire the page lock to serialise THP migrations but avoid dropping
1522 * page_table_lock if at all possible
1523 */
1524 page_locked = trylock_page(page);
1525 target_nid = mpol_misplaced(page, vma, haddr);
1526 if (target_nid == NUMA_NO_NODE) {
1527 /* If the page was locked, there are no parallel migrations */
1528 if (page_locked)
1529 goto clear_pmdnuma;
1530 }
1531
1532 /* Migration could have started since the pmd_trans_migrating check */
1533 if (!page_locked) {
1534 page_nid = NUMA_NO_NODE;
1535 if (!get_page_unless_zero(page))
1536 goto out_unlock;
1537 spin_unlock(vmf->ptl);
1538 put_and_wait_on_page_locked(page);
1539 goto out;
1540 }
1541
1542 /*
1543 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1544 * to serialises splits
1545 */
1546 get_page(page);
1547 spin_unlock(vmf->ptl);
1548 anon_vma = page_lock_anon_vma_read(page);
1549
1550 /* Confirm the PMD did not change while page_table_lock was released */
1551 spin_lock(vmf->ptl);
1552 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1553 unlock_page(page);
1554 put_page(page);
1555 page_nid = NUMA_NO_NODE;
1556 goto out_unlock;
1557 }
1558
1559 /* Bail if we fail to protect against THP splits for any reason */
1560 if (unlikely(!anon_vma)) {
1561 put_page(page);
1562 page_nid = NUMA_NO_NODE;
1563 goto clear_pmdnuma;
1564 }
1565
1566 /*
1567 * Since we took the NUMA fault, we must have observed the !accessible
1568 * bit. Make sure all other CPUs agree with that, to avoid them
1569 * modifying the page we're about to migrate.
1570 *
1571 * Must be done under PTL such that we'll observe the relevant
1572 * inc_tlb_flush_pending().
1573 *
1574 * We are not sure a pending tlb flush here is for a huge page
1575 * mapping or not. Hence use the tlb range variant
1576 */
1577 if (mm_tlb_flush_pending(vma->vm_mm)) {
1578 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1579 /*
1580 * change_huge_pmd() released the pmd lock before
1581 * invalidating the secondary MMUs sharing the primary
1582 * MMU pagetables (with ->invalidate_range()). The
1583 * mmu_notifier_invalidate_range_end() (which
1584 * internally calls ->invalidate_range()) in
1585 * change_pmd_range() will run after us, so we can't
1586 * rely on it here and we need an explicit invalidate.
1587 */
1588 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1589 haddr + HPAGE_PMD_SIZE);
1590 }
1591
1592 /*
1593 * Migrate the THP to the requested node, returns with page unlocked
1594 * and access rights restored.
1595 */
1596 spin_unlock(vmf->ptl);
1597
1598 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1599 vmf->pmd, pmd, vmf->address, page, target_nid);
1600 if (migrated) {
1601 flags |= TNF_MIGRATED;
1602 page_nid = target_nid;
1603 } else
1604 flags |= TNF_MIGRATE_FAIL;
1605
1606 goto out;
1607clear_pmdnuma:
1608 BUG_ON(!PageLocked(page));
1609 was_writable = pmd_savedwrite(pmd);
1610 pmd = pmd_modify(pmd, vma->vm_page_prot);
1611 pmd = pmd_mkyoung(pmd);
1612 if (was_writable)
1613 pmd = pmd_mkwrite(pmd);
1614 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1615 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1616 unlock_page(page);
1617out_unlock:
1618 spin_unlock(vmf->ptl);
1619
1620out:
1621 if (anon_vma)
1622 page_unlock_anon_vma_read(anon_vma);
1623
1624 if (page_nid != NUMA_NO_NODE)
1625 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1626 flags);
1627
1628 return 0;
1629}
1630
1631/*
1632 * Return true if we do MADV_FREE successfully on entire pmd page.
1633 * Otherwise, return false.
1634 */
1635bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1636 pmd_t *pmd, unsigned long addr, unsigned long next)
1637{
1638 spinlock_t *ptl;
1639 pmd_t orig_pmd;
1640 struct page *page;
1641 struct mm_struct *mm = tlb->mm;
1642 bool ret = false;
1643
1644 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1645
1646 ptl = pmd_trans_huge_lock(pmd, vma);
1647 if (!ptl)
1648 goto out_unlocked;
1649
1650 orig_pmd = *pmd;
1651 if (is_huge_zero_pmd(orig_pmd))
1652 goto out;
1653
1654 if (unlikely(!pmd_present(orig_pmd))) {
1655 VM_BUG_ON(thp_migration_supported() &&
1656 !is_pmd_migration_entry(orig_pmd));
1657 goto out;
1658 }
1659
1660 page = pmd_page(orig_pmd);
1661 /*
1662 * If other processes are mapping this page, we couldn't discard
1663 * the page unless they all do MADV_FREE so let's skip the page.
1664 */
1665 if (page_mapcount(page) != 1)
1666 goto out;
1667
1668 if (!trylock_page(page))
1669 goto out;
1670
1671 /*
1672 * If user want to discard part-pages of THP, split it so MADV_FREE
1673 * will deactivate only them.
1674 */
1675 if (next - addr != HPAGE_PMD_SIZE) {
1676 get_page(page);
1677 spin_unlock(ptl);
1678 split_huge_page(page);
1679 unlock_page(page);
1680 put_page(page);
1681 goto out_unlocked;
1682 }
1683
1684 if (PageDirty(page))
1685 ClearPageDirty(page);
1686 unlock_page(page);
1687
1688 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1689 pmdp_invalidate(vma, addr, pmd);
1690 orig_pmd = pmd_mkold(orig_pmd);
1691 orig_pmd = pmd_mkclean(orig_pmd);
1692
1693 set_pmd_at(mm, addr, pmd, orig_pmd);
1694 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1695 }
1696
1697 mark_page_lazyfree(page);
1698 ret = true;
1699out:
1700 spin_unlock(ptl);
1701out_unlocked:
1702 return ret;
1703}
1704
1705static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1706{
1707 pgtable_t pgtable;
1708
1709 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1710 pte_free(mm, pgtable);
1711 mm_dec_nr_ptes(mm);
1712}
1713
1714int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1715 pmd_t *pmd, unsigned long addr)
1716{
1717 pmd_t orig_pmd;
1718 spinlock_t *ptl;
1719
1720 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1721
1722 ptl = __pmd_trans_huge_lock(pmd, vma);
1723 if (!ptl)
1724 return 0;
1725 /*
1726 * For architectures like ppc64 we look at deposited pgtable
1727 * when calling pmdp_huge_get_and_clear. So do the
1728 * pgtable_trans_huge_withdraw after finishing pmdp related
1729 * operations.
1730 */
1731 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1732 tlb->fullmm);
1733 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1734 if (vma_is_dax(vma)) {
1735 if (arch_needs_pgtable_deposit())
1736 zap_deposited_table(tlb->mm, pmd);
1737 spin_unlock(ptl);
1738 if (is_huge_zero_pmd(orig_pmd))
1739 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1740 } else if (is_huge_zero_pmd(orig_pmd)) {
1741 zap_deposited_table(tlb->mm, pmd);
1742 spin_unlock(ptl);
1743 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1744 } else {
1745 struct page *page = NULL;
1746 int flush_needed = 1;
1747
1748 if (pmd_present(orig_pmd)) {
1749 page = pmd_page(orig_pmd);
1750 page_remove_rmap(page, true);
1751 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1752 VM_BUG_ON_PAGE(!PageHead(page), page);
1753 } else if (thp_migration_supported()) {
1754 swp_entry_t entry;
1755
1756 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1757 entry = pmd_to_swp_entry(orig_pmd);
1758 page = pfn_to_page(swp_offset(entry));
1759 flush_needed = 0;
1760 } else
1761 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1762
1763 if (PageAnon(page)) {
1764 zap_deposited_table(tlb->mm, pmd);
1765 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1766 } else {
1767 if (arch_needs_pgtable_deposit())
1768 zap_deposited_table(tlb->mm, pmd);
1769 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1770 }
1771
1772 spin_unlock(ptl);
1773 if (flush_needed)
1774 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1775 }
1776 return 1;
1777}
1778
1779#ifndef pmd_move_must_withdraw
1780static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1781 spinlock_t *old_pmd_ptl,
1782 struct vm_area_struct *vma)
1783{
1784 /*
1785 * With split pmd lock we also need to move preallocated
1786 * PTE page table if new_pmd is on different PMD page table.
1787 *
1788 * We also don't deposit and withdraw tables for file pages.
1789 */
1790 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1791}
1792#endif
1793
1794static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1795{
1796#ifdef CONFIG_MEM_SOFT_DIRTY
1797 if (unlikely(is_pmd_migration_entry(pmd)))
1798 pmd = pmd_swp_mksoft_dirty(pmd);
1799 else if (pmd_present(pmd))
1800 pmd = pmd_mksoft_dirty(pmd);
1801#endif
1802 return pmd;
1803}
1804
1805bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1806 unsigned long new_addr, unsigned long old_end,
1807 pmd_t *old_pmd, pmd_t *new_pmd)
1808{
1809 spinlock_t *old_ptl, *new_ptl;
1810 pmd_t pmd;
1811 struct mm_struct *mm = vma->vm_mm;
1812 bool force_flush = false;
1813
1814 if ((old_addr & ~HPAGE_PMD_MASK) ||
1815 (new_addr & ~HPAGE_PMD_MASK) ||
1816 old_end - old_addr < HPAGE_PMD_SIZE)
1817 return false;
1818
1819 /*
1820 * The destination pmd shouldn't be established, free_pgtables()
1821 * should have release it.
1822 */
1823 if (WARN_ON(!pmd_none(*new_pmd))) {
1824 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1825 return false;
1826 }
1827
1828 /*
1829 * We don't have to worry about the ordering of src and dst
1830 * ptlocks because exclusive mmap_sem prevents deadlock.
1831 */
1832 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1833 if (old_ptl) {
1834 new_ptl = pmd_lockptr(mm, new_pmd);
1835 if (new_ptl != old_ptl)
1836 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1837 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1838 if (pmd_present(pmd))
1839 force_flush = true;
1840 VM_BUG_ON(!pmd_none(*new_pmd));
1841
1842 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1843 pgtable_t pgtable;
1844 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1845 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1846 }
1847 pmd = move_soft_dirty_pmd(pmd);
1848 set_pmd_at(mm, new_addr, new_pmd, pmd);
1849 if (force_flush)
1850 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1851 if (new_ptl != old_ptl)
1852 spin_unlock(new_ptl);
1853 spin_unlock(old_ptl);
1854 return true;
1855 }
1856 return false;
1857}
1858
1859/*
1860 * Returns
1861 * - 0 if PMD could not be locked
1862 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1863 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1864 */
1865int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1866 unsigned long addr, pgprot_t newprot, int prot_numa)
1867{
1868 struct mm_struct *mm = vma->vm_mm;
1869 spinlock_t *ptl;
1870 pmd_t entry;
1871 bool preserve_write;
1872 int ret;
1873
1874 ptl = __pmd_trans_huge_lock(pmd, vma);
1875 if (!ptl)
1876 return 0;
1877
1878 preserve_write = prot_numa && pmd_write(*pmd);
1879 ret = 1;
1880
1881#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1882 if (is_swap_pmd(*pmd)) {
1883 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1884
1885 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1886 if (is_write_migration_entry(entry)) {
1887 pmd_t newpmd;
1888 /*
1889 * A protection check is difficult so
1890 * just be safe and disable write
1891 */
1892 make_migration_entry_read(&entry);
1893 newpmd = swp_entry_to_pmd(entry);
1894 if (pmd_swp_soft_dirty(*pmd))
1895 newpmd = pmd_swp_mksoft_dirty(newpmd);
1896 set_pmd_at(mm, addr, pmd, newpmd);
1897 }
1898 goto unlock;
1899 }
1900#endif
1901
1902 /*
1903 * Avoid trapping faults against the zero page. The read-only
1904 * data is likely to be read-cached on the local CPU and
1905 * local/remote hits to the zero page are not interesting.
1906 */
1907 if (prot_numa && is_huge_zero_pmd(*pmd))
1908 goto unlock;
1909
1910 if (prot_numa && pmd_protnone(*pmd))
1911 goto unlock;
1912
1913 /*
1914 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1915 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1916 * which is also under down_read(mmap_sem):
1917 *
1918 * CPU0: CPU1:
1919 * change_huge_pmd(prot_numa=1)
1920 * pmdp_huge_get_and_clear_notify()
1921 * madvise_dontneed()
1922 * zap_pmd_range()
1923 * pmd_trans_huge(*pmd) == 0 (without ptl)
1924 * // skip the pmd
1925 * set_pmd_at();
1926 * // pmd is re-established
1927 *
1928 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1929 * which may break userspace.
1930 *
1931 * pmdp_invalidate() is required to make sure we don't miss
1932 * dirty/young flags set by hardware.
1933 */
1934 entry = pmdp_invalidate(vma, addr, pmd);
1935
1936 entry = pmd_modify(entry, newprot);
1937 if (preserve_write)
1938 entry = pmd_mk_savedwrite(entry);
1939 ret = HPAGE_PMD_NR;
1940 set_pmd_at(mm, addr, pmd, entry);
1941 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1942unlock:
1943 spin_unlock(ptl);
1944 return ret;
1945}
1946
1947/*
1948 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1949 *
1950 * Note that if it returns page table lock pointer, this routine returns without
1951 * unlocking page table lock. So callers must unlock it.
1952 */
1953spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1954{
1955 spinlock_t *ptl;
1956 ptl = pmd_lock(vma->vm_mm, pmd);
1957 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1958 pmd_devmap(*pmd)))
1959 return ptl;
1960 spin_unlock(ptl);
1961 return NULL;
1962}
1963
1964/*
1965 * Returns true if a given pud maps a thp, false otherwise.
1966 *
1967 * Note that if it returns true, this routine returns without unlocking page
1968 * table lock. So callers must unlock it.
1969 */
1970spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1971{
1972 spinlock_t *ptl;
1973
1974 ptl = pud_lock(vma->vm_mm, pud);
1975 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1976 return ptl;
1977 spin_unlock(ptl);
1978 return NULL;
1979}
1980
1981#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1982int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1983 pud_t *pud, unsigned long addr)
1984{
1985 spinlock_t *ptl;
1986
1987 ptl = __pud_trans_huge_lock(pud, vma);
1988 if (!ptl)
1989 return 0;
1990 /*
1991 * For architectures like ppc64 we look at deposited pgtable
1992 * when calling pudp_huge_get_and_clear. So do the
1993 * pgtable_trans_huge_withdraw after finishing pudp related
1994 * operations.
1995 */
1996 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1997 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1998 if (vma_is_dax(vma)) {
1999 spin_unlock(ptl);
2000 /* No zero page support yet */
2001 } else {
2002 /* No support for anonymous PUD pages yet */
2003 BUG();
2004 }
2005 return 1;
2006}
2007
2008static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2009 unsigned long haddr)
2010{
2011 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2012 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2013 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2014 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2015
2016 count_vm_event(THP_SPLIT_PUD);
2017
2018 pudp_huge_clear_flush_notify(vma, haddr, pud);
2019}
2020
2021void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2022 unsigned long address)
2023{
2024 spinlock_t *ptl;
2025 struct mmu_notifier_range range;
2026
2027 mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PUD_MASK,
2028 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2029 mmu_notifier_invalidate_range_start(&range);
2030 ptl = pud_lock(vma->vm_mm, pud);
2031 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2032 goto out;
2033 __split_huge_pud_locked(vma, pud, range.start);
2034
2035out:
2036 spin_unlock(ptl);
2037 /*
2038 * No need to double call mmu_notifier->invalidate_range() callback as
2039 * the above pudp_huge_clear_flush_notify() did already call it.
2040 */
2041 mmu_notifier_invalidate_range_only_end(&range);
2042}
2043#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2044
2045static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2046 unsigned long haddr, pmd_t *pmd)
2047{
2048 struct mm_struct *mm = vma->vm_mm;
2049 pgtable_t pgtable;
2050 pmd_t _pmd;
2051 int i;
2052
2053 /*
2054 * Leave pmd empty until pte is filled note that it is fine to delay
2055 * notification until mmu_notifier_invalidate_range_end() as we are
2056 * replacing a zero pmd write protected page with a zero pte write
2057 * protected page.
2058 *
2059 * See Documentation/vm/mmu_notifier.rst
2060 */
2061 pmdp_huge_clear_flush(vma, haddr, pmd);
2062
2063 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2064 pmd_populate(mm, &_pmd, pgtable);
2065
2066 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2067 pte_t *pte, entry;
2068 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2069 entry = pte_mkspecial(entry);
2070 pte = pte_offset_map(&_pmd, haddr);
2071 VM_BUG_ON(!pte_none(*pte));
2072 set_pte_at(mm, haddr, pte, entry);
2073 pte_unmap(pte);
2074 }
2075 smp_wmb(); /* make pte visible before pmd */
2076 pmd_populate(mm, pmd, pgtable);
2077}
2078
2079static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2080 unsigned long haddr, bool freeze)
2081{
2082 struct mm_struct *mm = vma->vm_mm;
2083 struct page *page;
2084 pgtable_t pgtable;
2085 pmd_t old_pmd, _pmd;
2086 bool young, write, soft_dirty, pmd_migration = false;
2087 unsigned long addr;
2088 int i;
2089
2090 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2091 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2092 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2093 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2094 && !pmd_devmap(*pmd));
2095
2096 count_vm_event(THP_SPLIT_PMD);
2097
2098 if (!vma_is_anonymous(vma)) {
2099 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2100 /*
2101 * We are going to unmap this huge page. So
2102 * just go ahead and zap it
2103 */
2104 if (arch_needs_pgtable_deposit())
2105 zap_deposited_table(mm, pmd);
2106 if (vma_is_dax(vma))
2107 return;
2108 page = pmd_page(_pmd);
2109 if (!PageDirty(page) && pmd_dirty(_pmd))
2110 set_page_dirty(page);
2111 if (!PageReferenced(page) && pmd_young(_pmd))
2112 SetPageReferenced(page);
2113 page_remove_rmap(page, true);
2114 put_page(page);
2115 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2116 return;
2117 } else if (is_huge_zero_pmd(*pmd)) {
2118 /*
2119 * FIXME: Do we want to invalidate secondary mmu by calling
2120 * mmu_notifier_invalidate_range() see comments below inside
2121 * __split_huge_pmd() ?
2122 *
2123 * We are going from a zero huge page write protected to zero
2124 * small page also write protected so it does not seems useful
2125 * to invalidate secondary mmu at this time.
2126 */
2127 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2128 }
2129
2130 /*
2131 * Up to this point the pmd is present and huge and userland has the
2132 * whole access to the hugepage during the split (which happens in
2133 * place). If we overwrite the pmd with the not-huge version pointing
2134 * to the pte here (which of course we could if all CPUs were bug
2135 * free), userland could trigger a small page size TLB miss on the
2136 * small sized TLB while the hugepage TLB entry is still established in
2137 * the huge TLB. Some CPU doesn't like that.
2138 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2139 * 383 on page 93. Intel should be safe but is also warns that it's
2140 * only safe if the permission and cache attributes of the two entries
2141 * loaded in the two TLB is identical (which should be the case here).
2142 * But it is generally safer to never allow small and huge TLB entries
2143 * for the same virtual address to be loaded simultaneously. So instead
2144 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2145 * current pmd notpresent (atomically because here the pmd_trans_huge
2146 * must remain set at all times on the pmd until the split is complete
2147 * for this pmd), then we flush the SMP TLB and finally we write the
2148 * non-huge version of the pmd entry with pmd_populate.
2149 */
2150 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2151
2152 pmd_migration = is_pmd_migration_entry(old_pmd);
2153 if (unlikely(pmd_migration)) {
2154 swp_entry_t entry;
2155
2156 entry = pmd_to_swp_entry(old_pmd);
2157 page = pfn_to_page(swp_offset(entry));
2158 write = is_write_migration_entry(entry);
2159 young = false;
2160 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2161 } else {
2162 page = pmd_page(old_pmd);
2163 if (pmd_dirty(old_pmd))
2164 SetPageDirty(page);
2165 write = pmd_write(old_pmd);
2166 young = pmd_young(old_pmd);
2167 soft_dirty = pmd_soft_dirty(old_pmd);
2168 }
2169 VM_BUG_ON_PAGE(!page_count(page), page);
2170 page_ref_add(page, HPAGE_PMD_NR - 1);
2171
2172 /*
2173 * Withdraw the table only after we mark the pmd entry invalid.
2174 * This's critical for some architectures (Power).
2175 */
2176 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2177 pmd_populate(mm, &_pmd, pgtable);
2178
2179 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2180 pte_t entry, *pte;
2181 /*
2182 * Note that NUMA hinting access restrictions are not
2183 * transferred to avoid any possibility of altering
2184 * permissions across VMAs.
2185 */
2186 if (freeze || pmd_migration) {
2187 swp_entry_t swp_entry;
2188 swp_entry = make_migration_entry(page + i, write);
2189 entry = swp_entry_to_pte(swp_entry);
2190 if (soft_dirty)
2191 entry = pte_swp_mksoft_dirty(entry);
2192 } else {
2193 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2194 entry = maybe_mkwrite(entry, vma);
2195 if (!write)
2196 entry = pte_wrprotect(entry);
2197 if (!young)
2198 entry = pte_mkold(entry);
2199 if (soft_dirty)
2200 entry = pte_mksoft_dirty(entry);
2201 }
2202 pte = pte_offset_map(&_pmd, addr);
2203 BUG_ON(!pte_none(*pte));
2204 set_pte_at(mm, addr, pte, entry);
2205 atomic_inc(&page[i]._mapcount);
2206 pte_unmap(pte);
2207 }
2208
2209 /*
2210 * Set PG_double_map before dropping compound_mapcount to avoid
2211 * false-negative page_mapped().
2212 */
2213 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2214 for (i = 0; i < HPAGE_PMD_NR; i++)
2215 atomic_inc(&page[i]._mapcount);
2216 }
2217
2218 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2219 /* Last compound_mapcount is gone. */
2220 __dec_node_page_state(page, NR_ANON_THPS);
2221 if (TestClearPageDoubleMap(page)) {
2222 /* No need in mapcount reference anymore */
2223 for (i = 0; i < HPAGE_PMD_NR; i++)
2224 atomic_dec(&page[i]._mapcount);
2225 }
2226 }
2227
2228 smp_wmb(); /* make pte visible before pmd */
2229 pmd_populate(mm, pmd, pgtable);
2230
2231 if (freeze) {
2232 for (i = 0; i < HPAGE_PMD_NR; i++) {
2233 page_remove_rmap(page + i, false);
2234 put_page(page + i);
2235 }
2236 }
2237}
2238
2239void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2240 unsigned long address, bool freeze, struct page *page)
2241{
2242 spinlock_t *ptl;
2243 struct mmu_notifier_range range;
2244
2245 mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PMD_MASK,
2246 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2247 mmu_notifier_invalidate_range_start(&range);
2248 ptl = pmd_lock(vma->vm_mm, pmd);
2249
2250 /*
2251 * If caller asks to setup a migration entries, we need a page to check
2252 * pmd against. Otherwise we can end up replacing wrong page.
2253 */
2254 VM_BUG_ON(freeze && !page);
2255 if (page && page != pmd_page(*pmd))
2256 goto out;
2257
2258 if (pmd_trans_huge(*pmd)) {
2259 page = pmd_page(*pmd);
2260 if (PageMlocked(page))
2261 clear_page_mlock(page);
2262 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2263 goto out;
2264 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2265out:
2266 spin_unlock(ptl);
2267 /*
2268 * No need to double call mmu_notifier->invalidate_range() callback.
2269 * They are 3 cases to consider inside __split_huge_pmd_locked():
2270 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2271 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2272 * fault will trigger a flush_notify before pointing to a new page
2273 * (it is fine if the secondary mmu keeps pointing to the old zero
2274 * page in the meantime)
2275 * 3) Split a huge pmd into pte pointing to the same page. No need
2276 * to invalidate secondary tlb entry they are all still valid.
2277 * any further changes to individual pte will notify. So no need
2278 * to call mmu_notifier->invalidate_range()
2279 */
2280 mmu_notifier_invalidate_range_only_end(&range);
2281}
2282
2283void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2284 bool freeze, struct page *page)
2285{
2286 pgd_t *pgd;
2287 p4d_t *p4d;
2288 pud_t *pud;
2289 pmd_t *pmd;
2290
2291 pgd = pgd_offset(vma->vm_mm, address);
2292 if (!pgd_present(*pgd))
2293 return;
2294
2295 p4d = p4d_offset(pgd, address);
2296 if (!p4d_present(*p4d))
2297 return;
2298
2299 pud = pud_offset(p4d, address);
2300 if (!pud_present(*pud))
2301 return;
2302
2303 pmd = pmd_offset(pud, address);
2304
2305 __split_huge_pmd(vma, pmd, address, freeze, page);
2306}
2307
2308void vma_adjust_trans_huge(struct vm_area_struct *vma,
2309 unsigned long start,
2310 unsigned long end,
2311 long adjust_next)
2312{
2313 /*
2314 * If the new start address isn't hpage aligned and it could
2315 * previously contain an hugepage: check if we need to split
2316 * an huge pmd.
2317 */
2318 if (start & ~HPAGE_PMD_MASK &&
2319 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2320 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2321 split_huge_pmd_address(vma, start, false, NULL);
2322
2323 /*
2324 * If the new end address isn't hpage aligned and it could
2325 * previously contain an hugepage: check if we need to split
2326 * an huge pmd.
2327 */
2328 if (end & ~HPAGE_PMD_MASK &&
2329 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2330 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2331 split_huge_pmd_address(vma, end, false, NULL);
2332
2333 /*
2334 * If we're also updating the vma->vm_next->vm_start, if the new
2335 * vm_next->vm_start isn't page aligned and it could previously
2336 * contain an hugepage: check if we need to split an huge pmd.
2337 */
2338 if (adjust_next > 0) {
2339 struct vm_area_struct *next = vma->vm_next;
2340 unsigned long nstart = next->vm_start;
2341 nstart += adjust_next << PAGE_SHIFT;
2342 if (nstart & ~HPAGE_PMD_MASK &&
2343 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2344 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2345 split_huge_pmd_address(next, nstart, false, NULL);
2346 }
2347}
2348
2349static void unmap_page(struct page *page)
2350{
2351 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2352 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2353 bool unmap_success;
2354
2355 VM_BUG_ON_PAGE(!PageHead(page), page);
2356
2357 if (PageAnon(page))
2358 ttu_flags |= TTU_SPLIT_FREEZE;
2359
2360 unmap_success = try_to_unmap(page, ttu_flags);
2361 VM_BUG_ON_PAGE(!unmap_success, page);
2362}
2363
2364static void remap_page(struct page *page)
2365{
2366 int i;
2367 if (PageTransHuge(page)) {
2368 remove_migration_ptes(page, page, true);
2369 } else {
2370 for (i = 0; i < HPAGE_PMD_NR; i++)
2371 remove_migration_ptes(page + i, page + i, true);
2372 }
2373}
2374
2375static void __split_huge_page_tail(struct page *head, int tail,
2376 struct lruvec *lruvec, struct list_head *list)
2377{
2378 struct page *page_tail = head + tail;
2379
2380 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2381
2382 /*
2383 * Clone page flags before unfreezing refcount.
2384 *
2385 * After successful get_page_unless_zero() might follow flags change,
2386 * for exmaple lock_page() which set PG_waiters.
2387 */
2388 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2389 page_tail->flags |= (head->flags &
2390 ((1L << PG_referenced) |
2391 (1L << PG_swapbacked) |
2392 (1L << PG_swapcache) |
2393 (1L << PG_mlocked) |
2394 (1L << PG_uptodate) |
2395 (1L << PG_active) |
2396 (1L << PG_workingset) |
2397 (1L << PG_locked) |
2398 (1L << PG_unevictable) |
2399 (1L << PG_dirty)));
2400
2401 /* ->mapping in first tail page is compound_mapcount */
2402 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2403 page_tail);
2404 page_tail->mapping = head->mapping;
2405 page_tail->index = head->index + tail;
2406
2407 /* Page flags must be visible before we make the page non-compound. */
2408 smp_wmb();
2409
2410 /*
2411 * Clear PageTail before unfreezing page refcount.
2412 *
2413 * After successful get_page_unless_zero() might follow put_page()
2414 * which needs correct compound_head().
2415 */
2416 clear_compound_head(page_tail);
2417
2418 /* Finally unfreeze refcount. Additional reference from page cache. */
2419 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2420 PageSwapCache(head)));
2421
2422 if (page_is_young(head))
2423 set_page_young(page_tail);
2424 if (page_is_idle(head))
2425 set_page_idle(page_tail);
2426
2427 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2428
2429 /*
2430 * always add to the tail because some iterators expect new
2431 * pages to show after the currently processed elements - e.g.
2432 * migrate_pages
2433 */
2434 lru_add_page_tail(head, page_tail, lruvec, list);
2435}
2436
2437static void __split_huge_page(struct page *page, struct list_head *list,
2438 pgoff_t end, unsigned long flags)
2439{
2440 struct page *head = compound_head(page);
2441 pg_data_t *pgdat = page_pgdat(head);
2442 struct lruvec *lruvec;
2443 int i;
2444
2445 lruvec = mem_cgroup_page_lruvec(head, pgdat);
2446
2447 /* complete memcg works before add pages to LRU */
2448 mem_cgroup_split_huge_fixup(head);
2449
2450 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2451 __split_huge_page_tail(head, i, lruvec, list);
2452 /* Some pages can be beyond i_size: drop them from page cache */
2453 if (head[i].index >= end) {
2454 ClearPageDirty(head + i);
2455 __delete_from_page_cache(head + i, NULL);
2456 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2457 shmem_uncharge(head->mapping->host, 1);
2458 put_page(head + i);
2459 }
2460 }
2461
2462 ClearPageCompound(head);
2463 /* See comment in __split_huge_page_tail() */
2464 if (PageAnon(head)) {
2465 /* Additional pin to swap cache */
2466 if (PageSwapCache(head))
2467 page_ref_add(head, 2);
2468 else
2469 page_ref_inc(head);
2470 } else {
2471 /* Additional pin to page cache */
2472 page_ref_add(head, 2);
2473 xa_unlock(&head->mapping->i_pages);
2474 }
2475
2476 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2477
2478 remap_page(head);
2479
2480 for (i = 0; i < HPAGE_PMD_NR; i++) {
2481 struct page *subpage = head + i;
2482 if (subpage == page)
2483 continue;
2484 unlock_page(subpage);
2485
2486 /*
2487 * Subpages may be freed if there wasn't any mapping
2488 * like if add_to_swap() is running on a lru page that
2489 * had its mapping zapped. And freeing these pages
2490 * requires taking the lru_lock so we do the put_page
2491 * of the tail pages after the split is complete.
2492 */
2493 put_page(subpage);
2494 }
2495}
2496
2497int total_mapcount(struct page *page)
2498{
2499 int i, compound, ret;
2500
2501 VM_BUG_ON_PAGE(PageTail(page), page);
2502
2503 if (likely(!PageCompound(page)))
2504 return atomic_read(&page->_mapcount) + 1;
2505
2506 compound = compound_mapcount(page);
2507 if (PageHuge(page))
2508 return compound;
2509 ret = compound;
2510 for (i = 0; i < HPAGE_PMD_NR; i++)
2511 ret += atomic_read(&page[i]._mapcount) + 1;
2512 /* File pages has compound_mapcount included in _mapcount */
2513 if (!PageAnon(page))
2514 return ret - compound * HPAGE_PMD_NR;
2515 if (PageDoubleMap(page))
2516 ret -= HPAGE_PMD_NR;
2517 return ret;
2518}
2519
2520/*
2521 * This calculates accurately how many mappings a transparent hugepage
2522 * has (unlike page_mapcount() which isn't fully accurate). This full
2523 * accuracy is primarily needed to know if copy-on-write faults can
2524 * reuse the page and change the mapping to read-write instead of
2525 * copying them. At the same time this returns the total_mapcount too.
2526 *
2527 * The function returns the highest mapcount any one of the subpages
2528 * has. If the return value is one, even if different processes are
2529 * mapping different subpages of the transparent hugepage, they can
2530 * all reuse it, because each process is reusing a different subpage.
2531 *
2532 * The total_mapcount is instead counting all virtual mappings of the
2533 * subpages. If the total_mapcount is equal to "one", it tells the
2534 * caller all mappings belong to the same "mm" and in turn the
2535 * anon_vma of the transparent hugepage can become the vma->anon_vma
2536 * local one as no other process may be mapping any of the subpages.
2537 *
2538 * It would be more accurate to replace page_mapcount() with
2539 * page_trans_huge_mapcount(), however we only use
2540 * page_trans_huge_mapcount() in the copy-on-write faults where we
2541 * need full accuracy to avoid breaking page pinning, because
2542 * page_trans_huge_mapcount() is slower than page_mapcount().
2543 */
2544int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2545{
2546 int i, ret, _total_mapcount, mapcount;
2547
2548 /* hugetlbfs shouldn't call it */
2549 VM_BUG_ON_PAGE(PageHuge(page), page);
2550
2551 if (likely(!PageTransCompound(page))) {
2552 mapcount = atomic_read(&page->_mapcount) + 1;
2553 if (total_mapcount)
2554 *total_mapcount = mapcount;
2555 return mapcount;
2556 }
2557
2558 page = compound_head(page);
2559
2560 _total_mapcount = ret = 0;
2561 for (i = 0; i < HPAGE_PMD_NR; i++) {
2562 mapcount = atomic_read(&page[i]._mapcount) + 1;
2563 ret = max(ret, mapcount);
2564 _total_mapcount += mapcount;
2565 }
2566 if (PageDoubleMap(page)) {
2567 ret -= 1;
2568 _total_mapcount -= HPAGE_PMD_NR;
2569 }
2570 mapcount = compound_mapcount(page);
2571 ret += mapcount;
2572 _total_mapcount += mapcount;
2573 if (total_mapcount)
2574 *total_mapcount = _total_mapcount;
2575 return ret;
2576}
2577
2578/* Racy check whether the huge page can be split */
2579bool can_split_huge_page(struct page *page, int *pextra_pins)
2580{
2581 int extra_pins;
2582
2583 /* Additional pins from page cache */
2584 if (PageAnon(page))
2585 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2586 else
2587 extra_pins = HPAGE_PMD_NR;
2588 if (pextra_pins)
2589 *pextra_pins = extra_pins;
2590 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2591}
2592
2593/*
2594 * This function splits huge page into normal pages. @page can point to any
2595 * subpage of huge page to split. Split doesn't change the position of @page.
2596 *
2597 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2598 * The huge page must be locked.
2599 *
2600 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2601 *
2602 * Both head page and tail pages will inherit mapping, flags, and so on from
2603 * the hugepage.
2604 *
2605 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2606 * they are not mapped.
2607 *
2608 * Returns 0 if the hugepage is split successfully.
2609 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2610 * us.
2611 */
2612int split_huge_page_to_list(struct page *page, struct list_head *list)
2613{
2614 struct page *head = compound_head(page);
2615 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2616 struct anon_vma *anon_vma = NULL;
2617 struct address_space *mapping = NULL;
2618 int count, mapcount, extra_pins, ret;
2619 bool mlocked;
2620 unsigned long flags;
2621 pgoff_t end;
2622
2623 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2624 VM_BUG_ON_PAGE(!PageLocked(page), page);
2625 VM_BUG_ON_PAGE(!PageCompound(page), page);
2626
2627 if (PageWriteback(page))
2628 return -EBUSY;
2629
2630 if (PageAnon(head)) {
2631 /*
2632 * The caller does not necessarily hold an mmap_sem that would
2633 * prevent the anon_vma disappearing so we first we take a
2634 * reference to it and then lock the anon_vma for write. This
2635 * is similar to page_lock_anon_vma_read except the write lock
2636 * is taken to serialise against parallel split or collapse
2637 * operations.
2638 */
2639 anon_vma = page_get_anon_vma(head);
2640 if (!anon_vma) {
2641 ret = -EBUSY;
2642 goto out;
2643 }
2644 end = -1;
2645 mapping = NULL;
2646 anon_vma_lock_write(anon_vma);
2647 } else {
2648 mapping = head->mapping;
2649
2650 /* Truncated ? */
2651 if (!mapping) {
2652 ret = -EBUSY;
2653 goto out;
2654 }
2655
2656 anon_vma = NULL;
2657 i_mmap_lock_read(mapping);
2658
2659 /*
2660 *__split_huge_page() may need to trim off pages beyond EOF:
2661 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2662 * which cannot be nested inside the page tree lock. So note
2663 * end now: i_size itself may be changed at any moment, but
2664 * head page lock is good enough to serialize the trimming.
2665 */
2666 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2667 }
2668
2669 /*
2670 * Racy check if we can split the page, before unmap_page() will
2671 * split PMDs
2672 */
2673 if (!can_split_huge_page(head, &extra_pins)) {
2674 ret = -EBUSY;
2675 goto out_unlock;
2676 }
2677
2678 mlocked = PageMlocked(page);
2679 unmap_page(head);
2680 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2681
2682 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2683 if (mlocked)
2684 lru_add_drain();
2685
2686 /* prevent PageLRU to go away from under us, and freeze lru stats */
2687 spin_lock_irqsave(&pgdata->lru_lock, flags);
2688
2689 if (mapping) {
2690 XA_STATE(xas, &mapping->i_pages, page_index(head));
2691
2692 /*
2693 * Check if the head page is present in page cache.
2694 * We assume all tail are present too, if head is there.
2695 */
2696 xa_lock(&mapping->i_pages);
2697 if (xas_load(&xas) != head)
2698 goto fail;
2699 }
2700
2701 /* Prevent deferred_split_scan() touching ->_refcount */
2702 spin_lock(&pgdata->split_queue_lock);
2703 count = page_count(head);
2704 mapcount = total_mapcount(head);
2705 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2706 if (!list_empty(page_deferred_list(head))) {
2707 pgdata->split_queue_len--;
2708 list_del(page_deferred_list(head));
2709 }
2710 if (mapping)
2711 __dec_node_page_state(page, NR_SHMEM_THPS);
2712 spin_unlock(&pgdata->split_queue_lock);
2713 __split_huge_page(page, list, end, flags);
2714 if (PageSwapCache(head)) {
2715 swp_entry_t entry = { .val = page_private(head) };
2716
2717 ret = split_swap_cluster(entry);
2718 } else
2719 ret = 0;
2720 } else {
2721 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2722 pr_alert("total_mapcount: %u, page_count(): %u\n",
2723 mapcount, count);
2724 if (PageTail(page))
2725 dump_page(head, NULL);
2726 dump_page(page, "total_mapcount(head) > 0");
2727 BUG();
2728 }
2729 spin_unlock(&pgdata->split_queue_lock);
2730fail: if (mapping)
2731 xa_unlock(&mapping->i_pages);
2732 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2733 remap_page(head);
2734 ret = -EBUSY;
2735 }
2736
2737out_unlock:
2738 if (anon_vma) {
2739 anon_vma_unlock_write(anon_vma);
2740 put_anon_vma(anon_vma);
2741 }
2742 if (mapping)
2743 i_mmap_unlock_read(mapping);
2744out:
2745 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2746 return ret;
2747}
2748
2749void free_transhuge_page(struct page *page)
2750{
2751 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2752 unsigned long flags;
2753
2754 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2755 if (!list_empty(page_deferred_list(page))) {
2756 pgdata->split_queue_len--;
2757 list_del(page_deferred_list(page));
2758 }
2759 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2760 free_compound_page(page);
2761}
2762
2763void deferred_split_huge_page(struct page *page)
2764{
2765 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2766 unsigned long flags;
2767
2768 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2769
2770 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2771 if (list_empty(page_deferred_list(page))) {
2772 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2773 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2774 pgdata->split_queue_len++;
2775 }
2776 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2777}
2778
2779static unsigned long deferred_split_count(struct shrinker *shrink,
2780 struct shrink_control *sc)
2781{
2782 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2783 return READ_ONCE(pgdata->split_queue_len);
2784}
2785
2786static unsigned long deferred_split_scan(struct shrinker *shrink,
2787 struct shrink_control *sc)
2788{
2789 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2790 unsigned long flags;
2791 LIST_HEAD(list), *pos, *next;
2792 struct page *page;
2793 int split = 0;
2794
2795 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2796 /* Take pin on all head pages to avoid freeing them under us */
2797 list_for_each_safe(pos, next, &pgdata->split_queue) {
2798 page = list_entry((void *)pos, struct page, mapping);
2799 page = compound_head(page);
2800 if (get_page_unless_zero(page)) {
2801 list_move(page_deferred_list(page), &list);
2802 } else {
2803 /* We lost race with put_compound_page() */
2804 list_del_init(page_deferred_list(page));
2805 pgdata->split_queue_len--;
2806 }
2807 if (!--sc->nr_to_scan)
2808 break;
2809 }
2810 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2811
2812 list_for_each_safe(pos, next, &list) {
2813 page = list_entry((void *)pos, struct page, mapping);
2814 if (!trylock_page(page))
2815 goto next;
2816 /* split_huge_page() removes page from list on success */
2817 if (!split_huge_page(page))
2818 split++;
2819 unlock_page(page);
2820next:
2821 put_page(page);
2822 }
2823
2824 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2825 list_splice_tail(&list, &pgdata->split_queue);
2826 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2827
2828 /*
2829 * Stop shrinker if we didn't split any page, but the queue is empty.
2830 * This can happen if pages were freed under us.
2831 */
2832 if (!split && list_empty(&pgdata->split_queue))
2833 return SHRINK_STOP;
2834 return split;
2835}
2836
2837static struct shrinker deferred_split_shrinker = {
2838 .count_objects = deferred_split_count,
2839 .scan_objects = deferred_split_scan,
2840 .seeks = DEFAULT_SEEKS,
2841 .flags = SHRINKER_NUMA_AWARE,
2842};
2843
2844#ifdef CONFIG_DEBUG_FS
2845static int split_huge_pages_set(void *data, u64 val)
2846{
2847 struct zone *zone;
2848 struct page *page;
2849 unsigned long pfn, max_zone_pfn;
2850 unsigned long total = 0, split = 0;
2851
2852 if (val != 1)
2853 return -EINVAL;
2854
2855 for_each_populated_zone(zone) {
2856 max_zone_pfn = zone_end_pfn(zone);
2857 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2858 if (!pfn_valid(pfn))
2859 continue;
2860
2861 page = pfn_to_page(pfn);
2862 if (!get_page_unless_zero(page))
2863 continue;
2864
2865 if (zone != page_zone(page))
2866 goto next;
2867
2868 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2869 goto next;
2870
2871 total++;
2872 lock_page(page);
2873 if (!split_huge_page(page))
2874 split++;
2875 unlock_page(page);
2876next:
2877 put_page(page);
2878 }
2879 }
2880
2881 pr_info("%lu of %lu THP split\n", split, total);
2882
2883 return 0;
2884}
2885DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2886 "%llu\n");
2887
2888static int __init split_huge_pages_debugfs(void)
2889{
2890 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2891 &split_huge_pages_fops);
2892 return 0;
2893}
2894late_initcall(split_huge_pages_debugfs);
2895#endif
2896
2897#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2898void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2899 struct page *page)
2900{
2901 struct vm_area_struct *vma = pvmw->vma;
2902 struct mm_struct *mm = vma->vm_mm;
2903 unsigned long address = pvmw->address;
2904 pmd_t pmdval;
2905 swp_entry_t entry;
2906 pmd_t pmdswp;
2907
2908 if (!(pvmw->pmd && !pvmw->pte))
2909 return;
2910
2911 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2912 pmdval = *pvmw->pmd;
2913 pmdp_invalidate(vma, address, pvmw->pmd);
2914 if (pmd_dirty(pmdval))
2915 set_page_dirty(page);
2916 entry = make_migration_entry(page, pmd_write(pmdval));
2917 pmdswp = swp_entry_to_pmd(entry);
2918 if (pmd_soft_dirty(pmdval))
2919 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2920 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2921 page_remove_rmap(page, true);
2922 put_page(page);
2923}
2924
2925void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2926{
2927 struct vm_area_struct *vma = pvmw->vma;
2928 struct mm_struct *mm = vma->vm_mm;
2929 unsigned long address = pvmw->address;
2930 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2931 pmd_t pmde;
2932 swp_entry_t entry;
2933
2934 if (!(pvmw->pmd && !pvmw->pte))
2935 return;
2936
2937 entry = pmd_to_swp_entry(*pvmw->pmd);
2938 get_page(new);
2939 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2940 if (pmd_swp_soft_dirty(*pvmw->pmd))
2941 pmde = pmd_mksoft_dirty(pmde);
2942 if (is_write_migration_entry(entry))
2943 pmde = maybe_pmd_mkwrite(pmde, vma);
2944
2945 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2946 if (PageAnon(new))
2947 page_add_anon_rmap(new, vma, mmun_start, true);
2948 else
2949 page_add_file_rmap(new, true);
2950 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2951 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2952 mlock_vma_page(new);
2953 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2954}
2955#endif
2956