Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2009 Red Hat, Inc.
4 */
5
6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/sched/mm.h>
11#include <linux/sched/coredump.h>
12#include <linux/sched/numa_balancing.h>
13#include <linux/highmem.h>
14#include <linux/hugetlb.h>
15#include <linux/mmu_notifier.h>
16#include <linux/rmap.h>
17#include <linux/swap.h>
18#include <linux/shrinker.h>
19#include <linux/mm_inline.h>
20#include <linux/swapops.h>
21#include <linux/dax.h>
22#include <linux/khugepaged.h>
23#include <linux/freezer.h>
24#include <linux/pfn_t.h>
25#include <linux/mman.h>
26#include <linux/memremap.h>
27#include <linux/pagemap.h>
28#include <linux/debugfs.h>
29#include <linux/migrate.h>
30#include <linux/hashtable.h>
31#include <linux/userfaultfd_k.h>
32#include <linux/page_idle.h>
33#include <linux/shmem_fs.h>
34#include <linux/oom.h>
35#include <linux/numa.h>
36#include <linux/page_owner.h>
37
38#include <asm/tlb.h>
39#include <asm/pgalloc.h>
40#include "internal.h"
41
42/*
43 * By default, transparent hugepage support is disabled in order to avoid
44 * risking an increased memory footprint for applications that are not
45 * guaranteed to benefit from it. When transparent hugepage support is
46 * enabled, it is for all mappings, and khugepaged scans all mappings.
47 * Defrag is invoked by khugepaged hugepage allocations and by page faults
48 * for all hugepage allocations.
49 */
50unsigned long transparent_hugepage_flags __read_mostly =
51#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
52 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
53#endif
54#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
55 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
56#endif
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
59 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
60
61static struct shrinker deferred_split_shrinker;
62
63static atomic_t huge_zero_refcount;
64struct page *huge_zero_page __read_mostly;
65unsigned long huge_zero_pfn __read_mostly = ~0UL;
66
67static inline bool file_thp_enabled(struct vm_area_struct *vma)
68{
69 return transhuge_vma_enabled(vma, vma->vm_flags) && vma->vm_file &&
70 !inode_is_open_for_write(vma->vm_file->f_inode) &&
71 (vma->vm_flags & VM_EXEC);
72}
73
74bool transparent_hugepage_active(struct vm_area_struct *vma)
75{
76 /* The addr is used to check if the vma size fits */
77 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
78
79 if (!transhuge_vma_suitable(vma, addr))
80 return false;
81 if (vma_is_anonymous(vma))
82 return __transparent_hugepage_enabled(vma);
83 if (vma_is_shmem(vma))
84 return shmem_huge_enabled(vma);
85 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
86 return file_thp_enabled(vma);
87
88 return false;
89}
90
91static bool get_huge_zero_page(void)
92{
93 struct page *zero_page;
94retry:
95 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
96 return true;
97
98 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
99 HPAGE_PMD_ORDER);
100 if (!zero_page) {
101 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
102 return false;
103 }
104 count_vm_event(THP_ZERO_PAGE_ALLOC);
105 preempt_disable();
106 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
107 preempt_enable();
108 __free_pages(zero_page, compound_order(zero_page));
109 goto retry;
110 }
111 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
112
113 /* We take additional reference here. It will be put back by shrinker */
114 atomic_set(&huge_zero_refcount, 2);
115 preempt_enable();
116 return true;
117}
118
119static void put_huge_zero_page(void)
120{
121 /*
122 * Counter should never go to zero here. Only shrinker can put
123 * last reference.
124 */
125 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
126}
127
128struct page *mm_get_huge_zero_page(struct mm_struct *mm)
129{
130 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
131 return READ_ONCE(huge_zero_page);
132
133 if (!get_huge_zero_page())
134 return NULL;
135
136 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
137 put_huge_zero_page();
138
139 return READ_ONCE(huge_zero_page);
140}
141
142void mm_put_huge_zero_page(struct mm_struct *mm)
143{
144 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
145 put_huge_zero_page();
146}
147
148static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
149 struct shrink_control *sc)
150{
151 /* we can free zero page only if last reference remains */
152 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
153}
154
155static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
156 struct shrink_control *sc)
157{
158 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
159 struct page *zero_page = xchg(&huge_zero_page, NULL);
160 BUG_ON(zero_page == NULL);
161 WRITE_ONCE(huge_zero_pfn, ~0UL);
162 __free_pages(zero_page, compound_order(zero_page));
163 return HPAGE_PMD_NR;
164 }
165
166 return 0;
167}
168
169static struct shrinker huge_zero_page_shrinker = {
170 .count_objects = shrink_huge_zero_page_count,
171 .scan_objects = shrink_huge_zero_page_scan,
172 .seeks = DEFAULT_SEEKS,
173};
174
175#ifdef CONFIG_SYSFS
176static ssize_t enabled_show(struct kobject *kobj,
177 struct kobj_attribute *attr, char *buf)
178{
179 const char *output;
180
181 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
182 output = "[always] madvise never";
183 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
184 &transparent_hugepage_flags))
185 output = "always [madvise] never";
186 else
187 output = "always madvise [never]";
188
189 return sysfs_emit(buf, "%s\n", output);
190}
191
192static ssize_t enabled_store(struct kobject *kobj,
193 struct kobj_attribute *attr,
194 const char *buf, size_t count)
195{
196 ssize_t ret = count;
197
198 if (sysfs_streq(buf, "always")) {
199 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
200 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
201 } else if (sysfs_streq(buf, "madvise")) {
202 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
203 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
204 } else if (sysfs_streq(buf, "never")) {
205 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
206 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
207 } else
208 ret = -EINVAL;
209
210 if (ret > 0) {
211 int err = start_stop_khugepaged();
212 if (err)
213 ret = err;
214 }
215 return ret;
216}
217static struct kobj_attribute enabled_attr =
218 __ATTR(enabled, 0644, enabled_show, enabled_store);
219
220ssize_t single_hugepage_flag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf,
222 enum transparent_hugepage_flag flag)
223{
224 return sysfs_emit(buf, "%d\n",
225 !!test_bit(flag, &transparent_hugepage_flags));
226}
227
228ssize_t single_hugepage_flag_store(struct kobject *kobj,
229 struct kobj_attribute *attr,
230 const char *buf, size_t count,
231 enum transparent_hugepage_flag flag)
232{
233 unsigned long value;
234 int ret;
235
236 ret = kstrtoul(buf, 10, &value);
237 if (ret < 0)
238 return ret;
239 if (value > 1)
240 return -EINVAL;
241
242 if (value)
243 set_bit(flag, &transparent_hugepage_flags);
244 else
245 clear_bit(flag, &transparent_hugepage_flags);
246
247 return count;
248}
249
250static ssize_t defrag_show(struct kobject *kobj,
251 struct kobj_attribute *attr, char *buf)
252{
253 const char *output;
254
255 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
256 &transparent_hugepage_flags))
257 output = "[always] defer defer+madvise madvise never";
258 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
259 &transparent_hugepage_flags))
260 output = "always [defer] defer+madvise madvise never";
261 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
262 &transparent_hugepage_flags))
263 output = "always defer [defer+madvise] madvise never";
264 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
265 &transparent_hugepage_flags))
266 output = "always defer defer+madvise [madvise] never";
267 else
268 output = "always defer defer+madvise madvise [never]";
269
270 return sysfs_emit(buf, "%s\n", output);
271}
272
273static ssize_t defrag_store(struct kobject *kobj,
274 struct kobj_attribute *attr,
275 const char *buf, size_t count)
276{
277 if (sysfs_streq(buf, "always")) {
278 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
279 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
280 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
281 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
282 } else if (sysfs_streq(buf, "defer+madvise")) {
283 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
284 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
285 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
286 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
287 } else if (sysfs_streq(buf, "defer")) {
288 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
289 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
290 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
291 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
292 } else if (sysfs_streq(buf, "madvise")) {
293 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
294 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
295 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
296 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
297 } else if (sysfs_streq(buf, "never")) {
298 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
299 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
300 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
301 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
302 } else
303 return -EINVAL;
304
305 return count;
306}
307static struct kobj_attribute defrag_attr =
308 __ATTR(defrag, 0644, defrag_show, defrag_store);
309
310static ssize_t use_zero_page_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf)
312{
313 return single_hugepage_flag_show(kobj, attr, buf,
314 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
315}
316static ssize_t use_zero_page_store(struct kobject *kobj,
317 struct kobj_attribute *attr, const char *buf, size_t count)
318{
319 return single_hugepage_flag_store(kobj, attr, buf, count,
320 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
321}
322static struct kobj_attribute use_zero_page_attr =
323 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
324
325static ssize_t hpage_pmd_size_show(struct kobject *kobj,
326 struct kobj_attribute *attr, char *buf)
327{
328 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
329}
330static struct kobj_attribute hpage_pmd_size_attr =
331 __ATTR_RO(hpage_pmd_size);
332
333static struct attribute *hugepage_attr[] = {
334 &enabled_attr.attr,
335 &defrag_attr.attr,
336 &use_zero_page_attr.attr,
337 &hpage_pmd_size_attr.attr,
338#ifdef CONFIG_SHMEM
339 &shmem_enabled_attr.attr,
340#endif
341 NULL,
342};
343
344static const struct attribute_group hugepage_attr_group = {
345 .attrs = hugepage_attr,
346};
347
348static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
349{
350 int err;
351
352 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
353 if (unlikely(!*hugepage_kobj)) {
354 pr_err("failed to create transparent hugepage kobject\n");
355 return -ENOMEM;
356 }
357
358 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
359 if (err) {
360 pr_err("failed to register transparent hugepage group\n");
361 goto delete_obj;
362 }
363
364 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
365 if (err) {
366 pr_err("failed to register transparent hugepage group\n");
367 goto remove_hp_group;
368 }
369
370 return 0;
371
372remove_hp_group:
373 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
374delete_obj:
375 kobject_put(*hugepage_kobj);
376 return err;
377}
378
379static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
380{
381 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
382 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
383 kobject_put(hugepage_kobj);
384}
385#else
386static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
387{
388 return 0;
389}
390
391static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
392{
393}
394#endif /* CONFIG_SYSFS */
395
396static int __init hugepage_init(void)
397{
398 int err;
399 struct kobject *hugepage_kobj;
400
401 if (!has_transparent_hugepage()) {
402 /*
403 * Hardware doesn't support hugepages, hence disable
404 * DAX PMD support.
405 */
406 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
407 return -EINVAL;
408 }
409
410 /*
411 * hugepages can't be allocated by the buddy allocator
412 */
413 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
414 /*
415 * we use page->mapping and page->index in second tail page
416 * as list_head: assuming THP order >= 2
417 */
418 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
419
420 err = hugepage_init_sysfs(&hugepage_kobj);
421 if (err)
422 goto err_sysfs;
423
424 err = khugepaged_init();
425 if (err)
426 goto err_slab;
427
428 err = register_shrinker(&huge_zero_page_shrinker);
429 if (err)
430 goto err_hzp_shrinker;
431 err = register_shrinker(&deferred_split_shrinker);
432 if (err)
433 goto err_split_shrinker;
434
435 /*
436 * By default disable transparent hugepages on smaller systems,
437 * where the extra memory used could hurt more than TLB overhead
438 * is likely to save. The admin can still enable it through /sys.
439 */
440 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
441 transparent_hugepage_flags = 0;
442 return 0;
443 }
444
445 err = start_stop_khugepaged();
446 if (err)
447 goto err_khugepaged;
448
449 return 0;
450err_khugepaged:
451 unregister_shrinker(&deferred_split_shrinker);
452err_split_shrinker:
453 unregister_shrinker(&huge_zero_page_shrinker);
454err_hzp_shrinker:
455 khugepaged_destroy();
456err_slab:
457 hugepage_exit_sysfs(hugepage_kobj);
458err_sysfs:
459 return err;
460}
461subsys_initcall(hugepage_init);
462
463static int __init setup_transparent_hugepage(char *str)
464{
465 int ret = 0;
466 if (!str)
467 goto out;
468 if (!strcmp(str, "always")) {
469 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
470 &transparent_hugepage_flags);
471 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
472 &transparent_hugepage_flags);
473 ret = 1;
474 } else if (!strcmp(str, "madvise")) {
475 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
476 &transparent_hugepage_flags);
477 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
478 &transparent_hugepage_flags);
479 ret = 1;
480 } else if (!strcmp(str, "never")) {
481 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
482 &transparent_hugepage_flags);
483 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
484 &transparent_hugepage_flags);
485 ret = 1;
486 }
487out:
488 if (!ret)
489 pr_warn("transparent_hugepage= cannot parse, ignored\n");
490 return ret;
491}
492__setup("transparent_hugepage=", setup_transparent_hugepage);
493
494pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
495{
496 if (likely(vma->vm_flags & VM_WRITE))
497 pmd = pmd_mkwrite(pmd);
498 return pmd;
499}
500
501#ifdef CONFIG_MEMCG
502static inline struct deferred_split *get_deferred_split_queue(struct page *page)
503{
504 struct mem_cgroup *memcg = page_memcg(compound_head(page));
505 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
506
507 if (memcg)
508 return &memcg->deferred_split_queue;
509 else
510 return &pgdat->deferred_split_queue;
511}
512#else
513static inline struct deferred_split *get_deferred_split_queue(struct page *page)
514{
515 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
516
517 return &pgdat->deferred_split_queue;
518}
519#endif
520
521void prep_transhuge_page(struct page *page)
522{
523 /*
524 * we use page->mapping and page->indexlru in second tail page
525 * as list_head: assuming THP order >= 2
526 */
527
528 INIT_LIST_HEAD(page_deferred_list(page));
529 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
530}
531
532bool is_transparent_hugepage(struct page *page)
533{
534 if (!PageCompound(page))
535 return false;
536
537 page = compound_head(page);
538 return is_huge_zero_page(page) ||
539 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
540}
541EXPORT_SYMBOL_GPL(is_transparent_hugepage);
542
543static unsigned long __thp_get_unmapped_area(struct file *filp,
544 unsigned long addr, unsigned long len,
545 loff_t off, unsigned long flags, unsigned long size)
546{
547 loff_t off_end = off + len;
548 loff_t off_align = round_up(off, size);
549 unsigned long len_pad, ret;
550
551 if (off_end <= off_align || (off_end - off_align) < size)
552 return 0;
553
554 len_pad = len + size;
555 if (len_pad < len || (off + len_pad) < off)
556 return 0;
557
558 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
559 off >> PAGE_SHIFT, flags);
560
561 /*
562 * The failure might be due to length padding. The caller will retry
563 * without the padding.
564 */
565 if (IS_ERR_VALUE(ret))
566 return 0;
567
568 /*
569 * Do not try to align to THP boundary if allocation at the address
570 * hint succeeds.
571 */
572 if (ret == addr)
573 return addr;
574
575 ret += (off - ret) & (size - 1);
576 return ret;
577}
578
579unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
580 unsigned long len, unsigned long pgoff, unsigned long flags)
581{
582 unsigned long ret;
583 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
584
585 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
586 goto out;
587
588 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
589 if (ret)
590 return ret;
591out:
592 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
593}
594EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
595
596static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
597 struct page *page, gfp_t gfp)
598{
599 struct vm_area_struct *vma = vmf->vma;
600 pgtable_t pgtable;
601 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
602 vm_fault_t ret = 0;
603
604 VM_BUG_ON_PAGE(!PageCompound(page), page);
605
606 if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
607 put_page(page);
608 count_vm_event(THP_FAULT_FALLBACK);
609 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
610 return VM_FAULT_FALLBACK;
611 }
612 cgroup_throttle_swaprate(page, gfp);
613
614 pgtable = pte_alloc_one(vma->vm_mm);
615 if (unlikely(!pgtable)) {
616 ret = VM_FAULT_OOM;
617 goto release;
618 }
619
620 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
621 /*
622 * The memory barrier inside __SetPageUptodate makes sure that
623 * clear_huge_page writes become visible before the set_pmd_at()
624 * write.
625 */
626 __SetPageUptodate(page);
627
628 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
629 if (unlikely(!pmd_none(*vmf->pmd))) {
630 goto unlock_release;
631 } else {
632 pmd_t entry;
633
634 ret = check_stable_address_space(vma->vm_mm);
635 if (ret)
636 goto unlock_release;
637
638 /* Deliver the page fault to userland */
639 if (userfaultfd_missing(vma)) {
640 spin_unlock(vmf->ptl);
641 put_page(page);
642 pte_free(vma->vm_mm, pgtable);
643 ret = handle_userfault(vmf, VM_UFFD_MISSING);
644 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
645 return ret;
646 }
647
648 entry = mk_huge_pmd(page, vma->vm_page_prot);
649 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
650 page_add_new_anon_rmap(page, vma, haddr, true);
651 lru_cache_add_inactive_or_unevictable(page, vma);
652 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
653 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
654 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
655 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
656 mm_inc_nr_ptes(vma->vm_mm);
657 spin_unlock(vmf->ptl);
658 count_vm_event(THP_FAULT_ALLOC);
659 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
660 }
661
662 return 0;
663unlock_release:
664 spin_unlock(vmf->ptl);
665release:
666 if (pgtable)
667 pte_free(vma->vm_mm, pgtable);
668 put_page(page);
669 return ret;
670
671}
672
673/*
674 * always: directly stall for all thp allocations
675 * defer: wake kswapd and fail if not immediately available
676 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
677 * fail if not immediately available
678 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
679 * available
680 * never: never stall for any thp allocation
681 */
682gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
683{
684 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
685
686 /* Always do synchronous compaction */
687 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
688 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
689
690 /* Kick kcompactd and fail quickly */
691 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
692 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
693
694 /* Synchronous compaction if madvised, otherwise kick kcompactd */
695 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
696 return GFP_TRANSHUGE_LIGHT |
697 (vma_madvised ? __GFP_DIRECT_RECLAIM :
698 __GFP_KSWAPD_RECLAIM);
699
700 /* Only do synchronous compaction if madvised */
701 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
702 return GFP_TRANSHUGE_LIGHT |
703 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
704
705 return GFP_TRANSHUGE_LIGHT;
706}
707
708/* Caller must hold page table lock. */
709static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
710 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
711 struct page *zero_page)
712{
713 pmd_t entry;
714 if (!pmd_none(*pmd))
715 return;
716 entry = mk_pmd(zero_page, vma->vm_page_prot);
717 entry = pmd_mkhuge(entry);
718 if (pgtable)
719 pgtable_trans_huge_deposit(mm, pmd, pgtable);
720 set_pmd_at(mm, haddr, pmd, entry);
721 mm_inc_nr_ptes(mm);
722}
723
724vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
725{
726 struct vm_area_struct *vma = vmf->vma;
727 gfp_t gfp;
728 struct page *page;
729 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
730
731 if (!transhuge_vma_suitable(vma, haddr))
732 return VM_FAULT_FALLBACK;
733 if (unlikely(anon_vma_prepare(vma)))
734 return VM_FAULT_OOM;
735 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
736 return VM_FAULT_OOM;
737 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
738 !mm_forbids_zeropage(vma->vm_mm) &&
739 transparent_hugepage_use_zero_page()) {
740 pgtable_t pgtable;
741 struct page *zero_page;
742 vm_fault_t ret;
743 pgtable = pte_alloc_one(vma->vm_mm);
744 if (unlikely(!pgtable))
745 return VM_FAULT_OOM;
746 zero_page = mm_get_huge_zero_page(vma->vm_mm);
747 if (unlikely(!zero_page)) {
748 pte_free(vma->vm_mm, pgtable);
749 count_vm_event(THP_FAULT_FALLBACK);
750 return VM_FAULT_FALLBACK;
751 }
752 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
753 ret = 0;
754 if (pmd_none(*vmf->pmd)) {
755 ret = check_stable_address_space(vma->vm_mm);
756 if (ret) {
757 spin_unlock(vmf->ptl);
758 pte_free(vma->vm_mm, pgtable);
759 } else if (userfaultfd_missing(vma)) {
760 spin_unlock(vmf->ptl);
761 pte_free(vma->vm_mm, pgtable);
762 ret = handle_userfault(vmf, VM_UFFD_MISSING);
763 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
764 } else {
765 set_huge_zero_page(pgtable, vma->vm_mm, vma,
766 haddr, vmf->pmd, zero_page);
767 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
768 spin_unlock(vmf->ptl);
769 }
770 } else {
771 spin_unlock(vmf->ptl);
772 pte_free(vma->vm_mm, pgtable);
773 }
774 return ret;
775 }
776 gfp = vma_thp_gfp_mask(vma);
777 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
778 if (unlikely(!page)) {
779 count_vm_event(THP_FAULT_FALLBACK);
780 return VM_FAULT_FALLBACK;
781 }
782 prep_transhuge_page(page);
783 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
784}
785
786static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
787 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
788 pgtable_t pgtable)
789{
790 struct mm_struct *mm = vma->vm_mm;
791 pmd_t entry;
792 spinlock_t *ptl;
793
794 ptl = pmd_lock(mm, pmd);
795 if (!pmd_none(*pmd)) {
796 if (write) {
797 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
798 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
799 goto out_unlock;
800 }
801 entry = pmd_mkyoung(*pmd);
802 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
803 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
804 update_mmu_cache_pmd(vma, addr, pmd);
805 }
806
807 goto out_unlock;
808 }
809
810 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
811 if (pfn_t_devmap(pfn))
812 entry = pmd_mkdevmap(entry);
813 if (write) {
814 entry = pmd_mkyoung(pmd_mkdirty(entry));
815 entry = maybe_pmd_mkwrite(entry, vma);
816 }
817
818 if (pgtable) {
819 pgtable_trans_huge_deposit(mm, pmd, pgtable);
820 mm_inc_nr_ptes(mm);
821 pgtable = NULL;
822 }
823
824 set_pmd_at(mm, addr, pmd, entry);
825 update_mmu_cache_pmd(vma, addr, pmd);
826
827out_unlock:
828 spin_unlock(ptl);
829 if (pgtable)
830 pte_free(mm, pgtable);
831}
832
833/**
834 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
835 * @vmf: Structure describing the fault
836 * @pfn: pfn to insert
837 * @pgprot: page protection to use
838 * @write: whether it's a write fault
839 *
840 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
841 * also consult the vmf_insert_mixed_prot() documentation when
842 * @pgprot != @vmf->vma->vm_page_prot.
843 *
844 * Return: vm_fault_t value.
845 */
846vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
847 pgprot_t pgprot, bool write)
848{
849 unsigned long addr = vmf->address & PMD_MASK;
850 struct vm_area_struct *vma = vmf->vma;
851 pgtable_t pgtable = NULL;
852
853 /*
854 * If we had pmd_special, we could avoid all these restrictions,
855 * but we need to be consistent with PTEs and architectures that
856 * can't support a 'special' bit.
857 */
858 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
859 !pfn_t_devmap(pfn));
860 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
861 (VM_PFNMAP|VM_MIXEDMAP));
862 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
863
864 if (addr < vma->vm_start || addr >= vma->vm_end)
865 return VM_FAULT_SIGBUS;
866
867 if (arch_needs_pgtable_deposit()) {
868 pgtable = pte_alloc_one(vma->vm_mm);
869 if (!pgtable)
870 return VM_FAULT_OOM;
871 }
872
873 track_pfn_insert(vma, &pgprot, pfn);
874
875 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
876 return VM_FAULT_NOPAGE;
877}
878EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
879
880#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
881static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
882{
883 if (likely(vma->vm_flags & VM_WRITE))
884 pud = pud_mkwrite(pud);
885 return pud;
886}
887
888static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
889 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
890{
891 struct mm_struct *mm = vma->vm_mm;
892 pud_t entry;
893 spinlock_t *ptl;
894
895 ptl = pud_lock(mm, pud);
896 if (!pud_none(*pud)) {
897 if (write) {
898 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
899 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
900 goto out_unlock;
901 }
902 entry = pud_mkyoung(*pud);
903 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
904 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
905 update_mmu_cache_pud(vma, addr, pud);
906 }
907 goto out_unlock;
908 }
909
910 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
911 if (pfn_t_devmap(pfn))
912 entry = pud_mkdevmap(entry);
913 if (write) {
914 entry = pud_mkyoung(pud_mkdirty(entry));
915 entry = maybe_pud_mkwrite(entry, vma);
916 }
917 set_pud_at(mm, addr, pud, entry);
918 update_mmu_cache_pud(vma, addr, pud);
919
920out_unlock:
921 spin_unlock(ptl);
922}
923
924/**
925 * vmf_insert_pfn_pud_prot - insert a pud size pfn
926 * @vmf: Structure describing the fault
927 * @pfn: pfn to insert
928 * @pgprot: page protection to use
929 * @write: whether it's a write fault
930 *
931 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
932 * also consult the vmf_insert_mixed_prot() documentation when
933 * @pgprot != @vmf->vma->vm_page_prot.
934 *
935 * Return: vm_fault_t value.
936 */
937vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
938 pgprot_t pgprot, bool write)
939{
940 unsigned long addr = vmf->address & PUD_MASK;
941 struct vm_area_struct *vma = vmf->vma;
942
943 /*
944 * If we had pud_special, we could avoid all these restrictions,
945 * but we need to be consistent with PTEs and architectures that
946 * can't support a 'special' bit.
947 */
948 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
949 !pfn_t_devmap(pfn));
950 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
951 (VM_PFNMAP|VM_MIXEDMAP));
952 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
953
954 if (addr < vma->vm_start || addr >= vma->vm_end)
955 return VM_FAULT_SIGBUS;
956
957 track_pfn_insert(vma, &pgprot, pfn);
958
959 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
960 return VM_FAULT_NOPAGE;
961}
962EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
963#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
964
965static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
966 pmd_t *pmd, int flags)
967{
968 pmd_t _pmd;
969
970 _pmd = pmd_mkyoung(*pmd);
971 if (flags & FOLL_WRITE)
972 _pmd = pmd_mkdirty(_pmd);
973 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
974 pmd, _pmd, flags & FOLL_WRITE))
975 update_mmu_cache_pmd(vma, addr, pmd);
976}
977
978struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
979 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
980{
981 unsigned long pfn = pmd_pfn(*pmd);
982 struct mm_struct *mm = vma->vm_mm;
983 struct page *page;
984
985 assert_spin_locked(pmd_lockptr(mm, pmd));
986
987 /*
988 * When we COW a devmap PMD entry, we split it into PTEs, so we should
989 * not be in this function with `flags & FOLL_COW` set.
990 */
991 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
992
993 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
994 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
995 (FOLL_PIN | FOLL_GET)))
996 return NULL;
997
998 if (flags & FOLL_WRITE && !pmd_write(*pmd))
999 return NULL;
1000
1001 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1002 /* pass */;
1003 else
1004 return NULL;
1005
1006 if (flags & FOLL_TOUCH)
1007 touch_pmd(vma, addr, pmd, flags);
1008
1009 /*
1010 * device mapped pages can only be returned if the
1011 * caller will manage the page reference count.
1012 */
1013 if (!(flags & (FOLL_GET | FOLL_PIN)))
1014 return ERR_PTR(-EEXIST);
1015
1016 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1017 *pgmap = get_dev_pagemap(pfn, *pgmap);
1018 if (!*pgmap)
1019 return ERR_PTR(-EFAULT);
1020 page = pfn_to_page(pfn);
1021 if (!try_grab_page(page, flags))
1022 page = ERR_PTR(-ENOMEM);
1023
1024 return page;
1025}
1026
1027int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1028 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1029 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1030{
1031 spinlock_t *dst_ptl, *src_ptl;
1032 struct page *src_page;
1033 pmd_t pmd;
1034 pgtable_t pgtable = NULL;
1035 int ret = -ENOMEM;
1036
1037 /* Skip if can be re-fill on fault */
1038 if (!vma_is_anonymous(dst_vma))
1039 return 0;
1040
1041 pgtable = pte_alloc_one(dst_mm);
1042 if (unlikely(!pgtable))
1043 goto out;
1044
1045 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1046 src_ptl = pmd_lockptr(src_mm, src_pmd);
1047 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1048
1049 ret = -EAGAIN;
1050 pmd = *src_pmd;
1051
1052#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1053 if (unlikely(is_swap_pmd(pmd))) {
1054 swp_entry_t entry = pmd_to_swp_entry(pmd);
1055
1056 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1057 if (is_writable_migration_entry(entry)) {
1058 entry = make_readable_migration_entry(
1059 swp_offset(entry));
1060 pmd = swp_entry_to_pmd(entry);
1061 if (pmd_swp_soft_dirty(*src_pmd))
1062 pmd = pmd_swp_mksoft_dirty(pmd);
1063 if (pmd_swp_uffd_wp(*src_pmd))
1064 pmd = pmd_swp_mkuffd_wp(pmd);
1065 set_pmd_at(src_mm, addr, src_pmd, pmd);
1066 }
1067 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1068 mm_inc_nr_ptes(dst_mm);
1069 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1070 if (!userfaultfd_wp(dst_vma))
1071 pmd = pmd_swp_clear_uffd_wp(pmd);
1072 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1073 ret = 0;
1074 goto out_unlock;
1075 }
1076#endif
1077
1078 if (unlikely(!pmd_trans_huge(pmd))) {
1079 pte_free(dst_mm, pgtable);
1080 goto out_unlock;
1081 }
1082 /*
1083 * When page table lock is held, the huge zero pmd should not be
1084 * under splitting since we don't split the page itself, only pmd to
1085 * a page table.
1086 */
1087 if (is_huge_zero_pmd(pmd)) {
1088 /*
1089 * get_huge_zero_page() will never allocate a new page here,
1090 * since we already have a zero page to copy. It just takes a
1091 * reference.
1092 */
1093 mm_get_huge_zero_page(dst_mm);
1094 goto out_zero_page;
1095 }
1096
1097 src_page = pmd_page(pmd);
1098 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1099
1100 /*
1101 * If this page is a potentially pinned page, split and retry the fault
1102 * with smaller page size. Normally this should not happen because the
1103 * userspace should use MADV_DONTFORK upon pinned regions. This is a
1104 * best effort that the pinned pages won't be replaced by another
1105 * random page during the coming copy-on-write.
1106 */
1107 if (unlikely(page_needs_cow_for_dma(src_vma, src_page))) {
1108 pte_free(dst_mm, pgtable);
1109 spin_unlock(src_ptl);
1110 spin_unlock(dst_ptl);
1111 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1112 return -EAGAIN;
1113 }
1114
1115 get_page(src_page);
1116 page_dup_rmap(src_page, true);
1117 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118out_zero_page:
1119 mm_inc_nr_ptes(dst_mm);
1120 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1121 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1122 if (!userfaultfd_wp(dst_vma))
1123 pmd = pmd_clear_uffd_wp(pmd);
1124 pmd = pmd_mkold(pmd_wrprotect(pmd));
1125 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1126
1127 ret = 0;
1128out_unlock:
1129 spin_unlock(src_ptl);
1130 spin_unlock(dst_ptl);
1131out:
1132 return ret;
1133}
1134
1135#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1136static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1137 pud_t *pud, int flags)
1138{
1139 pud_t _pud;
1140
1141 _pud = pud_mkyoung(*pud);
1142 if (flags & FOLL_WRITE)
1143 _pud = pud_mkdirty(_pud);
1144 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1145 pud, _pud, flags & FOLL_WRITE))
1146 update_mmu_cache_pud(vma, addr, pud);
1147}
1148
1149struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1150 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1151{
1152 unsigned long pfn = pud_pfn(*pud);
1153 struct mm_struct *mm = vma->vm_mm;
1154 struct page *page;
1155
1156 assert_spin_locked(pud_lockptr(mm, pud));
1157
1158 if (flags & FOLL_WRITE && !pud_write(*pud))
1159 return NULL;
1160
1161 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1162 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1163 (FOLL_PIN | FOLL_GET)))
1164 return NULL;
1165
1166 if (pud_present(*pud) && pud_devmap(*pud))
1167 /* pass */;
1168 else
1169 return NULL;
1170
1171 if (flags & FOLL_TOUCH)
1172 touch_pud(vma, addr, pud, flags);
1173
1174 /*
1175 * device mapped pages can only be returned if the
1176 * caller will manage the page reference count.
1177 *
1178 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1179 */
1180 if (!(flags & (FOLL_GET | FOLL_PIN)))
1181 return ERR_PTR(-EEXIST);
1182
1183 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1184 *pgmap = get_dev_pagemap(pfn, *pgmap);
1185 if (!*pgmap)
1186 return ERR_PTR(-EFAULT);
1187 page = pfn_to_page(pfn);
1188 if (!try_grab_page(page, flags))
1189 page = ERR_PTR(-ENOMEM);
1190
1191 return page;
1192}
1193
1194int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1195 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1196 struct vm_area_struct *vma)
1197{
1198 spinlock_t *dst_ptl, *src_ptl;
1199 pud_t pud;
1200 int ret;
1201
1202 dst_ptl = pud_lock(dst_mm, dst_pud);
1203 src_ptl = pud_lockptr(src_mm, src_pud);
1204 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1205
1206 ret = -EAGAIN;
1207 pud = *src_pud;
1208 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1209 goto out_unlock;
1210
1211 /*
1212 * When page table lock is held, the huge zero pud should not be
1213 * under splitting since we don't split the page itself, only pud to
1214 * a page table.
1215 */
1216 if (is_huge_zero_pud(pud)) {
1217 /* No huge zero pud yet */
1218 }
1219
1220 /* Please refer to comments in copy_huge_pmd() */
1221 if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1222 spin_unlock(src_ptl);
1223 spin_unlock(dst_ptl);
1224 __split_huge_pud(vma, src_pud, addr);
1225 return -EAGAIN;
1226 }
1227
1228 pudp_set_wrprotect(src_mm, addr, src_pud);
1229 pud = pud_mkold(pud_wrprotect(pud));
1230 set_pud_at(dst_mm, addr, dst_pud, pud);
1231
1232 ret = 0;
1233out_unlock:
1234 spin_unlock(src_ptl);
1235 spin_unlock(dst_ptl);
1236 return ret;
1237}
1238
1239void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1240{
1241 pud_t entry;
1242 unsigned long haddr;
1243 bool write = vmf->flags & FAULT_FLAG_WRITE;
1244
1245 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1246 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1247 goto unlock;
1248
1249 entry = pud_mkyoung(orig_pud);
1250 if (write)
1251 entry = pud_mkdirty(entry);
1252 haddr = vmf->address & HPAGE_PUD_MASK;
1253 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1254 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1255
1256unlock:
1257 spin_unlock(vmf->ptl);
1258}
1259#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1260
1261void huge_pmd_set_accessed(struct vm_fault *vmf)
1262{
1263 pmd_t entry;
1264 unsigned long haddr;
1265 bool write = vmf->flags & FAULT_FLAG_WRITE;
1266 pmd_t orig_pmd = vmf->orig_pmd;
1267
1268 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1269 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1270 goto unlock;
1271
1272 entry = pmd_mkyoung(orig_pmd);
1273 if (write)
1274 entry = pmd_mkdirty(entry);
1275 haddr = vmf->address & HPAGE_PMD_MASK;
1276 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1277 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1278
1279unlock:
1280 spin_unlock(vmf->ptl);
1281}
1282
1283vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1284{
1285 struct vm_area_struct *vma = vmf->vma;
1286 struct page *page;
1287 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1288 pmd_t orig_pmd = vmf->orig_pmd;
1289
1290 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1291 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1292
1293 if (is_huge_zero_pmd(orig_pmd))
1294 goto fallback;
1295
1296 spin_lock(vmf->ptl);
1297
1298 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1299 spin_unlock(vmf->ptl);
1300 return 0;
1301 }
1302
1303 page = pmd_page(orig_pmd);
1304 VM_BUG_ON_PAGE(!PageHead(page), page);
1305
1306 /* Lock page for reuse_swap_page() */
1307 if (!trylock_page(page)) {
1308 get_page(page);
1309 spin_unlock(vmf->ptl);
1310 lock_page(page);
1311 spin_lock(vmf->ptl);
1312 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1313 spin_unlock(vmf->ptl);
1314 unlock_page(page);
1315 put_page(page);
1316 return 0;
1317 }
1318 put_page(page);
1319 }
1320
1321 /*
1322 * We can only reuse the page if nobody else maps the huge page or it's
1323 * part.
1324 */
1325 if (reuse_swap_page(page, NULL)) {
1326 pmd_t entry;
1327 entry = pmd_mkyoung(orig_pmd);
1328 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1329 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1330 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1331 unlock_page(page);
1332 spin_unlock(vmf->ptl);
1333 return VM_FAULT_WRITE;
1334 }
1335
1336 unlock_page(page);
1337 spin_unlock(vmf->ptl);
1338fallback:
1339 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1340 return VM_FAULT_FALLBACK;
1341}
1342
1343/*
1344 * FOLL_FORCE can write to even unwritable pmd's, but only
1345 * after we've gone through a COW cycle and they are dirty.
1346 */
1347static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1348{
1349 return pmd_write(pmd) ||
1350 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1351}
1352
1353struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1354 unsigned long addr,
1355 pmd_t *pmd,
1356 unsigned int flags)
1357{
1358 struct mm_struct *mm = vma->vm_mm;
1359 struct page *page = NULL;
1360
1361 assert_spin_locked(pmd_lockptr(mm, pmd));
1362
1363 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1364 goto out;
1365
1366 /* Avoid dumping huge zero page */
1367 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1368 return ERR_PTR(-EFAULT);
1369
1370 /* Full NUMA hinting faults to serialise migration in fault paths */
1371 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1372 goto out;
1373
1374 page = pmd_page(*pmd);
1375 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1376
1377 if (!try_grab_page(page, flags))
1378 return ERR_PTR(-ENOMEM);
1379
1380 if (flags & FOLL_TOUCH)
1381 touch_pmd(vma, addr, pmd, flags);
1382
1383 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1384 /*
1385 * We don't mlock() pte-mapped THPs. This way we can avoid
1386 * leaking mlocked pages into non-VM_LOCKED VMAs.
1387 *
1388 * For anon THP:
1389 *
1390 * In most cases the pmd is the only mapping of the page as we
1391 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1392 * writable private mappings in populate_vma_page_range().
1393 *
1394 * The only scenario when we have the page shared here is if we
1395 * mlocking read-only mapping shared over fork(). We skip
1396 * mlocking such pages.
1397 *
1398 * For file THP:
1399 *
1400 * We can expect PageDoubleMap() to be stable under page lock:
1401 * for file pages we set it in page_add_file_rmap(), which
1402 * requires page to be locked.
1403 */
1404
1405 if (PageAnon(page) && compound_mapcount(page) != 1)
1406 goto skip_mlock;
1407 if (PageDoubleMap(page) || !page->mapping)
1408 goto skip_mlock;
1409 if (!trylock_page(page))
1410 goto skip_mlock;
1411 if (page->mapping && !PageDoubleMap(page))
1412 mlock_vma_page(page);
1413 unlock_page(page);
1414 }
1415skip_mlock:
1416 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1417 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1418
1419out:
1420 return page;
1421}
1422
1423/* NUMA hinting page fault entry point for trans huge pmds */
1424vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1425{
1426 struct vm_area_struct *vma = vmf->vma;
1427 pmd_t oldpmd = vmf->orig_pmd;
1428 pmd_t pmd;
1429 struct page *page;
1430 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1431 int page_nid = NUMA_NO_NODE;
1432 int target_nid, last_cpupid = -1;
1433 bool migrated = false;
1434 bool was_writable = pmd_savedwrite(oldpmd);
1435 int flags = 0;
1436
1437 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1438 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1439 spin_unlock(vmf->ptl);
1440 goto out;
1441 }
1442
1443 /*
1444 * Since we took the NUMA fault, we must have observed the !accessible
1445 * bit. Make sure all other CPUs agree with that, to avoid them
1446 * modifying the page we're about to migrate.
1447 *
1448 * Must be done under PTL such that we'll observe the relevant
1449 * inc_tlb_flush_pending().
1450 *
1451 * We are not sure a pending tlb flush here is for a huge page
1452 * mapping or not. Hence use the tlb range variant
1453 */
1454 if (mm_tlb_flush_pending(vma->vm_mm)) {
1455 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1456 /*
1457 * change_huge_pmd() released the pmd lock before
1458 * invalidating the secondary MMUs sharing the primary
1459 * MMU pagetables (with ->invalidate_range()). The
1460 * mmu_notifier_invalidate_range_end() (which
1461 * internally calls ->invalidate_range()) in
1462 * change_pmd_range() will run after us, so we can't
1463 * rely on it here and we need an explicit invalidate.
1464 */
1465 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1466 haddr + HPAGE_PMD_SIZE);
1467 }
1468
1469 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1470 page = vm_normal_page_pmd(vma, haddr, pmd);
1471 if (!page)
1472 goto out_map;
1473
1474 /* See similar comment in do_numa_page for explanation */
1475 if (!was_writable)
1476 flags |= TNF_NO_GROUP;
1477
1478 page_nid = page_to_nid(page);
1479 last_cpupid = page_cpupid_last(page);
1480 target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1481 &flags);
1482
1483 if (target_nid == NUMA_NO_NODE) {
1484 put_page(page);
1485 goto out_map;
1486 }
1487
1488 spin_unlock(vmf->ptl);
1489
1490 migrated = migrate_misplaced_page(page, vma, target_nid);
1491 if (migrated) {
1492 flags |= TNF_MIGRATED;
1493 page_nid = target_nid;
1494 } else {
1495 flags |= TNF_MIGRATE_FAIL;
1496 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1497 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1498 spin_unlock(vmf->ptl);
1499 goto out;
1500 }
1501 goto out_map;
1502 }
1503
1504out:
1505 if (page_nid != NUMA_NO_NODE)
1506 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1507 flags);
1508
1509 return 0;
1510
1511out_map:
1512 /* Restore the PMD */
1513 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1514 pmd = pmd_mkyoung(pmd);
1515 if (was_writable)
1516 pmd = pmd_mkwrite(pmd);
1517 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1518 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1519 spin_unlock(vmf->ptl);
1520 goto out;
1521}
1522
1523/*
1524 * Return true if we do MADV_FREE successfully on entire pmd page.
1525 * Otherwise, return false.
1526 */
1527bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1528 pmd_t *pmd, unsigned long addr, unsigned long next)
1529{
1530 spinlock_t *ptl;
1531 pmd_t orig_pmd;
1532 struct page *page;
1533 struct mm_struct *mm = tlb->mm;
1534 bool ret = false;
1535
1536 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1537
1538 ptl = pmd_trans_huge_lock(pmd, vma);
1539 if (!ptl)
1540 goto out_unlocked;
1541
1542 orig_pmd = *pmd;
1543 if (is_huge_zero_pmd(orig_pmd))
1544 goto out;
1545
1546 if (unlikely(!pmd_present(orig_pmd))) {
1547 VM_BUG_ON(thp_migration_supported() &&
1548 !is_pmd_migration_entry(orig_pmd));
1549 goto out;
1550 }
1551
1552 page = pmd_page(orig_pmd);
1553 /*
1554 * If other processes are mapping this page, we couldn't discard
1555 * the page unless they all do MADV_FREE so let's skip the page.
1556 */
1557 if (total_mapcount(page) != 1)
1558 goto out;
1559
1560 if (!trylock_page(page))
1561 goto out;
1562
1563 /*
1564 * If user want to discard part-pages of THP, split it so MADV_FREE
1565 * will deactivate only them.
1566 */
1567 if (next - addr != HPAGE_PMD_SIZE) {
1568 get_page(page);
1569 spin_unlock(ptl);
1570 split_huge_page(page);
1571 unlock_page(page);
1572 put_page(page);
1573 goto out_unlocked;
1574 }
1575
1576 if (PageDirty(page))
1577 ClearPageDirty(page);
1578 unlock_page(page);
1579
1580 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1581 pmdp_invalidate(vma, addr, pmd);
1582 orig_pmd = pmd_mkold(orig_pmd);
1583 orig_pmd = pmd_mkclean(orig_pmd);
1584
1585 set_pmd_at(mm, addr, pmd, orig_pmd);
1586 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1587 }
1588
1589 mark_page_lazyfree(page);
1590 ret = true;
1591out:
1592 spin_unlock(ptl);
1593out_unlocked:
1594 return ret;
1595}
1596
1597static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1598{
1599 pgtable_t pgtable;
1600
1601 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1602 pte_free(mm, pgtable);
1603 mm_dec_nr_ptes(mm);
1604}
1605
1606int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1607 pmd_t *pmd, unsigned long addr)
1608{
1609 pmd_t orig_pmd;
1610 spinlock_t *ptl;
1611
1612 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1613
1614 ptl = __pmd_trans_huge_lock(pmd, vma);
1615 if (!ptl)
1616 return 0;
1617 /*
1618 * For architectures like ppc64 we look at deposited pgtable
1619 * when calling pmdp_huge_get_and_clear. So do the
1620 * pgtable_trans_huge_withdraw after finishing pmdp related
1621 * operations.
1622 */
1623 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1624 tlb->fullmm);
1625 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1626 if (vma_is_special_huge(vma)) {
1627 if (arch_needs_pgtable_deposit())
1628 zap_deposited_table(tlb->mm, pmd);
1629 spin_unlock(ptl);
1630 } else if (is_huge_zero_pmd(orig_pmd)) {
1631 zap_deposited_table(tlb->mm, pmd);
1632 spin_unlock(ptl);
1633 } else {
1634 struct page *page = NULL;
1635 int flush_needed = 1;
1636
1637 if (pmd_present(orig_pmd)) {
1638 page = pmd_page(orig_pmd);
1639 page_remove_rmap(page, true);
1640 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1641 VM_BUG_ON_PAGE(!PageHead(page), page);
1642 } else if (thp_migration_supported()) {
1643 swp_entry_t entry;
1644
1645 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1646 entry = pmd_to_swp_entry(orig_pmd);
1647 page = pfn_swap_entry_to_page(entry);
1648 flush_needed = 0;
1649 } else
1650 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1651
1652 if (PageAnon(page)) {
1653 zap_deposited_table(tlb->mm, pmd);
1654 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1655 } else {
1656 if (arch_needs_pgtable_deposit())
1657 zap_deposited_table(tlb->mm, pmd);
1658 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1659 }
1660
1661 spin_unlock(ptl);
1662 if (flush_needed)
1663 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1664 }
1665 return 1;
1666}
1667
1668#ifndef pmd_move_must_withdraw
1669static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1670 spinlock_t *old_pmd_ptl,
1671 struct vm_area_struct *vma)
1672{
1673 /*
1674 * With split pmd lock we also need to move preallocated
1675 * PTE page table if new_pmd is on different PMD page table.
1676 *
1677 * We also don't deposit and withdraw tables for file pages.
1678 */
1679 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1680}
1681#endif
1682
1683static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1684{
1685#ifdef CONFIG_MEM_SOFT_DIRTY
1686 if (unlikely(is_pmd_migration_entry(pmd)))
1687 pmd = pmd_swp_mksoft_dirty(pmd);
1688 else if (pmd_present(pmd))
1689 pmd = pmd_mksoft_dirty(pmd);
1690#endif
1691 return pmd;
1692}
1693
1694bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1695 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1696{
1697 spinlock_t *old_ptl, *new_ptl;
1698 pmd_t pmd;
1699 struct mm_struct *mm = vma->vm_mm;
1700 bool force_flush = false;
1701
1702 /*
1703 * The destination pmd shouldn't be established, free_pgtables()
1704 * should have release it.
1705 */
1706 if (WARN_ON(!pmd_none(*new_pmd))) {
1707 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1708 return false;
1709 }
1710
1711 /*
1712 * We don't have to worry about the ordering of src and dst
1713 * ptlocks because exclusive mmap_lock prevents deadlock.
1714 */
1715 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1716 if (old_ptl) {
1717 new_ptl = pmd_lockptr(mm, new_pmd);
1718 if (new_ptl != old_ptl)
1719 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1720 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1721 if (pmd_present(pmd))
1722 force_flush = true;
1723 VM_BUG_ON(!pmd_none(*new_pmd));
1724
1725 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1726 pgtable_t pgtable;
1727 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1728 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1729 }
1730 pmd = move_soft_dirty_pmd(pmd);
1731 set_pmd_at(mm, new_addr, new_pmd, pmd);
1732 if (force_flush)
1733 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1734 if (new_ptl != old_ptl)
1735 spin_unlock(new_ptl);
1736 spin_unlock(old_ptl);
1737 return true;
1738 }
1739 return false;
1740}
1741
1742/*
1743 * Returns
1744 * - 0 if PMD could not be locked
1745 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1746 * or if prot_numa but THP migration is not supported
1747 * - HPAGE_PMD_NR if protections changed and TLB flush necessary
1748 */
1749int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1750 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1751{
1752 struct mm_struct *mm = vma->vm_mm;
1753 spinlock_t *ptl;
1754 pmd_t entry;
1755 bool preserve_write;
1756 int ret;
1757 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1758 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1759 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1760
1761 if (prot_numa && !thp_migration_supported())
1762 return 1;
1763
1764 ptl = __pmd_trans_huge_lock(pmd, vma);
1765 if (!ptl)
1766 return 0;
1767
1768 preserve_write = prot_numa && pmd_write(*pmd);
1769 ret = 1;
1770
1771#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1772 if (is_swap_pmd(*pmd)) {
1773 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1774
1775 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1776 if (is_writable_migration_entry(entry)) {
1777 pmd_t newpmd;
1778 /*
1779 * A protection check is difficult so
1780 * just be safe and disable write
1781 */
1782 entry = make_readable_migration_entry(
1783 swp_offset(entry));
1784 newpmd = swp_entry_to_pmd(entry);
1785 if (pmd_swp_soft_dirty(*pmd))
1786 newpmd = pmd_swp_mksoft_dirty(newpmd);
1787 if (pmd_swp_uffd_wp(*pmd))
1788 newpmd = pmd_swp_mkuffd_wp(newpmd);
1789 set_pmd_at(mm, addr, pmd, newpmd);
1790 }
1791 goto unlock;
1792 }
1793#endif
1794
1795 /*
1796 * Avoid trapping faults against the zero page. The read-only
1797 * data is likely to be read-cached on the local CPU and
1798 * local/remote hits to the zero page are not interesting.
1799 */
1800 if (prot_numa && is_huge_zero_pmd(*pmd))
1801 goto unlock;
1802
1803 if (prot_numa && pmd_protnone(*pmd))
1804 goto unlock;
1805
1806 /*
1807 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1808 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1809 * which is also under mmap_read_lock(mm):
1810 *
1811 * CPU0: CPU1:
1812 * change_huge_pmd(prot_numa=1)
1813 * pmdp_huge_get_and_clear_notify()
1814 * madvise_dontneed()
1815 * zap_pmd_range()
1816 * pmd_trans_huge(*pmd) == 0 (without ptl)
1817 * // skip the pmd
1818 * set_pmd_at();
1819 * // pmd is re-established
1820 *
1821 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1822 * which may break userspace.
1823 *
1824 * pmdp_invalidate() is required to make sure we don't miss
1825 * dirty/young flags set by hardware.
1826 */
1827 entry = pmdp_invalidate(vma, addr, pmd);
1828
1829 entry = pmd_modify(entry, newprot);
1830 if (preserve_write)
1831 entry = pmd_mk_savedwrite(entry);
1832 if (uffd_wp) {
1833 entry = pmd_wrprotect(entry);
1834 entry = pmd_mkuffd_wp(entry);
1835 } else if (uffd_wp_resolve) {
1836 /*
1837 * Leave the write bit to be handled by PF interrupt
1838 * handler, then things like COW could be properly
1839 * handled.
1840 */
1841 entry = pmd_clear_uffd_wp(entry);
1842 }
1843 ret = HPAGE_PMD_NR;
1844 set_pmd_at(mm, addr, pmd, entry);
1845 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1846unlock:
1847 spin_unlock(ptl);
1848 return ret;
1849}
1850
1851/*
1852 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1853 *
1854 * Note that if it returns page table lock pointer, this routine returns without
1855 * unlocking page table lock. So callers must unlock it.
1856 */
1857spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1858{
1859 spinlock_t *ptl;
1860 ptl = pmd_lock(vma->vm_mm, pmd);
1861 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1862 pmd_devmap(*pmd)))
1863 return ptl;
1864 spin_unlock(ptl);
1865 return NULL;
1866}
1867
1868/*
1869 * Returns true if a given pud maps a thp, false otherwise.
1870 *
1871 * Note that if it returns true, this routine returns without unlocking page
1872 * table lock. So callers must unlock it.
1873 */
1874spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1875{
1876 spinlock_t *ptl;
1877
1878 ptl = pud_lock(vma->vm_mm, pud);
1879 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1880 return ptl;
1881 spin_unlock(ptl);
1882 return NULL;
1883}
1884
1885#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1886int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1887 pud_t *pud, unsigned long addr)
1888{
1889 spinlock_t *ptl;
1890
1891 ptl = __pud_trans_huge_lock(pud, vma);
1892 if (!ptl)
1893 return 0;
1894 /*
1895 * For architectures like ppc64 we look at deposited pgtable
1896 * when calling pudp_huge_get_and_clear. So do the
1897 * pgtable_trans_huge_withdraw after finishing pudp related
1898 * operations.
1899 */
1900 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1901 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1902 if (vma_is_special_huge(vma)) {
1903 spin_unlock(ptl);
1904 /* No zero page support yet */
1905 } else {
1906 /* No support for anonymous PUD pages yet */
1907 BUG();
1908 }
1909 return 1;
1910}
1911
1912static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1913 unsigned long haddr)
1914{
1915 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1916 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1917 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1918 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1919
1920 count_vm_event(THP_SPLIT_PUD);
1921
1922 pudp_huge_clear_flush_notify(vma, haddr, pud);
1923}
1924
1925void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1926 unsigned long address)
1927{
1928 spinlock_t *ptl;
1929 struct mmu_notifier_range range;
1930
1931 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1932 address & HPAGE_PUD_MASK,
1933 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1934 mmu_notifier_invalidate_range_start(&range);
1935 ptl = pud_lock(vma->vm_mm, pud);
1936 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1937 goto out;
1938 __split_huge_pud_locked(vma, pud, range.start);
1939
1940out:
1941 spin_unlock(ptl);
1942 /*
1943 * No need to double call mmu_notifier->invalidate_range() callback as
1944 * the above pudp_huge_clear_flush_notify() did already call it.
1945 */
1946 mmu_notifier_invalidate_range_only_end(&range);
1947}
1948#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1949
1950static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1951 unsigned long haddr, pmd_t *pmd)
1952{
1953 struct mm_struct *mm = vma->vm_mm;
1954 pgtable_t pgtable;
1955 pmd_t _pmd;
1956 int i;
1957
1958 /*
1959 * Leave pmd empty until pte is filled note that it is fine to delay
1960 * notification until mmu_notifier_invalidate_range_end() as we are
1961 * replacing a zero pmd write protected page with a zero pte write
1962 * protected page.
1963 *
1964 * See Documentation/vm/mmu_notifier.rst
1965 */
1966 pmdp_huge_clear_flush(vma, haddr, pmd);
1967
1968 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1969 pmd_populate(mm, &_pmd, pgtable);
1970
1971 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1972 pte_t *pte, entry;
1973 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1974 entry = pte_mkspecial(entry);
1975 pte = pte_offset_map(&_pmd, haddr);
1976 VM_BUG_ON(!pte_none(*pte));
1977 set_pte_at(mm, haddr, pte, entry);
1978 pte_unmap(pte);
1979 }
1980 smp_wmb(); /* make pte visible before pmd */
1981 pmd_populate(mm, pmd, pgtable);
1982}
1983
1984static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1985 unsigned long haddr, bool freeze)
1986{
1987 struct mm_struct *mm = vma->vm_mm;
1988 struct page *page;
1989 pgtable_t pgtable;
1990 pmd_t old_pmd, _pmd;
1991 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
1992 unsigned long addr;
1993 int i;
1994
1995 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1996 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1997 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1998 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
1999 && !pmd_devmap(*pmd));
2000
2001 count_vm_event(THP_SPLIT_PMD);
2002
2003 if (!vma_is_anonymous(vma)) {
2004 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2005 /*
2006 * We are going to unmap this huge page. So
2007 * just go ahead and zap it
2008 */
2009 if (arch_needs_pgtable_deposit())
2010 zap_deposited_table(mm, pmd);
2011 if (vma_is_special_huge(vma))
2012 return;
2013 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2014 swp_entry_t entry;
2015
2016 entry = pmd_to_swp_entry(old_pmd);
2017 page = pfn_swap_entry_to_page(entry);
2018 } else {
2019 page = pmd_page(old_pmd);
2020 if (!PageDirty(page) && pmd_dirty(old_pmd))
2021 set_page_dirty(page);
2022 if (!PageReferenced(page) && pmd_young(old_pmd))
2023 SetPageReferenced(page);
2024 page_remove_rmap(page, true);
2025 put_page(page);
2026 }
2027 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2028 return;
2029 }
2030
2031 if (is_huge_zero_pmd(*pmd)) {
2032 /*
2033 * FIXME: Do we want to invalidate secondary mmu by calling
2034 * mmu_notifier_invalidate_range() see comments below inside
2035 * __split_huge_pmd() ?
2036 *
2037 * We are going from a zero huge page write protected to zero
2038 * small page also write protected so it does not seems useful
2039 * to invalidate secondary mmu at this time.
2040 */
2041 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2042 }
2043
2044 /*
2045 * Up to this point the pmd is present and huge and userland has the
2046 * whole access to the hugepage during the split (which happens in
2047 * place). If we overwrite the pmd with the not-huge version pointing
2048 * to the pte here (which of course we could if all CPUs were bug
2049 * free), userland could trigger a small page size TLB miss on the
2050 * small sized TLB while the hugepage TLB entry is still established in
2051 * the huge TLB. Some CPU doesn't like that.
2052 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2053 * 383 on page 105. Intel should be safe but is also warns that it's
2054 * only safe if the permission and cache attributes of the two entries
2055 * loaded in the two TLB is identical (which should be the case here).
2056 * But it is generally safer to never allow small and huge TLB entries
2057 * for the same virtual address to be loaded simultaneously. So instead
2058 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2059 * current pmd notpresent (atomically because here the pmd_trans_huge
2060 * must remain set at all times on the pmd until the split is complete
2061 * for this pmd), then we flush the SMP TLB and finally we write the
2062 * non-huge version of the pmd entry with pmd_populate.
2063 */
2064 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2065
2066 pmd_migration = is_pmd_migration_entry(old_pmd);
2067 if (unlikely(pmd_migration)) {
2068 swp_entry_t entry;
2069
2070 entry = pmd_to_swp_entry(old_pmd);
2071 page = pfn_swap_entry_to_page(entry);
2072 write = is_writable_migration_entry(entry);
2073 young = false;
2074 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2075 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2076 } else {
2077 page = pmd_page(old_pmd);
2078 if (pmd_dirty(old_pmd))
2079 SetPageDirty(page);
2080 write = pmd_write(old_pmd);
2081 young = pmd_young(old_pmd);
2082 soft_dirty = pmd_soft_dirty(old_pmd);
2083 uffd_wp = pmd_uffd_wp(old_pmd);
2084 }
2085 VM_BUG_ON_PAGE(!page_count(page), page);
2086 page_ref_add(page, HPAGE_PMD_NR - 1);
2087
2088 /*
2089 * Withdraw the table only after we mark the pmd entry invalid.
2090 * This's critical for some architectures (Power).
2091 */
2092 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2093 pmd_populate(mm, &_pmd, pgtable);
2094
2095 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2096 pte_t entry, *pte;
2097 /*
2098 * Note that NUMA hinting access restrictions are not
2099 * transferred to avoid any possibility of altering
2100 * permissions across VMAs.
2101 */
2102 if (freeze || pmd_migration) {
2103 swp_entry_t swp_entry;
2104 if (write)
2105 swp_entry = make_writable_migration_entry(
2106 page_to_pfn(page + i));
2107 else
2108 swp_entry = make_readable_migration_entry(
2109 page_to_pfn(page + i));
2110 entry = swp_entry_to_pte(swp_entry);
2111 if (soft_dirty)
2112 entry = pte_swp_mksoft_dirty(entry);
2113 if (uffd_wp)
2114 entry = pte_swp_mkuffd_wp(entry);
2115 } else {
2116 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2117 entry = maybe_mkwrite(entry, vma);
2118 if (!write)
2119 entry = pte_wrprotect(entry);
2120 if (!young)
2121 entry = pte_mkold(entry);
2122 if (soft_dirty)
2123 entry = pte_mksoft_dirty(entry);
2124 if (uffd_wp)
2125 entry = pte_mkuffd_wp(entry);
2126 }
2127 pte = pte_offset_map(&_pmd, addr);
2128 BUG_ON(!pte_none(*pte));
2129 set_pte_at(mm, addr, pte, entry);
2130 if (!pmd_migration)
2131 atomic_inc(&page[i]._mapcount);
2132 pte_unmap(pte);
2133 }
2134
2135 if (!pmd_migration) {
2136 /*
2137 * Set PG_double_map before dropping compound_mapcount to avoid
2138 * false-negative page_mapped().
2139 */
2140 if (compound_mapcount(page) > 1 &&
2141 !TestSetPageDoubleMap(page)) {
2142 for (i = 0; i < HPAGE_PMD_NR; i++)
2143 atomic_inc(&page[i]._mapcount);
2144 }
2145
2146 lock_page_memcg(page);
2147 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2148 /* Last compound_mapcount is gone. */
2149 __mod_lruvec_page_state(page, NR_ANON_THPS,
2150 -HPAGE_PMD_NR);
2151 if (TestClearPageDoubleMap(page)) {
2152 /* No need in mapcount reference anymore */
2153 for (i = 0; i < HPAGE_PMD_NR; i++)
2154 atomic_dec(&page[i]._mapcount);
2155 }
2156 }
2157 unlock_page_memcg(page);
2158 }
2159
2160 smp_wmb(); /* make pte visible before pmd */
2161 pmd_populate(mm, pmd, pgtable);
2162
2163 if (freeze) {
2164 for (i = 0; i < HPAGE_PMD_NR; i++) {
2165 page_remove_rmap(page + i, false);
2166 put_page(page + i);
2167 }
2168 }
2169}
2170
2171void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2172 unsigned long address, bool freeze, struct page *page)
2173{
2174 spinlock_t *ptl;
2175 struct mmu_notifier_range range;
2176 bool do_unlock_page = false;
2177 pmd_t _pmd;
2178
2179 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2180 address & HPAGE_PMD_MASK,
2181 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2182 mmu_notifier_invalidate_range_start(&range);
2183 ptl = pmd_lock(vma->vm_mm, pmd);
2184
2185 /*
2186 * If caller asks to setup a migration entries, we need a page to check
2187 * pmd against. Otherwise we can end up replacing wrong page.
2188 */
2189 VM_BUG_ON(freeze && !page);
2190 if (page) {
2191 VM_WARN_ON_ONCE(!PageLocked(page));
2192 if (page != pmd_page(*pmd))
2193 goto out;
2194 }
2195
2196repeat:
2197 if (pmd_trans_huge(*pmd)) {
2198 if (!page) {
2199 page = pmd_page(*pmd);
2200 /*
2201 * An anonymous page must be locked, to ensure that a
2202 * concurrent reuse_swap_page() sees stable mapcount;
2203 * but reuse_swap_page() is not used on shmem or file,
2204 * and page lock must not be taken when zap_pmd_range()
2205 * calls __split_huge_pmd() while i_mmap_lock is held.
2206 */
2207 if (PageAnon(page)) {
2208 if (unlikely(!trylock_page(page))) {
2209 get_page(page);
2210 _pmd = *pmd;
2211 spin_unlock(ptl);
2212 lock_page(page);
2213 spin_lock(ptl);
2214 if (unlikely(!pmd_same(*pmd, _pmd))) {
2215 unlock_page(page);
2216 put_page(page);
2217 page = NULL;
2218 goto repeat;
2219 }
2220 put_page(page);
2221 }
2222 do_unlock_page = true;
2223 }
2224 }
2225 if (PageMlocked(page))
2226 clear_page_mlock(page);
2227 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2228 goto out;
2229 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2230out:
2231 spin_unlock(ptl);
2232 if (do_unlock_page)
2233 unlock_page(page);
2234 /*
2235 * No need to double call mmu_notifier->invalidate_range() callback.
2236 * They are 3 cases to consider inside __split_huge_pmd_locked():
2237 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2238 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2239 * fault will trigger a flush_notify before pointing to a new page
2240 * (it is fine if the secondary mmu keeps pointing to the old zero
2241 * page in the meantime)
2242 * 3) Split a huge pmd into pte pointing to the same page. No need
2243 * to invalidate secondary tlb entry they are all still valid.
2244 * any further changes to individual pte will notify. So no need
2245 * to call mmu_notifier->invalidate_range()
2246 */
2247 mmu_notifier_invalidate_range_only_end(&range);
2248}
2249
2250void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2251 bool freeze, struct page *page)
2252{
2253 pgd_t *pgd;
2254 p4d_t *p4d;
2255 pud_t *pud;
2256 pmd_t *pmd;
2257
2258 pgd = pgd_offset(vma->vm_mm, address);
2259 if (!pgd_present(*pgd))
2260 return;
2261
2262 p4d = p4d_offset(pgd, address);
2263 if (!p4d_present(*p4d))
2264 return;
2265
2266 pud = pud_offset(p4d, address);
2267 if (!pud_present(*pud))
2268 return;
2269
2270 pmd = pmd_offset(pud, address);
2271
2272 __split_huge_pmd(vma, pmd, address, freeze, page);
2273}
2274
2275static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2276{
2277 /*
2278 * If the new address isn't hpage aligned and it could previously
2279 * contain an hugepage: check if we need to split an huge pmd.
2280 */
2281 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2282 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2283 ALIGN(address, HPAGE_PMD_SIZE)))
2284 split_huge_pmd_address(vma, address, false, NULL);
2285}
2286
2287void vma_adjust_trans_huge(struct vm_area_struct *vma,
2288 unsigned long start,
2289 unsigned long end,
2290 long adjust_next)
2291{
2292 /* Check if we need to split start first. */
2293 split_huge_pmd_if_needed(vma, start);
2294
2295 /* Check if we need to split end next. */
2296 split_huge_pmd_if_needed(vma, end);
2297
2298 /*
2299 * If we're also updating the vma->vm_next->vm_start,
2300 * check if we need to split it.
2301 */
2302 if (adjust_next > 0) {
2303 struct vm_area_struct *next = vma->vm_next;
2304 unsigned long nstart = next->vm_start;
2305 nstart += adjust_next;
2306 split_huge_pmd_if_needed(next, nstart);
2307 }
2308}
2309
2310static void unmap_page(struct page *page)
2311{
2312 enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2313 TTU_SYNC;
2314
2315 VM_BUG_ON_PAGE(!PageHead(page), page);
2316
2317 /*
2318 * Anon pages need migration entries to preserve them, but file
2319 * pages can simply be left unmapped, then faulted back on demand.
2320 * If that is ever changed (perhaps for mlock), update remap_page().
2321 */
2322 if (PageAnon(page))
2323 try_to_migrate(page, ttu_flags);
2324 else
2325 try_to_unmap(page, ttu_flags | TTU_IGNORE_MLOCK);
2326
2327 VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2328}
2329
2330static void remap_page(struct page *page, unsigned int nr)
2331{
2332 int i;
2333
2334 /* If unmap_page() uses try_to_migrate() on file, remove this check */
2335 if (!PageAnon(page))
2336 return;
2337 if (PageTransHuge(page)) {
2338 remove_migration_ptes(page, page, true);
2339 } else {
2340 for (i = 0; i < nr; i++)
2341 remove_migration_ptes(page + i, page + i, true);
2342 }
2343}
2344
2345static void lru_add_page_tail(struct page *head, struct page *tail,
2346 struct lruvec *lruvec, struct list_head *list)
2347{
2348 VM_BUG_ON_PAGE(!PageHead(head), head);
2349 VM_BUG_ON_PAGE(PageCompound(tail), head);
2350 VM_BUG_ON_PAGE(PageLRU(tail), head);
2351 lockdep_assert_held(&lruvec->lru_lock);
2352
2353 if (list) {
2354 /* page reclaim is reclaiming a huge page */
2355 VM_WARN_ON(PageLRU(head));
2356 get_page(tail);
2357 list_add_tail(&tail->lru, list);
2358 } else {
2359 /* head is still on lru (and we have it frozen) */
2360 VM_WARN_ON(!PageLRU(head));
2361 SetPageLRU(tail);
2362 list_add_tail(&tail->lru, &head->lru);
2363 }
2364}
2365
2366static void __split_huge_page_tail(struct page *head, int tail,
2367 struct lruvec *lruvec, struct list_head *list)
2368{
2369 struct page *page_tail = head + tail;
2370
2371 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2372
2373 /*
2374 * Clone page flags before unfreezing refcount.
2375 *
2376 * After successful get_page_unless_zero() might follow flags change,
2377 * for example lock_page() which set PG_waiters.
2378 */
2379 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2380 page_tail->flags |= (head->flags &
2381 ((1L << PG_referenced) |
2382 (1L << PG_swapbacked) |
2383 (1L << PG_swapcache) |
2384 (1L << PG_mlocked) |
2385 (1L << PG_uptodate) |
2386 (1L << PG_active) |
2387 (1L << PG_workingset) |
2388 (1L << PG_locked) |
2389 (1L << PG_unevictable) |
2390#ifdef CONFIG_64BIT
2391 (1L << PG_arch_2) |
2392#endif
2393 (1L << PG_dirty)));
2394
2395 /* ->mapping in first tail page is compound_mapcount */
2396 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2397 page_tail);
2398 page_tail->mapping = head->mapping;
2399 page_tail->index = head->index + tail;
2400
2401 /* Page flags must be visible before we make the page non-compound. */
2402 smp_wmb();
2403
2404 /*
2405 * Clear PageTail before unfreezing page refcount.
2406 *
2407 * After successful get_page_unless_zero() might follow put_page()
2408 * which needs correct compound_head().
2409 */
2410 clear_compound_head(page_tail);
2411
2412 /* Finally unfreeze refcount. Additional reference from page cache. */
2413 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2414 PageSwapCache(head)));
2415
2416 if (page_is_young(head))
2417 set_page_young(page_tail);
2418 if (page_is_idle(head))
2419 set_page_idle(page_tail);
2420
2421 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2422
2423 /*
2424 * always add to the tail because some iterators expect new
2425 * pages to show after the currently processed elements - e.g.
2426 * migrate_pages
2427 */
2428 lru_add_page_tail(head, page_tail, lruvec, list);
2429}
2430
2431static void __split_huge_page(struct page *page, struct list_head *list,
2432 pgoff_t end)
2433{
2434 struct page *head = compound_head(page);
2435 struct lruvec *lruvec;
2436 struct address_space *swap_cache = NULL;
2437 unsigned long offset = 0;
2438 unsigned int nr = thp_nr_pages(head);
2439 int i;
2440
2441 /* complete memcg works before add pages to LRU */
2442 split_page_memcg(head, nr);
2443
2444 if (PageAnon(head) && PageSwapCache(head)) {
2445 swp_entry_t entry = { .val = page_private(head) };
2446
2447 offset = swp_offset(entry);
2448 swap_cache = swap_address_space(entry);
2449 xa_lock(&swap_cache->i_pages);
2450 }
2451
2452 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2453 lruvec = lock_page_lruvec(head);
2454
2455 for (i = nr - 1; i >= 1; i--) {
2456 __split_huge_page_tail(head, i, lruvec, list);
2457 /* Some pages can be beyond i_size: drop them from page cache */
2458 if (head[i].index >= end) {
2459 ClearPageDirty(head + i);
2460 __delete_from_page_cache(head + i, NULL);
2461 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2462 shmem_uncharge(head->mapping->host, 1);
2463 put_page(head + i);
2464 } else if (!PageAnon(page)) {
2465 __xa_store(&head->mapping->i_pages, head[i].index,
2466 head + i, 0);
2467 } else if (swap_cache) {
2468 __xa_store(&swap_cache->i_pages, offset + i,
2469 head + i, 0);
2470 }
2471 }
2472
2473 ClearPageCompound(head);
2474 unlock_page_lruvec(lruvec);
2475 /* Caller disabled irqs, so they are still disabled here */
2476
2477 split_page_owner(head, nr);
2478
2479 /* See comment in __split_huge_page_tail() */
2480 if (PageAnon(head)) {
2481 /* Additional pin to swap cache */
2482 if (PageSwapCache(head)) {
2483 page_ref_add(head, 2);
2484 xa_unlock(&swap_cache->i_pages);
2485 } else {
2486 page_ref_inc(head);
2487 }
2488 } else {
2489 /* Additional pin to page cache */
2490 page_ref_add(head, 2);
2491 xa_unlock(&head->mapping->i_pages);
2492 }
2493 local_irq_enable();
2494
2495 remap_page(head, nr);
2496
2497 if (PageSwapCache(head)) {
2498 swp_entry_t entry = { .val = page_private(head) };
2499
2500 split_swap_cluster(entry);
2501 }
2502
2503 for (i = 0; i < nr; i++) {
2504 struct page *subpage = head + i;
2505 if (subpage == page)
2506 continue;
2507 unlock_page(subpage);
2508
2509 /*
2510 * Subpages may be freed if there wasn't any mapping
2511 * like if add_to_swap() is running on a lru page that
2512 * had its mapping zapped. And freeing these pages
2513 * requires taking the lru_lock so we do the put_page
2514 * of the tail pages after the split is complete.
2515 */
2516 put_page(subpage);
2517 }
2518}
2519
2520int total_mapcount(struct page *page)
2521{
2522 int i, compound, nr, ret;
2523
2524 VM_BUG_ON_PAGE(PageTail(page), page);
2525
2526 if (likely(!PageCompound(page)))
2527 return atomic_read(&page->_mapcount) + 1;
2528
2529 compound = compound_mapcount(page);
2530 nr = compound_nr(page);
2531 if (PageHuge(page))
2532 return compound;
2533 ret = compound;
2534 for (i = 0; i < nr; i++)
2535 ret += atomic_read(&page[i]._mapcount) + 1;
2536 /* File pages has compound_mapcount included in _mapcount */
2537 if (!PageAnon(page))
2538 return ret - compound * nr;
2539 if (PageDoubleMap(page))
2540 ret -= nr;
2541 return ret;
2542}
2543
2544/*
2545 * This calculates accurately how many mappings a transparent hugepage
2546 * has (unlike page_mapcount() which isn't fully accurate). This full
2547 * accuracy is primarily needed to know if copy-on-write faults can
2548 * reuse the page and change the mapping to read-write instead of
2549 * copying them. At the same time this returns the total_mapcount too.
2550 *
2551 * The function returns the highest mapcount any one of the subpages
2552 * has. If the return value is one, even if different processes are
2553 * mapping different subpages of the transparent hugepage, they can
2554 * all reuse it, because each process is reusing a different subpage.
2555 *
2556 * The total_mapcount is instead counting all virtual mappings of the
2557 * subpages. If the total_mapcount is equal to "one", it tells the
2558 * caller all mappings belong to the same "mm" and in turn the
2559 * anon_vma of the transparent hugepage can become the vma->anon_vma
2560 * local one as no other process may be mapping any of the subpages.
2561 *
2562 * It would be more accurate to replace page_mapcount() with
2563 * page_trans_huge_mapcount(), however we only use
2564 * page_trans_huge_mapcount() in the copy-on-write faults where we
2565 * need full accuracy to avoid breaking page pinning, because
2566 * page_trans_huge_mapcount() is slower than page_mapcount().
2567 */
2568int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2569{
2570 int i, ret, _total_mapcount, mapcount;
2571
2572 /* hugetlbfs shouldn't call it */
2573 VM_BUG_ON_PAGE(PageHuge(page), page);
2574
2575 if (likely(!PageTransCompound(page))) {
2576 mapcount = atomic_read(&page->_mapcount) + 1;
2577 if (total_mapcount)
2578 *total_mapcount = mapcount;
2579 return mapcount;
2580 }
2581
2582 page = compound_head(page);
2583
2584 _total_mapcount = ret = 0;
2585 for (i = 0; i < thp_nr_pages(page); i++) {
2586 mapcount = atomic_read(&page[i]._mapcount) + 1;
2587 ret = max(ret, mapcount);
2588 _total_mapcount += mapcount;
2589 }
2590 if (PageDoubleMap(page)) {
2591 ret -= 1;
2592 _total_mapcount -= thp_nr_pages(page);
2593 }
2594 mapcount = compound_mapcount(page);
2595 ret += mapcount;
2596 _total_mapcount += mapcount;
2597 if (total_mapcount)
2598 *total_mapcount = _total_mapcount;
2599 return ret;
2600}
2601
2602/* Racy check whether the huge page can be split */
2603bool can_split_huge_page(struct page *page, int *pextra_pins)
2604{
2605 int extra_pins;
2606
2607 /* Additional pins from page cache */
2608 if (PageAnon(page))
2609 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2610 else
2611 extra_pins = thp_nr_pages(page);
2612 if (pextra_pins)
2613 *pextra_pins = extra_pins;
2614 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2615}
2616
2617/*
2618 * This function splits huge page into normal pages. @page can point to any
2619 * subpage of huge page to split. Split doesn't change the position of @page.
2620 *
2621 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2622 * The huge page must be locked.
2623 *
2624 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2625 *
2626 * Both head page and tail pages will inherit mapping, flags, and so on from
2627 * the hugepage.
2628 *
2629 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2630 * they are not mapped.
2631 *
2632 * Returns 0 if the hugepage is split successfully.
2633 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2634 * us.
2635 */
2636int split_huge_page_to_list(struct page *page, struct list_head *list)
2637{
2638 struct page *head = compound_head(page);
2639 struct deferred_split *ds_queue = get_deferred_split_queue(head);
2640 struct anon_vma *anon_vma = NULL;
2641 struct address_space *mapping = NULL;
2642 int extra_pins, ret;
2643 pgoff_t end;
2644
2645 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2646 VM_BUG_ON_PAGE(!PageLocked(head), head);
2647 VM_BUG_ON_PAGE(!PageCompound(head), head);
2648
2649 if (PageWriteback(head))
2650 return -EBUSY;
2651
2652 if (PageAnon(head)) {
2653 /*
2654 * The caller does not necessarily hold an mmap_lock that would
2655 * prevent the anon_vma disappearing so we first we take a
2656 * reference to it and then lock the anon_vma for write. This
2657 * is similar to page_lock_anon_vma_read except the write lock
2658 * is taken to serialise against parallel split or collapse
2659 * operations.
2660 */
2661 anon_vma = page_get_anon_vma(head);
2662 if (!anon_vma) {
2663 ret = -EBUSY;
2664 goto out;
2665 }
2666 end = -1;
2667 mapping = NULL;
2668 anon_vma_lock_write(anon_vma);
2669 } else {
2670 mapping = head->mapping;
2671
2672 /* Truncated ? */
2673 if (!mapping) {
2674 ret = -EBUSY;
2675 goto out;
2676 }
2677
2678 anon_vma = NULL;
2679 i_mmap_lock_read(mapping);
2680
2681 /*
2682 *__split_huge_page() may need to trim off pages beyond EOF:
2683 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2684 * which cannot be nested inside the page tree lock. So note
2685 * end now: i_size itself may be changed at any moment, but
2686 * head page lock is good enough to serialize the trimming.
2687 */
2688 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2689 }
2690
2691 /*
2692 * Racy check if we can split the page, before unmap_page() will
2693 * split PMDs
2694 */
2695 if (!can_split_huge_page(head, &extra_pins)) {
2696 ret = -EBUSY;
2697 goto out_unlock;
2698 }
2699
2700 unmap_page(head);
2701
2702 /* block interrupt reentry in xa_lock and spinlock */
2703 local_irq_disable();
2704 if (mapping) {
2705 XA_STATE(xas, &mapping->i_pages, page_index(head));
2706
2707 /*
2708 * Check if the head page is present in page cache.
2709 * We assume all tail are present too, if head is there.
2710 */
2711 xa_lock(&mapping->i_pages);
2712 if (xas_load(&xas) != head)
2713 goto fail;
2714 }
2715
2716 /* Prevent deferred_split_scan() touching ->_refcount */
2717 spin_lock(&ds_queue->split_queue_lock);
2718 if (page_ref_freeze(head, 1 + extra_pins)) {
2719 if (!list_empty(page_deferred_list(head))) {
2720 ds_queue->split_queue_len--;
2721 list_del(page_deferred_list(head));
2722 }
2723 spin_unlock(&ds_queue->split_queue_lock);
2724 if (mapping) {
2725 int nr = thp_nr_pages(head);
2726
2727 if (PageSwapBacked(head)) {
2728 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2729 -nr);
2730 } else {
2731 __mod_lruvec_page_state(head, NR_FILE_THPS,
2732 -nr);
2733 filemap_nr_thps_dec(mapping);
2734 }
2735 }
2736
2737 __split_huge_page(page, list, end);
2738 ret = 0;
2739 } else {
2740 spin_unlock(&ds_queue->split_queue_lock);
2741fail:
2742 if (mapping)
2743 xa_unlock(&mapping->i_pages);
2744 local_irq_enable();
2745 remap_page(head, thp_nr_pages(head));
2746 ret = -EBUSY;
2747 }
2748
2749out_unlock:
2750 if (anon_vma) {
2751 anon_vma_unlock_write(anon_vma);
2752 put_anon_vma(anon_vma);
2753 }
2754 if (mapping)
2755 i_mmap_unlock_read(mapping);
2756out:
2757 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2758 return ret;
2759}
2760
2761void free_transhuge_page(struct page *page)
2762{
2763 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2764 unsigned long flags;
2765
2766 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2767 if (!list_empty(page_deferred_list(page))) {
2768 ds_queue->split_queue_len--;
2769 list_del(page_deferred_list(page));
2770 }
2771 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2772 free_compound_page(page);
2773}
2774
2775void deferred_split_huge_page(struct page *page)
2776{
2777 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2778#ifdef CONFIG_MEMCG
2779 struct mem_cgroup *memcg = page_memcg(compound_head(page));
2780#endif
2781 unsigned long flags;
2782
2783 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2784
2785 /*
2786 * The try_to_unmap() in page reclaim path might reach here too,
2787 * this may cause a race condition to corrupt deferred split queue.
2788 * And, if page reclaim is already handling the same page, it is
2789 * unnecessary to handle it again in shrinker.
2790 *
2791 * Check PageSwapCache to determine if the page is being
2792 * handled by page reclaim since THP swap would add the page into
2793 * swap cache before calling try_to_unmap().
2794 */
2795 if (PageSwapCache(page))
2796 return;
2797
2798 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2799 if (list_empty(page_deferred_list(page))) {
2800 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2801 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2802 ds_queue->split_queue_len++;
2803#ifdef CONFIG_MEMCG
2804 if (memcg)
2805 set_shrinker_bit(memcg, page_to_nid(page),
2806 deferred_split_shrinker.id);
2807#endif
2808 }
2809 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2810}
2811
2812static unsigned long deferred_split_count(struct shrinker *shrink,
2813 struct shrink_control *sc)
2814{
2815 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2816 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2817
2818#ifdef CONFIG_MEMCG
2819 if (sc->memcg)
2820 ds_queue = &sc->memcg->deferred_split_queue;
2821#endif
2822 return READ_ONCE(ds_queue->split_queue_len);
2823}
2824
2825static unsigned long deferred_split_scan(struct shrinker *shrink,
2826 struct shrink_control *sc)
2827{
2828 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2829 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2830 unsigned long flags;
2831 LIST_HEAD(list), *pos, *next;
2832 struct page *page;
2833 int split = 0;
2834
2835#ifdef CONFIG_MEMCG
2836 if (sc->memcg)
2837 ds_queue = &sc->memcg->deferred_split_queue;
2838#endif
2839
2840 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2841 /* Take pin on all head pages to avoid freeing them under us */
2842 list_for_each_safe(pos, next, &ds_queue->split_queue) {
2843 page = list_entry((void *)pos, struct page, deferred_list);
2844 page = compound_head(page);
2845 if (get_page_unless_zero(page)) {
2846 list_move(page_deferred_list(page), &list);
2847 } else {
2848 /* We lost race with put_compound_page() */
2849 list_del_init(page_deferred_list(page));
2850 ds_queue->split_queue_len--;
2851 }
2852 if (!--sc->nr_to_scan)
2853 break;
2854 }
2855 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2856
2857 list_for_each_safe(pos, next, &list) {
2858 page = list_entry((void *)pos, struct page, deferred_list);
2859 if (!trylock_page(page))
2860 goto next;
2861 /* split_huge_page() removes page from list on success */
2862 if (!split_huge_page(page))
2863 split++;
2864 unlock_page(page);
2865next:
2866 put_page(page);
2867 }
2868
2869 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2870 list_splice_tail(&list, &ds_queue->split_queue);
2871 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2872
2873 /*
2874 * Stop shrinker if we didn't split any page, but the queue is empty.
2875 * This can happen if pages were freed under us.
2876 */
2877 if (!split && list_empty(&ds_queue->split_queue))
2878 return SHRINK_STOP;
2879 return split;
2880}
2881
2882static struct shrinker deferred_split_shrinker = {
2883 .count_objects = deferred_split_count,
2884 .scan_objects = deferred_split_scan,
2885 .seeks = DEFAULT_SEEKS,
2886 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2887 SHRINKER_NONSLAB,
2888};
2889
2890#ifdef CONFIG_DEBUG_FS
2891static void split_huge_pages_all(void)
2892{
2893 struct zone *zone;
2894 struct page *page;
2895 unsigned long pfn, max_zone_pfn;
2896 unsigned long total = 0, split = 0;
2897
2898 pr_debug("Split all THPs\n");
2899 for_each_populated_zone(zone) {
2900 max_zone_pfn = zone_end_pfn(zone);
2901 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2902 if (!pfn_valid(pfn))
2903 continue;
2904
2905 page = pfn_to_page(pfn);
2906 if (!get_page_unless_zero(page))
2907 continue;
2908
2909 if (zone != page_zone(page))
2910 goto next;
2911
2912 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2913 goto next;
2914
2915 total++;
2916 lock_page(page);
2917 if (!split_huge_page(page))
2918 split++;
2919 unlock_page(page);
2920next:
2921 put_page(page);
2922 cond_resched();
2923 }
2924 }
2925
2926 pr_debug("%lu of %lu THP split\n", split, total);
2927}
2928
2929static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2930{
2931 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2932 is_vm_hugetlb_page(vma);
2933}
2934
2935static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2936 unsigned long vaddr_end)
2937{
2938 int ret = 0;
2939 struct task_struct *task;
2940 struct mm_struct *mm;
2941 unsigned long total = 0, split = 0;
2942 unsigned long addr;
2943
2944 vaddr_start &= PAGE_MASK;
2945 vaddr_end &= PAGE_MASK;
2946
2947 /* Find the task_struct from pid */
2948 rcu_read_lock();
2949 task = find_task_by_vpid(pid);
2950 if (!task) {
2951 rcu_read_unlock();
2952 ret = -ESRCH;
2953 goto out;
2954 }
2955 get_task_struct(task);
2956 rcu_read_unlock();
2957
2958 /* Find the mm_struct */
2959 mm = get_task_mm(task);
2960 put_task_struct(task);
2961
2962 if (!mm) {
2963 ret = -EINVAL;
2964 goto out;
2965 }
2966
2967 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2968 pid, vaddr_start, vaddr_end);
2969
2970 mmap_read_lock(mm);
2971 /*
2972 * always increase addr by PAGE_SIZE, since we could have a PTE page
2973 * table filled with PTE-mapped THPs, each of which is distinct.
2974 */
2975 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2976 struct vm_area_struct *vma = find_vma(mm, addr);
2977 unsigned int follflags;
2978 struct page *page;
2979
2980 if (!vma || addr < vma->vm_start)
2981 break;
2982
2983 /* skip special VMA and hugetlb VMA */
2984 if (vma_not_suitable_for_thp_split(vma)) {
2985 addr = vma->vm_end;
2986 continue;
2987 }
2988
2989 /* FOLL_DUMP to ignore special (like zero) pages */
2990 follflags = FOLL_GET | FOLL_DUMP;
2991 page = follow_page(vma, addr, follflags);
2992
2993 if (IS_ERR(page))
2994 continue;
2995 if (!page)
2996 continue;
2997
2998 if (!is_transparent_hugepage(page))
2999 goto next;
3000
3001 total++;
3002 if (!can_split_huge_page(compound_head(page), NULL))
3003 goto next;
3004
3005 if (!trylock_page(page))
3006 goto next;
3007
3008 if (!split_huge_page(page))
3009 split++;
3010
3011 unlock_page(page);
3012next:
3013 put_page(page);
3014 cond_resched();
3015 }
3016 mmap_read_unlock(mm);
3017 mmput(mm);
3018
3019 pr_debug("%lu of %lu THP split\n", split, total);
3020
3021out:
3022 return ret;
3023}
3024
3025static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3026 pgoff_t off_end)
3027{
3028 struct filename *file;
3029 struct file *candidate;
3030 struct address_space *mapping;
3031 int ret = -EINVAL;
3032 pgoff_t index;
3033 int nr_pages = 1;
3034 unsigned long total = 0, split = 0;
3035
3036 file = getname_kernel(file_path);
3037 if (IS_ERR(file))
3038 return ret;
3039
3040 candidate = file_open_name(file, O_RDONLY, 0);
3041 if (IS_ERR(candidate))
3042 goto out;
3043
3044 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3045 file_path, off_start, off_end);
3046
3047 mapping = candidate->f_mapping;
3048
3049 for (index = off_start; index < off_end; index += nr_pages) {
3050 struct page *fpage = pagecache_get_page(mapping, index,
3051 FGP_ENTRY | FGP_HEAD, 0);
3052
3053 nr_pages = 1;
3054 if (xa_is_value(fpage) || !fpage)
3055 continue;
3056
3057 if (!is_transparent_hugepage(fpage))
3058 goto next;
3059
3060 total++;
3061 nr_pages = thp_nr_pages(fpage);
3062
3063 if (!trylock_page(fpage))
3064 goto next;
3065
3066 if (!split_huge_page(fpage))
3067 split++;
3068
3069 unlock_page(fpage);
3070next:
3071 put_page(fpage);
3072 cond_resched();
3073 }
3074
3075 filp_close(candidate, NULL);
3076 ret = 0;
3077
3078 pr_debug("%lu of %lu file-backed THP split\n", split, total);
3079out:
3080 putname(file);
3081 return ret;
3082}
3083
3084#define MAX_INPUT_BUF_SZ 255
3085
3086static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3087 size_t count, loff_t *ppops)
3088{
3089 static DEFINE_MUTEX(split_debug_mutex);
3090 ssize_t ret;
3091 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3092 char input_buf[MAX_INPUT_BUF_SZ];
3093 int pid;
3094 unsigned long vaddr_start, vaddr_end;
3095
3096 ret = mutex_lock_interruptible(&split_debug_mutex);
3097 if (ret)
3098 return ret;
3099
3100 ret = -EFAULT;
3101
3102 memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3103 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3104 goto out;
3105
3106 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3107
3108 if (input_buf[0] == '/') {
3109 char *tok;
3110 char *buf = input_buf;
3111 char file_path[MAX_INPUT_BUF_SZ];
3112 pgoff_t off_start = 0, off_end = 0;
3113 size_t input_len = strlen(input_buf);
3114
3115 tok = strsep(&buf, ",");
3116 if (tok) {
3117 strcpy(file_path, tok);
3118 } else {
3119 ret = -EINVAL;
3120 goto out;
3121 }
3122
3123 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3124 if (ret != 2) {
3125 ret = -EINVAL;
3126 goto out;
3127 }
3128 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3129 if (!ret)
3130 ret = input_len;
3131
3132 goto out;
3133 }
3134
3135 ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3136 if (ret == 1 && pid == 1) {
3137 split_huge_pages_all();
3138 ret = strlen(input_buf);
3139 goto out;
3140 } else if (ret != 3) {
3141 ret = -EINVAL;
3142 goto out;
3143 }
3144
3145 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3146 if (!ret)
3147 ret = strlen(input_buf);
3148out:
3149 mutex_unlock(&split_debug_mutex);
3150 return ret;
3151
3152}
3153
3154static const struct file_operations split_huge_pages_fops = {
3155 .owner = THIS_MODULE,
3156 .write = split_huge_pages_write,
3157 .llseek = no_llseek,
3158};
3159
3160static int __init split_huge_pages_debugfs(void)
3161{
3162 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3163 &split_huge_pages_fops);
3164 return 0;
3165}
3166late_initcall(split_huge_pages_debugfs);
3167#endif
3168
3169#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3170void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3171 struct page *page)
3172{
3173 struct vm_area_struct *vma = pvmw->vma;
3174 struct mm_struct *mm = vma->vm_mm;
3175 unsigned long address = pvmw->address;
3176 pmd_t pmdval;
3177 swp_entry_t entry;
3178 pmd_t pmdswp;
3179
3180 if (!(pvmw->pmd && !pvmw->pte))
3181 return;
3182
3183 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3184 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3185 if (pmd_dirty(pmdval))
3186 set_page_dirty(page);
3187 if (pmd_write(pmdval))
3188 entry = make_writable_migration_entry(page_to_pfn(page));
3189 else
3190 entry = make_readable_migration_entry(page_to_pfn(page));
3191 pmdswp = swp_entry_to_pmd(entry);
3192 if (pmd_soft_dirty(pmdval))
3193 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3194 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3195 page_remove_rmap(page, true);
3196 put_page(page);
3197}
3198
3199void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3200{
3201 struct vm_area_struct *vma = pvmw->vma;
3202 struct mm_struct *mm = vma->vm_mm;
3203 unsigned long address = pvmw->address;
3204 unsigned long mmun_start = address & HPAGE_PMD_MASK;
3205 pmd_t pmde;
3206 swp_entry_t entry;
3207
3208 if (!(pvmw->pmd && !pvmw->pte))
3209 return;
3210
3211 entry = pmd_to_swp_entry(*pvmw->pmd);
3212 get_page(new);
3213 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3214 if (pmd_swp_soft_dirty(*pvmw->pmd))
3215 pmde = pmd_mksoft_dirty(pmde);
3216 if (is_writable_migration_entry(entry))
3217 pmde = maybe_pmd_mkwrite(pmde, vma);
3218 if (pmd_swp_uffd_wp(*pvmw->pmd))
3219 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3220
3221 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3222 if (PageAnon(new))
3223 page_add_anon_rmap(new, vma, mmun_start, true);
3224 else
3225 page_add_file_rmap(new, true);
3226 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3227 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3228 mlock_vma_page(new);
3229 update_mmu_cache_pmd(vma, address, pvmw->pmd);
3230}
3231#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2009 Red Hat, Inc.
4 */
5
6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/sched/mm.h>
11#include <linux/sched/coredump.h>
12#include <linux/sched/numa_balancing.h>
13#include <linux/highmem.h>
14#include <linux/hugetlb.h>
15#include <linux/mmu_notifier.h>
16#include <linux/rmap.h>
17#include <linux/swap.h>
18#include <linux/shrinker.h>
19#include <linux/mm_inline.h>
20#include <linux/swapops.h>
21#include <linux/backing-dev.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#include <linux/page_owner.h>
38#include <linux/sched/sysctl.h>
39#include <linux/memory-tiers.h>
40
41#include <asm/tlb.h>
42#include <asm/pgalloc.h>
43#include "internal.h"
44#include "swap.h"
45
46#define CREATE_TRACE_POINTS
47#include <trace/events/thp.h>
48
49/*
50 * By default, transparent hugepage support is disabled in order to avoid
51 * risking an increased memory footprint for applications that are not
52 * guaranteed to benefit from it. When transparent hugepage support is
53 * enabled, it is for all mappings, and khugepaged scans all mappings.
54 * Defrag is invoked by khugepaged hugepage allocations and by page faults
55 * for all hugepage allocations.
56 */
57unsigned long transparent_hugepage_flags __read_mostly =
58#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
59 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
60#endif
61#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
62 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
63#endif
64 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
65 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
66 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
67
68static struct shrinker deferred_split_shrinker;
69
70static atomic_t huge_zero_refcount;
71struct page *huge_zero_page __read_mostly;
72unsigned long huge_zero_pfn __read_mostly = ~0UL;
73
74bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
75 bool smaps, bool in_pf, bool enforce_sysfs)
76{
77 if (!vma->vm_mm) /* vdso */
78 return false;
79
80 /*
81 * Explicitly disabled through madvise or prctl, or some
82 * architectures may disable THP for some mappings, for
83 * example, s390 kvm.
84 * */
85 if ((vm_flags & VM_NOHUGEPAGE) ||
86 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
87 return false;
88 /*
89 * If the hardware/firmware marked hugepage support disabled.
90 */
91 if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
92 return false;
93
94 /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
95 if (vma_is_dax(vma))
96 return in_pf;
97
98 /*
99 * Special VMA and hugetlb VMA.
100 * Must be checked after dax since some dax mappings may have
101 * VM_MIXEDMAP set.
102 */
103 if (vm_flags & VM_NO_KHUGEPAGED)
104 return false;
105
106 /*
107 * Check alignment for file vma and size for both file and anon vma.
108 *
109 * Skip the check for page fault. Huge fault does the check in fault
110 * handlers. And this check is not suitable for huge PUD fault.
111 */
112 if (!in_pf &&
113 !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
114 return false;
115
116 /*
117 * Enabled via shmem mount options or sysfs settings.
118 * Must be done before hugepage flags check since shmem has its
119 * own flags.
120 */
121 if (!in_pf && shmem_file(vma->vm_file))
122 return shmem_huge_enabled(vma, !enforce_sysfs);
123
124 /* Enforce sysfs THP requirements as necessary */
125 if (enforce_sysfs &&
126 (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
127 !hugepage_flags_always())))
128 return false;
129
130 /* Only regular file is valid */
131 if (!in_pf && file_thp_enabled(vma))
132 return true;
133
134 if (!vma_is_anonymous(vma))
135 return false;
136
137 if (vma_is_temporary_stack(vma))
138 return false;
139
140 /*
141 * THPeligible bit of smaps should show 1 for proper VMAs even
142 * though anon_vma is not initialized yet.
143 *
144 * Allow page fault since anon_vma may be not initialized until
145 * the first page fault.
146 */
147 if (!vma->anon_vma)
148 return (smaps || in_pf);
149
150 return true;
151}
152
153static bool get_huge_zero_page(void)
154{
155 struct page *zero_page;
156retry:
157 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
158 return true;
159
160 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
161 HPAGE_PMD_ORDER);
162 if (!zero_page) {
163 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
164 return false;
165 }
166 preempt_disable();
167 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
168 preempt_enable();
169 __free_pages(zero_page, compound_order(zero_page));
170 goto retry;
171 }
172 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
173
174 /* We take additional reference here. It will be put back by shrinker */
175 atomic_set(&huge_zero_refcount, 2);
176 preempt_enable();
177 count_vm_event(THP_ZERO_PAGE_ALLOC);
178 return true;
179}
180
181static void put_huge_zero_page(void)
182{
183 /*
184 * Counter should never go to zero here. Only shrinker can put
185 * last reference.
186 */
187 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
188}
189
190struct page *mm_get_huge_zero_page(struct mm_struct *mm)
191{
192 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
193 return READ_ONCE(huge_zero_page);
194
195 if (!get_huge_zero_page())
196 return NULL;
197
198 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
199 put_huge_zero_page();
200
201 return READ_ONCE(huge_zero_page);
202}
203
204void mm_put_huge_zero_page(struct mm_struct *mm)
205{
206 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
207 put_huge_zero_page();
208}
209
210static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
211 struct shrink_control *sc)
212{
213 /* we can free zero page only if last reference remains */
214 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
215}
216
217static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
218 struct shrink_control *sc)
219{
220 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
221 struct page *zero_page = xchg(&huge_zero_page, NULL);
222 BUG_ON(zero_page == NULL);
223 WRITE_ONCE(huge_zero_pfn, ~0UL);
224 __free_pages(zero_page, compound_order(zero_page));
225 return HPAGE_PMD_NR;
226 }
227
228 return 0;
229}
230
231static struct shrinker huge_zero_page_shrinker = {
232 .count_objects = shrink_huge_zero_page_count,
233 .scan_objects = shrink_huge_zero_page_scan,
234 .seeks = DEFAULT_SEEKS,
235};
236
237#ifdef CONFIG_SYSFS
238static ssize_t enabled_show(struct kobject *kobj,
239 struct kobj_attribute *attr, char *buf)
240{
241 const char *output;
242
243 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
244 output = "[always] madvise never";
245 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
246 &transparent_hugepage_flags))
247 output = "always [madvise] never";
248 else
249 output = "always madvise [never]";
250
251 return sysfs_emit(buf, "%s\n", output);
252}
253
254static ssize_t enabled_store(struct kobject *kobj,
255 struct kobj_attribute *attr,
256 const char *buf, size_t count)
257{
258 ssize_t ret = count;
259
260 if (sysfs_streq(buf, "always")) {
261 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
262 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
263 } else if (sysfs_streq(buf, "madvise")) {
264 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
265 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
266 } else if (sysfs_streq(buf, "never")) {
267 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
268 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
269 } else
270 ret = -EINVAL;
271
272 if (ret > 0) {
273 int err = start_stop_khugepaged();
274 if (err)
275 ret = err;
276 }
277 return ret;
278}
279
280static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
281
282ssize_t single_hugepage_flag_show(struct kobject *kobj,
283 struct kobj_attribute *attr, char *buf,
284 enum transparent_hugepage_flag flag)
285{
286 return sysfs_emit(buf, "%d\n",
287 !!test_bit(flag, &transparent_hugepage_flags));
288}
289
290ssize_t single_hugepage_flag_store(struct kobject *kobj,
291 struct kobj_attribute *attr,
292 const char *buf, size_t count,
293 enum transparent_hugepage_flag flag)
294{
295 unsigned long value;
296 int ret;
297
298 ret = kstrtoul(buf, 10, &value);
299 if (ret < 0)
300 return ret;
301 if (value > 1)
302 return -EINVAL;
303
304 if (value)
305 set_bit(flag, &transparent_hugepage_flags);
306 else
307 clear_bit(flag, &transparent_hugepage_flags);
308
309 return count;
310}
311
312static ssize_t defrag_show(struct kobject *kobj,
313 struct kobj_attribute *attr, char *buf)
314{
315 const char *output;
316
317 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
318 &transparent_hugepage_flags))
319 output = "[always] defer defer+madvise madvise never";
320 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
321 &transparent_hugepage_flags))
322 output = "always [defer] defer+madvise madvise never";
323 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
324 &transparent_hugepage_flags))
325 output = "always defer [defer+madvise] madvise never";
326 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
327 &transparent_hugepage_flags))
328 output = "always defer defer+madvise [madvise] never";
329 else
330 output = "always defer defer+madvise madvise [never]";
331
332 return sysfs_emit(buf, "%s\n", output);
333}
334
335static ssize_t defrag_store(struct kobject *kobj,
336 struct kobj_attribute *attr,
337 const char *buf, size_t count)
338{
339 if (sysfs_streq(buf, "always")) {
340 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
341 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
342 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
343 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
344 } else if (sysfs_streq(buf, "defer+madvise")) {
345 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
346 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
347 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
348 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
349 } else if (sysfs_streq(buf, "defer")) {
350 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
351 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
352 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
353 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
354 } else if (sysfs_streq(buf, "madvise")) {
355 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
356 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
357 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
358 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
359 } else if (sysfs_streq(buf, "never")) {
360 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
361 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
362 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
363 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
364 } else
365 return -EINVAL;
366
367 return count;
368}
369static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
370
371static ssize_t use_zero_page_show(struct kobject *kobj,
372 struct kobj_attribute *attr, char *buf)
373{
374 return single_hugepage_flag_show(kobj, attr, buf,
375 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
376}
377static ssize_t use_zero_page_store(struct kobject *kobj,
378 struct kobj_attribute *attr, const char *buf, size_t count)
379{
380 return single_hugepage_flag_store(kobj, attr, buf, count,
381 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
382}
383static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
384
385static ssize_t hpage_pmd_size_show(struct kobject *kobj,
386 struct kobj_attribute *attr, char *buf)
387{
388 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
389}
390static struct kobj_attribute hpage_pmd_size_attr =
391 __ATTR_RO(hpage_pmd_size);
392
393static struct attribute *hugepage_attr[] = {
394 &enabled_attr.attr,
395 &defrag_attr.attr,
396 &use_zero_page_attr.attr,
397 &hpage_pmd_size_attr.attr,
398#ifdef CONFIG_SHMEM
399 &shmem_enabled_attr.attr,
400#endif
401 NULL,
402};
403
404static const struct attribute_group hugepage_attr_group = {
405 .attrs = hugepage_attr,
406};
407
408static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
409{
410 int err;
411
412 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
413 if (unlikely(!*hugepage_kobj)) {
414 pr_err("failed to create transparent hugepage kobject\n");
415 return -ENOMEM;
416 }
417
418 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
419 if (err) {
420 pr_err("failed to register transparent hugepage group\n");
421 goto delete_obj;
422 }
423
424 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
425 if (err) {
426 pr_err("failed to register transparent hugepage group\n");
427 goto remove_hp_group;
428 }
429
430 return 0;
431
432remove_hp_group:
433 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
434delete_obj:
435 kobject_put(*hugepage_kobj);
436 return err;
437}
438
439static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
440{
441 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
442 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
443 kobject_put(hugepage_kobj);
444}
445#else
446static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
447{
448 return 0;
449}
450
451static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
452{
453}
454#endif /* CONFIG_SYSFS */
455
456static int __init hugepage_init(void)
457{
458 int err;
459 struct kobject *hugepage_kobj;
460
461 if (!has_transparent_hugepage()) {
462 /*
463 * Hardware doesn't support hugepages, hence disable
464 * DAX PMD support.
465 */
466 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
467 return -EINVAL;
468 }
469
470 /*
471 * hugepages can't be allocated by the buddy allocator
472 */
473 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
474 /*
475 * we use page->mapping and page->index in second tail page
476 * as list_head: assuming THP order >= 2
477 */
478 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
479
480 err = hugepage_init_sysfs(&hugepage_kobj);
481 if (err)
482 goto err_sysfs;
483
484 err = khugepaged_init();
485 if (err)
486 goto err_slab;
487
488 err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
489 if (err)
490 goto err_hzp_shrinker;
491 err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
492 if (err)
493 goto err_split_shrinker;
494
495 /*
496 * By default disable transparent hugepages on smaller systems,
497 * where the extra memory used could hurt more than TLB overhead
498 * is likely to save. The admin can still enable it through /sys.
499 */
500 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
501 transparent_hugepage_flags = 0;
502 return 0;
503 }
504
505 err = start_stop_khugepaged();
506 if (err)
507 goto err_khugepaged;
508
509 return 0;
510err_khugepaged:
511 unregister_shrinker(&deferred_split_shrinker);
512err_split_shrinker:
513 unregister_shrinker(&huge_zero_page_shrinker);
514err_hzp_shrinker:
515 khugepaged_destroy();
516err_slab:
517 hugepage_exit_sysfs(hugepage_kobj);
518err_sysfs:
519 return err;
520}
521subsys_initcall(hugepage_init);
522
523static int __init setup_transparent_hugepage(char *str)
524{
525 int ret = 0;
526 if (!str)
527 goto out;
528 if (!strcmp(str, "always")) {
529 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
530 &transparent_hugepage_flags);
531 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
532 &transparent_hugepage_flags);
533 ret = 1;
534 } else if (!strcmp(str, "madvise")) {
535 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
536 &transparent_hugepage_flags);
537 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
538 &transparent_hugepage_flags);
539 ret = 1;
540 } else if (!strcmp(str, "never")) {
541 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
542 &transparent_hugepage_flags);
543 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
544 &transparent_hugepage_flags);
545 ret = 1;
546 }
547out:
548 if (!ret)
549 pr_warn("transparent_hugepage= cannot parse, ignored\n");
550 return ret;
551}
552__setup("transparent_hugepage=", setup_transparent_hugepage);
553
554pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
555{
556 if (likely(vma->vm_flags & VM_WRITE))
557 pmd = pmd_mkwrite(pmd);
558 return pmd;
559}
560
561#ifdef CONFIG_MEMCG
562static inline struct deferred_split *get_deferred_split_queue(struct page *page)
563{
564 struct mem_cgroup *memcg = page_memcg(compound_head(page));
565 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
566
567 if (memcg)
568 return &memcg->deferred_split_queue;
569 else
570 return &pgdat->deferred_split_queue;
571}
572#else
573static inline struct deferred_split *get_deferred_split_queue(struct page *page)
574{
575 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
576
577 return &pgdat->deferred_split_queue;
578}
579#endif
580
581void prep_transhuge_page(struct page *page)
582{
583 /*
584 * we use page->mapping and page->index in second tail page
585 * as list_head: assuming THP order >= 2
586 */
587
588 INIT_LIST_HEAD(page_deferred_list(page));
589 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
590}
591
592static inline bool is_transparent_hugepage(struct page *page)
593{
594 if (!PageCompound(page))
595 return false;
596
597 page = compound_head(page);
598 return is_huge_zero_page(page) ||
599 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
600}
601
602static unsigned long __thp_get_unmapped_area(struct file *filp,
603 unsigned long addr, unsigned long len,
604 loff_t off, unsigned long flags, unsigned long size)
605{
606 loff_t off_end = off + len;
607 loff_t off_align = round_up(off, size);
608 unsigned long len_pad, ret;
609
610 if (off_end <= off_align || (off_end - off_align) < size)
611 return 0;
612
613 len_pad = len + size;
614 if (len_pad < len || (off + len_pad) < off)
615 return 0;
616
617 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
618 off >> PAGE_SHIFT, flags);
619
620 /*
621 * The failure might be due to length padding. The caller will retry
622 * without the padding.
623 */
624 if (IS_ERR_VALUE(ret))
625 return 0;
626
627 /*
628 * Do not try to align to THP boundary if allocation at the address
629 * hint succeeds.
630 */
631 if (ret == addr)
632 return addr;
633
634 ret += (off - ret) & (size - 1);
635 return ret;
636}
637
638unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
639 unsigned long len, unsigned long pgoff, unsigned long flags)
640{
641 unsigned long ret;
642 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
643
644 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
645 if (ret)
646 return ret;
647
648 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
649}
650EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
651
652static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
653 struct page *page, gfp_t gfp)
654{
655 struct vm_area_struct *vma = vmf->vma;
656 pgtable_t pgtable;
657 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
658 vm_fault_t ret = 0;
659
660 VM_BUG_ON_PAGE(!PageCompound(page), page);
661
662 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
663 put_page(page);
664 count_vm_event(THP_FAULT_FALLBACK);
665 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
666 return VM_FAULT_FALLBACK;
667 }
668 cgroup_throttle_swaprate(page, gfp);
669
670 pgtable = pte_alloc_one(vma->vm_mm);
671 if (unlikely(!pgtable)) {
672 ret = VM_FAULT_OOM;
673 goto release;
674 }
675
676 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
677 /*
678 * The memory barrier inside __SetPageUptodate makes sure that
679 * clear_huge_page writes become visible before the set_pmd_at()
680 * write.
681 */
682 __SetPageUptodate(page);
683
684 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
685 if (unlikely(!pmd_none(*vmf->pmd))) {
686 goto unlock_release;
687 } else {
688 pmd_t entry;
689
690 ret = check_stable_address_space(vma->vm_mm);
691 if (ret)
692 goto unlock_release;
693
694 /* Deliver the page fault to userland */
695 if (userfaultfd_missing(vma)) {
696 spin_unlock(vmf->ptl);
697 put_page(page);
698 pte_free(vma->vm_mm, pgtable);
699 ret = handle_userfault(vmf, VM_UFFD_MISSING);
700 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
701 return ret;
702 }
703
704 entry = mk_huge_pmd(page, vma->vm_page_prot);
705 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
706 page_add_new_anon_rmap(page, vma, haddr);
707 lru_cache_add_inactive_or_unevictable(page, vma);
708 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
709 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
710 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
711 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
712 mm_inc_nr_ptes(vma->vm_mm);
713 spin_unlock(vmf->ptl);
714 count_vm_event(THP_FAULT_ALLOC);
715 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
716 }
717
718 return 0;
719unlock_release:
720 spin_unlock(vmf->ptl);
721release:
722 if (pgtable)
723 pte_free(vma->vm_mm, pgtable);
724 put_page(page);
725 return ret;
726
727}
728
729/*
730 * always: directly stall for all thp allocations
731 * defer: wake kswapd and fail if not immediately available
732 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
733 * fail if not immediately available
734 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
735 * available
736 * never: never stall for any thp allocation
737 */
738gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
739{
740 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
741
742 /* Always do synchronous compaction */
743 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
744 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
745
746 /* Kick kcompactd and fail quickly */
747 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
748 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
749
750 /* Synchronous compaction if madvised, otherwise kick kcompactd */
751 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
752 return GFP_TRANSHUGE_LIGHT |
753 (vma_madvised ? __GFP_DIRECT_RECLAIM :
754 __GFP_KSWAPD_RECLAIM);
755
756 /* Only do synchronous compaction if madvised */
757 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
758 return GFP_TRANSHUGE_LIGHT |
759 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
760
761 return GFP_TRANSHUGE_LIGHT;
762}
763
764/* Caller must hold page table lock. */
765static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
766 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
767 struct page *zero_page)
768{
769 pmd_t entry;
770 if (!pmd_none(*pmd))
771 return;
772 entry = mk_pmd(zero_page, vma->vm_page_prot);
773 entry = pmd_mkhuge(entry);
774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
775 set_pmd_at(mm, haddr, pmd, entry);
776 mm_inc_nr_ptes(mm);
777}
778
779vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
780{
781 struct vm_area_struct *vma = vmf->vma;
782 gfp_t gfp;
783 struct folio *folio;
784 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
785
786 if (!transhuge_vma_suitable(vma, haddr))
787 return VM_FAULT_FALLBACK;
788 if (unlikely(anon_vma_prepare(vma)))
789 return VM_FAULT_OOM;
790 khugepaged_enter_vma(vma, vma->vm_flags);
791
792 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
793 !mm_forbids_zeropage(vma->vm_mm) &&
794 transparent_hugepage_use_zero_page()) {
795 pgtable_t pgtable;
796 struct page *zero_page;
797 vm_fault_t ret;
798 pgtable = pte_alloc_one(vma->vm_mm);
799 if (unlikely(!pgtable))
800 return VM_FAULT_OOM;
801 zero_page = mm_get_huge_zero_page(vma->vm_mm);
802 if (unlikely(!zero_page)) {
803 pte_free(vma->vm_mm, pgtable);
804 count_vm_event(THP_FAULT_FALLBACK);
805 return VM_FAULT_FALLBACK;
806 }
807 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
808 ret = 0;
809 if (pmd_none(*vmf->pmd)) {
810 ret = check_stable_address_space(vma->vm_mm);
811 if (ret) {
812 spin_unlock(vmf->ptl);
813 pte_free(vma->vm_mm, pgtable);
814 } else if (userfaultfd_missing(vma)) {
815 spin_unlock(vmf->ptl);
816 pte_free(vma->vm_mm, pgtable);
817 ret = handle_userfault(vmf, VM_UFFD_MISSING);
818 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
819 } else {
820 set_huge_zero_page(pgtable, vma->vm_mm, vma,
821 haddr, vmf->pmd, zero_page);
822 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
823 spin_unlock(vmf->ptl);
824 }
825 } else {
826 spin_unlock(vmf->ptl);
827 pte_free(vma->vm_mm, pgtable);
828 }
829 return ret;
830 }
831 gfp = vma_thp_gfp_mask(vma);
832 folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
833 if (unlikely(!folio)) {
834 count_vm_event(THP_FAULT_FALLBACK);
835 return VM_FAULT_FALLBACK;
836 }
837 return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
838}
839
840static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
841 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
842 pgtable_t pgtable)
843{
844 struct mm_struct *mm = vma->vm_mm;
845 pmd_t entry;
846 spinlock_t *ptl;
847
848 ptl = pmd_lock(mm, pmd);
849 if (!pmd_none(*pmd)) {
850 if (write) {
851 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
852 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
853 goto out_unlock;
854 }
855 entry = pmd_mkyoung(*pmd);
856 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
857 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
858 update_mmu_cache_pmd(vma, addr, pmd);
859 }
860
861 goto out_unlock;
862 }
863
864 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
865 if (pfn_t_devmap(pfn))
866 entry = pmd_mkdevmap(entry);
867 if (write) {
868 entry = pmd_mkyoung(pmd_mkdirty(entry));
869 entry = maybe_pmd_mkwrite(entry, vma);
870 }
871
872 if (pgtable) {
873 pgtable_trans_huge_deposit(mm, pmd, pgtable);
874 mm_inc_nr_ptes(mm);
875 pgtable = NULL;
876 }
877
878 set_pmd_at(mm, addr, pmd, entry);
879 update_mmu_cache_pmd(vma, addr, pmd);
880
881out_unlock:
882 spin_unlock(ptl);
883 if (pgtable)
884 pte_free(mm, pgtable);
885}
886
887/**
888 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
889 * @vmf: Structure describing the fault
890 * @pfn: pfn to insert
891 * @pgprot: page protection to use
892 * @write: whether it's a write fault
893 *
894 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
895 * also consult the vmf_insert_mixed_prot() documentation when
896 * @pgprot != @vmf->vma->vm_page_prot.
897 *
898 * Return: vm_fault_t value.
899 */
900vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
901 pgprot_t pgprot, bool write)
902{
903 unsigned long addr = vmf->address & PMD_MASK;
904 struct vm_area_struct *vma = vmf->vma;
905 pgtable_t pgtable = NULL;
906
907 /*
908 * If we had pmd_special, we could avoid all these restrictions,
909 * but we need to be consistent with PTEs and architectures that
910 * can't support a 'special' bit.
911 */
912 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
913 !pfn_t_devmap(pfn));
914 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
915 (VM_PFNMAP|VM_MIXEDMAP));
916 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
917
918 if (addr < vma->vm_start || addr >= vma->vm_end)
919 return VM_FAULT_SIGBUS;
920
921 if (arch_needs_pgtable_deposit()) {
922 pgtable = pte_alloc_one(vma->vm_mm);
923 if (!pgtable)
924 return VM_FAULT_OOM;
925 }
926
927 track_pfn_insert(vma, &pgprot, pfn);
928
929 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
930 return VM_FAULT_NOPAGE;
931}
932EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
933
934#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
935static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
936{
937 if (likely(vma->vm_flags & VM_WRITE))
938 pud = pud_mkwrite(pud);
939 return pud;
940}
941
942static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
943 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
944{
945 struct mm_struct *mm = vma->vm_mm;
946 pud_t entry;
947 spinlock_t *ptl;
948
949 ptl = pud_lock(mm, pud);
950 if (!pud_none(*pud)) {
951 if (write) {
952 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
953 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
954 goto out_unlock;
955 }
956 entry = pud_mkyoung(*pud);
957 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
958 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
959 update_mmu_cache_pud(vma, addr, pud);
960 }
961 goto out_unlock;
962 }
963
964 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
965 if (pfn_t_devmap(pfn))
966 entry = pud_mkdevmap(entry);
967 if (write) {
968 entry = pud_mkyoung(pud_mkdirty(entry));
969 entry = maybe_pud_mkwrite(entry, vma);
970 }
971 set_pud_at(mm, addr, pud, entry);
972 update_mmu_cache_pud(vma, addr, pud);
973
974out_unlock:
975 spin_unlock(ptl);
976}
977
978/**
979 * vmf_insert_pfn_pud_prot - insert a pud size pfn
980 * @vmf: Structure describing the fault
981 * @pfn: pfn to insert
982 * @pgprot: page protection to use
983 * @write: whether it's a write fault
984 *
985 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
986 * also consult the vmf_insert_mixed_prot() documentation when
987 * @pgprot != @vmf->vma->vm_page_prot.
988 *
989 * Return: vm_fault_t value.
990 */
991vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
992 pgprot_t pgprot, bool write)
993{
994 unsigned long addr = vmf->address & PUD_MASK;
995 struct vm_area_struct *vma = vmf->vma;
996
997 /*
998 * If we had pud_special, we could avoid all these restrictions,
999 * but we need to be consistent with PTEs and architectures that
1000 * can't support a 'special' bit.
1001 */
1002 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1003 !pfn_t_devmap(pfn));
1004 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1005 (VM_PFNMAP|VM_MIXEDMAP));
1006 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1007
1008 if (addr < vma->vm_start || addr >= vma->vm_end)
1009 return VM_FAULT_SIGBUS;
1010
1011 track_pfn_insert(vma, &pgprot, pfn);
1012
1013 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1014 return VM_FAULT_NOPAGE;
1015}
1016EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1017#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1018
1019static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1020 pmd_t *pmd, bool write)
1021{
1022 pmd_t _pmd;
1023
1024 _pmd = pmd_mkyoung(*pmd);
1025 if (write)
1026 _pmd = pmd_mkdirty(_pmd);
1027 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1028 pmd, _pmd, write))
1029 update_mmu_cache_pmd(vma, addr, pmd);
1030}
1031
1032struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1033 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1034{
1035 unsigned long pfn = pmd_pfn(*pmd);
1036 struct mm_struct *mm = vma->vm_mm;
1037 struct page *page;
1038 int ret;
1039
1040 assert_spin_locked(pmd_lockptr(mm, pmd));
1041
1042 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1043 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1044 (FOLL_PIN | FOLL_GET)))
1045 return NULL;
1046
1047 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1048 return NULL;
1049
1050 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1051 /* pass */;
1052 else
1053 return NULL;
1054
1055 if (flags & FOLL_TOUCH)
1056 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1057
1058 /*
1059 * device mapped pages can only be returned if the
1060 * caller will manage the page reference count.
1061 */
1062 if (!(flags & (FOLL_GET | FOLL_PIN)))
1063 return ERR_PTR(-EEXIST);
1064
1065 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1066 *pgmap = get_dev_pagemap(pfn, *pgmap);
1067 if (!*pgmap)
1068 return ERR_PTR(-EFAULT);
1069 page = pfn_to_page(pfn);
1070 ret = try_grab_page(page, flags);
1071 if (ret)
1072 page = ERR_PTR(ret);
1073
1074 return page;
1075}
1076
1077int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1078 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1079 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1080{
1081 spinlock_t *dst_ptl, *src_ptl;
1082 struct page *src_page;
1083 pmd_t pmd;
1084 pgtable_t pgtable = NULL;
1085 int ret = -ENOMEM;
1086
1087 /* Skip if can be re-fill on fault */
1088 if (!vma_is_anonymous(dst_vma))
1089 return 0;
1090
1091 pgtable = pte_alloc_one(dst_mm);
1092 if (unlikely(!pgtable))
1093 goto out;
1094
1095 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1096 src_ptl = pmd_lockptr(src_mm, src_pmd);
1097 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1098
1099 ret = -EAGAIN;
1100 pmd = *src_pmd;
1101
1102#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1103 if (unlikely(is_swap_pmd(pmd))) {
1104 swp_entry_t entry = pmd_to_swp_entry(pmd);
1105
1106 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1107 if (!is_readable_migration_entry(entry)) {
1108 entry = make_readable_migration_entry(
1109 swp_offset(entry));
1110 pmd = swp_entry_to_pmd(entry);
1111 if (pmd_swp_soft_dirty(*src_pmd))
1112 pmd = pmd_swp_mksoft_dirty(pmd);
1113 if (pmd_swp_uffd_wp(*src_pmd))
1114 pmd = pmd_swp_mkuffd_wp(pmd);
1115 set_pmd_at(src_mm, addr, src_pmd, pmd);
1116 }
1117 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118 mm_inc_nr_ptes(dst_mm);
1119 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1120 if (!userfaultfd_wp(dst_vma))
1121 pmd = pmd_swp_clear_uffd_wp(pmd);
1122 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1123 ret = 0;
1124 goto out_unlock;
1125 }
1126#endif
1127
1128 if (unlikely(!pmd_trans_huge(pmd))) {
1129 pte_free(dst_mm, pgtable);
1130 goto out_unlock;
1131 }
1132 /*
1133 * When page table lock is held, the huge zero pmd should not be
1134 * under splitting since we don't split the page itself, only pmd to
1135 * a page table.
1136 */
1137 if (is_huge_zero_pmd(pmd)) {
1138 /*
1139 * get_huge_zero_page() will never allocate a new page here,
1140 * since we already have a zero page to copy. It just takes a
1141 * reference.
1142 */
1143 mm_get_huge_zero_page(dst_mm);
1144 goto out_zero_page;
1145 }
1146
1147 src_page = pmd_page(pmd);
1148 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1149
1150 get_page(src_page);
1151 if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1152 /* Page maybe pinned: split and retry the fault on PTEs. */
1153 put_page(src_page);
1154 pte_free(dst_mm, pgtable);
1155 spin_unlock(src_ptl);
1156 spin_unlock(dst_ptl);
1157 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1158 return -EAGAIN;
1159 }
1160 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1161out_zero_page:
1162 mm_inc_nr_ptes(dst_mm);
1163 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1164 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1165 if (!userfaultfd_wp(dst_vma))
1166 pmd = pmd_clear_uffd_wp(pmd);
1167 pmd = pmd_mkold(pmd_wrprotect(pmd));
1168 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1169
1170 ret = 0;
1171out_unlock:
1172 spin_unlock(src_ptl);
1173 spin_unlock(dst_ptl);
1174out:
1175 return ret;
1176}
1177
1178#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1179static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1180 pud_t *pud, bool write)
1181{
1182 pud_t _pud;
1183
1184 _pud = pud_mkyoung(*pud);
1185 if (write)
1186 _pud = pud_mkdirty(_pud);
1187 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1188 pud, _pud, write))
1189 update_mmu_cache_pud(vma, addr, pud);
1190}
1191
1192struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1193 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1194{
1195 unsigned long pfn = pud_pfn(*pud);
1196 struct mm_struct *mm = vma->vm_mm;
1197 struct page *page;
1198 int ret;
1199
1200 assert_spin_locked(pud_lockptr(mm, pud));
1201
1202 if (flags & FOLL_WRITE && !pud_write(*pud))
1203 return NULL;
1204
1205 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1206 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1207 (FOLL_PIN | FOLL_GET)))
1208 return NULL;
1209
1210 if (pud_present(*pud) && pud_devmap(*pud))
1211 /* pass */;
1212 else
1213 return NULL;
1214
1215 if (flags & FOLL_TOUCH)
1216 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1217
1218 /*
1219 * device mapped pages can only be returned if the
1220 * caller will manage the page reference count.
1221 *
1222 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1223 */
1224 if (!(flags & (FOLL_GET | FOLL_PIN)))
1225 return ERR_PTR(-EEXIST);
1226
1227 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1228 *pgmap = get_dev_pagemap(pfn, *pgmap);
1229 if (!*pgmap)
1230 return ERR_PTR(-EFAULT);
1231 page = pfn_to_page(pfn);
1232
1233 ret = try_grab_page(page, flags);
1234 if (ret)
1235 page = ERR_PTR(ret);
1236
1237 return page;
1238}
1239
1240int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1241 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1242 struct vm_area_struct *vma)
1243{
1244 spinlock_t *dst_ptl, *src_ptl;
1245 pud_t pud;
1246 int ret;
1247
1248 dst_ptl = pud_lock(dst_mm, dst_pud);
1249 src_ptl = pud_lockptr(src_mm, src_pud);
1250 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1251
1252 ret = -EAGAIN;
1253 pud = *src_pud;
1254 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1255 goto out_unlock;
1256
1257 /*
1258 * When page table lock is held, the huge zero pud should not be
1259 * under splitting since we don't split the page itself, only pud to
1260 * a page table.
1261 */
1262 if (is_huge_zero_pud(pud)) {
1263 /* No huge zero pud yet */
1264 }
1265
1266 /*
1267 * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1268 * and split if duplicating fails.
1269 */
1270 pudp_set_wrprotect(src_mm, addr, src_pud);
1271 pud = pud_mkold(pud_wrprotect(pud));
1272 set_pud_at(dst_mm, addr, dst_pud, pud);
1273
1274 ret = 0;
1275out_unlock:
1276 spin_unlock(src_ptl);
1277 spin_unlock(dst_ptl);
1278 return ret;
1279}
1280
1281void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1282{
1283 bool write = vmf->flags & FAULT_FLAG_WRITE;
1284
1285 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1286 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1287 goto unlock;
1288
1289 touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1290unlock:
1291 spin_unlock(vmf->ptl);
1292}
1293#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1294
1295void huge_pmd_set_accessed(struct vm_fault *vmf)
1296{
1297 bool write = vmf->flags & FAULT_FLAG_WRITE;
1298
1299 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1300 if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1301 goto unlock;
1302
1303 touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1304
1305unlock:
1306 spin_unlock(vmf->ptl);
1307}
1308
1309vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1310{
1311 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1312 struct vm_area_struct *vma = vmf->vma;
1313 struct folio *folio;
1314 struct page *page;
1315 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1316 pmd_t orig_pmd = vmf->orig_pmd;
1317
1318 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1319 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1320
1321 if (is_huge_zero_pmd(orig_pmd))
1322 goto fallback;
1323
1324 spin_lock(vmf->ptl);
1325
1326 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1327 spin_unlock(vmf->ptl);
1328 return 0;
1329 }
1330
1331 page = pmd_page(orig_pmd);
1332 folio = page_folio(page);
1333 VM_BUG_ON_PAGE(!PageHead(page), page);
1334
1335 /* Early check when only holding the PT lock. */
1336 if (PageAnonExclusive(page))
1337 goto reuse;
1338
1339 if (!folio_trylock(folio)) {
1340 folio_get(folio);
1341 spin_unlock(vmf->ptl);
1342 folio_lock(folio);
1343 spin_lock(vmf->ptl);
1344 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345 spin_unlock(vmf->ptl);
1346 folio_unlock(folio);
1347 folio_put(folio);
1348 return 0;
1349 }
1350 folio_put(folio);
1351 }
1352
1353 /* Recheck after temporarily dropping the PT lock. */
1354 if (PageAnonExclusive(page)) {
1355 folio_unlock(folio);
1356 goto reuse;
1357 }
1358
1359 /*
1360 * See do_wp_page(): we can only reuse the folio exclusively if
1361 * there are no additional references. Note that we always drain
1362 * the LRU pagevecs immediately after adding a THP.
1363 */
1364 if (folio_ref_count(folio) >
1365 1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1366 goto unlock_fallback;
1367 if (folio_test_swapcache(folio))
1368 folio_free_swap(folio);
1369 if (folio_ref_count(folio) == 1) {
1370 pmd_t entry;
1371
1372 page_move_anon_rmap(page, vma);
1373 folio_unlock(folio);
1374reuse:
1375 if (unlikely(unshare)) {
1376 spin_unlock(vmf->ptl);
1377 return 0;
1378 }
1379 entry = pmd_mkyoung(orig_pmd);
1380 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1381 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1382 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1383 spin_unlock(vmf->ptl);
1384 return 0;
1385 }
1386
1387unlock_fallback:
1388 folio_unlock(folio);
1389 spin_unlock(vmf->ptl);
1390fallback:
1391 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1392 return VM_FAULT_FALLBACK;
1393}
1394
1395static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1396 unsigned long addr, pmd_t pmd)
1397{
1398 struct page *page;
1399
1400 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1401 return false;
1402
1403 /* Don't touch entries that are not even readable (NUMA hinting). */
1404 if (pmd_protnone(pmd))
1405 return false;
1406
1407 /* Do we need write faults for softdirty tracking? */
1408 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1409 return false;
1410
1411 /* Do we need write faults for uffd-wp tracking? */
1412 if (userfaultfd_huge_pmd_wp(vma, pmd))
1413 return false;
1414
1415 if (!(vma->vm_flags & VM_SHARED)) {
1416 /* See can_change_pte_writable(). */
1417 page = vm_normal_page_pmd(vma, addr, pmd);
1418 return page && PageAnon(page) && PageAnonExclusive(page);
1419 }
1420
1421 /* See can_change_pte_writable(). */
1422 return pmd_dirty(pmd);
1423}
1424
1425/* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1426static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1427 struct vm_area_struct *vma,
1428 unsigned int flags)
1429{
1430 /* If the pmd is writable, we can write to the page. */
1431 if (pmd_write(pmd))
1432 return true;
1433
1434 /* Maybe FOLL_FORCE is set to override it? */
1435 if (!(flags & FOLL_FORCE))
1436 return false;
1437
1438 /* But FOLL_FORCE has no effect on shared mappings */
1439 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1440 return false;
1441
1442 /* ... or read-only private ones */
1443 if (!(vma->vm_flags & VM_MAYWRITE))
1444 return false;
1445
1446 /* ... or already writable ones that just need to take a write fault */
1447 if (vma->vm_flags & VM_WRITE)
1448 return false;
1449
1450 /*
1451 * See can_change_pte_writable(): we broke COW and could map the page
1452 * writable if we have an exclusive anonymous page ...
1453 */
1454 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1455 return false;
1456
1457 /* ... and a write-fault isn't required for other reasons. */
1458 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1459 return false;
1460 return !userfaultfd_huge_pmd_wp(vma, pmd);
1461}
1462
1463struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1464 unsigned long addr,
1465 pmd_t *pmd,
1466 unsigned int flags)
1467{
1468 struct mm_struct *mm = vma->vm_mm;
1469 struct page *page;
1470 int ret;
1471
1472 assert_spin_locked(pmd_lockptr(mm, pmd));
1473
1474 page = pmd_page(*pmd);
1475 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1476
1477 if ((flags & FOLL_WRITE) &&
1478 !can_follow_write_pmd(*pmd, page, vma, flags))
1479 return NULL;
1480
1481 /* Avoid dumping huge zero page */
1482 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1483 return ERR_PTR(-EFAULT);
1484
1485 /* Full NUMA hinting faults to serialise migration in fault paths */
1486 if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1487 return NULL;
1488
1489 if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1490 return ERR_PTR(-EMLINK);
1491
1492 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1493 !PageAnonExclusive(page), page);
1494
1495 ret = try_grab_page(page, flags);
1496 if (ret)
1497 return ERR_PTR(ret);
1498
1499 if (flags & FOLL_TOUCH)
1500 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1501
1502 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1503 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1504
1505 return page;
1506}
1507
1508/* NUMA hinting page fault entry point for trans huge pmds */
1509vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1510{
1511 struct vm_area_struct *vma = vmf->vma;
1512 pmd_t oldpmd = vmf->orig_pmd;
1513 pmd_t pmd;
1514 struct page *page;
1515 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1516 int page_nid = NUMA_NO_NODE;
1517 int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1518 bool migrated = false, writable = false;
1519 int flags = 0;
1520
1521 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1522 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1523 spin_unlock(vmf->ptl);
1524 goto out;
1525 }
1526
1527 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1528
1529 /*
1530 * Detect now whether the PMD could be writable; this information
1531 * is only valid while holding the PT lock.
1532 */
1533 writable = pmd_write(pmd);
1534 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1535 can_change_pmd_writable(vma, vmf->address, pmd))
1536 writable = true;
1537
1538 page = vm_normal_page_pmd(vma, haddr, pmd);
1539 if (!page)
1540 goto out_map;
1541
1542 /* See similar comment in do_numa_page for explanation */
1543 if (!writable)
1544 flags |= TNF_NO_GROUP;
1545
1546 page_nid = page_to_nid(page);
1547 /*
1548 * For memory tiering mode, cpupid of slow memory page is used
1549 * to record page access time. So use default value.
1550 */
1551 if (node_is_toptier(page_nid))
1552 last_cpupid = page_cpupid_last(page);
1553 target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1554 &flags);
1555
1556 if (target_nid == NUMA_NO_NODE) {
1557 put_page(page);
1558 goto out_map;
1559 }
1560
1561 spin_unlock(vmf->ptl);
1562 writable = false;
1563
1564 migrated = migrate_misplaced_page(page, vma, target_nid);
1565 if (migrated) {
1566 flags |= TNF_MIGRATED;
1567 page_nid = target_nid;
1568 } else {
1569 flags |= TNF_MIGRATE_FAIL;
1570 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1571 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1572 spin_unlock(vmf->ptl);
1573 goto out;
1574 }
1575 goto out_map;
1576 }
1577
1578out:
1579 if (page_nid != NUMA_NO_NODE)
1580 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1581 flags);
1582
1583 return 0;
1584
1585out_map:
1586 /* Restore the PMD */
1587 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1588 pmd = pmd_mkyoung(pmd);
1589 if (writable)
1590 pmd = pmd_mkwrite(pmd);
1591 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1592 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1593 spin_unlock(vmf->ptl);
1594 goto out;
1595}
1596
1597/*
1598 * Return true if we do MADV_FREE successfully on entire pmd page.
1599 * Otherwise, return false.
1600 */
1601bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1602 pmd_t *pmd, unsigned long addr, unsigned long next)
1603{
1604 spinlock_t *ptl;
1605 pmd_t orig_pmd;
1606 struct page *page;
1607 struct mm_struct *mm = tlb->mm;
1608 bool ret = false;
1609
1610 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1611
1612 ptl = pmd_trans_huge_lock(pmd, vma);
1613 if (!ptl)
1614 goto out_unlocked;
1615
1616 orig_pmd = *pmd;
1617 if (is_huge_zero_pmd(orig_pmd))
1618 goto out;
1619
1620 if (unlikely(!pmd_present(orig_pmd))) {
1621 VM_BUG_ON(thp_migration_supported() &&
1622 !is_pmd_migration_entry(orig_pmd));
1623 goto out;
1624 }
1625
1626 page = pmd_page(orig_pmd);
1627 /*
1628 * If other processes are mapping this page, we couldn't discard
1629 * the page unless they all do MADV_FREE so let's skip the page.
1630 */
1631 if (total_mapcount(page) != 1)
1632 goto out;
1633
1634 if (!trylock_page(page))
1635 goto out;
1636
1637 /*
1638 * If user want to discard part-pages of THP, split it so MADV_FREE
1639 * will deactivate only them.
1640 */
1641 if (next - addr != HPAGE_PMD_SIZE) {
1642 get_page(page);
1643 spin_unlock(ptl);
1644 split_huge_page(page);
1645 unlock_page(page);
1646 put_page(page);
1647 goto out_unlocked;
1648 }
1649
1650 if (PageDirty(page))
1651 ClearPageDirty(page);
1652 unlock_page(page);
1653
1654 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1655 pmdp_invalidate(vma, addr, pmd);
1656 orig_pmd = pmd_mkold(orig_pmd);
1657 orig_pmd = pmd_mkclean(orig_pmd);
1658
1659 set_pmd_at(mm, addr, pmd, orig_pmd);
1660 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1661 }
1662
1663 mark_page_lazyfree(page);
1664 ret = true;
1665out:
1666 spin_unlock(ptl);
1667out_unlocked:
1668 return ret;
1669}
1670
1671static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1672{
1673 pgtable_t pgtable;
1674
1675 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1676 pte_free(mm, pgtable);
1677 mm_dec_nr_ptes(mm);
1678}
1679
1680int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1681 pmd_t *pmd, unsigned long addr)
1682{
1683 pmd_t orig_pmd;
1684 spinlock_t *ptl;
1685
1686 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1687
1688 ptl = __pmd_trans_huge_lock(pmd, vma);
1689 if (!ptl)
1690 return 0;
1691 /*
1692 * For architectures like ppc64 we look at deposited pgtable
1693 * when calling pmdp_huge_get_and_clear. So do the
1694 * pgtable_trans_huge_withdraw after finishing pmdp related
1695 * operations.
1696 */
1697 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1698 tlb->fullmm);
1699 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1700 if (vma_is_special_huge(vma)) {
1701 if (arch_needs_pgtable_deposit())
1702 zap_deposited_table(tlb->mm, pmd);
1703 spin_unlock(ptl);
1704 } else if (is_huge_zero_pmd(orig_pmd)) {
1705 zap_deposited_table(tlb->mm, pmd);
1706 spin_unlock(ptl);
1707 } else {
1708 struct page *page = NULL;
1709 int flush_needed = 1;
1710
1711 if (pmd_present(orig_pmd)) {
1712 page = pmd_page(orig_pmd);
1713 page_remove_rmap(page, vma, true);
1714 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1715 VM_BUG_ON_PAGE(!PageHead(page), page);
1716 } else if (thp_migration_supported()) {
1717 swp_entry_t entry;
1718
1719 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1720 entry = pmd_to_swp_entry(orig_pmd);
1721 page = pfn_swap_entry_to_page(entry);
1722 flush_needed = 0;
1723 } else
1724 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1725
1726 if (PageAnon(page)) {
1727 zap_deposited_table(tlb->mm, pmd);
1728 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1729 } else {
1730 if (arch_needs_pgtable_deposit())
1731 zap_deposited_table(tlb->mm, pmd);
1732 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1733 }
1734
1735 spin_unlock(ptl);
1736 if (flush_needed)
1737 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1738 }
1739 return 1;
1740}
1741
1742#ifndef pmd_move_must_withdraw
1743static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1744 spinlock_t *old_pmd_ptl,
1745 struct vm_area_struct *vma)
1746{
1747 /*
1748 * With split pmd lock we also need to move preallocated
1749 * PTE page table if new_pmd is on different PMD page table.
1750 *
1751 * We also don't deposit and withdraw tables for file pages.
1752 */
1753 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1754}
1755#endif
1756
1757static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1758{
1759#ifdef CONFIG_MEM_SOFT_DIRTY
1760 if (unlikely(is_pmd_migration_entry(pmd)))
1761 pmd = pmd_swp_mksoft_dirty(pmd);
1762 else if (pmd_present(pmd))
1763 pmd = pmd_mksoft_dirty(pmd);
1764#endif
1765 return pmd;
1766}
1767
1768bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1769 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1770{
1771 spinlock_t *old_ptl, *new_ptl;
1772 pmd_t pmd;
1773 struct mm_struct *mm = vma->vm_mm;
1774 bool force_flush = false;
1775
1776 /*
1777 * The destination pmd shouldn't be established, free_pgtables()
1778 * should have release it.
1779 */
1780 if (WARN_ON(!pmd_none(*new_pmd))) {
1781 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1782 return false;
1783 }
1784
1785 /*
1786 * We don't have to worry about the ordering of src and dst
1787 * ptlocks because exclusive mmap_lock prevents deadlock.
1788 */
1789 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1790 if (old_ptl) {
1791 new_ptl = pmd_lockptr(mm, new_pmd);
1792 if (new_ptl != old_ptl)
1793 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1794 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1795 if (pmd_present(pmd))
1796 force_flush = true;
1797 VM_BUG_ON(!pmd_none(*new_pmd));
1798
1799 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1800 pgtable_t pgtable;
1801 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1802 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1803 }
1804 pmd = move_soft_dirty_pmd(pmd);
1805 set_pmd_at(mm, new_addr, new_pmd, pmd);
1806 if (force_flush)
1807 flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1808 if (new_ptl != old_ptl)
1809 spin_unlock(new_ptl);
1810 spin_unlock(old_ptl);
1811 return true;
1812 }
1813 return false;
1814}
1815
1816/*
1817 * Returns
1818 * - 0 if PMD could not be locked
1819 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1820 * or if prot_numa but THP migration is not supported
1821 * - HPAGE_PMD_NR if protections changed and TLB flush necessary
1822 */
1823int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1824 pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1825 unsigned long cp_flags)
1826{
1827 struct mm_struct *mm = vma->vm_mm;
1828 spinlock_t *ptl;
1829 pmd_t oldpmd, entry;
1830 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1831 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1832 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1833 int ret = 1;
1834
1835 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1836
1837 if (prot_numa && !thp_migration_supported())
1838 return 1;
1839
1840 ptl = __pmd_trans_huge_lock(pmd, vma);
1841 if (!ptl)
1842 return 0;
1843
1844#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1845 if (is_swap_pmd(*pmd)) {
1846 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1847 struct page *page = pfn_swap_entry_to_page(entry);
1848
1849 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1850 if (is_writable_migration_entry(entry)) {
1851 pmd_t newpmd;
1852 /*
1853 * A protection check is difficult so
1854 * just be safe and disable write
1855 */
1856 if (PageAnon(page))
1857 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1858 else
1859 entry = make_readable_migration_entry(swp_offset(entry));
1860 newpmd = swp_entry_to_pmd(entry);
1861 if (pmd_swp_soft_dirty(*pmd))
1862 newpmd = pmd_swp_mksoft_dirty(newpmd);
1863 if (pmd_swp_uffd_wp(*pmd))
1864 newpmd = pmd_swp_mkuffd_wp(newpmd);
1865 set_pmd_at(mm, addr, pmd, newpmd);
1866 }
1867 goto unlock;
1868 }
1869#endif
1870
1871 if (prot_numa) {
1872 struct page *page;
1873 bool toptier;
1874 /*
1875 * Avoid trapping faults against the zero page. The read-only
1876 * data is likely to be read-cached on the local CPU and
1877 * local/remote hits to the zero page are not interesting.
1878 */
1879 if (is_huge_zero_pmd(*pmd))
1880 goto unlock;
1881
1882 if (pmd_protnone(*pmd))
1883 goto unlock;
1884
1885 page = pmd_page(*pmd);
1886 toptier = node_is_toptier(page_to_nid(page));
1887 /*
1888 * Skip scanning top tier node if normal numa
1889 * balancing is disabled
1890 */
1891 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1892 toptier)
1893 goto unlock;
1894
1895 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1896 !toptier)
1897 xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1898 }
1899 /*
1900 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1901 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1902 * which is also under mmap_read_lock(mm):
1903 *
1904 * CPU0: CPU1:
1905 * change_huge_pmd(prot_numa=1)
1906 * pmdp_huge_get_and_clear_notify()
1907 * madvise_dontneed()
1908 * zap_pmd_range()
1909 * pmd_trans_huge(*pmd) == 0 (without ptl)
1910 * // skip the pmd
1911 * set_pmd_at();
1912 * // pmd is re-established
1913 *
1914 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1915 * which may break userspace.
1916 *
1917 * pmdp_invalidate_ad() is required to make sure we don't miss
1918 * dirty/young flags set by hardware.
1919 */
1920 oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1921
1922 entry = pmd_modify(oldpmd, newprot);
1923 if (uffd_wp) {
1924 entry = pmd_wrprotect(entry);
1925 entry = pmd_mkuffd_wp(entry);
1926 } else if (uffd_wp_resolve) {
1927 /*
1928 * Leave the write bit to be handled by PF interrupt
1929 * handler, then things like COW could be properly
1930 * handled.
1931 */
1932 entry = pmd_clear_uffd_wp(entry);
1933 }
1934
1935 /* See change_pte_range(). */
1936 if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1937 can_change_pmd_writable(vma, addr, entry))
1938 entry = pmd_mkwrite(entry);
1939
1940 ret = HPAGE_PMD_NR;
1941 set_pmd_at(mm, addr, pmd, entry);
1942
1943 if (huge_pmd_needs_flush(oldpmd, entry))
1944 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1945unlock:
1946 spin_unlock(ptl);
1947 return ret;
1948}
1949
1950/*
1951 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1952 *
1953 * Note that if it returns page table lock pointer, this routine returns without
1954 * unlocking page table lock. So callers must unlock it.
1955 */
1956spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1957{
1958 spinlock_t *ptl;
1959 ptl = pmd_lock(vma->vm_mm, pmd);
1960 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1961 pmd_devmap(*pmd)))
1962 return ptl;
1963 spin_unlock(ptl);
1964 return NULL;
1965}
1966
1967/*
1968 * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1969 *
1970 * Note that if it returns page table lock pointer, this routine returns without
1971 * unlocking page table lock. So callers must unlock it.
1972 */
1973spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1974{
1975 spinlock_t *ptl;
1976
1977 ptl = pud_lock(vma->vm_mm, pud);
1978 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1979 return ptl;
1980 spin_unlock(ptl);
1981 return NULL;
1982}
1983
1984#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1985int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1986 pud_t *pud, unsigned long addr)
1987{
1988 spinlock_t *ptl;
1989
1990 ptl = __pud_trans_huge_lock(pud, vma);
1991 if (!ptl)
1992 return 0;
1993
1994 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1995 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1996 if (vma_is_special_huge(vma)) {
1997 spin_unlock(ptl);
1998 /* No zero page support yet */
1999 } else {
2000 /* No support for anonymous PUD pages yet */
2001 BUG();
2002 }
2003 return 1;
2004}
2005
2006static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2007 unsigned long haddr)
2008{
2009 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2010 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2011 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2012 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2013
2014 count_vm_event(THP_SPLIT_PUD);
2015
2016 pudp_huge_clear_flush_notify(vma, haddr, pud);
2017}
2018
2019void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2020 unsigned long address)
2021{
2022 spinlock_t *ptl;
2023 struct mmu_notifier_range range;
2024
2025 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2026 address & HPAGE_PUD_MASK,
2027 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2028 mmu_notifier_invalidate_range_start(&range);
2029 ptl = pud_lock(vma->vm_mm, pud);
2030 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2031 goto out;
2032 __split_huge_pud_locked(vma, pud, range.start);
2033
2034out:
2035 spin_unlock(ptl);
2036 /*
2037 * No need to double call mmu_notifier->invalidate_range() callback as
2038 * the above pudp_huge_clear_flush_notify() did already call it.
2039 */
2040 mmu_notifier_invalidate_range_only_end(&range);
2041}
2042#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2043
2044static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2045 unsigned long haddr, pmd_t *pmd)
2046{
2047 struct mm_struct *mm = vma->vm_mm;
2048 pgtable_t pgtable;
2049 pmd_t _pmd;
2050 int i;
2051
2052 /*
2053 * Leave pmd empty until pte is filled note that it is fine to delay
2054 * notification until mmu_notifier_invalidate_range_end() as we are
2055 * replacing a zero pmd write protected page with a zero pte write
2056 * protected page.
2057 *
2058 * See Documentation/mm/mmu_notifier.rst
2059 */
2060 pmdp_huge_clear_flush(vma, haddr, pmd);
2061
2062 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2063 pmd_populate(mm, &_pmd, pgtable);
2064
2065 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2066 pte_t *pte, entry;
2067 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2068 entry = pte_mkspecial(entry);
2069 pte = pte_offset_map(&_pmd, haddr);
2070 VM_BUG_ON(!pte_none(*pte));
2071 set_pte_at(mm, haddr, pte, entry);
2072 pte_unmap(pte);
2073 }
2074 smp_wmb(); /* make pte visible before pmd */
2075 pmd_populate(mm, pmd, pgtable);
2076}
2077
2078static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2079 unsigned long haddr, bool freeze)
2080{
2081 struct mm_struct *mm = vma->vm_mm;
2082 struct page *page;
2083 pgtable_t pgtable;
2084 pmd_t old_pmd, _pmd;
2085 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2086 bool anon_exclusive = false, dirty = 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 old_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_special_huge(vma))
2107 return;
2108 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2109 swp_entry_t entry;
2110
2111 entry = pmd_to_swp_entry(old_pmd);
2112 page = pfn_swap_entry_to_page(entry);
2113 } else {
2114 page = pmd_page(old_pmd);
2115 if (!PageDirty(page) && pmd_dirty(old_pmd))
2116 set_page_dirty(page);
2117 if (!PageReferenced(page) && pmd_young(old_pmd))
2118 SetPageReferenced(page);
2119 page_remove_rmap(page, vma, true);
2120 put_page(page);
2121 }
2122 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2123 return;
2124 }
2125
2126 if (is_huge_zero_pmd(*pmd)) {
2127 /*
2128 * FIXME: Do we want to invalidate secondary mmu by calling
2129 * mmu_notifier_invalidate_range() see comments below inside
2130 * __split_huge_pmd() ?
2131 *
2132 * We are going from a zero huge page write protected to zero
2133 * small page also write protected so it does not seems useful
2134 * to invalidate secondary mmu at this time.
2135 */
2136 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2137 }
2138
2139 /*
2140 * Up to this point the pmd is present and huge and userland has the
2141 * whole access to the hugepage during the split (which happens in
2142 * place). If we overwrite the pmd with the not-huge version pointing
2143 * to the pte here (which of course we could if all CPUs were bug
2144 * free), userland could trigger a small page size TLB miss on the
2145 * small sized TLB while the hugepage TLB entry is still established in
2146 * the huge TLB. Some CPU doesn't like that.
2147 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2148 * 383 on page 105. Intel should be safe but is also warns that it's
2149 * only safe if the permission and cache attributes of the two entries
2150 * loaded in the two TLB is identical (which should be the case here).
2151 * But it is generally safer to never allow small and huge TLB entries
2152 * for the same virtual address to be loaded simultaneously. So instead
2153 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2154 * current pmd notpresent (atomically because here the pmd_trans_huge
2155 * must remain set at all times on the pmd until the split is complete
2156 * for this pmd), then we flush the SMP TLB and finally we write the
2157 * non-huge version of the pmd entry with pmd_populate.
2158 */
2159 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2160
2161 pmd_migration = is_pmd_migration_entry(old_pmd);
2162 if (unlikely(pmd_migration)) {
2163 swp_entry_t entry;
2164
2165 entry = pmd_to_swp_entry(old_pmd);
2166 page = pfn_swap_entry_to_page(entry);
2167 write = is_writable_migration_entry(entry);
2168 if (PageAnon(page))
2169 anon_exclusive = is_readable_exclusive_migration_entry(entry);
2170 young = is_migration_entry_young(entry);
2171 dirty = is_migration_entry_dirty(entry);
2172 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2173 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2174 } else {
2175 page = pmd_page(old_pmd);
2176 if (pmd_dirty(old_pmd)) {
2177 dirty = true;
2178 SetPageDirty(page);
2179 }
2180 write = pmd_write(old_pmd);
2181 young = pmd_young(old_pmd);
2182 soft_dirty = pmd_soft_dirty(old_pmd);
2183 uffd_wp = pmd_uffd_wp(old_pmd);
2184
2185 VM_BUG_ON_PAGE(!page_count(page), page);
2186
2187 /*
2188 * Without "freeze", we'll simply split the PMD, propagating the
2189 * PageAnonExclusive() flag for each PTE by setting it for
2190 * each subpage -- no need to (temporarily) clear.
2191 *
2192 * With "freeze" we want to replace mapped pages by
2193 * migration entries right away. This is only possible if we
2194 * managed to clear PageAnonExclusive() -- see
2195 * set_pmd_migration_entry().
2196 *
2197 * In case we cannot clear PageAnonExclusive(), split the PMD
2198 * only and let try_to_migrate_one() fail later.
2199 *
2200 * See page_try_share_anon_rmap(): invalidate PMD first.
2201 */
2202 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2203 if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2204 freeze = false;
2205 if (!freeze)
2206 page_ref_add(page, HPAGE_PMD_NR - 1);
2207 }
2208
2209 /*
2210 * Withdraw the table only after we mark the pmd entry invalid.
2211 * This's critical for some architectures (Power).
2212 */
2213 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2214 pmd_populate(mm, &_pmd, pgtable);
2215
2216 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2217 pte_t entry, *pte;
2218 /*
2219 * Note that NUMA hinting access restrictions are not
2220 * transferred to avoid any possibility of altering
2221 * permissions across VMAs.
2222 */
2223 if (freeze || pmd_migration) {
2224 swp_entry_t swp_entry;
2225 if (write)
2226 swp_entry = make_writable_migration_entry(
2227 page_to_pfn(page + i));
2228 else if (anon_exclusive)
2229 swp_entry = make_readable_exclusive_migration_entry(
2230 page_to_pfn(page + i));
2231 else
2232 swp_entry = make_readable_migration_entry(
2233 page_to_pfn(page + i));
2234 if (young)
2235 swp_entry = make_migration_entry_young(swp_entry);
2236 if (dirty)
2237 swp_entry = make_migration_entry_dirty(swp_entry);
2238 entry = swp_entry_to_pte(swp_entry);
2239 if (soft_dirty)
2240 entry = pte_swp_mksoft_dirty(entry);
2241 if (uffd_wp)
2242 entry = pte_swp_mkuffd_wp(entry);
2243 } else {
2244 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2245 entry = maybe_mkwrite(entry, vma);
2246 if (anon_exclusive)
2247 SetPageAnonExclusive(page + i);
2248 if (!young)
2249 entry = pte_mkold(entry);
2250 /* NOTE: this may set soft-dirty too on some archs */
2251 if (dirty)
2252 entry = pte_mkdirty(entry);
2253 /*
2254 * NOTE: this needs to happen after pte_mkdirty,
2255 * because some archs (sparc64, loongarch) could
2256 * set hw write bit when mkdirty.
2257 */
2258 if (!write)
2259 entry = pte_wrprotect(entry);
2260 if (soft_dirty)
2261 entry = pte_mksoft_dirty(entry);
2262 if (uffd_wp)
2263 entry = pte_mkuffd_wp(entry);
2264 page_add_anon_rmap(page + i, vma, addr, false);
2265 }
2266 pte = pte_offset_map(&_pmd, addr);
2267 BUG_ON(!pte_none(*pte));
2268 set_pte_at(mm, addr, pte, entry);
2269 pte_unmap(pte);
2270 }
2271
2272 if (!pmd_migration)
2273 page_remove_rmap(page, vma, true);
2274 if (freeze)
2275 put_page(page);
2276
2277 smp_wmb(); /* make pte visible before pmd */
2278 pmd_populate(mm, pmd, pgtable);
2279}
2280
2281void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2282 unsigned long address, bool freeze, struct folio *folio)
2283{
2284 spinlock_t *ptl;
2285 struct mmu_notifier_range range;
2286
2287 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2288 address & HPAGE_PMD_MASK,
2289 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2290 mmu_notifier_invalidate_range_start(&range);
2291 ptl = pmd_lock(vma->vm_mm, pmd);
2292
2293 /*
2294 * If caller asks to setup a migration entry, we need a folio to check
2295 * pmd against. Otherwise we can end up replacing wrong folio.
2296 */
2297 VM_BUG_ON(freeze && !folio);
2298 VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2299
2300 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2301 is_pmd_migration_entry(*pmd)) {
2302 /*
2303 * It's safe to call pmd_page when folio is set because it's
2304 * guaranteed that pmd is present.
2305 */
2306 if (folio && folio != page_folio(pmd_page(*pmd)))
2307 goto out;
2308 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2309 }
2310
2311out:
2312 spin_unlock(ptl);
2313 /*
2314 * No need to double call mmu_notifier->invalidate_range() callback.
2315 * They are 3 cases to consider inside __split_huge_pmd_locked():
2316 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2317 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2318 * fault will trigger a flush_notify before pointing to a new page
2319 * (it is fine if the secondary mmu keeps pointing to the old zero
2320 * page in the meantime)
2321 * 3) Split a huge pmd into pte pointing to the same page. No need
2322 * to invalidate secondary tlb entry they are all still valid.
2323 * any further changes to individual pte will notify. So no need
2324 * to call mmu_notifier->invalidate_range()
2325 */
2326 mmu_notifier_invalidate_range_only_end(&range);
2327}
2328
2329void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2330 bool freeze, struct folio *folio)
2331{
2332 pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2333
2334 if (!pmd)
2335 return;
2336
2337 __split_huge_pmd(vma, pmd, address, freeze, folio);
2338}
2339
2340static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2341{
2342 /*
2343 * If the new address isn't hpage aligned and it could previously
2344 * contain an hugepage: check if we need to split an huge pmd.
2345 */
2346 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2347 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2348 ALIGN(address, HPAGE_PMD_SIZE)))
2349 split_huge_pmd_address(vma, address, false, NULL);
2350}
2351
2352void vma_adjust_trans_huge(struct vm_area_struct *vma,
2353 unsigned long start,
2354 unsigned long end,
2355 long adjust_next)
2356{
2357 /* Check if we need to split start first. */
2358 split_huge_pmd_if_needed(vma, start);
2359
2360 /* Check if we need to split end next. */
2361 split_huge_pmd_if_needed(vma, end);
2362
2363 /*
2364 * If we're also updating the next vma vm_start,
2365 * check if we need to split it.
2366 */
2367 if (adjust_next > 0) {
2368 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2369 unsigned long nstart = next->vm_start;
2370 nstart += adjust_next;
2371 split_huge_pmd_if_needed(next, nstart);
2372 }
2373}
2374
2375static void unmap_folio(struct folio *folio)
2376{
2377 enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2378 TTU_SYNC;
2379
2380 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2381
2382 /*
2383 * Anon pages need migration entries to preserve them, but file
2384 * pages can simply be left unmapped, then faulted back on demand.
2385 * If that is ever changed (perhaps for mlock), update remap_page().
2386 */
2387 if (folio_test_anon(folio))
2388 try_to_migrate(folio, ttu_flags);
2389 else
2390 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2391}
2392
2393static void remap_page(struct folio *folio, unsigned long nr)
2394{
2395 int i = 0;
2396
2397 /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2398 if (!folio_test_anon(folio))
2399 return;
2400 for (;;) {
2401 remove_migration_ptes(folio, folio, true);
2402 i += folio_nr_pages(folio);
2403 if (i >= nr)
2404 break;
2405 folio = folio_next(folio);
2406 }
2407}
2408
2409static void lru_add_page_tail(struct page *head, struct page *tail,
2410 struct lruvec *lruvec, struct list_head *list)
2411{
2412 VM_BUG_ON_PAGE(!PageHead(head), head);
2413 VM_BUG_ON_PAGE(PageCompound(tail), head);
2414 VM_BUG_ON_PAGE(PageLRU(tail), head);
2415 lockdep_assert_held(&lruvec->lru_lock);
2416
2417 if (list) {
2418 /* page reclaim is reclaiming a huge page */
2419 VM_WARN_ON(PageLRU(head));
2420 get_page(tail);
2421 list_add_tail(&tail->lru, list);
2422 } else {
2423 /* head is still on lru (and we have it frozen) */
2424 VM_WARN_ON(!PageLRU(head));
2425 if (PageUnevictable(tail))
2426 tail->mlock_count = 0;
2427 else
2428 list_add_tail(&tail->lru, &head->lru);
2429 SetPageLRU(tail);
2430 }
2431}
2432
2433static void __split_huge_page_tail(struct page *head, int tail,
2434 struct lruvec *lruvec, struct list_head *list)
2435{
2436 struct page *page_tail = head + tail;
2437
2438 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2439
2440 /*
2441 * Clone page flags before unfreezing refcount.
2442 *
2443 * After successful get_page_unless_zero() might follow flags change,
2444 * for example lock_page() which set PG_waiters.
2445 *
2446 * Note that for mapped sub-pages of an anonymous THP,
2447 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2448 * the migration entry instead from where remap_page() will restore it.
2449 * We can still have PG_anon_exclusive set on effectively unmapped and
2450 * unreferenced sub-pages of an anonymous THP: we can simply drop
2451 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2452 */
2453 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2454 page_tail->flags |= (head->flags &
2455 ((1L << PG_referenced) |
2456 (1L << PG_swapbacked) |
2457 (1L << PG_swapcache) |
2458 (1L << PG_mlocked) |
2459 (1L << PG_uptodate) |
2460 (1L << PG_active) |
2461 (1L << PG_workingset) |
2462 (1L << PG_locked) |
2463 (1L << PG_unevictable) |
2464#ifdef CONFIG_ARCH_USES_PG_ARCH_X
2465 (1L << PG_arch_2) |
2466 (1L << PG_arch_3) |
2467#endif
2468 (1L << PG_dirty) |
2469 LRU_GEN_MASK | LRU_REFS_MASK));
2470
2471 /* ->mapping in first and second tail page is replaced by other uses */
2472 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2473 page_tail);
2474 page_tail->mapping = head->mapping;
2475 page_tail->index = head->index + tail;
2476
2477 /*
2478 * page->private should not be set in tail pages with the exception
2479 * of swap cache pages that store the swp_entry_t in tail pages.
2480 * Fix up and warn once if private is unexpectedly set.
2481 *
2482 * What of 32-bit systems, on which head[1].compound_pincount overlays
2483 * head[1].private? No problem: THP_SWAP is not enabled on 32-bit, and
2484 * compound_pincount must be 0 for folio_ref_freeze() to have succeeded.
2485 */
2486 if (!folio_test_swapcache(page_folio(head))) {
2487 VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2488 page_tail->private = 0;
2489 }
2490
2491 /* Page flags must be visible before we make the page non-compound. */
2492 smp_wmb();
2493
2494 /*
2495 * Clear PageTail before unfreezing page refcount.
2496 *
2497 * After successful get_page_unless_zero() might follow put_page()
2498 * which needs correct compound_head().
2499 */
2500 clear_compound_head(page_tail);
2501
2502 /* Finally unfreeze refcount. Additional reference from page cache. */
2503 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2504 PageSwapCache(head)));
2505
2506 if (page_is_young(head))
2507 set_page_young(page_tail);
2508 if (page_is_idle(head))
2509 set_page_idle(page_tail);
2510
2511 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2512
2513 /*
2514 * always add to the tail because some iterators expect new
2515 * pages to show after the currently processed elements - e.g.
2516 * migrate_pages
2517 */
2518 lru_add_page_tail(head, page_tail, lruvec, list);
2519}
2520
2521static void __split_huge_page(struct page *page, struct list_head *list,
2522 pgoff_t end)
2523{
2524 struct folio *folio = page_folio(page);
2525 struct page *head = &folio->page;
2526 struct lruvec *lruvec;
2527 struct address_space *swap_cache = NULL;
2528 unsigned long offset = 0;
2529 unsigned int nr = thp_nr_pages(head);
2530 int i;
2531
2532 /* complete memcg works before add pages to LRU */
2533 split_page_memcg(head, nr);
2534
2535 if (PageAnon(head) && PageSwapCache(head)) {
2536 swp_entry_t entry = { .val = page_private(head) };
2537
2538 offset = swp_offset(entry);
2539 swap_cache = swap_address_space(entry);
2540 xa_lock(&swap_cache->i_pages);
2541 }
2542
2543 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2544 lruvec = folio_lruvec_lock(folio);
2545
2546 ClearPageHasHWPoisoned(head);
2547
2548 for (i = nr - 1; i >= 1; i--) {
2549 __split_huge_page_tail(head, i, lruvec, list);
2550 /* Some pages can be beyond EOF: drop them from page cache */
2551 if (head[i].index >= end) {
2552 struct folio *tail = page_folio(head + i);
2553
2554 if (shmem_mapping(head->mapping))
2555 shmem_uncharge(head->mapping->host, 1);
2556 else if (folio_test_clear_dirty(tail))
2557 folio_account_cleaned(tail,
2558 inode_to_wb(folio->mapping->host));
2559 __filemap_remove_folio(tail, NULL);
2560 folio_put(tail);
2561 } else if (!PageAnon(page)) {
2562 __xa_store(&head->mapping->i_pages, head[i].index,
2563 head + i, 0);
2564 } else if (swap_cache) {
2565 __xa_store(&swap_cache->i_pages, offset + i,
2566 head + i, 0);
2567 }
2568 }
2569
2570 ClearPageCompound(head);
2571 unlock_page_lruvec(lruvec);
2572 /* Caller disabled irqs, so they are still disabled here */
2573
2574 split_page_owner(head, nr);
2575
2576 /* See comment in __split_huge_page_tail() */
2577 if (PageAnon(head)) {
2578 /* Additional pin to swap cache */
2579 if (PageSwapCache(head)) {
2580 page_ref_add(head, 2);
2581 xa_unlock(&swap_cache->i_pages);
2582 } else {
2583 page_ref_inc(head);
2584 }
2585 } else {
2586 /* Additional pin to page cache */
2587 page_ref_add(head, 2);
2588 xa_unlock(&head->mapping->i_pages);
2589 }
2590 local_irq_enable();
2591
2592 remap_page(folio, nr);
2593
2594 if (PageSwapCache(head)) {
2595 swp_entry_t entry = { .val = page_private(head) };
2596
2597 split_swap_cluster(entry);
2598 }
2599
2600 for (i = 0; i < nr; i++) {
2601 struct page *subpage = head + i;
2602 if (subpage == page)
2603 continue;
2604 unlock_page(subpage);
2605
2606 /*
2607 * Subpages may be freed if there wasn't any mapping
2608 * like if add_to_swap() is running on a lru page that
2609 * had its mapping zapped. And freeing these pages
2610 * requires taking the lru_lock so we do the put_page
2611 * of the tail pages after the split is complete.
2612 */
2613 free_page_and_swap_cache(subpage);
2614 }
2615}
2616
2617/* Racy check whether the huge page can be split */
2618bool can_split_folio(struct folio *folio, int *pextra_pins)
2619{
2620 int extra_pins;
2621
2622 /* Additional pins from page cache */
2623 if (folio_test_anon(folio))
2624 extra_pins = folio_test_swapcache(folio) ?
2625 folio_nr_pages(folio) : 0;
2626 else
2627 extra_pins = folio_nr_pages(folio);
2628 if (pextra_pins)
2629 *pextra_pins = extra_pins;
2630 return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2631}
2632
2633/*
2634 * This function splits huge page into normal pages. @page can point to any
2635 * subpage of huge page to split. Split doesn't change the position of @page.
2636 *
2637 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2638 * The huge page must be locked.
2639 *
2640 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2641 *
2642 * Both head page and tail pages will inherit mapping, flags, and so on from
2643 * the hugepage.
2644 *
2645 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2646 * they are not mapped.
2647 *
2648 * Returns 0 if the hugepage is split successfully.
2649 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2650 * us.
2651 */
2652int split_huge_page_to_list(struct page *page, struct list_head *list)
2653{
2654 struct folio *folio = page_folio(page);
2655 struct deferred_split *ds_queue = get_deferred_split_queue(&folio->page);
2656 XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2657 struct anon_vma *anon_vma = NULL;
2658 struct address_space *mapping = NULL;
2659 int extra_pins, ret;
2660 pgoff_t end;
2661 bool is_hzp;
2662
2663 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2664 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2665
2666 is_hzp = is_huge_zero_page(&folio->page);
2667 VM_WARN_ON_ONCE_FOLIO(is_hzp, folio);
2668 if (is_hzp)
2669 return -EBUSY;
2670
2671 if (folio_test_writeback(folio))
2672 return -EBUSY;
2673
2674 if (folio_test_anon(folio)) {
2675 /*
2676 * The caller does not necessarily hold an mmap_lock that would
2677 * prevent the anon_vma disappearing so we first we take a
2678 * reference to it and then lock the anon_vma for write. This
2679 * is similar to folio_lock_anon_vma_read except the write lock
2680 * is taken to serialise against parallel split or collapse
2681 * operations.
2682 */
2683 anon_vma = folio_get_anon_vma(folio);
2684 if (!anon_vma) {
2685 ret = -EBUSY;
2686 goto out;
2687 }
2688 end = -1;
2689 mapping = NULL;
2690 anon_vma_lock_write(anon_vma);
2691 } else {
2692 gfp_t gfp;
2693
2694 mapping = folio->mapping;
2695
2696 /* Truncated ? */
2697 if (!mapping) {
2698 ret = -EBUSY;
2699 goto out;
2700 }
2701
2702 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2703 GFP_RECLAIM_MASK);
2704
2705 if (folio_test_private(folio) &&
2706 !filemap_release_folio(folio, gfp)) {
2707 ret = -EBUSY;
2708 goto out;
2709 }
2710
2711 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2712 if (xas_error(&xas)) {
2713 ret = xas_error(&xas);
2714 goto out;
2715 }
2716
2717 anon_vma = NULL;
2718 i_mmap_lock_read(mapping);
2719
2720 /*
2721 *__split_huge_page() may need to trim off pages beyond EOF:
2722 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2723 * which cannot be nested inside the page tree lock. So note
2724 * end now: i_size itself may be changed at any moment, but
2725 * folio lock is good enough to serialize the trimming.
2726 */
2727 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2728 if (shmem_mapping(mapping))
2729 end = shmem_fallocend(mapping->host, end);
2730 }
2731
2732 /*
2733 * Racy check if we can split the page, before unmap_folio() will
2734 * split PMDs
2735 */
2736 if (!can_split_folio(folio, &extra_pins)) {
2737 ret = -EAGAIN;
2738 goto out_unlock;
2739 }
2740
2741 unmap_folio(folio);
2742
2743 /* block interrupt reentry in xa_lock and spinlock */
2744 local_irq_disable();
2745 if (mapping) {
2746 /*
2747 * Check if the folio is present in page cache.
2748 * We assume all tail are present too, if folio is there.
2749 */
2750 xas_lock(&xas);
2751 xas_reset(&xas);
2752 if (xas_load(&xas) != folio)
2753 goto fail;
2754 }
2755
2756 /* Prevent deferred_split_scan() touching ->_refcount */
2757 spin_lock(&ds_queue->split_queue_lock);
2758 if (folio_ref_freeze(folio, 1 + extra_pins)) {
2759 if (!list_empty(page_deferred_list(&folio->page))) {
2760 ds_queue->split_queue_len--;
2761 list_del(page_deferred_list(&folio->page));
2762 }
2763 spin_unlock(&ds_queue->split_queue_lock);
2764 if (mapping) {
2765 int nr = folio_nr_pages(folio);
2766
2767 xas_split(&xas, folio, folio_order(folio));
2768 if (folio_test_swapbacked(folio)) {
2769 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2770 -nr);
2771 } else {
2772 __lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2773 -nr);
2774 filemap_nr_thps_dec(mapping);
2775 }
2776 }
2777
2778 __split_huge_page(page, list, end);
2779 ret = 0;
2780 } else {
2781 spin_unlock(&ds_queue->split_queue_lock);
2782fail:
2783 if (mapping)
2784 xas_unlock(&xas);
2785 local_irq_enable();
2786 remap_page(folio, folio_nr_pages(folio));
2787 ret = -EAGAIN;
2788 }
2789
2790out_unlock:
2791 if (anon_vma) {
2792 anon_vma_unlock_write(anon_vma);
2793 put_anon_vma(anon_vma);
2794 }
2795 if (mapping)
2796 i_mmap_unlock_read(mapping);
2797out:
2798 xas_destroy(&xas);
2799 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2800 return ret;
2801}
2802
2803void free_transhuge_page(struct page *page)
2804{
2805 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2806 unsigned long flags;
2807
2808 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2809 if (!list_empty(page_deferred_list(page))) {
2810 ds_queue->split_queue_len--;
2811 list_del(page_deferred_list(page));
2812 }
2813 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2814 free_compound_page(page);
2815}
2816
2817void deferred_split_huge_page(struct page *page)
2818{
2819 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2820#ifdef CONFIG_MEMCG
2821 struct mem_cgroup *memcg = page_memcg(compound_head(page));
2822#endif
2823 unsigned long flags;
2824
2825 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2826
2827 /*
2828 * The try_to_unmap() in page reclaim path might reach here too,
2829 * this may cause a race condition to corrupt deferred split queue.
2830 * And, if page reclaim is already handling the same page, it is
2831 * unnecessary to handle it again in shrinker.
2832 *
2833 * Check PageSwapCache to determine if the page is being
2834 * handled by page reclaim since THP swap would add the page into
2835 * swap cache before calling try_to_unmap().
2836 */
2837 if (PageSwapCache(page))
2838 return;
2839
2840 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2841 if (list_empty(page_deferred_list(page))) {
2842 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2843 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2844 ds_queue->split_queue_len++;
2845#ifdef CONFIG_MEMCG
2846 if (memcg)
2847 set_shrinker_bit(memcg, page_to_nid(page),
2848 deferred_split_shrinker.id);
2849#endif
2850 }
2851 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2852}
2853
2854static unsigned long deferred_split_count(struct shrinker *shrink,
2855 struct shrink_control *sc)
2856{
2857 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2858 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2859
2860#ifdef CONFIG_MEMCG
2861 if (sc->memcg)
2862 ds_queue = &sc->memcg->deferred_split_queue;
2863#endif
2864 return READ_ONCE(ds_queue->split_queue_len);
2865}
2866
2867static unsigned long deferred_split_scan(struct shrinker *shrink,
2868 struct shrink_control *sc)
2869{
2870 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2871 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2872 unsigned long flags;
2873 LIST_HEAD(list), *pos, *next;
2874 struct page *page;
2875 int split = 0;
2876
2877#ifdef CONFIG_MEMCG
2878 if (sc->memcg)
2879 ds_queue = &sc->memcg->deferred_split_queue;
2880#endif
2881
2882 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2883 /* Take pin on all head pages to avoid freeing them under us */
2884 list_for_each_safe(pos, next, &ds_queue->split_queue) {
2885 page = list_entry((void *)pos, struct page, deferred_list);
2886 page = compound_head(page);
2887 if (get_page_unless_zero(page)) {
2888 list_move(page_deferred_list(page), &list);
2889 } else {
2890 /* We lost race with put_compound_page() */
2891 list_del_init(page_deferred_list(page));
2892 ds_queue->split_queue_len--;
2893 }
2894 if (!--sc->nr_to_scan)
2895 break;
2896 }
2897 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2898
2899 list_for_each_safe(pos, next, &list) {
2900 page = list_entry((void *)pos, struct page, deferred_list);
2901 if (!trylock_page(page))
2902 goto next;
2903 /* split_huge_page() removes page from list on success */
2904 if (!split_huge_page(page))
2905 split++;
2906 unlock_page(page);
2907next:
2908 put_page(page);
2909 }
2910
2911 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2912 list_splice_tail(&list, &ds_queue->split_queue);
2913 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2914
2915 /*
2916 * Stop shrinker if we didn't split any page, but the queue is empty.
2917 * This can happen if pages were freed under us.
2918 */
2919 if (!split && list_empty(&ds_queue->split_queue))
2920 return SHRINK_STOP;
2921 return split;
2922}
2923
2924static struct shrinker deferred_split_shrinker = {
2925 .count_objects = deferred_split_count,
2926 .scan_objects = deferred_split_scan,
2927 .seeks = DEFAULT_SEEKS,
2928 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2929 SHRINKER_NONSLAB,
2930};
2931
2932#ifdef CONFIG_DEBUG_FS
2933static void split_huge_pages_all(void)
2934{
2935 struct zone *zone;
2936 struct page *page;
2937 unsigned long pfn, max_zone_pfn;
2938 unsigned long total = 0, split = 0;
2939
2940 pr_debug("Split all THPs\n");
2941 for_each_zone(zone) {
2942 if (!managed_zone(zone))
2943 continue;
2944 max_zone_pfn = zone_end_pfn(zone);
2945 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2946 int nr_pages;
2947
2948 page = pfn_to_online_page(pfn);
2949 if (!page || !get_page_unless_zero(page))
2950 continue;
2951
2952 if (zone != page_zone(page))
2953 goto next;
2954
2955 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2956 goto next;
2957
2958 total++;
2959 lock_page(page);
2960 nr_pages = thp_nr_pages(page);
2961 if (!split_huge_page(page))
2962 split++;
2963 pfn += nr_pages - 1;
2964 unlock_page(page);
2965next:
2966 put_page(page);
2967 cond_resched();
2968 }
2969 }
2970
2971 pr_debug("%lu of %lu THP split\n", split, total);
2972}
2973
2974static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2975{
2976 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2977 is_vm_hugetlb_page(vma);
2978}
2979
2980static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2981 unsigned long vaddr_end)
2982{
2983 int ret = 0;
2984 struct task_struct *task;
2985 struct mm_struct *mm;
2986 unsigned long total = 0, split = 0;
2987 unsigned long addr;
2988
2989 vaddr_start &= PAGE_MASK;
2990 vaddr_end &= PAGE_MASK;
2991
2992 /* Find the task_struct from pid */
2993 rcu_read_lock();
2994 task = find_task_by_vpid(pid);
2995 if (!task) {
2996 rcu_read_unlock();
2997 ret = -ESRCH;
2998 goto out;
2999 }
3000 get_task_struct(task);
3001 rcu_read_unlock();
3002
3003 /* Find the mm_struct */
3004 mm = get_task_mm(task);
3005 put_task_struct(task);
3006
3007 if (!mm) {
3008 ret = -EINVAL;
3009 goto out;
3010 }
3011
3012 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3013 pid, vaddr_start, vaddr_end);
3014
3015 mmap_read_lock(mm);
3016 /*
3017 * always increase addr by PAGE_SIZE, since we could have a PTE page
3018 * table filled with PTE-mapped THPs, each of which is distinct.
3019 */
3020 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3021 struct vm_area_struct *vma = vma_lookup(mm, addr);
3022 struct page *page;
3023
3024 if (!vma)
3025 break;
3026
3027 /* skip special VMA and hugetlb VMA */
3028 if (vma_not_suitable_for_thp_split(vma)) {
3029 addr = vma->vm_end;
3030 continue;
3031 }
3032
3033 /* FOLL_DUMP to ignore special (like zero) pages */
3034 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3035
3036 if (IS_ERR_OR_NULL(page))
3037 continue;
3038
3039 if (!is_transparent_hugepage(page))
3040 goto next;
3041
3042 total++;
3043 if (!can_split_folio(page_folio(page), NULL))
3044 goto next;
3045
3046 if (!trylock_page(page))
3047 goto next;
3048
3049 if (!split_huge_page(page))
3050 split++;
3051
3052 unlock_page(page);
3053next:
3054 put_page(page);
3055 cond_resched();
3056 }
3057 mmap_read_unlock(mm);
3058 mmput(mm);
3059
3060 pr_debug("%lu of %lu THP split\n", split, total);
3061
3062out:
3063 return ret;
3064}
3065
3066static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3067 pgoff_t off_end)
3068{
3069 struct filename *file;
3070 struct file *candidate;
3071 struct address_space *mapping;
3072 int ret = -EINVAL;
3073 pgoff_t index;
3074 int nr_pages = 1;
3075 unsigned long total = 0, split = 0;
3076
3077 file = getname_kernel(file_path);
3078 if (IS_ERR(file))
3079 return ret;
3080
3081 candidate = file_open_name(file, O_RDONLY, 0);
3082 if (IS_ERR(candidate))
3083 goto out;
3084
3085 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3086 file_path, off_start, off_end);
3087
3088 mapping = candidate->f_mapping;
3089
3090 for (index = off_start; index < off_end; index += nr_pages) {
3091 struct folio *folio = __filemap_get_folio(mapping, index,
3092 FGP_ENTRY, 0);
3093
3094 nr_pages = 1;
3095 if (xa_is_value(folio) || !folio)
3096 continue;
3097
3098 if (!folio_test_large(folio))
3099 goto next;
3100
3101 total++;
3102 nr_pages = folio_nr_pages(folio);
3103
3104 if (!folio_trylock(folio))
3105 goto next;
3106
3107 if (!split_folio(folio))
3108 split++;
3109
3110 folio_unlock(folio);
3111next:
3112 folio_put(folio);
3113 cond_resched();
3114 }
3115
3116 filp_close(candidate, NULL);
3117 ret = 0;
3118
3119 pr_debug("%lu of %lu file-backed THP split\n", split, total);
3120out:
3121 putname(file);
3122 return ret;
3123}
3124
3125#define MAX_INPUT_BUF_SZ 255
3126
3127static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3128 size_t count, loff_t *ppops)
3129{
3130 static DEFINE_MUTEX(split_debug_mutex);
3131 ssize_t ret;
3132 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3133 char input_buf[MAX_INPUT_BUF_SZ];
3134 int pid;
3135 unsigned long vaddr_start, vaddr_end;
3136
3137 ret = mutex_lock_interruptible(&split_debug_mutex);
3138 if (ret)
3139 return ret;
3140
3141 ret = -EFAULT;
3142
3143 memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3144 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3145 goto out;
3146
3147 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3148
3149 if (input_buf[0] == '/') {
3150 char *tok;
3151 char *buf = input_buf;
3152 char file_path[MAX_INPUT_BUF_SZ];
3153 pgoff_t off_start = 0, off_end = 0;
3154 size_t input_len = strlen(input_buf);
3155
3156 tok = strsep(&buf, ",");
3157 if (tok) {
3158 strcpy(file_path, tok);
3159 } else {
3160 ret = -EINVAL;
3161 goto out;
3162 }
3163
3164 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3165 if (ret != 2) {
3166 ret = -EINVAL;
3167 goto out;
3168 }
3169 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3170 if (!ret)
3171 ret = input_len;
3172
3173 goto out;
3174 }
3175
3176 ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3177 if (ret == 1 && pid == 1) {
3178 split_huge_pages_all();
3179 ret = strlen(input_buf);
3180 goto out;
3181 } else if (ret != 3) {
3182 ret = -EINVAL;
3183 goto out;
3184 }
3185
3186 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3187 if (!ret)
3188 ret = strlen(input_buf);
3189out:
3190 mutex_unlock(&split_debug_mutex);
3191 return ret;
3192
3193}
3194
3195static const struct file_operations split_huge_pages_fops = {
3196 .owner = THIS_MODULE,
3197 .write = split_huge_pages_write,
3198 .llseek = no_llseek,
3199};
3200
3201static int __init split_huge_pages_debugfs(void)
3202{
3203 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3204 &split_huge_pages_fops);
3205 return 0;
3206}
3207late_initcall(split_huge_pages_debugfs);
3208#endif
3209
3210#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3211int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3212 struct page *page)
3213{
3214 struct vm_area_struct *vma = pvmw->vma;
3215 struct mm_struct *mm = vma->vm_mm;
3216 unsigned long address = pvmw->address;
3217 bool anon_exclusive;
3218 pmd_t pmdval;
3219 swp_entry_t entry;
3220 pmd_t pmdswp;
3221
3222 if (!(pvmw->pmd && !pvmw->pte))
3223 return 0;
3224
3225 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3226 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3227
3228 /* See page_try_share_anon_rmap(): invalidate PMD first. */
3229 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3230 if (anon_exclusive && page_try_share_anon_rmap(page)) {
3231 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3232 return -EBUSY;
3233 }
3234
3235 if (pmd_dirty(pmdval))
3236 set_page_dirty(page);
3237 if (pmd_write(pmdval))
3238 entry = make_writable_migration_entry(page_to_pfn(page));
3239 else if (anon_exclusive)
3240 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3241 else
3242 entry = make_readable_migration_entry(page_to_pfn(page));
3243 if (pmd_young(pmdval))
3244 entry = make_migration_entry_young(entry);
3245 if (pmd_dirty(pmdval))
3246 entry = make_migration_entry_dirty(entry);
3247 pmdswp = swp_entry_to_pmd(entry);
3248 if (pmd_soft_dirty(pmdval))
3249 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3250 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3251 page_remove_rmap(page, vma, true);
3252 put_page(page);
3253 trace_set_migration_pmd(address, pmd_val(pmdswp));
3254
3255 return 0;
3256}
3257
3258void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3259{
3260 struct vm_area_struct *vma = pvmw->vma;
3261 struct mm_struct *mm = vma->vm_mm;
3262 unsigned long address = pvmw->address;
3263 unsigned long haddr = address & HPAGE_PMD_MASK;
3264 pmd_t pmde;
3265 swp_entry_t entry;
3266
3267 if (!(pvmw->pmd && !pvmw->pte))
3268 return;
3269
3270 entry = pmd_to_swp_entry(*pvmw->pmd);
3271 get_page(new);
3272 pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3273 if (pmd_swp_soft_dirty(*pvmw->pmd))
3274 pmde = pmd_mksoft_dirty(pmde);
3275 if (pmd_swp_uffd_wp(*pvmw->pmd))
3276 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3277 if (!is_migration_entry_young(entry))
3278 pmde = pmd_mkold(pmde);
3279 /* NOTE: this may contain setting soft-dirty on some archs */
3280 if (PageDirty(new) && is_migration_entry_dirty(entry))
3281 pmde = pmd_mkdirty(pmde);
3282 if (is_writable_migration_entry(entry))
3283 pmde = maybe_pmd_mkwrite(pmde, vma);
3284 else
3285 pmde = pmd_wrprotect(pmde);
3286
3287 if (PageAnon(new)) {
3288 rmap_t rmap_flags = RMAP_COMPOUND;
3289
3290 if (!is_readable_migration_entry(entry))
3291 rmap_flags |= RMAP_EXCLUSIVE;
3292
3293 page_add_anon_rmap(new, vma, haddr, rmap_flags);
3294 } else {
3295 page_add_file_rmap(new, vma, true);
3296 }
3297 VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3298 set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3299
3300 /* No need to invalidate - it was non-present before */
3301 update_mmu_cache_pmd(vma, address, pvmw->pmd);
3302 trace_remove_migration_pmd(address, pmd_val(pmde));
3303}
3304#endif