Loading...
1/*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/mm.h>
11#include <linux/sched.h>
12#include <linux/sched/coredump.h>
13#include <linux/sched/numa_balancing.h>
14#include <linux/highmem.h>
15#include <linux/hugetlb.h>
16#include <linux/mmu_notifier.h>
17#include <linux/rmap.h>
18#include <linux/swap.h>
19#include <linux/shrinker.h>
20#include <linux/mm_inline.h>
21#include <linux/swapops.h>
22#include <linux/dax.h>
23#include <linux/khugepaged.h>
24#include <linux/freezer.h>
25#include <linux/pfn_t.h>
26#include <linux/mman.h>
27#include <linux/memremap.h>
28#include <linux/pagemap.h>
29#include <linux/debugfs.h>
30#include <linux/migrate.h>
31#include <linux/hashtable.h>
32#include <linux/userfaultfd_k.h>
33#include <linux/page_idle.h>
34#include <linux/shmem_fs.h>
35#include <linux/oom.h>
36
37#include <asm/tlb.h>
38#include <asm/pgalloc.h>
39#include "internal.h"
40
41/*
42 * By default, transparent hugepage support is disabled in order to avoid
43 * risking an increased memory footprint for applications that are not
44 * guaranteed to benefit from it. When transparent hugepage support is
45 * enabled, it is for all mappings, and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
48 */
49unsigned long transparent_hugepage_flags __read_mostly =
50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
52#endif
53#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
55#endif
56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
59
60static struct shrinker deferred_split_shrinker;
61
62static atomic_t huge_zero_refcount;
63struct page *huge_zero_page __read_mostly;
64
65static struct page *get_huge_zero_page(void)
66{
67 struct page *zero_page;
68retry:
69 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
70 return READ_ONCE(huge_zero_page);
71
72 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
73 HPAGE_PMD_ORDER);
74 if (!zero_page) {
75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
76 return NULL;
77 }
78 count_vm_event(THP_ZERO_PAGE_ALLOC);
79 preempt_disable();
80 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
81 preempt_enable();
82 __free_pages(zero_page, compound_order(zero_page));
83 goto retry;
84 }
85
86 /* We take additional reference here. It will be put back by shrinker */
87 atomic_set(&huge_zero_refcount, 2);
88 preempt_enable();
89 return READ_ONCE(huge_zero_page);
90}
91
92static void put_huge_zero_page(void)
93{
94 /*
95 * Counter should never go to zero here. Only shrinker can put
96 * last reference.
97 */
98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
99}
100
101struct page *mm_get_huge_zero_page(struct mm_struct *mm)
102{
103 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
104 return READ_ONCE(huge_zero_page);
105
106 if (!get_huge_zero_page())
107 return NULL;
108
109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
110 put_huge_zero_page();
111
112 return READ_ONCE(huge_zero_page);
113}
114
115void mm_put_huge_zero_page(struct mm_struct *mm)
116{
117 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
118 put_huge_zero_page();
119}
120
121static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
122 struct shrink_control *sc)
123{
124 /* we can free zero page only if last reference remains */
125 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
126}
127
128static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
129 struct shrink_control *sc)
130{
131 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
132 struct page *zero_page = xchg(&huge_zero_page, NULL);
133 BUG_ON(zero_page == NULL);
134 __free_pages(zero_page, compound_order(zero_page));
135 return HPAGE_PMD_NR;
136 }
137
138 return 0;
139}
140
141static struct shrinker huge_zero_page_shrinker = {
142 .count_objects = shrink_huge_zero_page_count,
143 .scan_objects = shrink_huge_zero_page_scan,
144 .seeks = DEFAULT_SEEKS,
145};
146
147#ifdef CONFIG_SYSFS
148static ssize_t enabled_show(struct kobject *kobj,
149 struct kobj_attribute *attr, char *buf)
150{
151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
152 return sprintf(buf, "[always] madvise never\n");
153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
154 return sprintf(buf, "always [madvise] never\n");
155 else
156 return sprintf(buf, "always madvise [never]\n");
157}
158
159static ssize_t enabled_store(struct kobject *kobj,
160 struct kobj_attribute *attr,
161 const char *buf, size_t count)
162{
163 ssize_t ret = count;
164
165 if (!memcmp("always", buf,
166 min(sizeof("always")-1, count))) {
167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
168 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
169 } else if (!memcmp("madvise", buf,
170 min(sizeof("madvise")-1, count))) {
171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
173 } else if (!memcmp("never", buf,
174 min(sizeof("never")-1, count))) {
175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
177 } else
178 ret = -EINVAL;
179
180 if (ret > 0) {
181 int err = start_stop_khugepaged();
182 if (err)
183 ret = err;
184 }
185 return ret;
186}
187static struct kobj_attribute enabled_attr =
188 __ATTR(enabled, 0644, enabled_show, enabled_store);
189
190ssize_t single_hugepage_flag_show(struct kobject *kobj,
191 struct kobj_attribute *attr, char *buf,
192 enum transparent_hugepage_flag flag)
193{
194 return sprintf(buf, "%d\n",
195 !!test_bit(flag, &transparent_hugepage_flags));
196}
197
198ssize_t single_hugepage_flag_store(struct kobject *kobj,
199 struct kobj_attribute *attr,
200 const char *buf, size_t count,
201 enum transparent_hugepage_flag flag)
202{
203 unsigned long value;
204 int ret;
205
206 ret = kstrtoul(buf, 10, &value);
207 if (ret < 0)
208 return ret;
209 if (value > 1)
210 return -EINVAL;
211
212 if (value)
213 set_bit(flag, &transparent_hugepage_flags);
214 else
215 clear_bit(flag, &transparent_hugepage_flags);
216
217 return count;
218}
219
220static ssize_t defrag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf)
222{
223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
224 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
226 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
228 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
230 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
231 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
232}
233
234static ssize_t defrag_store(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 const char *buf, size_t count)
237{
238 if (!memcmp("always", buf,
239 min(sizeof("always")-1, count))) {
240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
244 } else if (!memcmp("defer+madvise", buf,
245 min(sizeof("defer+madvise")-1, count))) {
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
250 } else if (!memcmp("defer", buf,
251 min(sizeof("defer")-1, count))) {
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
256 } else if (!memcmp("madvise", buf,
257 min(sizeof("madvise")-1, count))) {
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
262 } else if (!memcmp("never", buf,
263 min(sizeof("never")-1, count))) {
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
268 } else
269 return -EINVAL;
270
271 return count;
272}
273static struct kobj_attribute defrag_attr =
274 __ATTR(defrag, 0644, defrag_show, defrag_store);
275
276static ssize_t use_zero_page_show(struct kobject *kobj,
277 struct kobj_attribute *attr, char *buf)
278{
279 return single_hugepage_flag_show(kobj, attr, buf,
280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
281}
282static ssize_t use_zero_page_store(struct kobject *kobj,
283 struct kobj_attribute *attr, const char *buf, size_t count)
284{
285 return single_hugepage_flag_store(kobj, attr, buf, count,
286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
287}
288static struct kobj_attribute use_zero_page_attr =
289 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
290
291static ssize_t hpage_pmd_size_show(struct kobject *kobj,
292 struct kobj_attribute *attr, char *buf)
293{
294 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
295}
296static struct kobj_attribute hpage_pmd_size_attr =
297 __ATTR_RO(hpage_pmd_size);
298
299#ifdef CONFIG_DEBUG_VM
300static ssize_t debug_cow_show(struct kobject *kobj,
301 struct kobj_attribute *attr, char *buf)
302{
303 return single_hugepage_flag_show(kobj, attr, buf,
304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
305}
306static ssize_t debug_cow_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
309{
310 return single_hugepage_flag_store(kobj, attr, buf, count,
311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312}
313static struct kobj_attribute debug_cow_attr =
314 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
315#endif /* CONFIG_DEBUG_VM */
316
317static struct attribute *hugepage_attr[] = {
318 &enabled_attr.attr,
319 &defrag_attr.attr,
320 &use_zero_page_attr.attr,
321 &hpage_pmd_size_attr.attr,
322#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
323 &shmem_enabled_attr.attr,
324#endif
325#ifdef CONFIG_DEBUG_VM
326 &debug_cow_attr.attr,
327#endif
328 NULL,
329};
330
331static const struct attribute_group hugepage_attr_group = {
332 .attrs = hugepage_attr,
333};
334
335static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
336{
337 int err;
338
339 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
340 if (unlikely(!*hugepage_kobj)) {
341 pr_err("failed to create transparent hugepage kobject\n");
342 return -ENOMEM;
343 }
344
345 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
346 if (err) {
347 pr_err("failed to register transparent hugepage group\n");
348 goto delete_obj;
349 }
350
351 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
352 if (err) {
353 pr_err("failed to register transparent hugepage group\n");
354 goto remove_hp_group;
355 }
356
357 return 0;
358
359remove_hp_group:
360 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
361delete_obj:
362 kobject_put(*hugepage_kobj);
363 return err;
364}
365
366static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
367{
368 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
369 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
370 kobject_put(hugepage_kobj);
371}
372#else
373static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
374{
375 return 0;
376}
377
378static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
379{
380}
381#endif /* CONFIG_SYSFS */
382
383static int __init hugepage_init(void)
384{
385 int err;
386 struct kobject *hugepage_kobj;
387
388 if (!has_transparent_hugepage()) {
389 transparent_hugepage_flags = 0;
390 return -EINVAL;
391 }
392
393 /*
394 * hugepages can't be allocated by the buddy allocator
395 */
396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
397 /*
398 * we use page->mapping and page->index in second tail page
399 * as list_head: assuming THP order >= 2
400 */
401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
402
403 err = hugepage_init_sysfs(&hugepage_kobj);
404 if (err)
405 goto err_sysfs;
406
407 err = khugepaged_init();
408 if (err)
409 goto err_slab;
410
411 err = register_shrinker(&huge_zero_page_shrinker);
412 if (err)
413 goto err_hzp_shrinker;
414 err = register_shrinker(&deferred_split_shrinker);
415 if (err)
416 goto err_split_shrinker;
417
418 /*
419 * By default disable transparent hugepages on smaller systems,
420 * where the extra memory used could hurt more than TLB overhead
421 * is likely to save. The admin can still enable it through /sys.
422 */
423 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
424 transparent_hugepage_flags = 0;
425 return 0;
426 }
427
428 err = start_stop_khugepaged();
429 if (err)
430 goto err_khugepaged;
431
432 return 0;
433err_khugepaged:
434 unregister_shrinker(&deferred_split_shrinker);
435err_split_shrinker:
436 unregister_shrinker(&huge_zero_page_shrinker);
437err_hzp_shrinker:
438 khugepaged_destroy();
439err_slab:
440 hugepage_exit_sysfs(hugepage_kobj);
441err_sysfs:
442 return err;
443}
444subsys_initcall(hugepage_init);
445
446static int __init setup_transparent_hugepage(char *str)
447{
448 int ret = 0;
449 if (!str)
450 goto out;
451 if (!strcmp(str, "always")) {
452 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
453 &transparent_hugepage_flags);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455 &transparent_hugepage_flags);
456 ret = 1;
457 } else if (!strcmp(str, "madvise")) {
458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
459 &transparent_hugepage_flags);
460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
461 &transparent_hugepage_flags);
462 ret = 1;
463 } else if (!strcmp(str, "never")) {
464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
465 &transparent_hugepage_flags);
466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
467 &transparent_hugepage_flags);
468 ret = 1;
469 }
470out:
471 if (!ret)
472 pr_warn("transparent_hugepage= cannot parse, ignored\n");
473 return ret;
474}
475__setup("transparent_hugepage=", setup_transparent_hugepage);
476
477pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
478{
479 if (likely(vma->vm_flags & VM_WRITE))
480 pmd = pmd_mkwrite(pmd);
481 return pmd;
482}
483
484static inline struct list_head *page_deferred_list(struct page *page)
485{
486 /*
487 * ->lru in the tail pages is occupied by compound_head.
488 * Let's use ->mapping + ->index in the second tail page as list_head.
489 */
490 return (struct list_head *)&page[2].mapping;
491}
492
493void prep_transhuge_page(struct page *page)
494{
495 /*
496 * we use page->mapping and page->indexlru in second tail page
497 * as list_head: assuming THP order >= 2
498 */
499
500 INIT_LIST_HEAD(page_deferred_list(page));
501 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
502}
503
504unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
505 loff_t off, unsigned long flags, unsigned long size)
506{
507 unsigned long addr;
508 loff_t off_end = off + len;
509 loff_t off_align = round_up(off, size);
510 unsigned long len_pad;
511
512 if (off_end <= off_align || (off_end - off_align) < size)
513 return 0;
514
515 len_pad = len + size;
516 if (len_pad < len || (off + len_pad) < off)
517 return 0;
518
519 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
520 off >> PAGE_SHIFT, flags);
521 if (IS_ERR_VALUE(addr))
522 return 0;
523
524 addr += (off - addr) & (size - 1);
525 return addr;
526}
527
528unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
529 unsigned long len, unsigned long pgoff, unsigned long flags)
530{
531 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
532
533 if (addr)
534 goto out;
535 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
536 goto out;
537
538 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
539 if (addr)
540 return addr;
541
542 out:
543 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
544}
545EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
546
547static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
548 gfp_t gfp)
549{
550 struct vm_area_struct *vma = vmf->vma;
551 struct mem_cgroup *memcg;
552 pgtable_t pgtable;
553 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
554 int ret = 0;
555
556 VM_BUG_ON_PAGE(!PageCompound(page), page);
557
558 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
559 put_page(page);
560 count_vm_event(THP_FAULT_FALLBACK);
561 return VM_FAULT_FALLBACK;
562 }
563
564 pgtable = pte_alloc_one(vma->vm_mm, haddr);
565 if (unlikely(!pgtable)) {
566 ret = VM_FAULT_OOM;
567 goto release;
568 }
569
570 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
571 /*
572 * The memory barrier inside __SetPageUptodate makes sure that
573 * clear_huge_page writes become visible before the set_pmd_at()
574 * write.
575 */
576 __SetPageUptodate(page);
577
578 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
579 if (unlikely(!pmd_none(*vmf->pmd))) {
580 goto unlock_release;
581 } else {
582 pmd_t entry;
583
584 ret = check_stable_address_space(vma->vm_mm);
585 if (ret)
586 goto unlock_release;
587
588 /* Deliver the page fault to userland */
589 if (userfaultfd_missing(vma)) {
590 int ret;
591
592 spin_unlock(vmf->ptl);
593 mem_cgroup_cancel_charge(page, memcg, true);
594 put_page(page);
595 pte_free(vma->vm_mm, pgtable);
596 ret = handle_userfault(vmf, VM_UFFD_MISSING);
597 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
598 return ret;
599 }
600
601 entry = mk_huge_pmd(page, vma->vm_page_prot);
602 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
603 page_add_new_anon_rmap(page, vma, haddr, true);
604 mem_cgroup_commit_charge(page, memcg, false, true);
605 lru_cache_add_active_or_unevictable(page, vma);
606 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
607 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
608 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
609 mm_inc_nr_ptes(vma->vm_mm);
610 spin_unlock(vmf->ptl);
611 count_vm_event(THP_FAULT_ALLOC);
612 }
613
614 return 0;
615unlock_release:
616 spin_unlock(vmf->ptl);
617release:
618 if (pgtable)
619 pte_free(vma->vm_mm, pgtable);
620 mem_cgroup_cancel_charge(page, memcg, true);
621 put_page(page);
622 return ret;
623
624}
625
626/*
627 * always: directly stall for all thp allocations
628 * defer: wake kswapd and fail if not immediately available
629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
630 * fail if not immediately available
631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
632 * available
633 * never: never stall for any thp allocation
634 */
635static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
636{
637 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
638
639 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
640 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
644 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
645 __GFP_KSWAPD_RECLAIM);
646 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
647 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
648 0);
649 return GFP_TRANSHUGE_LIGHT;
650}
651
652/* Caller must hold page table lock. */
653static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
654 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
655 struct page *zero_page)
656{
657 pmd_t entry;
658 if (!pmd_none(*pmd))
659 return false;
660 entry = mk_pmd(zero_page, vma->vm_page_prot);
661 entry = pmd_mkhuge(entry);
662 if (pgtable)
663 pgtable_trans_huge_deposit(mm, pmd, pgtable);
664 set_pmd_at(mm, haddr, pmd, entry);
665 mm_inc_nr_ptes(mm);
666 return true;
667}
668
669int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
670{
671 struct vm_area_struct *vma = vmf->vma;
672 gfp_t gfp;
673 struct page *page;
674 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
675
676 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
677 return VM_FAULT_FALLBACK;
678 if (unlikely(anon_vma_prepare(vma)))
679 return VM_FAULT_OOM;
680 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
681 return VM_FAULT_OOM;
682 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
683 !mm_forbids_zeropage(vma->vm_mm) &&
684 transparent_hugepage_use_zero_page()) {
685 pgtable_t pgtable;
686 struct page *zero_page;
687 bool set;
688 int ret;
689 pgtable = pte_alloc_one(vma->vm_mm, haddr);
690 if (unlikely(!pgtable))
691 return VM_FAULT_OOM;
692 zero_page = mm_get_huge_zero_page(vma->vm_mm);
693 if (unlikely(!zero_page)) {
694 pte_free(vma->vm_mm, pgtable);
695 count_vm_event(THP_FAULT_FALLBACK);
696 return VM_FAULT_FALLBACK;
697 }
698 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
699 ret = 0;
700 set = false;
701 if (pmd_none(*vmf->pmd)) {
702 ret = check_stable_address_space(vma->vm_mm);
703 if (ret) {
704 spin_unlock(vmf->ptl);
705 } else if (userfaultfd_missing(vma)) {
706 spin_unlock(vmf->ptl);
707 ret = handle_userfault(vmf, VM_UFFD_MISSING);
708 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
709 } else {
710 set_huge_zero_page(pgtable, vma->vm_mm, vma,
711 haddr, vmf->pmd, zero_page);
712 spin_unlock(vmf->ptl);
713 set = true;
714 }
715 } else
716 spin_unlock(vmf->ptl);
717 if (!set)
718 pte_free(vma->vm_mm, pgtable);
719 return ret;
720 }
721 gfp = alloc_hugepage_direct_gfpmask(vma);
722 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
723 if (unlikely(!page)) {
724 count_vm_event(THP_FAULT_FALLBACK);
725 return VM_FAULT_FALLBACK;
726 }
727 prep_transhuge_page(page);
728 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
729}
730
731static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
732 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
733 pgtable_t pgtable)
734{
735 struct mm_struct *mm = vma->vm_mm;
736 pmd_t entry;
737 spinlock_t *ptl;
738
739 ptl = pmd_lock(mm, pmd);
740 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
741 if (pfn_t_devmap(pfn))
742 entry = pmd_mkdevmap(entry);
743 if (write) {
744 entry = pmd_mkyoung(pmd_mkdirty(entry));
745 entry = maybe_pmd_mkwrite(entry, vma);
746 }
747
748 if (pgtable) {
749 pgtable_trans_huge_deposit(mm, pmd, pgtable);
750 mm_inc_nr_ptes(mm);
751 }
752
753 set_pmd_at(mm, addr, pmd, entry);
754 update_mmu_cache_pmd(vma, addr, pmd);
755 spin_unlock(ptl);
756}
757
758int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
759 pmd_t *pmd, pfn_t pfn, bool write)
760{
761 pgprot_t pgprot = vma->vm_page_prot;
762 pgtable_t pgtable = NULL;
763 /*
764 * If we had pmd_special, we could avoid all these restrictions,
765 * but we need to be consistent with PTEs and architectures that
766 * can't support a 'special' bit.
767 */
768 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
769 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
770 (VM_PFNMAP|VM_MIXEDMAP));
771 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
772 BUG_ON(!pfn_t_devmap(pfn));
773
774 if (addr < vma->vm_start || addr >= vma->vm_end)
775 return VM_FAULT_SIGBUS;
776
777 if (arch_needs_pgtable_deposit()) {
778 pgtable = pte_alloc_one(vma->vm_mm, addr);
779 if (!pgtable)
780 return VM_FAULT_OOM;
781 }
782
783 track_pfn_insert(vma, &pgprot, pfn);
784
785 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
786 return VM_FAULT_NOPAGE;
787}
788EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
789
790#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
791static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
792{
793 if (likely(vma->vm_flags & VM_WRITE))
794 pud = pud_mkwrite(pud);
795 return pud;
796}
797
798static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
799 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
800{
801 struct mm_struct *mm = vma->vm_mm;
802 pud_t entry;
803 spinlock_t *ptl;
804
805 ptl = pud_lock(mm, pud);
806 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
807 if (pfn_t_devmap(pfn))
808 entry = pud_mkdevmap(entry);
809 if (write) {
810 entry = pud_mkyoung(pud_mkdirty(entry));
811 entry = maybe_pud_mkwrite(entry, vma);
812 }
813 set_pud_at(mm, addr, pud, entry);
814 update_mmu_cache_pud(vma, addr, pud);
815 spin_unlock(ptl);
816}
817
818int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
819 pud_t *pud, pfn_t pfn, bool write)
820{
821 pgprot_t pgprot = vma->vm_page_prot;
822 /*
823 * If we had pud_special, we could avoid all these restrictions,
824 * but we need to be consistent with PTEs and architectures that
825 * can't support a 'special' bit.
826 */
827 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
828 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
829 (VM_PFNMAP|VM_MIXEDMAP));
830 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
831 BUG_ON(!pfn_t_devmap(pfn));
832
833 if (addr < vma->vm_start || addr >= vma->vm_end)
834 return VM_FAULT_SIGBUS;
835
836 track_pfn_insert(vma, &pgprot, pfn);
837
838 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
839 return VM_FAULT_NOPAGE;
840}
841EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
842#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
843
844static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
845 pmd_t *pmd, int flags)
846{
847 pmd_t _pmd;
848
849 _pmd = pmd_mkyoung(*pmd);
850 if (flags & FOLL_WRITE)
851 _pmd = pmd_mkdirty(_pmd);
852 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
853 pmd, _pmd, flags & FOLL_WRITE))
854 update_mmu_cache_pmd(vma, addr, pmd);
855}
856
857struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
858 pmd_t *pmd, int flags)
859{
860 unsigned long pfn = pmd_pfn(*pmd);
861 struct mm_struct *mm = vma->vm_mm;
862 struct dev_pagemap *pgmap;
863 struct page *page;
864
865 assert_spin_locked(pmd_lockptr(mm, pmd));
866
867 /*
868 * When we COW a devmap PMD entry, we split it into PTEs, so we should
869 * not be in this function with `flags & FOLL_COW` set.
870 */
871 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
872
873 if (flags & FOLL_WRITE && !pmd_write(*pmd))
874 return NULL;
875
876 if (pmd_present(*pmd) && pmd_devmap(*pmd))
877 /* pass */;
878 else
879 return NULL;
880
881 if (flags & FOLL_TOUCH)
882 touch_pmd(vma, addr, pmd, flags);
883
884 /*
885 * device mapped pages can only be returned if the
886 * caller will manage the page reference count.
887 */
888 if (!(flags & FOLL_GET))
889 return ERR_PTR(-EEXIST);
890
891 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
892 pgmap = get_dev_pagemap(pfn, NULL);
893 if (!pgmap)
894 return ERR_PTR(-EFAULT);
895 page = pfn_to_page(pfn);
896 get_page(page);
897 put_dev_pagemap(pgmap);
898
899 return page;
900}
901
902int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
903 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
904 struct vm_area_struct *vma)
905{
906 spinlock_t *dst_ptl, *src_ptl;
907 struct page *src_page;
908 pmd_t pmd;
909 pgtable_t pgtable = NULL;
910 int ret = -ENOMEM;
911
912 /* Skip if can be re-fill on fault */
913 if (!vma_is_anonymous(vma))
914 return 0;
915
916 pgtable = pte_alloc_one(dst_mm, addr);
917 if (unlikely(!pgtable))
918 goto out;
919
920 dst_ptl = pmd_lock(dst_mm, dst_pmd);
921 src_ptl = pmd_lockptr(src_mm, src_pmd);
922 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
923
924 ret = -EAGAIN;
925 pmd = *src_pmd;
926
927#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
928 if (unlikely(is_swap_pmd(pmd))) {
929 swp_entry_t entry = pmd_to_swp_entry(pmd);
930
931 VM_BUG_ON(!is_pmd_migration_entry(pmd));
932 if (is_write_migration_entry(entry)) {
933 make_migration_entry_read(&entry);
934 pmd = swp_entry_to_pmd(entry);
935 if (pmd_swp_soft_dirty(*src_pmd))
936 pmd = pmd_swp_mksoft_dirty(pmd);
937 set_pmd_at(src_mm, addr, src_pmd, pmd);
938 }
939 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
940 mm_inc_nr_ptes(dst_mm);
941 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
942 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
943 ret = 0;
944 goto out_unlock;
945 }
946#endif
947
948 if (unlikely(!pmd_trans_huge(pmd))) {
949 pte_free(dst_mm, pgtable);
950 goto out_unlock;
951 }
952 /*
953 * When page table lock is held, the huge zero pmd should not be
954 * under splitting since we don't split the page itself, only pmd to
955 * a page table.
956 */
957 if (is_huge_zero_pmd(pmd)) {
958 struct page *zero_page;
959 /*
960 * get_huge_zero_page() will never allocate a new page here,
961 * since we already have a zero page to copy. It just takes a
962 * reference.
963 */
964 zero_page = mm_get_huge_zero_page(dst_mm);
965 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
966 zero_page);
967 ret = 0;
968 goto out_unlock;
969 }
970
971 src_page = pmd_page(pmd);
972 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
973 get_page(src_page);
974 page_dup_rmap(src_page, true);
975 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
976 mm_inc_nr_ptes(dst_mm);
977 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
978
979 pmdp_set_wrprotect(src_mm, addr, src_pmd);
980 pmd = pmd_mkold(pmd_wrprotect(pmd));
981 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
982
983 ret = 0;
984out_unlock:
985 spin_unlock(src_ptl);
986 spin_unlock(dst_ptl);
987out:
988 return ret;
989}
990
991#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
992static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
993 pud_t *pud, int flags)
994{
995 pud_t _pud;
996
997 _pud = pud_mkyoung(*pud);
998 if (flags & FOLL_WRITE)
999 _pud = pud_mkdirty(_pud);
1000 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1001 pud, _pud, flags & FOLL_WRITE))
1002 update_mmu_cache_pud(vma, addr, pud);
1003}
1004
1005struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1006 pud_t *pud, int flags)
1007{
1008 unsigned long pfn = pud_pfn(*pud);
1009 struct mm_struct *mm = vma->vm_mm;
1010 struct dev_pagemap *pgmap;
1011 struct page *page;
1012
1013 assert_spin_locked(pud_lockptr(mm, pud));
1014
1015 if (flags & FOLL_WRITE && !pud_write(*pud))
1016 return NULL;
1017
1018 if (pud_present(*pud) && pud_devmap(*pud))
1019 /* pass */;
1020 else
1021 return NULL;
1022
1023 if (flags & FOLL_TOUCH)
1024 touch_pud(vma, addr, pud, flags);
1025
1026 /*
1027 * device mapped pages can only be returned if the
1028 * caller will manage the page reference count.
1029 */
1030 if (!(flags & FOLL_GET))
1031 return ERR_PTR(-EEXIST);
1032
1033 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1034 pgmap = get_dev_pagemap(pfn, NULL);
1035 if (!pgmap)
1036 return ERR_PTR(-EFAULT);
1037 page = pfn_to_page(pfn);
1038 get_page(page);
1039 put_dev_pagemap(pgmap);
1040
1041 return page;
1042}
1043
1044int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1046 struct vm_area_struct *vma)
1047{
1048 spinlock_t *dst_ptl, *src_ptl;
1049 pud_t pud;
1050 int ret;
1051
1052 dst_ptl = pud_lock(dst_mm, dst_pud);
1053 src_ptl = pud_lockptr(src_mm, src_pud);
1054 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1055
1056 ret = -EAGAIN;
1057 pud = *src_pud;
1058 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1059 goto out_unlock;
1060
1061 /*
1062 * When page table lock is held, the huge zero pud should not be
1063 * under splitting since we don't split the page itself, only pud to
1064 * a page table.
1065 */
1066 if (is_huge_zero_pud(pud)) {
1067 /* No huge zero pud yet */
1068 }
1069
1070 pudp_set_wrprotect(src_mm, addr, src_pud);
1071 pud = pud_mkold(pud_wrprotect(pud));
1072 set_pud_at(dst_mm, addr, dst_pud, pud);
1073
1074 ret = 0;
1075out_unlock:
1076 spin_unlock(src_ptl);
1077 spin_unlock(dst_ptl);
1078 return ret;
1079}
1080
1081void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1082{
1083 pud_t entry;
1084 unsigned long haddr;
1085 bool write = vmf->flags & FAULT_FLAG_WRITE;
1086
1087 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1088 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1089 goto unlock;
1090
1091 entry = pud_mkyoung(orig_pud);
1092 if (write)
1093 entry = pud_mkdirty(entry);
1094 haddr = vmf->address & HPAGE_PUD_MASK;
1095 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1096 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1097
1098unlock:
1099 spin_unlock(vmf->ptl);
1100}
1101#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1102
1103void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1104{
1105 pmd_t entry;
1106 unsigned long haddr;
1107 bool write = vmf->flags & FAULT_FLAG_WRITE;
1108
1109 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1110 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1111 goto unlock;
1112
1113 entry = pmd_mkyoung(orig_pmd);
1114 if (write)
1115 entry = pmd_mkdirty(entry);
1116 haddr = vmf->address & HPAGE_PMD_MASK;
1117 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1118 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1119
1120unlock:
1121 spin_unlock(vmf->ptl);
1122}
1123
1124static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1125 struct page *page)
1126{
1127 struct vm_area_struct *vma = vmf->vma;
1128 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1129 struct mem_cgroup *memcg;
1130 pgtable_t pgtable;
1131 pmd_t _pmd;
1132 int ret = 0, i;
1133 struct page **pages;
1134 unsigned long mmun_start; /* For mmu_notifiers */
1135 unsigned long mmun_end; /* For mmu_notifiers */
1136
1137 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1138 GFP_KERNEL);
1139 if (unlikely(!pages)) {
1140 ret |= VM_FAULT_OOM;
1141 goto out;
1142 }
1143
1144 for (i = 0; i < HPAGE_PMD_NR; i++) {
1145 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1146 vmf->address, page_to_nid(page));
1147 if (unlikely(!pages[i] ||
1148 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1149 GFP_KERNEL, &memcg, false))) {
1150 if (pages[i])
1151 put_page(pages[i]);
1152 while (--i >= 0) {
1153 memcg = (void *)page_private(pages[i]);
1154 set_page_private(pages[i], 0);
1155 mem_cgroup_cancel_charge(pages[i], memcg,
1156 false);
1157 put_page(pages[i]);
1158 }
1159 kfree(pages);
1160 ret |= VM_FAULT_OOM;
1161 goto out;
1162 }
1163 set_page_private(pages[i], (unsigned long)memcg);
1164 }
1165
1166 for (i = 0; i < HPAGE_PMD_NR; i++) {
1167 copy_user_highpage(pages[i], page + i,
1168 haddr + PAGE_SIZE * i, vma);
1169 __SetPageUptodate(pages[i]);
1170 cond_resched();
1171 }
1172
1173 mmun_start = haddr;
1174 mmun_end = haddr + HPAGE_PMD_SIZE;
1175 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1176
1177 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1178 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1179 goto out_free_pages;
1180 VM_BUG_ON_PAGE(!PageHead(page), page);
1181
1182 /*
1183 * Leave pmd empty until pte is filled note we must notify here as
1184 * concurrent CPU thread might write to new page before the call to
1185 * mmu_notifier_invalidate_range_end() happens which can lead to a
1186 * device seeing memory write in different order than CPU.
1187 *
1188 * See Documentation/vm/mmu_notifier.txt
1189 */
1190 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1191
1192 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1193 pmd_populate(vma->vm_mm, &_pmd, pgtable);
1194
1195 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1196 pte_t entry;
1197 entry = mk_pte(pages[i], vma->vm_page_prot);
1198 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1199 memcg = (void *)page_private(pages[i]);
1200 set_page_private(pages[i], 0);
1201 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1202 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1203 lru_cache_add_active_or_unevictable(pages[i], vma);
1204 vmf->pte = pte_offset_map(&_pmd, haddr);
1205 VM_BUG_ON(!pte_none(*vmf->pte));
1206 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1207 pte_unmap(vmf->pte);
1208 }
1209 kfree(pages);
1210
1211 smp_wmb(); /* make pte visible before pmd */
1212 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1213 page_remove_rmap(page, true);
1214 spin_unlock(vmf->ptl);
1215
1216 /*
1217 * No need to double call mmu_notifier->invalidate_range() callback as
1218 * the above pmdp_huge_clear_flush_notify() did already call it.
1219 */
1220 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1221 mmun_end);
1222
1223 ret |= VM_FAULT_WRITE;
1224 put_page(page);
1225
1226out:
1227 return ret;
1228
1229out_free_pages:
1230 spin_unlock(vmf->ptl);
1231 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1232 for (i = 0; i < HPAGE_PMD_NR; i++) {
1233 memcg = (void *)page_private(pages[i]);
1234 set_page_private(pages[i], 0);
1235 mem_cgroup_cancel_charge(pages[i], memcg, false);
1236 put_page(pages[i]);
1237 }
1238 kfree(pages);
1239 goto out;
1240}
1241
1242int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1243{
1244 struct vm_area_struct *vma = vmf->vma;
1245 struct page *page = NULL, *new_page;
1246 struct mem_cgroup *memcg;
1247 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1248 unsigned long mmun_start; /* For mmu_notifiers */
1249 unsigned long mmun_end; /* For mmu_notifiers */
1250 gfp_t huge_gfp; /* for allocation and charge */
1251 int ret = 0;
1252
1253 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1254 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1255 if (is_huge_zero_pmd(orig_pmd))
1256 goto alloc;
1257 spin_lock(vmf->ptl);
1258 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1259 goto out_unlock;
1260
1261 page = pmd_page(orig_pmd);
1262 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1263 /*
1264 * We can only reuse the page if nobody else maps the huge page or it's
1265 * part.
1266 */
1267 if (!trylock_page(page)) {
1268 get_page(page);
1269 spin_unlock(vmf->ptl);
1270 lock_page(page);
1271 spin_lock(vmf->ptl);
1272 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1273 unlock_page(page);
1274 put_page(page);
1275 goto out_unlock;
1276 }
1277 put_page(page);
1278 }
1279 if (reuse_swap_page(page, NULL)) {
1280 pmd_t entry;
1281 entry = pmd_mkyoung(orig_pmd);
1282 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1283 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1284 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1285 ret |= VM_FAULT_WRITE;
1286 unlock_page(page);
1287 goto out_unlock;
1288 }
1289 unlock_page(page);
1290 get_page(page);
1291 spin_unlock(vmf->ptl);
1292alloc:
1293 if (transparent_hugepage_enabled(vma) &&
1294 !transparent_hugepage_debug_cow()) {
1295 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1296 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1297 } else
1298 new_page = NULL;
1299
1300 if (likely(new_page)) {
1301 prep_transhuge_page(new_page);
1302 } else {
1303 if (!page) {
1304 split_huge_pmd(vma, vmf->pmd, vmf->address);
1305 ret |= VM_FAULT_FALLBACK;
1306 } else {
1307 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1308 if (ret & VM_FAULT_OOM) {
1309 split_huge_pmd(vma, vmf->pmd, vmf->address);
1310 ret |= VM_FAULT_FALLBACK;
1311 }
1312 put_page(page);
1313 }
1314 count_vm_event(THP_FAULT_FALLBACK);
1315 goto out;
1316 }
1317
1318 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1319 huge_gfp, &memcg, true))) {
1320 put_page(new_page);
1321 split_huge_pmd(vma, vmf->pmd, vmf->address);
1322 if (page)
1323 put_page(page);
1324 ret |= VM_FAULT_FALLBACK;
1325 count_vm_event(THP_FAULT_FALLBACK);
1326 goto out;
1327 }
1328
1329 count_vm_event(THP_FAULT_ALLOC);
1330
1331 if (!page)
1332 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1333 else
1334 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1335 __SetPageUptodate(new_page);
1336
1337 mmun_start = haddr;
1338 mmun_end = haddr + HPAGE_PMD_SIZE;
1339 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1340
1341 spin_lock(vmf->ptl);
1342 if (page)
1343 put_page(page);
1344 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345 spin_unlock(vmf->ptl);
1346 mem_cgroup_cancel_charge(new_page, memcg, true);
1347 put_page(new_page);
1348 goto out_mn;
1349 } else {
1350 pmd_t entry;
1351 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1352 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1353 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1354 page_add_new_anon_rmap(new_page, vma, haddr, true);
1355 mem_cgroup_commit_charge(new_page, memcg, false, true);
1356 lru_cache_add_active_or_unevictable(new_page, vma);
1357 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1358 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1359 if (!page) {
1360 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1361 } else {
1362 VM_BUG_ON_PAGE(!PageHead(page), page);
1363 page_remove_rmap(page, true);
1364 put_page(page);
1365 }
1366 ret |= VM_FAULT_WRITE;
1367 }
1368 spin_unlock(vmf->ptl);
1369out_mn:
1370 /*
1371 * No need to double call mmu_notifier->invalidate_range() callback as
1372 * the above pmdp_huge_clear_flush_notify() did already call it.
1373 */
1374 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1375 mmun_end);
1376out:
1377 return ret;
1378out_unlock:
1379 spin_unlock(vmf->ptl);
1380 return ret;
1381}
1382
1383/*
1384 * FOLL_FORCE can write to even unwritable pmd's, but only
1385 * after we've gone through a COW cycle and they are dirty.
1386 */
1387static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1388{
1389 return pmd_write(pmd) ||
1390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1391}
1392
1393struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1394 unsigned long addr,
1395 pmd_t *pmd,
1396 unsigned int flags)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct page *page = NULL;
1400
1401 assert_spin_locked(pmd_lockptr(mm, pmd));
1402
1403 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1404 goto out;
1405
1406 /* Avoid dumping huge zero page */
1407 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1408 return ERR_PTR(-EFAULT);
1409
1410 /* Full NUMA hinting faults to serialise migration in fault paths */
1411 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1412 goto out;
1413
1414 page = pmd_page(*pmd);
1415 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1416 if (flags & FOLL_TOUCH)
1417 touch_pmd(vma, addr, pmd, flags);
1418 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1419 /*
1420 * We don't mlock() pte-mapped THPs. This way we can avoid
1421 * leaking mlocked pages into non-VM_LOCKED VMAs.
1422 *
1423 * For anon THP:
1424 *
1425 * In most cases the pmd is the only mapping of the page as we
1426 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1427 * writable private mappings in populate_vma_page_range().
1428 *
1429 * The only scenario when we have the page shared here is if we
1430 * mlocking read-only mapping shared over fork(). We skip
1431 * mlocking such pages.
1432 *
1433 * For file THP:
1434 *
1435 * We can expect PageDoubleMap() to be stable under page lock:
1436 * for file pages we set it in page_add_file_rmap(), which
1437 * requires page to be locked.
1438 */
1439
1440 if (PageAnon(page) && compound_mapcount(page) != 1)
1441 goto skip_mlock;
1442 if (PageDoubleMap(page) || !page->mapping)
1443 goto skip_mlock;
1444 if (!trylock_page(page))
1445 goto skip_mlock;
1446 lru_add_drain();
1447 if (page->mapping && !PageDoubleMap(page))
1448 mlock_vma_page(page);
1449 unlock_page(page);
1450 }
1451skip_mlock:
1452 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1453 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1454 if (flags & FOLL_GET)
1455 get_page(page);
1456
1457out:
1458 return page;
1459}
1460
1461/* NUMA hinting page fault entry point for trans huge pmds */
1462int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1463{
1464 struct vm_area_struct *vma = vmf->vma;
1465 struct anon_vma *anon_vma = NULL;
1466 struct page *page;
1467 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1468 int page_nid = -1, this_nid = numa_node_id();
1469 int target_nid, last_cpupid = -1;
1470 bool page_locked;
1471 bool migrated = false;
1472 bool was_writable;
1473 int flags = 0;
1474
1475 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1476 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1477 goto out_unlock;
1478
1479 /*
1480 * If there are potential migrations, wait for completion and retry
1481 * without disrupting NUMA hinting information. Do not relock and
1482 * check_same as the page may no longer be mapped.
1483 */
1484 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1485 page = pmd_page(*vmf->pmd);
1486 if (!get_page_unless_zero(page))
1487 goto out_unlock;
1488 spin_unlock(vmf->ptl);
1489 wait_on_page_locked(page);
1490 put_page(page);
1491 goto out;
1492 }
1493
1494 page = pmd_page(pmd);
1495 BUG_ON(is_huge_zero_page(page));
1496 page_nid = page_to_nid(page);
1497 last_cpupid = page_cpupid_last(page);
1498 count_vm_numa_event(NUMA_HINT_FAULTS);
1499 if (page_nid == this_nid) {
1500 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1501 flags |= TNF_FAULT_LOCAL;
1502 }
1503
1504 /* See similar comment in do_numa_page for explanation */
1505 if (!pmd_savedwrite(pmd))
1506 flags |= TNF_NO_GROUP;
1507
1508 /*
1509 * Acquire the page lock to serialise THP migrations but avoid dropping
1510 * page_table_lock if at all possible
1511 */
1512 page_locked = trylock_page(page);
1513 target_nid = mpol_misplaced(page, vma, haddr);
1514 if (target_nid == -1) {
1515 /* If the page was locked, there are no parallel migrations */
1516 if (page_locked)
1517 goto clear_pmdnuma;
1518 }
1519
1520 /* Migration could have started since the pmd_trans_migrating check */
1521 if (!page_locked) {
1522 page_nid = -1;
1523 if (!get_page_unless_zero(page))
1524 goto out_unlock;
1525 spin_unlock(vmf->ptl);
1526 wait_on_page_locked(page);
1527 put_page(page);
1528 goto out;
1529 }
1530
1531 /*
1532 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1533 * to serialises splits
1534 */
1535 get_page(page);
1536 spin_unlock(vmf->ptl);
1537 anon_vma = page_lock_anon_vma_read(page);
1538
1539 /* Confirm the PMD did not change while page_table_lock was released */
1540 spin_lock(vmf->ptl);
1541 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1542 unlock_page(page);
1543 put_page(page);
1544 page_nid = -1;
1545 goto out_unlock;
1546 }
1547
1548 /* Bail if we fail to protect against THP splits for any reason */
1549 if (unlikely(!anon_vma)) {
1550 put_page(page);
1551 page_nid = -1;
1552 goto clear_pmdnuma;
1553 }
1554
1555 /*
1556 * Since we took the NUMA fault, we must have observed the !accessible
1557 * bit. Make sure all other CPUs agree with that, to avoid them
1558 * modifying the page we're about to migrate.
1559 *
1560 * Must be done under PTL such that we'll observe the relevant
1561 * inc_tlb_flush_pending().
1562 *
1563 * We are not sure a pending tlb flush here is for a huge page
1564 * mapping or not. Hence use the tlb range variant
1565 */
1566 if (mm_tlb_flush_pending(vma->vm_mm))
1567 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1568
1569 /*
1570 * Migrate the THP to the requested node, returns with page unlocked
1571 * and access rights restored.
1572 */
1573 spin_unlock(vmf->ptl);
1574
1575 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1576 vmf->pmd, pmd, vmf->address, page, target_nid);
1577 if (migrated) {
1578 flags |= TNF_MIGRATED;
1579 page_nid = target_nid;
1580 } else
1581 flags |= TNF_MIGRATE_FAIL;
1582
1583 goto out;
1584clear_pmdnuma:
1585 BUG_ON(!PageLocked(page));
1586 was_writable = pmd_savedwrite(pmd);
1587 pmd = pmd_modify(pmd, vma->vm_page_prot);
1588 pmd = pmd_mkyoung(pmd);
1589 if (was_writable)
1590 pmd = pmd_mkwrite(pmd);
1591 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1592 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1593 unlock_page(page);
1594out_unlock:
1595 spin_unlock(vmf->ptl);
1596
1597out:
1598 if (anon_vma)
1599 page_unlock_anon_vma_read(anon_vma);
1600
1601 if (page_nid != -1)
1602 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1603 flags);
1604
1605 return 0;
1606}
1607
1608/*
1609 * Return true if we do MADV_FREE successfully on entire pmd page.
1610 * Otherwise, return false.
1611 */
1612bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1613 pmd_t *pmd, unsigned long addr, unsigned long next)
1614{
1615 spinlock_t *ptl;
1616 pmd_t orig_pmd;
1617 struct page *page;
1618 struct mm_struct *mm = tlb->mm;
1619 bool ret = false;
1620
1621 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1622
1623 ptl = pmd_trans_huge_lock(pmd, vma);
1624 if (!ptl)
1625 goto out_unlocked;
1626
1627 orig_pmd = *pmd;
1628 if (is_huge_zero_pmd(orig_pmd))
1629 goto out;
1630
1631 if (unlikely(!pmd_present(orig_pmd))) {
1632 VM_BUG_ON(thp_migration_supported() &&
1633 !is_pmd_migration_entry(orig_pmd));
1634 goto out;
1635 }
1636
1637 page = pmd_page(orig_pmd);
1638 /*
1639 * If other processes are mapping this page, we couldn't discard
1640 * the page unless they all do MADV_FREE so let's skip the page.
1641 */
1642 if (page_mapcount(page) != 1)
1643 goto out;
1644
1645 if (!trylock_page(page))
1646 goto out;
1647
1648 /*
1649 * If user want to discard part-pages of THP, split it so MADV_FREE
1650 * will deactivate only them.
1651 */
1652 if (next - addr != HPAGE_PMD_SIZE) {
1653 get_page(page);
1654 spin_unlock(ptl);
1655 split_huge_page(page);
1656 unlock_page(page);
1657 put_page(page);
1658 goto out_unlocked;
1659 }
1660
1661 if (PageDirty(page))
1662 ClearPageDirty(page);
1663 unlock_page(page);
1664
1665 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1666 pmdp_invalidate(vma, addr, pmd);
1667 orig_pmd = pmd_mkold(orig_pmd);
1668 orig_pmd = pmd_mkclean(orig_pmd);
1669
1670 set_pmd_at(mm, addr, pmd, orig_pmd);
1671 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1672 }
1673
1674 mark_page_lazyfree(page);
1675 ret = true;
1676out:
1677 spin_unlock(ptl);
1678out_unlocked:
1679 return ret;
1680}
1681
1682static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1683{
1684 pgtable_t pgtable;
1685
1686 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1687 pte_free(mm, pgtable);
1688 mm_dec_nr_ptes(mm);
1689}
1690
1691int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1692 pmd_t *pmd, unsigned long addr)
1693{
1694 pmd_t orig_pmd;
1695 spinlock_t *ptl;
1696
1697 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1698
1699 ptl = __pmd_trans_huge_lock(pmd, vma);
1700 if (!ptl)
1701 return 0;
1702 /*
1703 * For architectures like ppc64 we look at deposited pgtable
1704 * when calling pmdp_huge_get_and_clear. So do the
1705 * pgtable_trans_huge_withdraw after finishing pmdp related
1706 * operations.
1707 */
1708 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1709 tlb->fullmm);
1710 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1711 if (vma_is_dax(vma)) {
1712 if (arch_needs_pgtable_deposit())
1713 zap_deposited_table(tlb->mm, pmd);
1714 spin_unlock(ptl);
1715 if (is_huge_zero_pmd(orig_pmd))
1716 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1717 } else if (is_huge_zero_pmd(orig_pmd)) {
1718 zap_deposited_table(tlb->mm, pmd);
1719 spin_unlock(ptl);
1720 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1721 } else {
1722 struct page *page = NULL;
1723 int flush_needed = 1;
1724
1725 if (pmd_present(orig_pmd)) {
1726 page = pmd_page(orig_pmd);
1727 page_remove_rmap(page, true);
1728 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1729 VM_BUG_ON_PAGE(!PageHead(page), page);
1730 } else if (thp_migration_supported()) {
1731 swp_entry_t entry;
1732
1733 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1734 entry = pmd_to_swp_entry(orig_pmd);
1735 page = pfn_to_page(swp_offset(entry));
1736 flush_needed = 0;
1737 } else
1738 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1739
1740 if (PageAnon(page)) {
1741 zap_deposited_table(tlb->mm, pmd);
1742 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1743 } else {
1744 if (arch_needs_pgtable_deposit())
1745 zap_deposited_table(tlb->mm, pmd);
1746 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1747 }
1748
1749 spin_unlock(ptl);
1750 if (flush_needed)
1751 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1752 }
1753 return 1;
1754}
1755
1756#ifndef pmd_move_must_withdraw
1757static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1758 spinlock_t *old_pmd_ptl,
1759 struct vm_area_struct *vma)
1760{
1761 /*
1762 * With split pmd lock we also need to move preallocated
1763 * PTE page table if new_pmd is on different PMD page table.
1764 *
1765 * We also don't deposit and withdraw tables for file pages.
1766 */
1767 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1768}
1769#endif
1770
1771static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1772{
1773#ifdef CONFIG_MEM_SOFT_DIRTY
1774 if (unlikely(is_pmd_migration_entry(pmd)))
1775 pmd = pmd_swp_mksoft_dirty(pmd);
1776 else if (pmd_present(pmd))
1777 pmd = pmd_mksoft_dirty(pmd);
1778#endif
1779 return pmd;
1780}
1781
1782bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1783 unsigned long new_addr, unsigned long old_end,
1784 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1785{
1786 spinlock_t *old_ptl, *new_ptl;
1787 pmd_t pmd;
1788 struct mm_struct *mm = vma->vm_mm;
1789 bool force_flush = false;
1790
1791 if ((old_addr & ~HPAGE_PMD_MASK) ||
1792 (new_addr & ~HPAGE_PMD_MASK) ||
1793 old_end - old_addr < HPAGE_PMD_SIZE)
1794 return false;
1795
1796 /*
1797 * The destination pmd shouldn't be established, free_pgtables()
1798 * should have release it.
1799 */
1800 if (WARN_ON(!pmd_none(*new_pmd))) {
1801 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1802 return false;
1803 }
1804
1805 /*
1806 * We don't have to worry about the ordering of src and dst
1807 * ptlocks because exclusive mmap_sem prevents deadlock.
1808 */
1809 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1810 if (old_ptl) {
1811 new_ptl = pmd_lockptr(mm, new_pmd);
1812 if (new_ptl != old_ptl)
1813 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1814 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1815 if (pmd_present(pmd) && pmd_dirty(pmd))
1816 force_flush = true;
1817 VM_BUG_ON(!pmd_none(*new_pmd));
1818
1819 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1820 pgtable_t pgtable;
1821 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1822 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1823 }
1824 pmd = move_soft_dirty_pmd(pmd);
1825 set_pmd_at(mm, new_addr, new_pmd, pmd);
1826 if (new_ptl != old_ptl)
1827 spin_unlock(new_ptl);
1828 if (force_flush)
1829 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1830 else
1831 *need_flush = true;
1832 spin_unlock(old_ptl);
1833 return true;
1834 }
1835 return false;
1836}
1837
1838/*
1839 * Returns
1840 * - 0 if PMD could not be locked
1841 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1842 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1843 */
1844int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1845 unsigned long addr, pgprot_t newprot, int prot_numa)
1846{
1847 struct mm_struct *mm = vma->vm_mm;
1848 spinlock_t *ptl;
1849 pmd_t entry;
1850 bool preserve_write;
1851 int ret;
1852
1853 ptl = __pmd_trans_huge_lock(pmd, vma);
1854 if (!ptl)
1855 return 0;
1856
1857 preserve_write = prot_numa && pmd_write(*pmd);
1858 ret = 1;
1859
1860#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1861 if (is_swap_pmd(*pmd)) {
1862 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1863
1864 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1865 if (is_write_migration_entry(entry)) {
1866 pmd_t newpmd;
1867 /*
1868 * A protection check is difficult so
1869 * just be safe and disable write
1870 */
1871 make_migration_entry_read(&entry);
1872 newpmd = swp_entry_to_pmd(entry);
1873 if (pmd_swp_soft_dirty(*pmd))
1874 newpmd = pmd_swp_mksoft_dirty(newpmd);
1875 set_pmd_at(mm, addr, pmd, newpmd);
1876 }
1877 goto unlock;
1878 }
1879#endif
1880
1881 /*
1882 * Avoid trapping faults against the zero page. The read-only
1883 * data is likely to be read-cached on the local CPU and
1884 * local/remote hits to the zero page are not interesting.
1885 */
1886 if (prot_numa && is_huge_zero_pmd(*pmd))
1887 goto unlock;
1888
1889 if (prot_numa && pmd_protnone(*pmd))
1890 goto unlock;
1891
1892 /*
1893 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1894 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1895 * which is also under down_read(mmap_sem):
1896 *
1897 * CPU0: CPU1:
1898 * change_huge_pmd(prot_numa=1)
1899 * pmdp_huge_get_and_clear_notify()
1900 * madvise_dontneed()
1901 * zap_pmd_range()
1902 * pmd_trans_huge(*pmd) == 0 (without ptl)
1903 * // skip the pmd
1904 * set_pmd_at();
1905 * // pmd is re-established
1906 *
1907 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1908 * which may break userspace.
1909 *
1910 * pmdp_invalidate() is required to make sure we don't miss
1911 * dirty/young flags set by hardware.
1912 */
1913 entry = pmdp_invalidate(vma, addr, pmd);
1914
1915 entry = pmd_modify(entry, newprot);
1916 if (preserve_write)
1917 entry = pmd_mk_savedwrite(entry);
1918 ret = HPAGE_PMD_NR;
1919 set_pmd_at(mm, addr, pmd, entry);
1920 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1921unlock:
1922 spin_unlock(ptl);
1923 return ret;
1924}
1925
1926/*
1927 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1928 *
1929 * Note that if it returns page table lock pointer, this routine returns without
1930 * unlocking page table lock. So callers must unlock it.
1931 */
1932spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1933{
1934 spinlock_t *ptl;
1935 ptl = pmd_lock(vma->vm_mm, pmd);
1936 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1937 pmd_devmap(*pmd)))
1938 return ptl;
1939 spin_unlock(ptl);
1940 return NULL;
1941}
1942
1943/*
1944 * Returns true if a given pud maps a thp, false otherwise.
1945 *
1946 * Note that if it returns true, this routine returns without unlocking page
1947 * table lock. So callers must unlock it.
1948 */
1949spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1950{
1951 spinlock_t *ptl;
1952
1953 ptl = pud_lock(vma->vm_mm, pud);
1954 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1955 return ptl;
1956 spin_unlock(ptl);
1957 return NULL;
1958}
1959
1960#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1961int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1962 pud_t *pud, unsigned long addr)
1963{
1964 pud_t orig_pud;
1965 spinlock_t *ptl;
1966
1967 ptl = __pud_trans_huge_lock(pud, vma);
1968 if (!ptl)
1969 return 0;
1970 /*
1971 * For architectures like ppc64 we look at deposited pgtable
1972 * when calling pudp_huge_get_and_clear. So do the
1973 * pgtable_trans_huge_withdraw after finishing pudp related
1974 * operations.
1975 */
1976 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1977 tlb->fullmm);
1978 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1979 if (vma_is_dax(vma)) {
1980 spin_unlock(ptl);
1981 /* No zero page support yet */
1982 } else {
1983 /* No support for anonymous PUD pages yet */
1984 BUG();
1985 }
1986 return 1;
1987}
1988
1989static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1990 unsigned long haddr)
1991{
1992 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1993 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1994 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1995 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1996
1997 count_vm_event(THP_SPLIT_PUD);
1998
1999 pudp_huge_clear_flush_notify(vma, haddr, pud);
2000}
2001
2002void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2003 unsigned long address)
2004{
2005 spinlock_t *ptl;
2006 struct mm_struct *mm = vma->vm_mm;
2007 unsigned long haddr = address & HPAGE_PUD_MASK;
2008
2009 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2010 ptl = pud_lock(mm, pud);
2011 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2012 goto out;
2013 __split_huge_pud_locked(vma, pud, haddr);
2014
2015out:
2016 spin_unlock(ptl);
2017 /*
2018 * No need to double call mmu_notifier->invalidate_range() callback as
2019 * the above pudp_huge_clear_flush_notify() did already call it.
2020 */
2021 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2022 HPAGE_PUD_SIZE);
2023}
2024#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2025
2026static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2027 unsigned long haddr, pmd_t *pmd)
2028{
2029 struct mm_struct *mm = vma->vm_mm;
2030 pgtable_t pgtable;
2031 pmd_t _pmd;
2032 int i;
2033
2034 /*
2035 * Leave pmd empty until pte is filled note that it is fine to delay
2036 * notification until mmu_notifier_invalidate_range_end() as we are
2037 * replacing a zero pmd write protected page with a zero pte write
2038 * protected page.
2039 *
2040 * See Documentation/vm/mmu_notifier.txt
2041 */
2042 pmdp_huge_clear_flush(vma, haddr, pmd);
2043
2044 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2045 pmd_populate(mm, &_pmd, pgtable);
2046
2047 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2048 pte_t *pte, entry;
2049 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2050 entry = pte_mkspecial(entry);
2051 pte = pte_offset_map(&_pmd, haddr);
2052 VM_BUG_ON(!pte_none(*pte));
2053 set_pte_at(mm, haddr, pte, entry);
2054 pte_unmap(pte);
2055 }
2056 smp_wmb(); /* make pte visible before pmd */
2057 pmd_populate(mm, pmd, pgtable);
2058}
2059
2060static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2061 unsigned long haddr, bool freeze)
2062{
2063 struct mm_struct *mm = vma->vm_mm;
2064 struct page *page;
2065 pgtable_t pgtable;
2066 pmd_t old_pmd, _pmd;
2067 bool young, write, soft_dirty, pmd_migration = false;
2068 unsigned long addr;
2069 int i;
2070
2071 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2072 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2073 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2074 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2075 && !pmd_devmap(*pmd));
2076
2077 count_vm_event(THP_SPLIT_PMD);
2078
2079 if (!vma_is_anonymous(vma)) {
2080 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2081 /*
2082 * We are going to unmap this huge page. So
2083 * just go ahead and zap it
2084 */
2085 if (arch_needs_pgtable_deposit())
2086 zap_deposited_table(mm, pmd);
2087 if (vma_is_dax(vma))
2088 return;
2089 page = pmd_page(_pmd);
2090 if (!PageReferenced(page) && pmd_young(_pmd))
2091 SetPageReferenced(page);
2092 page_remove_rmap(page, true);
2093 put_page(page);
2094 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2095 return;
2096 } else if (is_huge_zero_pmd(*pmd)) {
2097 /*
2098 * FIXME: Do we want to invalidate secondary mmu by calling
2099 * mmu_notifier_invalidate_range() see comments below inside
2100 * __split_huge_pmd() ?
2101 *
2102 * We are going from a zero huge page write protected to zero
2103 * small page also write protected so it does not seems useful
2104 * to invalidate secondary mmu at this time.
2105 */
2106 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2107 }
2108
2109 /*
2110 * Up to this point the pmd is present and huge and userland has the
2111 * whole access to the hugepage during the split (which happens in
2112 * place). If we overwrite the pmd with the not-huge version pointing
2113 * to the pte here (which of course we could if all CPUs were bug
2114 * free), userland could trigger a small page size TLB miss on the
2115 * small sized TLB while the hugepage TLB entry is still established in
2116 * the huge TLB. Some CPU doesn't like that.
2117 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2118 * 383 on page 93. Intel should be safe but is also warns that it's
2119 * only safe if the permission and cache attributes of the two entries
2120 * loaded in the two TLB is identical (which should be the case here).
2121 * But it is generally safer to never allow small and huge TLB entries
2122 * for the same virtual address to be loaded simultaneously. So instead
2123 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2124 * current pmd notpresent (atomically because here the pmd_trans_huge
2125 * must remain set at all times on the pmd until the split is complete
2126 * for this pmd), then we flush the SMP TLB and finally we write the
2127 * non-huge version of the pmd entry with pmd_populate.
2128 */
2129 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2130
2131#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2132 pmd_migration = is_pmd_migration_entry(old_pmd);
2133 if (pmd_migration) {
2134 swp_entry_t entry;
2135
2136 entry = pmd_to_swp_entry(old_pmd);
2137 page = pfn_to_page(swp_offset(entry));
2138 } else
2139#endif
2140 page = pmd_page(old_pmd);
2141 VM_BUG_ON_PAGE(!page_count(page), page);
2142 page_ref_add(page, HPAGE_PMD_NR - 1);
2143 if (pmd_dirty(old_pmd))
2144 SetPageDirty(page);
2145 write = pmd_write(old_pmd);
2146 young = pmd_young(old_pmd);
2147 soft_dirty = pmd_soft_dirty(old_pmd);
2148
2149 /*
2150 * Withdraw the table only after we mark the pmd entry invalid.
2151 * This's critical for some architectures (Power).
2152 */
2153 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2154 pmd_populate(mm, &_pmd, pgtable);
2155
2156 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2157 pte_t entry, *pte;
2158 /*
2159 * Note that NUMA hinting access restrictions are not
2160 * transferred to avoid any possibility of altering
2161 * permissions across VMAs.
2162 */
2163 if (freeze || pmd_migration) {
2164 swp_entry_t swp_entry;
2165 swp_entry = make_migration_entry(page + i, write);
2166 entry = swp_entry_to_pte(swp_entry);
2167 if (soft_dirty)
2168 entry = pte_swp_mksoft_dirty(entry);
2169 } else {
2170 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2171 entry = maybe_mkwrite(entry, vma);
2172 if (!write)
2173 entry = pte_wrprotect(entry);
2174 if (!young)
2175 entry = pte_mkold(entry);
2176 if (soft_dirty)
2177 entry = pte_mksoft_dirty(entry);
2178 }
2179 pte = pte_offset_map(&_pmd, addr);
2180 BUG_ON(!pte_none(*pte));
2181 set_pte_at(mm, addr, pte, entry);
2182 atomic_inc(&page[i]._mapcount);
2183 pte_unmap(pte);
2184 }
2185
2186 /*
2187 * Set PG_double_map before dropping compound_mapcount to avoid
2188 * false-negative page_mapped().
2189 */
2190 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2191 for (i = 0; i < HPAGE_PMD_NR; i++)
2192 atomic_inc(&page[i]._mapcount);
2193 }
2194
2195 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2196 /* Last compound_mapcount is gone. */
2197 __dec_node_page_state(page, NR_ANON_THPS);
2198 if (TestClearPageDoubleMap(page)) {
2199 /* No need in mapcount reference anymore */
2200 for (i = 0; i < HPAGE_PMD_NR; i++)
2201 atomic_dec(&page[i]._mapcount);
2202 }
2203 }
2204
2205 smp_wmb(); /* make pte visible before pmd */
2206 pmd_populate(mm, pmd, pgtable);
2207
2208 if (freeze) {
2209 for (i = 0; i < HPAGE_PMD_NR; i++) {
2210 page_remove_rmap(page + i, false);
2211 put_page(page + i);
2212 }
2213 }
2214}
2215
2216void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2217 unsigned long address, bool freeze, struct page *page)
2218{
2219 spinlock_t *ptl;
2220 struct mm_struct *mm = vma->vm_mm;
2221 unsigned long haddr = address & HPAGE_PMD_MASK;
2222
2223 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2224 ptl = pmd_lock(mm, pmd);
2225
2226 /*
2227 * If caller asks to setup a migration entries, we need a page to check
2228 * pmd against. Otherwise we can end up replacing wrong page.
2229 */
2230 VM_BUG_ON(freeze && !page);
2231 if (page && page != pmd_page(*pmd))
2232 goto out;
2233
2234 if (pmd_trans_huge(*pmd)) {
2235 page = pmd_page(*pmd);
2236 if (PageMlocked(page))
2237 clear_page_mlock(page);
2238 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2239 goto out;
2240 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2241out:
2242 spin_unlock(ptl);
2243 /*
2244 * No need to double call mmu_notifier->invalidate_range() callback.
2245 * They are 3 cases to consider inside __split_huge_pmd_locked():
2246 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2247 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2248 * fault will trigger a flush_notify before pointing to a new page
2249 * (it is fine if the secondary mmu keeps pointing to the old zero
2250 * page in the meantime)
2251 * 3) Split a huge pmd into pte pointing to the same page. No need
2252 * to invalidate secondary tlb entry they are all still valid.
2253 * any further changes to individual pte will notify. So no need
2254 * to call mmu_notifier->invalidate_range()
2255 */
2256 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2257 HPAGE_PMD_SIZE);
2258}
2259
2260void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2261 bool freeze, struct page *page)
2262{
2263 pgd_t *pgd;
2264 p4d_t *p4d;
2265 pud_t *pud;
2266 pmd_t *pmd;
2267
2268 pgd = pgd_offset(vma->vm_mm, address);
2269 if (!pgd_present(*pgd))
2270 return;
2271
2272 p4d = p4d_offset(pgd, address);
2273 if (!p4d_present(*p4d))
2274 return;
2275
2276 pud = pud_offset(p4d, address);
2277 if (!pud_present(*pud))
2278 return;
2279
2280 pmd = pmd_offset(pud, address);
2281
2282 __split_huge_pmd(vma, pmd, address, freeze, page);
2283}
2284
2285void vma_adjust_trans_huge(struct vm_area_struct *vma,
2286 unsigned long start,
2287 unsigned long end,
2288 long adjust_next)
2289{
2290 /*
2291 * If the new start address isn't hpage aligned and it could
2292 * previously contain an hugepage: check if we need to split
2293 * an huge pmd.
2294 */
2295 if (start & ~HPAGE_PMD_MASK &&
2296 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2297 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2298 split_huge_pmd_address(vma, start, false, NULL);
2299
2300 /*
2301 * If the new end address isn't hpage aligned and it could
2302 * previously contain an hugepage: check if we need to split
2303 * an huge pmd.
2304 */
2305 if (end & ~HPAGE_PMD_MASK &&
2306 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2307 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2308 split_huge_pmd_address(vma, end, false, NULL);
2309
2310 /*
2311 * If we're also updating the vma->vm_next->vm_start, if the new
2312 * vm_next->vm_start isn't page aligned and it could previously
2313 * contain an hugepage: check if we need to split an huge pmd.
2314 */
2315 if (adjust_next > 0) {
2316 struct vm_area_struct *next = vma->vm_next;
2317 unsigned long nstart = next->vm_start;
2318 nstart += adjust_next << PAGE_SHIFT;
2319 if (nstart & ~HPAGE_PMD_MASK &&
2320 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2321 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2322 split_huge_pmd_address(next, nstart, false, NULL);
2323 }
2324}
2325
2326static void freeze_page(struct page *page)
2327{
2328 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2329 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2330 bool unmap_success;
2331
2332 VM_BUG_ON_PAGE(!PageHead(page), page);
2333
2334 if (PageAnon(page))
2335 ttu_flags |= TTU_SPLIT_FREEZE;
2336
2337 unmap_success = try_to_unmap(page, ttu_flags);
2338 VM_BUG_ON_PAGE(!unmap_success, page);
2339}
2340
2341static void unfreeze_page(struct page *page)
2342{
2343 int i;
2344 if (PageTransHuge(page)) {
2345 remove_migration_ptes(page, page, true);
2346 } else {
2347 for (i = 0; i < HPAGE_PMD_NR; i++)
2348 remove_migration_ptes(page + i, page + i, true);
2349 }
2350}
2351
2352static void __split_huge_page_tail(struct page *head, int tail,
2353 struct lruvec *lruvec, struct list_head *list)
2354{
2355 struct page *page_tail = head + tail;
2356
2357 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2358
2359 /*
2360 * Clone page flags before unfreezing refcount.
2361 *
2362 * After successful get_page_unless_zero() might follow flags change,
2363 * for exmaple lock_page() which set PG_waiters.
2364 */
2365 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2366 page_tail->flags |= (head->flags &
2367 ((1L << PG_referenced) |
2368 (1L << PG_swapbacked) |
2369 (1L << PG_swapcache) |
2370 (1L << PG_mlocked) |
2371 (1L << PG_uptodate) |
2372 (1L << PG_active) |
2373 (1L << PG_locked) |
2374 (1L << PG_unevictable) |
2375 (1L << PG_dirty)));
2376
2377 /* Page flags must be visible before we make the page non-compound. */
2378 smp_wmb();
2379
2380 /*
2381 * Clear PageTail before unfreezing page refcount.
2382 *
2383 * After successful get_page_unless_zero() might follow put_page()
2384 * which needs correct compound_head().
2385 */
2386 clear_compound_head(page_tail);
2387
2388 /* Finally unfreeze refcount. Additional reference from page cache. */
2389 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2390 PageSwapCache(head)));
2391
2392 if (page_is_young(head))
2393 set_page_young(page_tail);
2394 if (page_is_idle(head))
2395 set_page_idle(page_tail);
2396
2397 /* ->mapping in first tail page is compound_mapcount */
2398 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2399 page_tail);
2400 page_tail->mapping = head->mapping;
2401
2402 page_tail->index = head->index + tail;
2403 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2404
2405 /*
2406 * always add to the tail because some iterators expect new
2407 * pages to show after the currently processed elements - e.g.
2408 * migrate_pages
2409 */
2410 lru_add_page_tail(head, page_tail, lruvec, list);
2411}
2412
2413static void __split_huge_page(struct page *page, struct list_head *list,
2414 unsigned long flags)
2415{
2416 struct page *head = compound_head(page);
2417 struct zone *zone = page_zone(head);
2418 struct lruvec *lruvec;
2419 pgoff_t end = -1;
2420 int i;
2421
2422 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2423
2424 /* complete memcg works before add pages to LRU */
2425 mem_cgroup_split_huge_fixup(head);
2426
2427 if (!PageAnon(page))
2428 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2429
2430 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2431 __split_huge_page_tail(head, i, lruvec, list);
2432 /* Some pages can be beyond i_size: drop them from page cache */
2433 if (head[i].index >= end) {
2434 ClearPageDirty(head + i);
2435 __delete_from_page_cache(head + i, NULL);
2436 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2437 shmem_uncharge(head->mapping->host, 1);
2438 put_page(head + i);
2439 }
2440 }
2441
2442 ClearPageCompound(head);
2443 /* See comment in __split_huge_page_tail() */
2444 if (PageAnon(head)) {
2445 /* Additional pin to radix tree of swap cache */
2446 if (PageSwapCache(head))
2447 page_ref_add(head, 2);
2448 else
2449 page_ref_inc(head);
2450 } else {
2451 /* Additional pin to radix tree */
2452 page_ref_add(head, 2);
2453 xa_unlock(&head->mapping->i_pages);
2454 }
2455
2456 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2457
2458 unfreeze_page(head);
2459
2460 for (i = 0; i < HPAGE_PMD_NR; i++) {
2461 struct page *subpage = head + i;
2462 if (subpage == page)
2463 continue;
2464 unlock_page(subpage);
2465
2466 /*
2467 * Subpages may be freed if there wasn't any mapping
2468 * like if add_to_swap() is running on a lru page that
2469 * had its mapping zapped. And freeing these pages
2470 * requires taking the lru_lock so we do the put_page
2471 * of the tail pages after the split is complete.
2472 */
2473 put_page(subpage);
2474 }
2475}
2476
2477int total_mapcount(struct page *page)
2478{
2479 int i, compound, ret;
2480
2481 VM_BUG_ON_PAGE(PageTail(page), page);
2482
2483 if (likely(!PageCompound(page)))
2484 return atomic_read(&page->_mapcount) + 1;
2485
2486 compound = compound_mapcount(page);
2487 if (PageHuge(page))
2488 return compound;
2489 ret = compound;
2490 for (i = 0; i < HPAGE_PMD_NR; i++)
2491 ret += atomic_read(&page[i]._mapcount) + 1;
2492 /* File pages has compound_mapcount included in _mapcount */
2493 if (!PageAnon(page))
2494 return ret - compound * HPAGE_PMD_NR;
2495 if (PageDoubleMap(page))
2496 ret -= HPAGE_PMD_NR;
2497 return ret;
2498}
2499
2500/*
2501 * This calculates accurately how many mappings a transparent hugepage
2502 * has (unlike page_mapcount() which isn't fully accurate). This full
2503 * accuracy is primarily needed to know if copy-on-write faults can
2504 * reuse the page and change the mapping to read-write instead of
2505 * copying them. At the same time this returns the total_mapcount too.
2506 *
2507 * The function returns the highest mapcount any one of the subpages
2508 * has. If the return value is one, even if different processes are
2509 * mapping different subpages of the transparent hugepage, they can
2510 * all reuse it, because each process is reusing a different subpage.
2511 *
2512 * The total_mapcount is instead counting all virtual mappings of the
2513 * subpages. If the total_mapcount is equal to "one", it tells the
2514 * caller all mappings belong to the same "mm" and in turn the
2515 * anon_vma of the transparent hugepage can become the vma->anon_vma
2516 * local one as no other process may be mapping any of the subpages.
2517 *
2518 * It would be more accurate to replace page_mapcount() with
2519 * page_trans_huge_mapcount(), however we only use
2520 * page_trans_huge_mapcount() in the copy-on-write faults where we
2521 * need full accuracy to avoid breaking page pinning, because
2522 * page_trans_huge_mapcount() is slower than page_mapcount().
2523 */
2524int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2525{
2526 int i, ret, _total_mapcount, mapcount;
2527
2528 /* hugetlbfs shouldn't call it */
2529 VM_BUG_ON_PAGE(PageHuge(page), page);
2530
2531 if (likely(!PageTransCompound(page))) {
2532 mapcount = atomic_read(&page->_mapcount) + 1;
2533 if (total_mapcount)
2534 *total_mapcount = mapcount;
2535 return mapcount;
2536 }
2537
2538 page = compound_head(page);
2539
2540 _total_mapcount = ret = 0;
2541 for (i = 0; i < HPAGE_PMD_NR; i++) {
2542 mapcount = atomic_read(&page[i]._mapcount) + 1;
2543 ret = max(ret, mapcount);
2544 _total_mapcount += mapcount;
2545 }
2546 if (PageDoubleMap(page)) {
2547 ret -= 1;
2548 _total_mapcount -= HPAGE_PMD_NR;
2549 }
2550 mapcount = compound_mapcount(page);
2551 ret += mapcount;
2552 _total_mapcount += mapcount;
2553 if (total_mapcount)
2554 *total_mapcount = _total_mapcount;
2555 return ret;
2556}
2557
2558/* Racy check whether the huge page can be split */
2559bool can_split_huge_page(struct page *page, int *pextra_pins)
2560{
2561 int extra_pins;
2562
2563 /* Additional pins from radix tree */
2564 if (PageAnon(page))
2565 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2566 else
2567 extra_pins = HPAGE_PMD_NR;
2568 if (pextra_pins)
2569 *pextra_pins = extra_pins;
2570 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2571}
2572
2573/*
2574 * This function splits huge page into normal pages. @page can point to any
2575 * subpage of huge page to split. Split doesn't change the position of @page.
2576 *
2577 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2578 * The huge page must be locked.
2579 *
2580 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2581 *
2582 * Both head page and tail pages will inherit mapping, flags, and so on from
2583 * the hugepage.
2584 *
2585 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2586 * they are not mapped.
2587 *
2588 * Returns 0 if the hugepage is split successfully.
2589 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2590 * us.
2591 */
2592int split_huge_page_to_list(struct page *page, struct list_head *list)
2593{
2594 struct page *head = compound_head(page);
2595 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2596 struct anon_vma *anon_vma = NULL;
2597 struct address_space *mapping = NULL;
2598 int count, mapcount, extra_pins, ret;
2599 bool mlocked;
2600 unsigned long flags;
2601
2602 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2603 VM_BUG_ON_PAGE(!PageLocked(page), page);
2604 VM_BUG_ON_PAGE(!PageCompound(page), page);
2605
2606 if (PageWriteback(page))
2607 return -EBUSY;
2608
2609 if (PageAnon(head)) {
2610 /*
2611 * The caller does not necessarily hold an mmap_sem that would
2612 * prevent the anon_vma disappearing so we first we take a
2613 * reference to it and then lock the anon_vma for write. This
2614 * is similar to page_lock_anon_vma_read except the write lock
2615 * is taken to serialise against parallel split or collapse
2616 * operations.
2617 */
2618 anon_vma = page_get_anon_vma(head);
2619 if (!anon_vma) {
2620 ret = -EBUSY;
2621 goto out;
2622 }
2623 mapping = NULL;
2624 anon_vma_lock_write(anon_vma);
2625 } else {
2626 mapping = head->mapping;
2627
2628 /* Truncated ? */
2629 if (!mapping) {
2630 ret = -EBUSY;
2631 goto out;
2632 }
2633
2634 anon_vma = NULL;
2635 i_mmap_lock_read(mapping);
2636 }
2637
2638 /*
2639 * Racy check if we can split the page, before freeze_page() will
2640 * split PMDs
2641 */
2642 if (!can_split_huge_page(head, &extra_pins)) {
2643 ret = -EBUSY;
2644 goto out_unlock;
2645 }
2646
2647 mlocked = PageMlocked(page);
2648 freeze_page(head);
2649 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2650
2651 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2652 if (mlocked)
2653 lru_add_drain();
2654
2655 /* prevent PageLRU to go away from under us, and freeze lru stats */
2656 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2657
2658 if (mapping) {
2659 void **pslot;
2660
2661 xa_lock(&mapping->i_pages);
2662 pslot = radix_tree_lookup_slot(&mapping->i_pages,
2663 page_index(head));
2664 /*
2665 * Check if the head page is present in radix tree.
2666 * We assume all tail are present too, if head is there.
2667 */
2668 if (radix_tree_deref_slot_protected(pslot,
2669 &mapping->i_pages.xa_lock) != head)
2670 goto fail;
2671 }
2672
2673 /* Prevent deferred_split_scan() touching ->_refcount */
2674 spin_lock(&pgdata->split_queue_lock);
2675 count = page_count(head);
2676 mapcount = total_mapcount(head);
2677 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2678 if (!list_empty(page_deferred_list(head))) {
2679 pgdata->split_queue_len--;
2680 list_del(page_deferred_list(head));
2681 }
2682 if (mapping)
2683 __dec_node_page_state(page, NR_SHMEM_THPS);
2684 spin_unlock(&pgdata->split_queue_lock);
2685 __split_huge_page(page, list, flags);
2686 if (PageSwapCache(head)) {
2687 swp_entry_t entry = { .val = page_private(head) };
2688
2689 ret = split_swap_cluster(entry);
2690 } else
2691 ret = 0;
2692 } else {
2693 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2694 pr_alert("total_mapcount: %u, page_count(): %u\n",
2695 mapcount, count);
2696 if (PageTail(page))
2697 dump_page(head, NULL);
2698 dump_page(page, "total_mapcount(head) > 0");
2699 BUG();
2700 }
2701 spin_unlock(&pgdata->split_queue_lock);
2702fail: if (mapping)
2703 xa_unlock(&mapping->i_pages);
2704 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2705 unfreeze_page(head);
2706 ret = -EBUSY;
2707 }
2708
2709out_unlock:
2710 if (anon_vma) {
2711 anon_vma_unlock_write(anon_vma);
2712 put_anon_vma(anon_vma);
2713 }
2714 if (mapping)
2715 i_mmap_unlock_read(mapping);
2716out:
2717 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2718 return ret;
2719}
2720
2721void free_transhuge_page(struct page *page)
2722{
2723 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2724 unsigned long flags;
2725
2726 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2727 if (!list_empty(page_deferred_list(page))) {
2728 pgdata->split_queue_len--;
2729 list_del(page_deferred_list(page));
2730 }
2731 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2732 free_compound_page(page);
2733}
2734
2735void deferred_split_huge_page(struct page *page)
2736{
2737 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2738 unsigned long flags;
2739
2740 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2741
2742 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2743 if (list_empty(page_deferred_list(page))) {
2744 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2745 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2746 pgdata->split_queue_len++;
2747 }
2748 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2749}
2750
2751static unsigned long deferred_split_count(struct shrinker *shrink,
2752 struct shrink_control *sc)
2753{
2754 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2755 return READ_ONCE(pgdata->split_queue_len);
2756}
2757
2758static unsigned long deferred_split_scan(struct shrinker *shrink,
2759 struct shrink_control *sc)
2760{
2761 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2762 unsigned long flags;
2763 LIST_HEAD(list), *pos, *next;
2764 struct page *page;
2765 int split = 0;
2766
2767 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2768 /* Take pin on all head pages to avoid freeing them under us */
2769 list_for_each_safe(pos, next, &pgdata->split_queue) {
2770 page = list_entry((void *)pos, struct page, mapping);
2771 page = compound_head(page);
2772 if (get_page_unless_zero(page)) {
2773 list_move(page_deferred_list(page), &list);
2774 } else {
2775 /* We lost race with put_compound_page() */
2776 list_del_init(page_deferred_list(page));
2777 pgdata->split_queue_len--;
2778 }
2779 if (!--sc->nr_to_scan)
2780 break;
2781 }
2782 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2783
2784 list_for_each_safe(pos, next, &list) {
2785 page = list_entry((void *)pos, struct page, mapping);
2786 if (!trylock_page(page))
2787 goto next;
2788 /* split_huge_page() removes page from list on success */
2789 if (!split_huge_page(page))
2790 split++;
2791 unlock_page(page);
2792next:
2793 put_page(page);
2794 }
2795
2796 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2797 list_splice_tail(&list, &pgdata->split_queue);
2798 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2799
2800 /*
2801 * Stop shrinker if we didn't split any page, but the queue is empty.
2802 * This can happen if pages were freed under us.
2803 */
2804 if (!split && list_empty(&pgdata->split_queue))
2805 return SHRINK_STOP;
2806 return split;
2807}
2808
2809static struct shrinker deferred_split_shrinker = {
2810 .count_objects = deferred_split_count,
2811 .scan_objects = deferred_split_scan,
2812 .seeks = DEFAULT_SEEKS,
2813 .flags = SHRINKER_NUMA_AWARE,
2814};
2815
2816#ifdef CONFIG_DEBUG_FS
2817static int split_huge_pages_set(void *data, u64 val)
2818{
2819 struct zone *zone;
2820 struct page *page;
2821 unsigned long pfn, max_zone_pfn;
2822 unsigned long total = 0, split = 0;
2823
2824 if (val != 1)
2825 return -EINVAL;
2826
2827 for_each_populated_zone(zone) {
2828 max_zone_pfn = zone_end_pfn(zone);
2829 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2830 if (!pfn_valid(pfn))
2831 continue;
2832
2833 page = pfn_to_page(pfn);
2834 if (!get_page_unless_zero(page))
2835 continue;
2836
2837 if (zone != page_zone(page))
2838 goto next;
2839
2840 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2841 goto next;
2842
2843 total++;
2844 lock_page(page);
2845 if (!split_huge_page(page))
2846 split++;
2847 unlock_page(page);
2848next:
2849 put_page(page);
2850 }
2851 }
2852
2853 pr_info("%lu of %lu THP split\n", split, total);
2854
2855 return 0;
2856}
2857DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2858 "%llu\n");
2859
2860static int __init split_huge_pages_debugfs(void)
2861{
2862 void *ret;
2863
2864 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2865 &split_huge_pages_fops);
2866 if (!ret)
2867 pr_warn("Failed to create split_huge_pages in debugfs");
2868 return 0;
2869}
2870late_initcall(split_huge_pages_debugfs);
2871#endif
2872
2873#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2874void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2875 struct page *page)
2876{
2877 struct vm_area_struct *vma = pvmw->vma;
2878 struct mm_struct *mm = vma->vm_mm;
2879 unsigned long address = pvmw->address;
2880 pmd_t pmdval;
2881 swp_entry_t entry;
2882 pmd_t pmdswp;
2883
2884 if (!(pvmw->pmd && !pvmw->pte))
2885 return;
2886
2887 mmu_notifier_invalidate_range_start(mm, address,
2888 address + HPAGE_PMD_SIZE);
2889
2890 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2891 pmdval = *pvmw->pmd;
2892 pmdp_invalidate(vma, address, pvmw->pmd);
2893 if (pmd_dirty(pmdval))
2894 set_page_dirty(page);
2895 entry = make_migration_entry(page, pmd_write(pmdval));
2896 pmdswp = swp_entry_to_pmd(entry);
2897 if (pmd_soft_dirty(pmdval))
2898 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2899 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2900 page_remove_rmap(page, true);
2901 put_page(page);
2902
2903 mmu_notifier_invalidate_range_end(mm, address,
2904 address + HPAGE_PMD_SIZE);
2905}
2906
2907void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2908{
2909 struct vm_area_struct *vma = pvmw->vma;
2910 struct mm_struct *mm = vma->vm_mm;
2911 unsigned long address = pvmw->address;
2912 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2913 pmd_t pmde;
2914 swp_entry_t entry;
2915
2916 if (!(pvmw->pmd && !pvmw->pte))
2917 return;
2918
2919 entry = pmd_to_swp_entry(*pvmw->pmd);
2920 get_page(new);
2921 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2922 if (pmd_swp_soft_dirty(*pvmw->pmd))
2923 pmde = pmd_mksoft_dirty(pmde);
2924 if (is_write_migration_entry(entry))
2925 pmde = maybe_pmd_mkwrite(pmde, vma);
2926
2927 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2928 if (PageAnon(new))
2929 page_add_anon_rmap(new, vma, mmun_start, true);
2930 else
2931 page_add_file_rmap(new, true);
2932 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2933 if (vma->vm_flags & VM_LOCKED)
2934 mlock_vma_page(new);
2935 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2936}
2937#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2009 Red Hat, Inc.
4 */
5
6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/sched/mm.h>
11#include <linux/sched/coredump.h>
12#include <linux/sched/numa_balancing.h>
13#include <linux/highmem.h>
14#include <linux/hugetlb.h>
15#include <linux/mmu_notifier.h>
16#include <linux/rmap.h>
17#include <linux/swap.h>
18#include <linux/shrinker.h>
19#include <linux/mm_inline.h>
20#include <linux/swapops.h>
21#include <linux/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#include <linux/compat.h>
41
42#include <asm/tlb.h>
43#include <asm/pgalloc.h>
44#include "internal.h"
45#include "swap.h"
46
47#define CREATE_TRACE_POINTS
48#include <trace/events/thp.h>
49
50/*
51 * By default, transparent hugepage support is disabled in order to avoid
52 * risking an increased memory footprint for applications that are not
53 * guaranteed to benefit from it. When transparent hugepage support is
54 * enabled, it is for all mappings, and khugepaged scans all mappings.
55 * Defrag is invoked by khugepaged hugepage allocations and by page faults
56 * for all hugepage allocations.
57 */
58unsigned long transparent_hugepage_flags __read_mostly =
59#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
60 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
61#endif
62#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
63 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
64#endif
65 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
66 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
67 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
68
69static struct shrinker *deferred_split_shrinker;
70static unsigned long deferred_split_count(struct shrinker *shrink,
71 struct shrink_control *sc);
72static unsigned long deferred_split_scan(struct shrinker *shrink,
73 struct shrink_control *sc);
74
75static atomic_t huge_zero_refcount;
76struct page *huge_zero_page __read_mostly;
77unsigned long huge_zero_pfn __read_mostly = ~0UL;
78unsigned long huge_anon_orders_always __read_mostly;
79unsigned long huge_anon_orders_madvise __read_mostly;
80unsigned long huge_anon_orders_inherit __read_mostly;
81
82unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
83 unsigned long vm_flags, bool smaps,
84 bool in_pf, bool enforce_sysfs,
85 unsigned long orders)
86{
87 /* Check the intersection of requested and supported orders. */
88 orders &= vma_is_anonymous(vma) ?
89 THP_ORDERS_ALL_ANON : THP_ORDERS_ALL_FILE;
90 if (!orders)
91 return 0;
92
93 if (!vma->vm_mm) /* vdso */
94 return 0;
95
96 /*
97 * Explicitly disabled through madvise or prctl, or some
98 * architectures may disable THP for some mappings, for
99 * example, s390 kvm.
100 * */
101 if ((vm_flags & VM_NOHUGEPAGE) ||
102 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
103 return 0;
104 /*
105 * If the hardware/firmware marked hugepage support disabled.
106 */
107 if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED))
108 return 0;
109
110 /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
111 if (vma_is_dax(vma))
112 return in_pf ? orders : 0;
113
114 /*
115 * khugepaged special VMA and hugetlb VMA.
116 * Must be checked after dax since some dax mappings may have
117 * VM_MIXEDMAP set.
118 */
119 if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED))
120 return 0;
121
122 /*
123 * Check alignment for file vma and size for both file and anon vma by
124 * filtering out the unsuitable orders.
125 *
126 * Skip the check for page fault. Huge fault does the check in fault
127 * handlers.
128 */
129 if (!in_pf) {
130 int order = highest_order(orders);
131 unsigned long addr;
132
133 while (orders) {
134 addr = vma->vm_end - (PAGE_SIZE << order);
135 if (thp_vma_suitable_order(vma, addr, order))
136 break;
137 order = next_order(&orders, order);
138 }
139
140 if (!orders)
141 return 0;
142 }
143
144 /*
145 * Enabled via shmem mount options or sysfs settings.
146 * Must be done before hugepage flags check since shmem has its
147 * own flags.
148 */
149 if (!in_pf && shmem_file(vma->vm_file))
150 return shmem_is_huge(file_inode(vma->vm_file), vma->vm_pgoff,
151 !enforce_sysfs, vma->vm_mm, vm_flags)
152 ? orders : 0;
153
154 if (!vma_is_anonymous(vma)) {
155 /*
156 * Enforce sysfs THP requirements as necessary. Anonymous vmas
157 * were already handled in thp_vma_allowable_orders().
158 */
159 if (enforce_sysfs &&
160 (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
161 !hugepage_global_always())))
162 return 0;
163
164 /*
165 * Trust that ->huge_fault() handlers know what they are doing
166 * in fault path.
167 */
168 if (((in_pf || smaps)) && vma->vm_ops->huge_fault)
169 return orders;
170 /* Only regular file is valid in collapse path */
171 if (((!in_pf || smaps)) && file_thp_enabled(vma))
172 return orders;
173 return 0;
174 }
175
176 if (vma_is_temporary_stack(vma))
177 return 0;
178
179 /*
180 * THPeligible bit of smaps should show 1 for proper VMAs even
181 * though anon_vma is not initialized yet.
182 *
183 * Allow page fault since anon_vma may be not initialized until
184 * the first page fault.
185 */
186 if (!vma->anon_vma)
187 return (smaps || in_pf) ? orders : 0;
188
189 return orders;
190}
191
192static bool get_huge_zero_page(void)
193{
194 struct page *zero_page;
195retry:
196 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
197 return true;
198
199 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
200 HPAGE_PMD_ORDER);
201 if (!zero_page) {
202 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
203 return false;
204 }
205 preempt_disable();
206 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
207 preempt_enable();
208 __free_pages(zero_page, compound_order(zero_page));
209 goto retry;
210 }
211 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
212
213 /* We take additional reference here. It will be put back by shrinker */
214 atomic_set(&huge_zero_refcount, 2);
215 preempt_enable();
216 count_vm_event(THP_ZERO_PAGE_ALLOC);
217 return true;
218}
219
220static void put_huge_zero_page(void)
221{
222 /*
223 * Counter should never go to zero here. Only shrinker can put
224 * last reference.
225 */
226 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
227}
228
229struct page *mm_get_huge_zero_page(struct mm_struct *mm)
230{
231 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
232 return READ_ONCE(huge_zero_page);
233
234 if (!get_huge_zero_page())
235 return NULL;
236
237 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
238 put_huge_zero_page();
239
240 return READ_ONCE(huge_zero_page);
241}
242
243void mm_put_huge_zero_page(struct mm_struct *mm)
244{
245 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
246 put_huge_zero_page();
247}
248
249static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
250 struct shrink_control *sc)
251{
252 /* we can free zero page only if last reference remains */
253 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
254}
255
256static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
257 struct shrink_control *sc)
258{
259 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
260 struct page *zero_page = xchg(&huge_zero_page, NULL);
261 BUG_ON(zero_page == NULL);
262 WRITE_ONCE(huge_zero_pfn, ~0UL);
263 __free_pages(zero_page, compound_order(zero_page));
264 return HPAGE_PMD_NR;
265 }
266
267 return 0;
268}
269
270static struct shrinker *huge_zero_page_shrinker;
271
272#ifdef CONFIG_SYSFS
273static ssize_t enabled_show(struct kobject *kobj,
274 struct kobj_attribute *attr, char *buf)
275{
276 const char *output;
277
278 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
279 output = "[always] madvise never";
280 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
281 &transparent_hugepage_flags))
282 output = "always [madvise] never";
283 else
284 output = "always madvise [never]";
285
286 return sysfs_emit(buf, "%s\n", output);
287}
288
289static ssize_t enabled_store(struct kobject *kobj,
290 struct kobj_attribute *attr,
291 const char *buf, size_t count)
292{
293 ssize_t ret = count;
294
295 if (sysfs_streq(buf, "always")) {
296 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
297 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
298 } else if (sysfs_streq(buf, "madvise")) {
299 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
300 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
301 } else if (sysfs_streq(buf, "never")) {
302 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
303 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
304 } else
305 ret = -EINVAL;
306
307 if (ret > 0) {
308 int err = start_stop_khugepaged();
309 if (err)
310 ret = err;
311 }
312 return ret;
313}
314
315static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
316
317ssize_t single_hugepage_flag_show(struct kobject *kobj,
318 struct kobj_attribute *attr, char *buf,
319 enum transparent_hugepage_flag flag)
320{
321 return sysfs_emit(buf, "%d\n",
322 !!test_bit(flag, &transparent_hugepage_flags));
323}
324
325ssize_t single_hugepage_flag_store(struct kobject *kobj,
326 struct kobj_attribute *attr,
327 const char *buf, size_t count,
328 enum transparent_hugepage_flag flag)
329{
330 unsigned long value;
331 int ret;
332
333 ret = kstrtoul(buf, 10, &value);
334 if (ret < 0)
335 return ret;
336 if (value > 1)
337 return -EINVAL;
338
339 if (value)
340 set_bit(flag, &transparent_hugepage_flags);
341 else
342 clear_bit(flag, &transparent_hugepage_flags);
343
344 return count;
345}
346
347static ssize_t defrag_show(struct kobject *kobj,
348 struct kobj_attribute *attr, char *buf)
349{
350 const char *output;
351
352 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
353 &transparent_hugepage_flags))
354 output = "[always] defer defer+madvise madvise never";
355 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
356 &transparent_hugepage_flags))
357 output = "always [defer] defer+madvise madvise never";
358 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
359 &transparent_hugepage_flags))
360 output = "always defer [defer+madvise] madvise never";
361 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
362 &transparent_hugepage_flags))
363 output = "always defer defer+madvise [madvise] never";
364 else
365 output = "always defer defer+madvise madvise [never]";
366
367 return sysfs_emit(buf, "%s\n", output);
368}
369
370static ssize_t defrag_store(struct kobject *kobj,
371 struct kobj_attribute *attr,
372 const char *buf, size_t count)
373{
374 if (sysfs_streq(buf, "always")) {
375 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
376 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
377 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
378 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
379 } else if (sysfs_streq(buf, "defer+madvise")) {
380 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
381 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
382 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
383 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
384 } else if (sysfs_streq(buf, "defer")) {
385 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
386 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
387 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
388 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
389 } else if (sysfs_streq(buf, "madvise")) {
390 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
391 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
392 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
393 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
394 } else if (sysfs_streq(buf, "never")) {
395 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
396 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
397 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
398 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
399 } else
400 return -EINVAL;
401
402 return count;
403}
404static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
405
406static ssize_t use_zero_page_show(struct kobject *kobj,
407 struct kobj_attribute *attr, char *buf)
408{
409 return single_hugepage_flag_show(kobj, attr, buf,
410 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
411}
412static ssize_t use_zero_page_store(struct kobject *kobj,
413 struct kobj_attribute *attr, const char *buf, size_t count)
414{
415 return single_hugepage_flag_store(kobj, attr, buf, count,
416 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
417}
418static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
419
420static ssize_t hpage_pmd_size_show(struct kobject *kobj,
421 struct kobj_attribute *attr, char *buf)
422{
423 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
424}
425static struct kobj_attribute hpage_pmd_size_attr =
426 __ATTR_RO(hpage_pmd_size);
427
428static struct attribute *hugepage_attr[] = {
429 &enabled_attr.attr,
430 &defrag_attr.attr,
431 &use_zero_page_attr.attr,
432 &hpage_pmd_size_attr.attr,
433#ifdef CONFIG_SHMEM
434 &shmem_enabled_attr.attr,
435#endif
436 NULL,
437};
438
439static const struct attribute_group hugepage_attr_group = {
440 .attrs = hugepage_attr,
441};
442
443static void hugepage_exit_sysfs(struct kobject *hugepage_kobj);
444static void thpsize_release(struct kobject *kobj);
445static DEFINE_SPINLOCK(huge_anon_orders_lock);
446static LIST_HEAD(thpsize_list);
447
448struct thpsize {
449 struct kobject kobj;
450 struct list_head node;
451 int order;
452};
453
454#define to_thpsize(kobj) container_of(kobj, struct thpsize, kobj)
455
456static ssize_t thpsize_enabled_show(struct kobject *kobj,
457 struct kobj_attribute *attr, char *buf)
458{
459 int order = to_thpsize(kobj)->order;
460 const char *output;
461
462 if (test_bit(order, &huge_anon_orders_always))
463 output = "[always] inherit madvise never";
464 else if (test_bit(order, &huge_anon_orders_inherit))
465 output = "always [inherit] madvise never";
466 else if (test_bit(order, &huge_anon_orders_madvise))
467 output = "always inherit [madvise] never";
468 else
469 output = "always inherit madvise [never]";
470
471 return sysfs_emit(buf, "%s\n", output);
472}
473
474static ssize_t thpsize_enabled_store(struct kobject *kobj,
475 struct kobj_attribute *attr,
476 const char *buf, size_t count)
477{
478 int order = to_thpsize(kobj)->order;
479 ssize_t ret = count;
480
481 if (sysfs_streq(buf, "always")) {
482 spin_lock(&huge_anon_orders_lock);
483 clear_bit(order, &huge_anon_orders_inherit);
484 clear_bit(order, &huge_anon_orders_madvise);
485 set_bit(order, &huge_anon_orders_always);
486 spin_unlock(&huge_anon_orders_lock);
487 } else if (sysfs_streq(buf, "inherit")) {
488 spin_lock(&huge_anon_orders_lock);
489 clear_bit(order, &huge_anon_orders_always);
490 clear_bit(order, &huge_anon_orders_madvise);
491 set_bit(order, &huge_anon_orders_inherit);
492 spin_unlock(&huge_anon_orders_lock);
493 } else if (sysfs_streq(buf, "madvise")) {
494 spin_lock(&huge_anon_orders_lock);
495 clear_bit(order, &huge_anon_orders_always);
496 clear_bit(order, &huge_anon_orders_inherit);
497 set_bit(order, &huge_anon_orders_madvise);
498 spin_unlock(&huge_anon_orders_lock);
499 } else if (sysfs_streq(buf, "never")) {
500 spin_lock(&huge_anon_orders_lock);
501 clear_bit(order, &huge_anon_orders_always);
502 clear_bit(order, &huge_anon_orders_inherit);
503 clear_bit(order, &huge_anon_orders_madvise);
504 spin_unlock(&huge_anon_orders_lock);
505 } else
506 ret = -EINVAL;
507
508 return ret;
509}
510
511static struct kobj_attribute thpsize_enabled_attr =
512 __ATTR(enabled, 0644, thpsize_enabled_show, thpsize_enabled_store);
513
514static struct attribute *thpsize_attrs[] = {
515 &thpsize_enabled_attr.attr,
516 NULL,
517};
518
519static const struct attribute_group thpsize_attr_group = {
520 .attrs = thpsize_attrs,
521};
522
523static const struct kobj_type thpsize_ktype = {
524 .release = &thpsize_release,
525 .sysfs_ops = &kobj_sysfs_ops,
526};
527
528static struct thpsize *thpsize_create(int order, struct kobject *parent)
529{
530 unsigned long size = (PAGE_SIZE << order) / SZ_1K;
531 struct thpsize *thpsize;
532 int ret;
533
534 thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL);
535 if (!thpsize)
536 return ERR_PTR(-ENOMEM);
537
538 ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent,
539 "hugepages-%lukB", size);
540 if (ret) {
541 kfree(thpsize);
542 return ERR_PTR(ret);
543 }
544
545 ret = sysfs_create_group(&thpsize->kobj, &thpsize_attr_group);
546 if (ret) {
547 kobject_put(&thpsize->kobj);
548 return ERR_PTR(ret);
549 }
550
551 thpsize->order = order;
552 return thpsize;
553}
554
555static void thpsize_release(struct kobject *kobj)
556{
557 kfree(to_thpsize(kobj));
558}
559
560static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
561{
562 int err;
563 struct thpsize *thpsize;
564 unsigned long orders;
565 int order;
566
567 /*
568 * Default to setting PMD-sized THP to inherit the global setting and
569 * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time
570 * constant so we have to do this here.
571 */
572 huge_anon_orders_inherit = BIT(PMD_ORDER);
573
574 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
575 if (unlikely(!*hugepage_kobj)) {
576 pr_err("failed to create transparent hugepage kobject\n");
577 return -ENOMEM;
578 }
579
580 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
581 if (err) {
582 pr_err("failed to register transparent hugepage group\n");
583 goto delete_obj;
584 }
585
586 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
587 if (err) {
588 pr_err("failed to register transparent hugepage group\n");
589 goto remove_hp_group;
590 }
591
592 orders = THP_ORDERS_ALL_ANON;
593 order = highest_order(orders);
594 while (orders) {
595 thpsize = thpsize_create(order, *hugepage_kobj);
596 if (IS_ERR(thpsize)) {
597 pr_err("failed to create thpsize for order %d\n", order);
598 err = PTR_ERR(thpsize);
599 goto remove_all;
600 }
601 list_add(&thpsize->node, &thpsize_list);
602 order = next_order(&orders, order);
603 }
604
605 return 0;
606
607remove_all:
608 hugepage_exit_sysfs(*hugepage_kobj);
609 return err;
610remove_hp_group:
611 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
612delete_obj:
613 kobject_put(*hugepage_kobj);
614 return err;
615}
616
617static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
618{
619 struct thpsize *thpsize, *tmp;
620
621 list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) {
622 list_del(&thpsize->node);
623 kobject_put(&thpsize->kobj);
624 }
625
626 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
627 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
628 kobject_put(hugepage_kobj);
629}
630#else
631static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
632{
633 return 0;
634}
635
636static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
637{
638}
639#endif /* CONFIG_SYSFS */
640
641static int __init thp_shrinker_init(void)
642{
643 huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero");
644 if (!huge_zero_page_shrinker)
645 return -ENOMEM;
646
647 deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
648 SHRINKER_MEMCG_AWARE |
649 SHRINKER_NONSLAB,
650 "thp-deferred_split");
651 if (!deferred_split_shrinker) {
652 shrinker_free(huge_zero_page_shrinker);
653 return -ENOMEM;
654 }
655
656 huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count;
657 huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan;
658 shrinker_register(huge_zero_page_shrinker);
659
660 deferred_split_shrinker->count_objects = deferred_split_count;
661 deferred_split_shrinker->scan_objects = deferred_split_scan;
662 shrinker_register(deferred_split_shrinker);
663
664 return 0;
665}
666
667static void __init thp_shrinker_exit(void)
668{
669 shrinker_free(huge_zero_page_shrinker);
670 shrinker_free(deferred_split_shrinker);
671}
672
673static int __init hugepage_init(void)
674{
675 int err;
676 struct kobject *hugepage_kobj;
677
678 if (!has_transparent_hugepage()) {
679 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
680 return -EINVAL;
681 }
682
683 /*
684 * hugepages can't be allocated by the buddy allocator
685 */
686 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER);
687 /*
688 * we use page->mapping and page->index in second tail page
689 * as list_head: assuming THP order >= 2
690 */
691 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
692
693 err = hugepage_init_sysfs(&hugepage_kobj);
694 if (err)
695 goto err_sysfs;
696
697 err = khugepaged_init();
698 if (err)
699 goto err_slab;
700
701 err = thp_shrinker_init();
702 if (err)
703 goto err_shrinker;
704
705 /*
706 * By default disable transparent hugepages on smaller systems,
707 * where the extra memory used could hurt more than TLB overhead
708 * is likely to save. The admin can still enable it through /sys.
709 */
710 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
711 transparent_hugepage_flags = 0;
712 return 0;
713 }
714
715 err = start_stop_khugepaged();
716 if (err)
717 goto err_khugepaged;
718
719 return 0;
720err_khugepaged:
721 thp_shrinker_exit();
722err_shrinker:
723 khugepaged_destroy();
724err_slab:
725 hugepage_exit_sysfs(hugepage_kobj);
726err_sysfs:
727 return err;
728}
729subsys_initcall(hugepage_init);
730
731static int __init setup_transparent_hugepage(char *str)
732{
733 int ret = 0;
734 if (!str)
735 goto out;
736 if (!strcmp(str, "always")) {
737 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
738 &transparent_hugepage_flags);
739 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
740 &transparent_hugepage_flags);
741 ret = 1;
742 } else if (!strcmp(str, "madvise")) {
743 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
744 &transparent_hugepage_flags);
745 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
746 &transparent_hugepage_flags);
747 ret = 1;
748 } else if (!strcmp(str, "never")) {
749 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
750 &transparent_hugepage_flags);
751 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
752 &transparent_hugepage_flags);
753 ret = 1;
754 }
755out:
756 if (!ret)
757 pr_warn("transparent_hugepage= cannot parse, ignored\n");
758 return ret;
759}
760__setup("transparent_hugepage=", setup_transparent_hugepage);
761
762pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
763{
764 if (likely(vma->vm_flags & VM_WRITE))
765 pmd = pmd_mkwrite(pmd, vma);
766 return pmd;
767}
768
769#ifdef CONFIG_MEMCG
770static inline
771struct deferred_split *get_deferred_split_queue(struct folio *folio)
772{
773 struct mem_cgroup *memcg = folio_memcg(folio);
774 struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
775
776 if (memcg)
777 return &memcg->deferred_split_queue;
778 else
779 return &pgdat->deferred_split_queue;
780}
781#else
782static inline
783struct deferred_split *get_deferred_split_queue(struct folio *folio)
784{
785 struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
786
787 return &pgdat->deferred_split_queue;
788}
789#endif
790
791void folio_prep_large_rmappable(struct folio *folio)
792{
793 VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
794 INIT_LIST_HEAD(&folio->_deferred_list);
795 folio_set_large_rmappable(folio);
796}
797
798static inline bool is_transparent_hugepage(struct folio *folio)
799{
800 if (!folio_test_large(folio))
801 return false;
802
803 return is_huge_zero_page(&folio->page) ||
804 folio_test_large_rmappable(folio);
805}
806
807static unsigned long __thp_get_unmapped_area(struct file *filp,
808 unsigned long addr, unsigned long len,
809 loff_t off, unsigned long flags, unsigned long size)
810{
811 loff_t off_end = off + len;
812 loff_t off_align = round_up(off, size);
813 unsigned long len_pad, ret, off_sub;
814
815 if (IS_ENABLED(CONFIG_32BIT) || in_compat_syscall())
816 return 0;
817
818 if (off_end <= off_align || (off_end - off_align) < size)
819 return 0;
820
821 len_pad = len + size;
822 if (len_pad < len || (off + len_pad) < off)
823 return 0;
824
825 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
826 off >> PAGE_SHIFT, flags);
827
828 /*
829 * The failure might be due to length padding. The caller will retry
830 * without the padding.
831 */
832 if (IS_ERR_VALUE(ret))
833 return 0;
834
835 /*
836 * Do not try to align to THP boundary if allocation at the address
837 * hint succeeds.
838 */
839 if (ret == addr)
840 return addr;
841
842 off_sub = (off - ret) & (size - 1);
843
844 if (current->mm->get_unmapped_area == arch_get_unmapped_area_topdown &&
845 !off_sub)
846 return ret + size;
847
848 ret += off_sub;
849 return ret;
850}
851
852unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
853 unsigned long len, unsigned long pgoff, unsigned long flags)
854{
855 unsigned long ret;
856 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
857
858 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
859 if (ret)
860 return ret;
861
862 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
863}
864EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
865
866static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
867 struct page *page, gfp_t gfp)
868{
869 struct vm_area_struct *vma = vmf->vma;
870 struct folio *folio = page_folio(page);
871 pgtable_t pgtable;
872 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
873 vm_fault_t ret = 0;
874
875 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
876
877 if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
878 folio_put(folio);
879 count_vm_event(THP_FAULT_FALLBACK);
880 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
881 return VM_FAULT_FALLBACK;
882 }
883 folio_throttle_swaprate(folio, gfp);
884
885 pgtable = pte_alloc_one(vma->vm_mm);
886 if (unlikely(!pgtable)) {
887 ret = VM_FAULT_OOM;
888 goto release;
889 }
890
891 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
892 /*
893 * The memory barrier inside __folio_mark_uptodate makes sure that
894 * clear_huge_page writes become visible before the set_pmd_at()
895 * write.
896 */
897 __folio_mark_uptodate(folio);
898
899 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
900 if (unlikely(!pmd_none(*vmf->pmd))) {
901 goto unlock_release;
902 } else {
903 pmd_t entry;
904
905 ret = check_stable_address_space(vma->vm_mm);
906 if (ret)
907 goto unlock_release;
908
909 /* Deliver the page fault to userland */
910 if (userfaultfd_missing(vma)) {
911 spin_unlock(vmf->ptl);
912 folio_put(folio);
913 pte_free(vma->vm_mm, pgtable);
914 ret = handle_userfault(vmf, VM_UFFD_MISSING);
915 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
916 return ret;
917 }
918
919 entry = mk_huge_pmd(page, vma->vm_page_prot);
920 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
921 folio_add_new_anon_rmap(folio, vma, haddr);
922 folio_add_lru_vma(folio, vma);
923 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
924 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
925 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
926 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
927 mm_inc_nr_ptes(vma->vm_mm);
928 spin_unlock(vmf->ptl);
929 count_vm_event(THP_FAULT_ALLOC);
930 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
931 }
932
933 return 0;
934unlock_release:
935 spin_unlock(vmf->ptl);
936release:
937 if (pgtable)
938 pte_free(vma->vm_mm, pgtable);
939 folio_put(folio);
940 return ret;
941
942}
943
944/*
945 * always: directly stall for all thp allocations
946 * defer: wake kswapd and fail if not immediately available
947 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
948 * fail if not immediately available
949 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
950 * available
951 * never: never stall for any thp allocation
952 */
953gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
954{
955 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
956
957 /* Always do synchronous compaction */
958 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
959 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
960
961 /* Kick kcompactd and fail quickly */
962 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
963 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
964
965 /* Synchronous compaction if madvised, otherwise kick kcompactd */
966 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
967 return GFP_TRANSHUGE_LIGHT |
968 (vma_madvised ? __GFP_DIRECT_RECLAIM :
969 __GFP_KSWAPD_RECLAIM);
970
971 /* Only do synchronous compaction if madvised */
972 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
973 return GFP_TRANSHUGE_LIGHT |
974 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
975
976 return GFP_TRANSHUGE_LIGHT;
977}
978
979/* Caller must hold page table lock. */
980static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
981 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
982 struct page *zero_page)
983{
984 pmd_t entry;
985 if (!pmd_none(*pmd))
986 return;
987 entry = mk_pmd(zero_page, vma->vm_page_prot);
988 entry = pmd_mkhuge(entry);
989 pgtable_trans_huge_deposit(mm, pmd, pgtable);
990 set_pmd_at(mm, haddr, pmd, entry);
991 mm_inc_nr_ptes(mm);
992}
993
994vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
995{
996 struct vm_area_struct *vma = vmf->vma;
997 gfp_t gfp;
998 struct folio *folio;
999 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1000
1001 if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
1002 return VM_FAULT_FALLBACK;
1003 if (unlikely(anon_vma_prepare(vma)))
1004 return VM_FAULT_OOM;
1005 khugepaged_enter_vma(vma, vma->vm_flags);
1006
1007 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
1008 !mm_forbids_zeropage(vma->vm_mm) &&
1009 transparent_hugepage_use_zero_page()) {
1010 pgtable_t pgtable;
1011 struct page *zero_page;
1012 vm_fault_t ret;
1013 pgtable = pte_alloc_one(vma->vm_mm);
1014 if (unlikely(!pgtable))
1015 return VM_FAULT_OOM;
1016 zero_page = mm_get_huge_zero_page(vma->vm_mm);
1017 if (unlikely(!zero_page)) {
1018 pte_free(vma->vm_mm, pgtable);
1019 count_vm_event(THP_FAULT_FALLBACK);
1020 return VM_FAULT_FALLBACK;
1021 }
1022 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1023 ret = 0;
1024 if (pmd_none(*vmf->pmd)) {
1025 ret = check_stable_address_space(vma->vm_mm);
1026 if (ret) {
1027 spin_unlock(vmf->ptl);
1028 pte_free(vma->vm_mm, pgtable);
1029 } else if (userfaultfd_missing(vma)) {
1030 spin_unlock(vmf->ptl);
1031 pte_free(vma->vm_mm, pgtable);
1032 ret = handle_userfault(vmf, VM_UFFD_MISSING);
1033 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
1034 } else {
1035 set_huge_zero_page(pgtable, vma->vm_mm, vma,
1036 haddr, vmf->pmd, zero_page);
1037 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1038 spin_unlock(vmf->ptl);
1039 }
1040 } else {
1041 spin_unlock(vmf->ptl);
1042 pte_free(vma->vm_mm, pgtable);
1043 }
1044 return ret;
1045 }
1046 gfp = vma_thp_gfp_mask(vma);
1047 folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
1048 if (unlikely(!folio)) {
1049 count_vm_event(THP_FAULT_FALLBACK);
1050 return VM_FAULT_FALLBACK;
1051 }
1052 return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
1053}
1054
1055static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
1056 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
1057 pgtable_t pgtable)
1058{
1059 struct mm_struct *mm = vma->vm_mm;
1060 pmd_t entry;
1061 spinlock_t *ptl;
1062
1063 ptl = pmd_lock(mm, pmd);
1064 if (!pmd_none(*pmd)) {
1065 if (write) {
1066 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
1067 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
1068 goto out_unlock;
1069 }
1070 entry = pmd_mkyoung(*pmd);
1071 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1072 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
1073 update_mmu_cache_pmd(vma, addr, pmd);
1074 }
1075
1076 goto out_unlock;
1077 }
1078
1079 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
1080 if (pfn_t_devmap(pfn))
1081 entry = pmd_mkdevmap(entry);
1082 if (write) {
1083 entry = pmd_mkyoung(pmd_mkdirty(entry));
1084 entry = maybe_pmd_mkwrite(entry, vma);
1085 }
1086
1087 if (pgtable) {
1088 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1089 mm_inc_nr_ptes(mm);
1090 pgtable = NULL;
1091 }
1092
1093 set_pmd_at(mm, addr, pmd, entry);
1094 update_mmu_cache_pmd(vma, addr, pmd);
1095
1096out_unlock:
1097 spin_unlock(ptl);
1098 if (pgtable)
1099 pte_free(mm, pgtable);
1100}
1101
1102/**
1103 * vmf_insert_pfn_pmd - insert a pmd size pfn
1104 * @vmf: Structure describing the fault
1105 * @pfn: pfn to insert
1106 * @write: whether it's a write fault
1107 *
1108 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
1109 *
1110 * Return: vm_fault_t value.
1111 */
1112vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
1113{
1114 unsigned long addr = vmf->address & PMD_MASK;
1115 struct vm_area_struct *vma = vmf->vma;
1116 pgprot_t pgprot = vma->vm_page_prot;
1117 pgtable_t pgtable = NULL;
1118
1119 /*
1120 * If we had pmd_special, we could avoid all these restrictions,
1121 * but we need to be consistent with PTEs and architectures that
1122 * can't support a 'special' bit.
1123 */
1124 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1125 !pfn_t_devmap(pfn));
1126 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1127 (VM_PFNMAP|VM_MIXEDMAP));
1128 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1129
1130 if (addr < vma->vm_start || addr >= vma->vm_end)
1131 return VM_FAULT_SIGBUS;
1132
1133 if (arch_needs_pgtable_deposit()) {
1134 pgtable = pte_alloc_one(vma->vm_mm);
1135 if (!pgtable)
1136 return VM_FAULT_OOM;
1137 }
1138
1139 track_pfn_insert(vma, &pgprot, pfn);
1140
1141 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
1142 return VM_FAULT_NOPAGE;
1143}
1144EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
1145
1146#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1147static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
1148{
1149 if (likely(vma->vm_flags & VM_WRITE))
1150 pud = pud_mkwrite(pud);
1151 return pud;
1152}
1153
1154static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
1155 pud_t *pud, pfn_t pfn, bool write)
1156{
1157 struct mm_struct *mm = vma->vm_mm;
1158 pgprot_t prot = vma->vm_page_prot;
1159 pud_t entry;
1160 spinlock_t *ptl;
1161
1162 ptl = pud_lock(mm, pud);
1163 if (!pud_none(*pud)) {
1164 if (write) {
1165 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
1166 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
1167 goto out_unlock;
1168 }
1169 entry = pud_mkyoung(*pud);
1170 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
1171 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
1172 update_mmu_cache_pud(vma, addr, pud);
1173 }
1174 goto out_unlock;
1175 }
1176
1177 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
1178 if (pfn_t_devmap(pfn))
1179 entry = pud_mkdevmap(entry);
1180 if (write) {
1181 entry = pud_mkyoung(pud_mkdirty(entry));
1182 entry = maybe_pud_mkwrite(entry, vma);
1183 }
1184 set_pud_at(mm, addr, pud, entry);
1185 update_mmu_cache_pud(vma, addr, pud);
1186
1187out_unlock:
1188 spin_unlock(ptl);
1189}
1190
1191/**
1192 * vmf_insert_pfn_pud - insert a pud size pfn
1193 * @vmf: Structure describing the fault
1194 * @pfn: pfn to insert
1195 * @write: whether it's a write fault
1196 *
1197 * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
1198 *
1199 * Return: vm_fault_t value.
1200 */
1201vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
1202{
1203 unsigned long addr = vmf->address & PUD_MASK;
1204 struct vm_area_struct *vma = vmf->vma;
1205 pgprot_t pgprot = vma->vm_page_prot;
1206
1207 /*
1208 * If we had pud_special, we could avoid all these restrictions,
1209 * but we need to be consistent with PTEs and architectures that
1210 * can't support a 'special' bit.
1211 */
1212 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1213 !pfn_t_devmap(pfn));
1214 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1215 (VM_PFNMAP|VM_MIXEDMAP));
1216 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1217
1218 if (addr < vma->vm_start || addr >= vma->vm_end)
1219 return VM_FAULT_SIGBUS;
1220
1221 track_pfn_insert(vma, &pgprot, pfn);
1222
1223 insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
1224 return VM_FAULT_NOPAGE;
1225}
1226EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1227#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1228
1229static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1230 pmd_t *pmd, bool write)
1231{
1232 pmd_t _pmd;
1233
1234 _pmd = pmd_mkyoung(*pmd);
1235 if (write)
1236 _pmd = pmd_mkdirty(_pmd);
1237 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1238 pmd, _pmd, write))
1239 update_mmu_cache_pmd(vma, addr, pmd);
1240}
1241
1242struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1243 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1244{
1245 unsigned long pfn = pmd_pfn(*pmd);
1246 struct mm_struct *mm = vma->vm_mm;
1247 struct page *page;
1248 int ret;
1249
1250 assert_spin_locked(pmd_lockptr(mm, pmd));
1251
1252 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1253 return NULL;
1254
1255 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1256 /* pass */;
1257 else
1258 return NULL;
1259
1260 if (flags & FOLL_TOUCH)
1261 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1262
1263 /*
1264 * device mapped pages can only be returned if the
1265 * caller will manage the page reference count.
1266 */
1267 if (!(flags & (FOLL_GET | FOLL_PIN)))
1268 return ERR_PTR(-EEXIST);
1269
1270 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1271 *pgmap = get_dev_pagemap(pfn, *pgmap);
1272 if (!*pgmap)
1273 return ERR_PTR(-EFAULT);
1274 page = pfn_to_page(pfn);
1275 ret = try_grab_page(page, flags);
1276 if (ret)
1277 page = ERR_PTR(ret);
1278
1279 return page;
1280}
1281
1282int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1283 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1284 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1285{
1286 spinlock_t *dst_ptl, *src_ptl;
1287 struct page *src_page;
1288 struct folio *src_folio;
1289 pmd_t pmd;
1290 pgtable_t pgtable = NULL;
1291 int ret = -ENOMEM;
1292
1293 /* Skip if can be re-fill on fault */
1294 if (!vma_is_anonymous(dst_vma))
1295 return 0;
1296
1297 pgtable = pte_alloc_one(dst_mm);
1298 if (unlikely(!pgtable))
1299 goto out;
1300
1301 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1302 src_ptl = pmd_lockptr(src_mm, src_pmd);
1303 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1304
1305 ret = -EAGAIN;
1306 pmd = *src_pmd;
1307
1308#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1309 if (unlikely(is_swap_pmd(pmd))) {
1310 swp_entry_t entry = pmd_to_swp_entry(pmd);
1311
1312 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1313 if (!is_readable_migration_entry(entry)) {
1314 entry = make_readable_migration_entry(
1315 swp_offset(entry));
1316 pmd = swp_entry_to_pmd(entry);
1317 if (pmd_swp_soft_dirty(*src_pmd))
1318 pmd = pmd_swp_mksoft_dirty(pmd);
1319 if (pmd_swp_uffd_wp(*src_pmd))
1320 pmd = pmd_swp_mkuffd_wp(pmd);
1321 set_pmd_at(src_mm, addr, src_pmd, pmd);
1322 }
1323 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1324 mm_inc_nr_ptes(dst_mm);
1325 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1326 if (!userfaultfd_wp(dst_vma))
1327 pmd = pmd_swp_clear_uffd_wp(pmd);
1328 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1329 ret = 0;
1330 goto out_unlock;
1331 }
1332#endif
1333
1334 if (unlikely(!pmd_trans_huge(pmd))) {
1335 pte_free(dst_mm, pgtable);
1336 goto out_unlock;
1337 }
1338 /*
1339 * When page table lock is held, the huge zero pmd should not be
1340 * under splitting since we don't split the page itself, only pmd to
1341 * a page table.
1342 */
1343 if (is_huge_zero_pmd(pmd)) {
1344 /*
1345 * get_huge_zero_page() will never allocate a new page here,
1346 * since we already have a zero page to copy. It just takes a
1347 * reference.
1348 */
1349 mm_get_huge_zero_page(dst_mm);
1350 goto out_zero_page;
1351 }
1352
1353 src_page = pmd_page(pmd);
1354 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1355 src_folio = page_folio(src_page);
1356
1357 folio_get(src_folio);
1358 if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, src_vma))) {
1359 /* Page maybe pinned: split and retry the fault on PTEs. */
1360 folio_put(src_folio);
1361 pte_free(dst_mm, pgtable);
1362 spin_unlock(src_ptl);
1363 spin_unlock(dst_ptl);
1364 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1365 return -EAGAIN;
1366 }
1367 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1368out_zero_page:
1369 mm_inc_nr_ptes(dst_mm);
1370 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1371 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1372 if (!userfaultfd_wp(dst_vma))
1373 pmd = pmd_clear_uffd_wp(pmd);
1374 pmd = pmd_mkold(pmd_wrprotect(pmd));
1375 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1376
1377 ret = 0;
1378out_unlock:
1379 spin_unlock(src_ptl);
1380 spin_unlock(dst_ptl);
1381out:
1382 return ret;
1383}
1384
1385#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1386static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1387 pud_t *pud, bool write)
1388{
1389 pud_t _pud;
1390
1391 _pud = pud_mkyoung(*pud);
1392 if (write)
1393 _pud = pud_mkdirty(_pud);
1394 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1395 pud, _pud, write))
1396 update_mmu_cache_pud(vma, addr, pud);
1397}
1398
1399struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1400 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1401{
1402 unsigned long pfn = pud_pfn(*pud);
1403 struct mm_struct *mm = vma->vm_mm;
1404 struct page *page;
1405 int ret;
1406
1407 assert_spin_locked(pud_lockptr(mm, pud));
1408
1409 if (flags & FOLL_WRITE && !pud_write(*pud))
1410 return NULL;
1411
1412 if (pud_present(*pud) && pud_devmap(*pud))
1413 /* pass */;
1414 else
1415 return NULL;
1416
1417 if (flags & FOLL_TOUCH)
1418 touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1419
1420 /*
1421 * device mapped pages can only be returned if the
1422 * caller will manage the page reference count.
1423 *
1424 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1425 */
1426 if (!(flags & (FOLL_GET | FOLL_PIN)))
1427 return ERR_PTR(-EEXIST);
1428
1429 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1430 *pgmap = get_dev_pagemap(pfn, *pgmap);
1431 if (!*pgmap)
1432 return ERR_PTR(-EFAULT);
1433 page = pfn_to_page(pfn);
1434
1435 ret = try_grab_page(page, flags);
1436 if (ret)
1437 page = ERR_PTR(ret);
1438
1439 return page;
1440}
1441
1442int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1443 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1444 struct vm_area_struct *vma)
1445{
1446 spinlock_t *dst_ptl, *src_ptl;
1447 pud_t pud;
1448 int ret;
1449
1450 dst_ptl = pud_lock(dst_mm, dst_pud);
1451 src_ptl = pud_lockptr(src_mm, src_pud);
1452 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1453
1454 ret = -EAGAIN;
1455 pud = *src_pud;
1456 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1457 goto out_unlock;
1458
1459 /*
1460 * When page table lock is held, the huge zero pud should not be
1461 * under splitting since we don't split the page itself, only pud to
1462 * a page table.
1463 */
1464 if (is_huge_zero_pud(pud)) {
1465 /* No huge zero pud yet */
1466 }
1467
1468 /*
1469 * TODO: once we support anonymous pages, use
1470 * folio_try_dup_anon_rmap_*() and split if duplicating fails.
1471 */
1472 pudp_set_wrprotect(src_mm, addr, src_pud);
1473 pud = pud_mkold(pud_wrprotect(pud));
1474 set_pud_at(dst_mm, addr, dst_pud, pud);
1475
1476 ret = 0;
1477out_unlock:
1478 spin_unlock(src_ptl);
1479 spin_unlock(dst_ptl);
1480 return ret;
1481}
1482
1483void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1484{
1485 bool write = vmf->flags & FAULT_FLAG_WRITE;
1486
1487 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1488 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1489 goto unlock;
1490
1491 touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1492unlock:
1493 spin_unlock(vmf->ptl);
1494}
1495#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1496
1497void huge_pmd_set_accessed(struct vm_fault *vmf)
1498{
1499 bool write = vmf->flags & FAULT_FLAG_WRITE;
1500
1501 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1502 if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1503 goto unlock;
1504
1505 touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1506
1507unlock:
1508 spin_unlock(vmf->ptl);
1509}
1510
1511vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1512{
1513 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1514 struct vm_area_struct *vma = vmf->vma;
1515 struct folio *folio;
1516 struct page *page;
1517 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1518 pmd_t orig_pmd = vmf->orig_pmd;
1519
1520 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1521 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1522
1523 if (is_huge_zero_pmd(orig_pmd))
1524 goto fallback;
1525
1526 spin_lock(vmf->ptl);
1527
1528 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1529 spin_unlock(vmf->ptl);
1530 return 0;
1531 }
1532
1533 page = pmd_page(orig_pmd);
1534 folio = page_folio(page);
1535 VM_BUG_ON_PAGE(!PageHead(page), page);
1536
1537 /* Early check when only holding the PT lock. */
1538 if (PageAnonExclusive(page))
1539 goto reuse;
1540
1541 if (!folio_trylock(folio)) {
1542 folio_get(folio);
1543 spin_unlock(vmf->ptl);
1544 folio_lock(folio);
1545 spin_lock(vmf->ptl);
1546 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1547 spin_unlock(vmf->ptl);
1548 folio_unlock(folio);
1549 folio_put(folio);
1550 return 0;
1551 }
1552 folio_put(folio);
1553 }
1554
1555 /* Recheck after temporarily dropping the PT lock. */
1556 if (PageAnonExclusive(page)) {
1557 folio_unlock(folio);
1558 goto reuse;
1559 }
1560
1561 /*
1562 * See do_wp_page(): we can only reuse the folio exclusively if
1563 * there are no additional references. Note that we always drain
1564 * the LRU cache immediately after adding a THP.
1565 */
1566 if (folio_ref_count(folio) >
1567 1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1568 goto unlock_fallback;
1569 if (folio_test_swapcache(folio))
1570 folio_free_swap(folio);
1571 if (folio_ref_count(folio) == 1) {
1572 pmd_t entry;
1573
1574 folio_move_anon_rmap(folio, vma);
1575 SetPageAnonExclusive(page);
1576 folio_unlock(folio);
1577reuse:
1578 if (unlikely(unshare)) {
1579 spin_unlock(vmf->ptl);
1580 return 0;
1581 }
1582 entry = pmd_mkyoung(orig_pmd);
1583 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1584 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1585 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1586 spin_unlock(vmf->ptl);
1587 return 0;
1588 }
1589
1590unlock_fallback:
1591 folio_unlock(folio);
1592 spin_unlock(vmf->ptl);
1593fallback:
1594 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1595 return VM_FAULT_FALLBACK;
1596}
1597
1598static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1599 unsigned long addr, pmd_t pmd)
1600{
1601 struct page *page;
1602
1603 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1604 return false;
1605
1606 /* Don't touch entries that are not even readable (NUMA hinting). */
1607 if (pmd_protnone(pmd))
1608 return false;
1609
1610 /* Do we need write faults for softdirty tracking? */
1611 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1612 return false;
1613
1614 /* Do we need write faults for uffd-wp tracking? */
1615 if (userfaultfd_huge_pmd_wp(vma, pmd))
1616 return false;
1617
1618 if (!(vma->vm_flags & VM_SHARED)) {
1619 /* See can_change_pte_writable(). */
1620 page = vm_normal_page_pmd(vma, addr, pmd);
1621 return page && PageAnon(page) && PageAnonExclusive(page);
1622 }
1623
1624 /* See can_change_pte_writable(). */
1625 return pmd_dirty(pmd);
1626}
1627
1628/* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1629static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1630 struct vm_area_struct *vma,
1631 unsigned int flags)
1632{
1633 /* If the pmd is writable, we can write to the page. */
1634 if (pmd_write(pmd))
1635 return true;
1636
1637 /* Maybe FOLL_FORCE is set to override it? */
1638 if (!(flags & FOLL_FORCE))
1639 return false;
1640
1641 /* But FOLL_FORCE has no effect on shared mappings */
1642 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1643 return false;
1644
1645 /* ... or read-only private ones */
1646 if (!(vma->vm_flags & VM_MAYWRITE))
1647 return false;
1648
1649 /* ... or already writable ones that just need to take a write fault */
1650 if (vma->vm_flags & VM_WRITE)
1651 return false;
1652
1653 /*
1654 * See can_change_pte_writable(): we broke COW and could map the page
1655 * writable if we have an exclusive anonymous page ...
1656 */
1657 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1658 return false;
1659
1660 /* ... and a write-fault isn't required for other reasons. */
1661 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1662 return false;
1663 return !userfaultfd_huge_pmd_wp(vma, pmd);
1664}
1665
1666struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1667 unsigned long addr,
1668 pmd_t *pmd,
1669 unsigned int flags)
1670{
1671 struct mm_struct *mm = vma->vm_mm;
1672 struct page *page;
1673 int ret;
1674
1675 assert_spin_locked(pmd_lockptr(mm, pmd));
1676
1677 page = pmd_page(*pmd);
1678 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1679
1680 if ((flags & FOLL_WRITE) &&
1681 !can_follow_write_pmd(*pmd, page, vma, flags))
1682 return NULL;
1683
1684 /* Avoid dumping huge zero page */
1685 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1686 return ERR_PTR(-EFAULT);
1687
1688 if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags))
1689 return NULL;
1690
1691 if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1692 return ERR_PTR(-EMLINK);
1693
1694 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1695 !PageAnonExclusive(page), page);
1696
1697 ret = try_grab_page(page, flags);
1698 if (ret)
1699 return ERR_PTR(ret);
1700
1701 if (flags & FOLL_TOUCH)
1702 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1703
1704 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1705 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1706
1707 return page;
1708}
1709
1710/* NUMA hinting page fault entry point for trans huge pmds */
1711vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1712{
1713 struct vm_area_struct *vma = vmf->vma;
1714 pmd_t oldpmd = vmf->orig_pmd;
1715 pmd_t pmd;
1716 struct folio *folio;
1717 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1718 int nid = NUMA_NO_NODE;
1719 int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1720 bool migrated = false, writable = false;
1721 int flags = 0;
1722
1723 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1724 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1725 spin_unlock(vmf->ptl);
1726 goto out;
1727 }
1728
1729 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1730
1731 /*
1732 * Detect now whether the PMD could be writable; this information
1733 * is only valid while holding the PT lock.
1734 */
1735 writable = pmd_write(pmd);
1736 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1737 can_change_pmd_writable(vma, vmf->address, pmd))
1738 writable = true;
1739
1740 folio = vm_normal_folio_pmd(vma, haddr, pmd);
1741 if (!folio)
1742 goto out_map;
1743
1744 /* See similar comment in do_numa_page for explanation */
1745 if (!writable)
1746 flags |= TNF_NO_GROUP;
1747
1748 nid = folio_nid(folio);
1749 /*
1750 * For memory tiering mode, cpupid of slow memory page is used
1751 * to record page access time. So use default value.
1752 */
1753 if (node_is_toptier(nid))
1754 last_cpupid = folio_last_cpupid(folio);
1755 target_nid = numa_migrate_prep(folio, vma, haddr, nid, &flags);
1756 if (target_nid == NUMA_NO_NODE) {
1757 folio_put(folio);
1758 goto out_map;
1759 }
1760
1761 spin_unlock(vmf->ptl);
1762 writable = false;
1763
1764 migrated = migrate_misplaced_folio(folio, vma, target_nid);
1765 if (migrated) {
1766 flags |= TNF_MIGRATED;
1767 nid = target_nid;
1768 } else {
1769 flags |= TNF_MIGRATE_FAIL;
1770 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1771 if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1772 spin_unlock(vmf->ptl);
1773 goto out;
1774 }
1775 goto out_map;
1776 }
1777
1778out:
1779 if (nid != NUMA_NO_NODE)
1780 task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
1781
1782 return 0;
1783
1784out_map:
1785 /* Restore the PMD */
1786 pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1787 pmd = pmd_mkyoung(pmd);
1788 if (writable)
1789 pmd = pmd_mkwrite(pmd, vma);
1790 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1791 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1792 spin_unlock(vmf->ptl);
1793 goto out;
1794}
1795
1796/*
1797 * Return true if we do MADV_FREE successfully on entire pmd page.
1798 * Otherwise, return false.
1799 */
1800bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1801 pmd_t *pmd, unsigned long addr, unsigned long next)
1802{
1803 spinlock_t *ptl;
1804 pmd_t orig_pmd;
1805 struct folio *folio;
1806 struct mm_struct *mm = tlb->mm;
1807 bool ret = false;
1808
1809 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1810
1811 ptl = pmd_trans_huge_lock(pmd, vma);
1812 if (!ptl)
1813 goto out_unlocked;
1814
1815 orig_pmd = *pmd;
1816 if (is_huge_zero_pmd(orig_pmd))
1817 goto out;
1818
1819 if (unlikely(!pmd_present(orig_pmd))) {
1820 VM_BUG_ON(thp_migration_supported() &&
1821 !is_pmd_migration_entry(orig_pmd));
1822 goto out;
1823 }
1824
1825 folio = pfn_folio(pmd_pfn(orig_pmd));
1826 /*
1827 * If other processes are mapping this folio, we couldn't discard
1828 * the folio unless they all do MADV_FREE so let's skip the folio.
1829 */
1830 if (folio_estimated_sharers(folio) != 1)
1831 goto out;
1832
1833 if (!folio_trylock(folio))
1834 goto out;
1835
1836 /*
1837 * If user want to discard part-pages of THP, split it so MADV_FREE
1838 * will deactivate only them.
1839 */
1840 if (next - addr != HPAGE_PMD_SIZE) {
1841 folio_get(folio);
1842 spin_unlock(ptl);
1843 split_folio(folio);
1844 folio_unlock(folio);
1845 folio_put(folio);
1846 goto out_unlocked;
1847 }
1848
1849 if (folio_test_dirty(folio))
1850 folio_clear_dirty(folio);
1851 folio_unlock(folio);
1852
1853 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1854 pmdp_invalidate(vma, addr, pmd);
1855 orig_pmd = pmd_mkold(orig_pmd);
1856 orig_pmd = pmd_mkclean(orig_pmd);
1857
1858 set_pmd_at(mm, addr, pmd, orig_pmd);
1859 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1860 }
1861
1862 folio_mark_lazyfree(folio);
1863 ret = true;
1864out:
1865 spin_unlock(ptl);
1866out_unlocked:
1867 return ret;
1868}
1869
1870static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1871{
1872 pgtable_t pgtable;
1873
1874 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1875 pte_free(mm, pgtable);
1876 mm_dec_nr_ptes(mm);
1877}
1878
1879int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1880 pmd_t *pmd, unsigned long addr)
1881{
1882 pmd_t orig_pmd;
1883 spinlock_t *ptl;
1884
1885 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1886
1887 ptl = __pmd_trans_huge_lock(pmd, vma);
1888 if (!ptl)
1889 return 0;
1890 /*
1891 * For architectures like ppc64 we look at deposited pgtable
1892 * when calling pmdp_huge_get_and_clear. So do the
1893 * pgtable_trans_huge_withdraw after finishing pmdp related
1894 * operations.
1895 */
1896 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1897 tlb->fullmm);
1898 arch_check_zapped_pmd(vma, orig_pmd);
1899 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1900 if (vma_is_special_huge(vma)) {
1901 if (arch_needs_pgtable_deposit())
1902 zap_deposited_table(tlb->mm, pmd);
1903 spin_unlock(ptl);
1904 } else if (is_huge_zero_pmd(orig_pmd)) {
1905 zap_deposited_table(tlb->mm, pmd);
1906 spin_unlock(ptl);
1907 } else {
1908 struct page *page = NULL;
1909 int flush_needed = 1;
1910
1911 if (pmd_present(orig_pmd)) {
1912 page = pmd_page(orig_pmd);
1913 folio_remove_rmap_pmd(page_folio(page), page, vma);
1914 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1915 VM_BUG_ON_PAGE(!PageHead(page), page);
1916 } else if (thp_migration_supported()) {
1917 swp_entry_t entry;
1918
1919 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1920 entry = pmd_to_swp_entry(orig_pmd);
1921 page = pfn_swap_entry_to_page(entry);
1922 flush_needed = 0;
1923 } else
1924 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1925
1926 if (PageAnon(page)) {
1927 zap_deposited_table(tlb->mm, pmd);
1928 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1929 } else {
1930 if (arch_needs_pgtable_deposit())
1931 zap_deposited_table(tlb->mm, pmd);
1932 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1933 }
1934
1935 spin_unlock(ptl);
1936 if (flush_needed)
1937 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1938 }
1939 return 1;
1940}
1941
1942#ifndef pmd_move_must_withdraw
1943static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1944 spinlock_t *old_pmd_ptl,
1945 struct vm_area_struct *vma)
1946{
1947 /*
1948 * With split pmd lock we also need to move preallocated
1949 * PTE page table if new_pmd is on different PMD page table.
1950 *
1951 * We also don't deposit and withdraw tables for file pages.
1952 */
1953 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1954}
1955#endif
1956
1957static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1958{
1959#ifdef CONFIG_MEM_SOFT_DIRTY
1960 if (unlikely(is_pmd_migration_entry(pmd)))
1961 pmd = pmd_swp_mksoft_dirty(pmd);
1962 else if (pmd_present(pmd))
1963 pmd = pmd_mksoft_dirty(pmd);
1964#endif
1965 return pmd;
1966}
1967
1968bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1969 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1970{
1971 spinlock_t *old_ptl, *new_ptl;
1972 pmd_t pmd;
1973 struct mm_struct *mm = vma->vm_mm;
1974 bool force_flush = false;
1975
1976 /*
1977 * The destination pmd shouldn't be established, free_pgtables()
1978 * should have released it; but move_page_tables() might have already
1979 * inserted a page table, if racing against shmem/file collapse.
1980 */
1981 if (!pmd_none(*new_pmd)) {
1982 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1983 return false;
1984 }
1985
1986 /*
1987 * We don't have to worry about the ordering of src and dst
1988 * ptlocks because exclusive mmap_lock prevents deadlock.
1989 */
1990 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1991 if (old_ptl) {
1992 new_ptl = pmd_lockptr(mm, new_pmd);
1993 if (new_ptl != old_ptl)
1994 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1995 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1996 if (pmd_present(pmd))
1997 force_flush = true;
1998 VM_BUG_ON(!pmd_none(*new_pmd));
1999
2000 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
2001 pgtable_t pgtable;
2002 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
2003 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
2004 }
2005 pmd = move_soft_dirty_pmd(pmd);
2006 set_pmd_at(mm, new_addr, new_pmd, pmd);
2007 if (force_flush)
2008 flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
2009 if (new_ptl != old_ptl)
2010 spin_unlock(new_ptl);
2011 spin_unlock(old_ptl);
2012 return true;
2013 }
2014 return false;
2015}
2016
2017/*
2018 * Returns
2019 * - 0 if PMD could not be locked
2020 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
2021 * or if prot_numa but THP migration is not supported
2022 * - HPAGE_PMD_NR if protections changed and TLB flush necessary
2023 */
2024int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2025 pmd_t *pmd, unsigned long addr, pgprot_t newprot,
2026 unsigned long cp_flags)
2027{
2028 struct mm_struct *mm = vma->vm_mm;
2029 spinlock_t *ptl;
2030 pmd_t oldpmd, entry;
2031 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
2032 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
2033 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
2034 int ret = 1;
2035
2036 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2037
2038 if (prot_numa && !thp_migration_supported())
2039 return 1;
2040
2041 ptl = __pmd_trans_huge_lock(pmd, vma);
2042 if (!ptl)
2043 return 0;
2044
2045#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2046 if (is_swap_pmd(*pmd)) {
2047 swp_entry_t entry = pmd_to_swp_entry(*pmd);
2048 struct folio *folio = page_folio(pfn_swap_entry_to_page(entry));
2049 pmd_t newpmd;
2050
2051 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2052 if (is_writable_migration_entry(entry)) {
2053 /*
2054 * A protection check is difficult so
2055 * just be safe and disable write
2056 */
2057 if (folio_test_anon(folio))
2058 entry = make_readable_exclusive_migration_entry(swp_offset(entry));
2059 else
2060 entry = make_readable_migration_entry(swp_offset(entry));
2061 newpmd = swp_entry_to_pmd(entry);
2062 if (pmd_swp_soft_dirty(*pmd))
2063 newpmd = pmd_swp_mksoft_dirty(newpmd);
2064 } else {
2065 newpmd = *pmd;
2066 }
2067
2068 if (uffd_wp)
2069 newpmd = pmd_swp_mkuffd_wp(newpmd);
2070 else if (uffd_wp_resolve)
2071 newpmd = pmd_swp_clear_uffd_wp(newpmd);
2072 if (!pmd_same(*pmd, newpmd))
2073 set_pmd_at(mm, addr, pmd, newpmd);
2074 goto unlock;
2075 }
2076#endif
2077
2078 if (prot_numa) {
2079 struct folio *folio;
2080 bool toptier;
2081 /*
2082 * Avoid trapping faults against the zero page. The read-only
2083 * data is likely to be read-cached on the local CPU and
2084 * local/remote hits to the zero page are not interesting.
2085 */
2086 if (is_huge_zero_pmd(*pmd))
2087 goto unlock;
2088
2089 if (pmd_protnone(*pmd))
2090 goto unlock;
2091
2092 folio = page_folio(pmd_page(*pmd));
2093 toptier = node_is_toptier(folio_nid(folio));
2094 /*
2095 * Skip scanning top tier node if normal numa
2096 * balancing is disabled
2097 */
2098 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
2099 toptier)
2100 goto unlock;
2101
2102 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
2103 !toptier)
2104 folio_xchg_access_time(folio,
2105 jiffies_to_msecs(jiffies));
2106 }
2107 /*
2108 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
2109 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2110 * which is also under mmap_read_lock(mm):
2111 *
2112 * CPU0: CPU1:
2113 * change_huge_pmd(prot_numa=1)
2114 * pmdp_huge_get_and_clear_notify()
2115 * madvise_dontneed()
2116 * zap_pmd_range()
2117 * pmd_trans_huge(*pmd) == 0 (without ptl)
2118 * // skip the pmd
2119 * set_pmd_at();
2120 * // pmd is re-established
2121 *
2122 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2123 * which may break userspace.
2124 *
2125 * pmdp_invalidate_ad() is required to make sure we don't miss
2126 * dirty/young flags set by hardware.
2127 */
2128 oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
2129
2130 entry = pmd_modify(oldpmd, newprot);
2131 if (uffd_wp)
2132 entry = pmd_mkuffd_wp(entry);
2133 else if (uffd_wp_resolve)
2134 /*
2135 * Leave the write bit to be handled by PF interrupt
2136 * handler, then things like COW could be properly
2137 * handled.
2138 */
2139 entry = pmd_clear_uffd_wp(entry);
2140
2141 /* See change_pte_range(). */
2142 if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
2143 can_change_pmd_writable(vma, addr, entry))
2144 entry = pmd_mkwrite(entry, vma);
2145
2146 ret = HPAGE_PMD_NR;
2147 set_pmd_at(mm, addr, pmd, entry);
2148
2149 if (huge_pmd_needs_flush(oldpmd, entry))
2150 tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
2151unlock:
2152 spin_unlock(ptl);
2153 return ret;
2154}
2155
2156#ifdef CONFIG_USERFAULTFD
2157/*
2158 * The PT lock for src_pmd and the mmap_lock for reading are held by
2159 * the caller, but it must return after releasing the page_table_lock.
2160 * Just move the page from src_pmd to dst_pmd if possible.
2161 * Return zero if succeeded in moving the page, -EAGAIN if it needs to be
2162 * repeated by the caller, or other errors in case of failure.
2163 */
2164int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval,
2165 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
2166 unsigned long dst_addr, unsigned long src_addr)
2167{
2168 pmd_t _dst_pmd, src_pmdval;
2169 struct page *src_page;
2170 struct folio *src_folio;
2171 struct anon_vma *src_anon_vma;
2172 spinlock_t *src_ptl, *dst_ptl;
2173 pgtable_t src_pgtable;
2174 struct mmu_notifier_range range;
2175 int err = 0;
2176
2177 src_pmdval = *src_pmd;
2178 src_ptl = pmd_lockptr(mm, src_pmd);
2179
2180 lockdep_assert_held(src_ptl);
2181 mmap_assert_locked(mm);
2182
2183 /* Sanity checks before the operation */
2184 if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) ||
2185 WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) {
2186 spin_unlock(src_ptl);
2187 return -EINVAL;
2188 }
2189
2190 if (!pmd_trans_huge(src_pmdval)) {
2191 spin_unlock(src_ptl);
2192 if (is_pmd_migration_entry(src_pmdval)) {
2193 pmd_migration_entry_wait(mm, &src_pmdval);
2194 return -EAGAIN;
2195 }
2196 return -ENOENT;
2197 }
2198
2199 src_page = pmd_page(src_pmdval);
2200 if (unlikely(!PageAnonExclusive(src_page))) {
2201 spin_unlock(src_ptl);
2202 return -EBUSY;
2203 }
2204
2205 src_folio = page_folio(src_page);
2206 folio_get(src_folio);
2207 spin_unlock(src_ptl);
2208
2209 flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE);
2210 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr,
2211 src_addr + HPAGE_PMD_SIZE);
2212 mmu_notifier_invalidate_range_start(&range);
2213
2214 folio_lock(src_folio);
2215
2216 /*
2217 * split_huge_page walks the anon_vma chain without the page
2218 * lock. Serialize against it with the anon_vma lock, the page
2219 * lock is not enough.
2220 */
2221 src_anon_vma = folio_get_anon_vma(src_folio);
2222 if (!src_anon_vma) {
2223 err = -EAGAIN;
2224 goto unlock_folio;
2225 }
2226 anon_vma_lock_write(src_anon_vma);
2227
2228 dst_ptl = pmd_lockptr(mm, dst_pmd);
2229 double_pt_lock(src_ptl, dst_ptl);
2230 if (unlikely(!pmd_same(*src_pmd, src_pmdval) ||
2231 !pmd_same(*dst_pmd, dst_pmdval))) {
2232 err = -EAGAIN;
2233 goto unlock_ptls;
2234 }
2235 if (folio_maybe_dma_pinned(src_folio) ||
2236 !PageAnonExclusive(&src_folio->page)) {
2237 err = -EBUSY;
2238 goto unlock_ptls;
2239 }
2240
2241 if (WARN_ON_ONCE(!folio_test_head(src_folio)) ||
2242 WARN_ON_ONCE(!folio_test_anon(src_folio))) {
2243 err = -EBUSY;
2244 goto unlock_ptls;
2245 }
2246
2247 folio_move_anon_rmap(src_folio, dst_vma);
2248 WRITE_ONCE(src_folio->index, linear_page_index(dst_vma, dst_addr));
2249
2250 src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2251 /* Folio got pinned from under us. Put it back and fail the move. */
2252 if (folio_maybe_dma_pinned(src_folio)) {
2253 set_pmd_at(mm, src_addr, src_pmd, src_pmdval);
2254 err = -EBUSY;
2255 goto unlock_ptls;
2256 }
2257
2258 _dst_pmd = mk_huge_pmd(&src_folio->page, dst_vma->vm_page_prot);
2259 /* Follow mremap() behavior and treat the entry dirty after the move */
2260 _dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma);
2261 set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd);
2262
2263 src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd);
2264 pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable);
2265unlock_ptls:
2266 double_pt_unlock(src_ptl, dst_ptl);
2267 anon_vma_unlock_write(src_anon_vma);
2268 put_anon_vma(src_anon_vma);
2269unlock_folio:
2270 /* unblock rmap walks */
2271 folio_unlock(src_folio);
2272 mmu_notifier_invalidate_range_end(&range);
2273 folio_put(src_folio);
2274 return err;
2275}
2276#endif /* CONFIG_USERFAULTFD */
2277
2278/*
2279 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2280 *
2281 * Note that if it returns page table lock pointer, this routine returns without
2282 * unlocking page table lock. So callers must unlock it.
2283 */
2284spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2285{
2286 spinlock_t *ptl;
2287 ptl = pmd_lock(vma->vm_mm, pmd);
2288 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2289 pmd_devmap(*pmd)))
2290 return ptl;
2291 spin_unlock(ptl);
2292 return NULL;
2293}
2294
2295/*
2296 * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
2297 *
2298 * Note that if it returns page table lock pointer, this routine returns without
2299 * unlocking page table lock. So callers must unlock it.
2300 */
2301spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2302{
2303 spinlock_t *ptl;
2304
2305 ptl = pud_lock(vma->vm_mm, pud);
2306 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2307 return ptl;
2308 spin_unlock(ptl);
2309 return NULL;
2310}
2311
2312#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2313int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2314 pud_t *pud, unsigned long addr)
2315{
2316 spinlock_t *ptl;
2317
2318 ptl = __pud_trans_huge_lock(pud, vma);
2319 if (!ptl)
2320 return 0;
2321
2322 pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);
2323 tlb_remove_pud_tlb_entry(tlb, pud, addr);
2324 if (vma_is_special_huge(vma)) {
2325 spin_unlock(ptl);
2326 /* No zero page support yet */
2327 } else {
2328 /* No support for anonymous PUD pages yet */
2329 BUG();
2330 }
2331 return 1;
2332}
2333
2334static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2335 unsigned long haddr)
2336{
2337 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2338 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2339 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2340 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2341
2342 count_vm_event(THP_SPLIT_PUD);
2343
2344 pudp_huge_clear_flush(vma, haddr, pud);
2345}
2346
2347void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2348 unsigned long address)
2349{
2350 spinlock_t *ptl;
2351 struct mmu_notifier_range range;
2352
2353 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2354 address & HPAGE_PUD_MASK,
2355 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2356 mmu_notifier_invalidate_range_start(&range);
2357 ptl = pud_lock(vma->vm_mm, pud);
2358 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2359 goto out;
2360 __split_huge_pud_locked(vma, pud, range.start);
2361
2362out:
2363 spin_unlock(ptl);
2364 mmu_notifier_invalidate_range_end(&range);
2365}
2366#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2367
2368static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2369 unsigned long haddr, pmd_t *pmd)
2370{
2371 struct mm_struct *mm = vma->vm_mm;
2372 pgtable_t pgtable;
2373 pmd_t _pmd, old_pmd;
2374 unsigned long addr;
2375 pte_t *pte;
2376 int i;
2377
2378 /*
2379 * Leave pmd empty until pte is filled note that it is fine to delay
2380 * notification until mmu_notifier_invalidate_range_end() as we are
2381 * replacing a zero pmd write protected page with a zero pte write
2382 * protected page.
2383 *
2384 * See Documentation/mm/mmu_notifier.rst
2385 */
2386 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2387
2388 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2389 pmd_populate(mm, &_pmd, pgtable);
2390
2391 pte = pte_offset_map(&_pmd, haddr);
2392 VM_BUG_ON(!pte);
2393 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2394 pte_t entry;
2395
2396 entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2397 entry = pte_mkspecial(entry);
2398 if (pmd_uffd_wp(old_pmd))
2399 entry = pte_mkuffd_wp(entry);
2400 VM_BUG_ON(!pte_none(ptep_get(pte)));
2401 set_pte_at(mm, addr, pte, entry);
2402 pte++;
2403 }
2404 pte_unmap(pte - 1);
2405 smp_wmb(); /* make pte visible before pmd */
2406 pmd_populate(mm, pmd, pgtable);
2407}
2408
2409static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2410 unsigned long haddr, bool freeze)
2411{
2412 struct mm_struct *mm = vma->vm_mm;
2413 struct folio *folio;
2414 struct page *page;
2415 pgtable_t pgtable;
2416 pmd_t old_pmd, _pmd;
2417 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2418 bool anon_exclusive = false, dirty = false;
2419 unsigned long addr;
2420 pte_t *pte;
2421 int i;
2422
2423 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2424 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2425 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2426 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2427 && !pmd_devmap(*pmd));
2428
2429 count_vm_event(THP_SPLIT_PMD);
2430
2431 if (!vma_is_anonymous(vma)) {
2432 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2433 /*
2434 * We are going to unmap this huge page. So
2435 * just go ahead and zap it
2436 */
2437 if (arch_needs_pgtable_deposit())
2438 zap_deposited_table(mm, pmd);
2439 if (vma_is_special_huge(vma))
2440 return;
2441 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2442 swp_entry_t entry;
2443
2444 entry = pmd_to_swp_entry(old_pmd);
2445 page = pfn_swap_entry_to_page(entry);
2446 } else {
2447 page = pmd_page(old_pmd);
2448 folio = page_folio(page);
2449 if (!folio_test_dirty(folio) && pmd_dirty(old_pmd))
2450 folio_mark_dirty(folio);
2451 if (!folio_test_referenced(folio) && pmd_young(old_pmd))
2452 folio_set_referenced(folio);
2453 folio_remove_rmap_pmd(folio, page, vma);
2454 folio_put(folio);
2455 }
2456 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2457 return;
2458 }
2459
2460 if (is_huge_zero_pmd(*pmd)) {
2461 /*
2462 * FIXME: Do we want to invalidate secondary mmu by calling
2463 * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below
2464 * inside __split_huge_pmd() ?
2465 *
2466 * We are going from a zero huge page write protected to zero
2467 * small page also write protected so it does not seems useful
2468 * to invalidate secondary mmu at this time.
2469 */
2470 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2471 }
2472
2473 /*
2474 * Up to this point the pmd is present and huge and userland has the
2475 * whole access to the hugepage during the split (which happens in
2476 * place). If we overwrite the pmd with the not-huge version pointing
2477 * to the pte here (which of course we could if all CPUs were bug
2478 * free), userland could trigger a small page size TLB miss on the
2479 * small sized TLB while the hugepage TLB entry is still established in
2480 * the huge TLB. Some CPU doesn't like that.
2481 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2482 * 383 on page 105. Intel should be safe but is also warns that it's
2483 * only safe if the permission and cache attributes of the two entries
2484 * loaded in the two TLB is identical (which should be the case here).
2485 * But it is generally safer to never allow small and huge TLB entries
2486 * for the same virtual address to be loaded simultaneously. So instead
2487 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2488 * current pmd notpresent (atomically because here the pmd_trans_huge
2489 * must remain set at all times on the pmd until the split is complete
2490 * for this pmd), then we flush the SMP TLB and finally we write the
2491 * non-huge version of the pmd entry with pmd_populate.
2492 */
2493 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2494
2495 pmd_migration = is_pmd_migration_entry(old_pmd);
2496 if (unlikely(pmd_migration)) {
2497 swp_entry_t entry;
2498
2499 entry = pmd_to_swp_entry(old_pmd);
2500 page = pfn_swap_entry_to_page(entry);
2501 write = is_writable_migration_entry(entry);
2502 if (PageAnon(page))
2503 anon_exclusive = is_readable_exclusive_migration_entry(entry);
2504 young = is_migration_entry_young(entry);
2505 dirty = is_migration_entry_dirty(entry);
2506 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2507 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2508 } else {
2509 page = pmd_page(old_pmd);
2510 folio = page_folio(page);
2511 if (pmd_dirty(old_pmd)) {
2512 dirty = true;
2513 folio_set_dirty(folio);
2514 }
2515 write = pmd_write(old_pmd);
2516 young = pmd_young(old_pmd);
2517 soft_dirty = pmd_soft_dirty(old_pmd);
2518 uffd_wp = pmd_uffd_wp(old_pmd);
2519
2520 VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio);
2521 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2522
2523 /*
2524 * Without "freeze", we'll simply split the PMD, propagating the
2525 * PageAnonExclusive() flag for each PTE by setting it for
2526 * each subpage -- no need to (temporarily) clear.
2527 *
2528 * With "freeze" we want to replace mapped pages by
2529 * migration entries right away. This is only possible if we
2530 * managed to clear PageAnonExclusive() -- see
2531 * set_pmd_migration_entry().
2532 *
2533 * In case we cannot clear PageAnonExclusive(), split the PMD
2534 * only and let try_to_migrate_one() fail later.
2535 *
2536 * See folio_try_share_anon_rmap_pmd(): invalidate PMD first.
2537 */
2538 anon_exclusive = PageAnonExclusive(page);
2539 if (freeze && anon_exclusive &&
2540 folio_try_share_anon_rmap_pmd(folio, page))
2541 freeze = false;
2542 if (!freeze) {
2543 rmap_t rmap_flags = RMAP_NONE;
2544
2545 folio_ref_add(folio, HPAGE_PMD_NR - 1);
2546 if (anon_exclusive)
2547 rmap_flags |= RMAP_EXCLUSIVE;
2548 folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,
2549 vma, haddr, rmap_flags);
2550 }
2551 }
2552
2553 /*
2554 * Withdraw the table only after we mark the pmd entry invalid.
2555 * This's critical for some architectures (Power).
2556 */
2557 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2558 pmd_populate(mm, &_pmd, pgtable);
2559
2560 pte = pte_offset_map(&_pmd, haddr);
2561 VM_BUG_ON(!pte);
2562 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2563 pte_t entry;
2564 /*
2565 * Note that NUMA hinting access restrictions are not
2566 * transferred to avoid any possibility of altering
2567 * permissions across VMAs.
2568 */
2569 if (freeze || pmd_migration) {
2570 swp_entry_t swp_entry;
2571 if (write)
2572 swp_entry = make_writable_migration_entry(
2573 page_to_pfn(page + i));
2574 else if (anon_exclusive)
2575 swp_entry = make_readable_exclusive_migration_entry(
2576 page_to_pfn(page + i));
2577 else
2578 swp_entry = make_readable_migration_entry(
2579 page_to_pfn(page + i));
2580 if (young)
2581 swp_entry = make_migration_entry_young(swp_entry);
2582 if (dirty)
2583 swp_entry = make_migration_entry_dirty(swp_entry);
2584 entry = swp_entry_to_pte(swp_entry);
2585 if (soft_dirty)
2586 entry = pte_swp_mksoft_dirty(entry);
2587 if (uffd_wp)
2588 entry = pte_swp_mkuffd_wp(entry);
2589 } else {
2590 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2591 if (write)
2592 entry = pte_mkwrite(entry, vma);
2593 if (!young)
2594 entry = pte_mkold(entry);
2595 /* NOTE: this may set soft-dirty too on some archs */
2596 if (dirty)
2597 entry = pte_mkdirty(entry);
2598 if (soft_dirty)
2599 entry = pte_mksoft_dirty(entry);
2600 if (uffd_wp)
2601 entry = pte_mkuffd_wp(entry);
2602 }
2603 VM_BUG_ON(!pte_none(ptep_get(pte)));
2604 set_pte_at(mm, addr, pte, entry);
2605 pte++;
2606 }
2607 pte_unmap(pte - 1);
2608
2609 if (!pmd_migration)
2610 folio_remove_rmap_pmd(folio, page, vma);
2611 if (freeze)
2612 put_page(page);
2613
2614 smp_wmb(); /* make pte visible before pmd */
2615 pmd_populate(mm, pmd, pgtable);
2616}
2617
2618void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2619 unsigned long address, bool freeze, struct folio *folio)
2620{
2621 spinlock_t *ptl;
2622 struct mmu_notifier_range range;
2623
2624 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2625 address & HPAGE_PMD_MASK,
2626 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2627 mmu_notifier_invalidate_range_start(&range);
2628 ptl = pmd_lock(vma->vm_mm, pmd);
2629
2630 /*
2631 * If caller asks to setup a migration entry, we need a folio to check
2632 * pmd against. Otherwise we can end up replacing wrong folio.
2633 */
2634 VM_BUG_ON(freeze && !folio);
2635 VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2636
2637 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2638 is_pmd_migration_entry(*pmd)) {
2639 /*
2640 * It's safe to call pmd_page when folio is set because it's
2641 * guaranteed that pmd is present.
2642 */
2643 if (folio && folio != page_folio(pmd_page(*pmd)))
2644 goto out;
2645 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2646 }
2647
2648out:
2649 spin_unlock(ptl);
2650 mmu_notifier_invalidate_range_end(&range);
2651}
2652
2653void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2654 bool freeze, struct folio *folio)
2655{
2656 pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2657
2658 if (!pmd)
2659 return;
2660
2661 __split_huge_pmd(vma, pmd, address, freeze, folio);
2662}
2663
2664static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2665{
2666 /*
2667 * If the new address isn't hpage aligned and it could previously
2668 * contain an hugepage: check if we need to split an huge pmd.
2669 */
2670 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2671 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2672 ALIGN(address, HPAGE_PMD_SIZE)))
2673 split_huge_pmd_address(vma, address, false, NULL);
2674}
2675
2676void vma_adjust_trans_huge(struct vm_area_struct *vma,
2677 unsigned long start,
2678 unsigned long end,
2679 long adjust_next)
2680{
2681 /* Check if we need to split start first. */
2682 split_huge_pmd_if_needed(vma, start);
2683
2684 /* Check if we need to split end next. */
2685 split_huge_pmd_if_needed(vma, end);
2686
2687 /*
2688 * If we're also updating the next vma vm_start,
2689 * check if we need to split it.
2690 */
2691 if (adjust_next > 0) {
2692 struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2693 unsigned long nstart = next->vm_start;
2694 nstart += adjust_next;
2695 split_huge_pmd_if_needed(next, nstart);
2696 }
2697}
2698
2699static void unmap_folio(struct folio *folio)
2700{
2701 enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2702 TTU_SYNC | TTU_BATCH_FLUSH;
2703
2704 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2705
2706 /*
2707 * Anon pages need migration entries to preserve them, but file
2708 * pages can simply be left unmapped, then faulted back on demand.
2709 * If that is ever changed (perhaps for mlock), update remap_page().
2710 */
2711 if (folio_test_anon(folio))
2712 try_to_migrate(folio, ttu_flags);
2713 else
2714 try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2715
2716 try_to_unmap_flush();
2717}
2718
2719static void remap_page(struct folio *folio, unsigned long nr)
2720{
2721 int i = 0;
2722
2723 /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2724 if (!folio_test_anon(folio))
2725 return;
2726 for (;;) {
2727 remove_migration_ptes(folio, folio, true);
2728 i += folio_nr_pages(folio);
2729 if (i >= nr)
2730 break;
2731 folio = folio_next(folio);
2732 }
2733}
2734
2735static void lru_add_page_tail(struct page *head, struct page *tail,
2736 struct lruvec *lruvec, struct list_head *list)
2737{
2738 VM_BUG_ON_PAGE(!PageHead(head), head);
2739 VM_BUG_ON_PAGE(PageCompound(tail), head);
2740 VM_BUG_ON_PAGE(PageLRU(tail), head);
2741 lockdep_assert_held(&lruvec->lru_lock);
2742
2743 if (list) {
2744 /* page reclaim is reclaiming a huge page */
2745 VM_WARN_ON(PageLRU(head));
2746 get_page(tail);
2747 list_add_tail(&tail->lru, list);
2748 } else {
2749 /* head is still on lru (and we have it frozen) */
2750 VM_WARN_ON(!PageLRU(head));
2751 if (PageUnevictable(tail))
2752 tail->mlock_count = 0;
2753 else
2754 list_add_tail(&tail->lru, &head->lru);
2755 SetPageLRU(tail);
2756 }
2757}
2758
2759static void __split_huge_page_tail(struct folio *folio, int tail,
2760 struct lruvec *lruvec, struct list_head *list)
2761{
2762 struct page *head = &folio->page;
2763 struct page *page_tail = head + tail;
2764 /*
2765 * Careful: new_folio is not a "real" folio before we cleared PageTail.
2766 * Don't pass it around before clear_compound_head().
2767 */
2768 struct folio *new_folio = (struct folio *)page_tail;
2769
2770 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2771
2772 /*
2773 * Clone page flags before unfreezing refcount.
2774 *
2775 * After successful get_page_unless_zero() might follow flags change,
2776 * for example lock_page() which set PG_waiters.
2777 *
2778 * Note that for mapped sub-pages of an anonymous THP,
2779 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2780 * the migration entry instead from where remap_page() will restore it.
2781 * We can still have PG_anon_exclusive set on effectively unmapped and
2782 * unreferenced sub-pages of an anonymous THP: we can simply drop
2783 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2784 */
2785 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2786 page_tail->flags |= (head->flags &
2787 ((1L << PG_referenced) |
2788 (1L << PG_swapbacked) |
2789 (1L << PG_swapcache) |
2790 (1L << PG_mlocked) |
2791 (1L << PG_uptodate) |
2792 (1L << PG_active) |
2793 (1L << PG_workingset) |
2794 (1L << PG_locked) |
2795 (1L << PG_unevictable) |
2796#ifdef CONFIG_ARCH_USES_PG_ARCH_X
2797 (1L << PG_arch_2) |
2798 (1L << PG_arch_3) |
2799#endif
2800 (1L << PG_dirty) |
2801 LRU_GEN_MASK | LRU_REFS_MASK));
2802
2803 /* ->mapping in first and second tail page is replaced by other uses */
2804 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2805 page_tail);
2806 page_tail->mapping = head->mapping;
2807 page_tail->index = head->index + tail;
2808
2809 /*
2810 * page->private should not be set in tail pages. Fix up and warn once
2811 * if private is unexpectedly set.
2812 */
2813 if (unlikely(page_tail->private)) {
2814 VM_WARN_ON_ONCE_PAGE(true, page_tail);
2815 page_tail->private = 0;
2816 }
2817 if (folio_test_swapcache(folio))
2818 new_folio->swap.val = folio->swap.val + tail;
2819
2820 /* Page flags must be visible before we make the page non-compound. */
2821 smp_wmb();
2822
2823 /*
2824 * Clear PageTail before unfreezing page refcount.
2825 *
2826 * After successful get_page_unless_zero() might follow put_page()
2827 * which needs correct compound_head().
2828 */
2829 clear_compound_head(page_tail);
2830
2831 /* Finally unfreeze refcount. Additional reference from page cache. */
2832 page_ref_unfreeze(page_tail, 1 + (!folio_test_anon(folio) ||
2833 folio_test_swapcache(folio)));
2834
2835 if (folio_test_young(folio))
2836 folio_set_young(new_folio);
2837 if (folio_test_idle(folio))
2838 folio_set_idle(new_folio);
2839
2840 folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));
2841
2842 /*
2843 * always add to the tail because some iterators expect new
2844 * pages to show after the currently processed elements - e.g.
2845 * migrate_pages
2846 */
2847 lru_add_page_tail(head, page_tail, lruvec, list);
2848}
2849
2850static void __split_huge_page(struct page *page, struct list_head *list,
2851 pgoff_t end)
2852{
2853 struct folio *folio = page_folio(page);
2854 struct page *head = &folio->page;
2855 struct lruvec *lruvec;
2856 struct address_space *swap_cache = NULL;
2857 unsigned long offset = 0;
2858 unsigned int nr = thp_nr_pages(head);
2859 int i, nr_dropped = 0;
2860
2861 /* complete memcg works before add pages to LRU */
2862 split_page_memcg(head, nr);
2863
2864 if (folio_test_anon(folio) && folio_test_swapcache(folio)) {
2865 offset = swp_offset(folio->swap);
2866 swap_cache = swap_address_space(folio->swap);
2867 xa_lock(&swap_cache->i_pages);
2868 }
2869
2870 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2871 lruvec = folio_lruvec_lock(folio);
2872
2873 ClearPageHasHWPoisoned(head);
2874
2875 for (i = nr - 1; i >= 1; i--) {
2876 __split_huge_page_tail(folio, i, lruvec, list);
2877 /* Some pages can be beyond EOF: drop them from page cache */
2878 if (head[i].index >= end) {
2879 struct folio *tail = page_folio(head + i);
2880
2881 if (shmem_mapping(head->mapping))
2882 nr_dropped++;
2883 else if (folio_test_clear_dirty(tail))
2884 folio_account_cleaned(tail,
2885 inode_to_wb(folio->mapping->host));
2886 __filemap_remove_folio(tail, NULL);
2887 folio_put(tail);
2888 } else if (!PageAnon(page)) {
2889 __xa_store(&head->mapping->i_pages, head[i].index,
2890 head + i, 0);
2891 } else if (swap_cache) {
2892 __xa_store(&swap_cache->i_pages, offset + i,
2893 head + i, 0);
2894 }
2895 }
2896
2897 ClearPageCompound(head);
2898 unlock_page_lruvec(lruvec);
2899 /* Caller disabled irqs, so they are still disabled here */
2900
2901 split_page_owner(head, nr);
2902
2903 /* See comment in __split_huge_page_tail() */
2904 if (PageAnon(head)) {
2905 /* Additional pin to swap cache */
2906 if (PageSwapCache(head)) {
2907 page_ref_add(head, 2);
2908 xa_unlock(&swap_cache->i_pages);
2909 } else {
2910 page_ref_inc(head);
2911 }
2912 } else {
2913 /* Additional pin to page cache */
2914 page_ref_add(head, 2);
2915 xa_unlock(&head->mapping->i_pages);
2916 }
2917 local_irq_enable();
2918
2919 if (nr_dropped)
2920 shmem_uncharge(head->mapping->host, nr_dropped);
2921 remap_page(folio, nr);
2922
2923 if (folio_test_swapcache(folio))
2924 split_swap_cluster(folio->swap);
2925
2926 for (i = 0; i < nr; i++) {
2927 struct page *subpage = head + i;
2928 if (subpage == page)
2929 continue;
2930 unlock_page(subpage);
2931
2932 /*
2933 * Subpages may be freed if there wasn't any mapping
2934 * like if add_to_swap() is running on a lru page that
2935 * had its mapping zapped. And freeing these pages
2936 * requires taking the lru_lock so we do the put_page
2937 * of the tail pages after the split is complete.
2938 */
2939 free_page_and_swap_cache(subpage);
2940 }
2941}
2942
2943/* Racy check whether the huge page can be split */
2944bool can_split_folio(struct folio *folio, int *pextra_pins)
2945{
2946 int extra_pins;
2947
2948 /* Additional pins from page cache */
2949 if (folio_test_anon(folio))
2950 extra_pins = folio_test_swapcache(folio) ?
2951 folio_nr_pages(folio) : 0;
2952 else
2953 extra_pins = folio_nr_pages(folio);
2954 if (pextra_pins)
2955 *pextra_pins = extra_pins;
2956 return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2957}
2958
2959/*
2960 * This function splits huge page into normal pages. @page can point to any
2961 * subpage of huge page to split. Split doesn't change the position of @page.
2962 *
2963 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2964 * The huge page must be locked.
2965 *
2966 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2967 *
2968 * Both head page and tail pages will inherit mapping, flags, and so on from
2969 * the hugepage.
2970 *
2971 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2972 * they are not mapped.
2973 *
2974 * Returns 0 if the hugepage is split successfully.
2975 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2976 * us.
2977 */
2978int split_huge_page_to_list(struct page *page, struct list_head *list)
2979{
2980 struct folio *folio = page_folio(page);
2981 struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2982 XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2983 struct anon_vma *anon_vma = NULL;
2984 struct address_space *mapping = NULL;
2985 int extra_pins, ret;
2986 pgoff_t end;
2987 bool is_hzp;
2988
2989 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2990 VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2991
2992 is_hzp = is_huge_zero_page(&folio->page);
2993 if (is_hzp) {
2994 pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
2995 return -EBUSY;
2996 }
2997
2998 if (folio_test_writeback(folio))
2999 return -EBUSY;
3000
3001 if (folio_test_anon(folio)) {
3002 /*
3003 * The caller does not necessarily hold an mmap_lock that would
3004 * prevent the anon_vma disappearing so we first we take a
3005 * reference to it and then lock the anon_vma for write. This
3006 * is similar to folio_lock_anon_vma_read except the write lock
3007 * is taken to serialise against parallel split or collapse
3008 * operations.
3009 */
3010 anon_vma = folio_get_anon_vma(folio);
3011 if (!anon_vma) {
3012 ret = -EBUSY;
3013 goto out;
3014 }
3015 end = -1;
3016 mapping = NULL;
3017 anon_vma_lock_write(anon_vma);
3018 } else {
3019 gfp_t gfp;
3020
3021 mapping = folio->mapping;
3022
3023 /* Truncated ? */
3024 if (!mapping) {
3025 ret = -EBUSY;
3026 goto out;
3027 }
3028
3029 gfp = current_gfp_context(mapping_gfp_mask(mapping) &
3030 GFP_RECLAIM_MASK);
3031
3032 if (!filemap_release_folio(folio, gfp)) {
3033 ret = -EBUSY;
3034 goto out;
3035 }
3036
3037 xas_split_alloc(&xas, folio, folio_order(folio), gfp);
3038 if (xas_error(&xas)) {
3039 ret = xas_error(&xas);
3040 goto out;
3041 }
3042
3043 anon_vma = NULL;
3044 i_mmap_lock_read(mapping);
3045
3046 /*
3047 *__split_huge_page() may need to trim off pages beyond EOF:
3048 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
3049 * which cannot be nested inside the page tree lock. So note
3050 * end now: i_size itself may be changed at any moment, but
3051 * folio lock is good enough to serialize the trimming.
3052 */
3053 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3054 if (shmem_mapping(mapping))
3055 end = shmem_fallocend(mapping->host, end);
3056 }
3057
3058 /*
3059 * Racy check if we can split the page, before unmap_folio() will
3060 * split PMDs
3061 */
3062 if (!can_split_folio(folio, &extra_pins)) {
3063 ret = -EAGAIN;
3064 goto out_unlock;
3065 }
3066
3067 unmap_folio(folio);
3068
3069 /* block interrupt reentry in xa_lock and spinlock */
3070 local_irq_disable();
3071 if (mapping) {
3072 /*
3073 * Check if the folio is present in page cache.
3074 * We assume all tail are present too, if folio is there.
3075 */
3076 xas_lock(&xas);
3077 xas_reset(&xas);
3078 if (xas_load(&xas) != folio)
3079 goto fail;
3080 }
3081
3082 /* Prevent deferred_split_scan() touching ->_refcount */
3083 spin_lock(&ds_queue->split_queue_lock);
3084 if (folio_ref_freeze(folio, 1 + extra_pins)) {
3085 if (!list_empty(&folio->_deferred_list)) {
3086 ds_queue->split_queue_len--;
3087 list_del(&folio->_deferred_list);
3088 }
3089 spin_unlock(&ds_queue->split_queue_lock);
3090 if (mapping) {
3091 int nr = folio_nr_pages(folio);
3092
3093 xas_split(&xas, folio, folio_order(folio));
3094 if (folio_test_pmd_mappable(folio)) {
3095 if (folio_test_swapbacked(folio)) {
3096 __lruvec_stat_mod_folio(folio,
3097 NR_SHMEM_THPS, -nr);
3098 } else {
3099 __lruvec_stat_mod_folio(folio,
3100 NR_FILE_THPS, -nr);
3101 filemap_nr_thps_dec(mapping);
3102 }
3103 }
3104 }
3105
3106 __split_huge_page(page, list, end);
3107 ret = 0;
3108 } else {
3109 spin_unlock(&ds_queue->split_queue_lock);
3110fail:
3111 if (mapping)
3112 xas_unlock(&xas);
3113 local_irq_enable();
3114 remap_page(folio, folio_nr_pages(folio));
3115 ret = -EAGAIN;
3116 }
3117
3118out_unlock:
3119 if (anon_vma) {
3120 anon_vma_unlock_write(anon_vma);
3121 put_anon_vma(anon_vma);
3122 }
3123 if (mapping)
3124 i_mmap_unlock_read(mapping);
3125out:
3126 xas_destroy(&xas);
3127 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3128 return ret;
3129}
3130
3131void folio_undo_large_rmappable(struct folio *folio)
3132{
3133 struct deferred_split *ds_queue;
3134 unsigned long flags;
3135
3136 /*
3137 * At this point, there is no one trying to add the folio to
3138 * deferred_list. If folio is not in deferred_list, it's safe
3139 * to check without acquiring the split_queue_lock.
3140 */
3141 if (data_race(list_empty(&folio->_deferred_list)))
3142 return;
3143
3144 ds_queue = get_deferred_split_queue(folio);
3145 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3146 if (!list_empty(&folio->_deferred_list)) {
3147 ds_queue->split_queue_len--;
3148 list_del_init(&folio->_deferred_list);
3149 }
3150 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3151}
3152
3153void deferred_split_folio(struct folio *folio)
3154{
3155 struct deferred_split *ds_queue = get_deferred_split_queue(folio);
3156#ifdef CONFIG_MEMCG
3157 struct mem_cgroup *memcg = folio_memcg(folio);
3158#endif
3159 unsigned long flags;
3160
3161 VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
3162
3163 /*
3164 * The try_to_unmap() in page reclaim path might reach here too,
3165 * this may cause a race condition to corrupt deferred split queue.
3166 * And, if page reclaim is already handling the same folio, it is
3167 * unnecessary to handle it again in shrinker.
3168 *
3169 * Check the swapcache flag to determine if the folio is being
3170 * handled by page reclaim since THP swap would add the folio into
3171 * swap cache before calling try_to_unmap().
3172 */
3173 if (folio_test_swapcache(folio))
3174 return;
3175
3176 if (!list_empty(&folio->_deferred_list))
3177 return;
3178
3179 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3180 if (list_empty(&folio->_deferred_list)) {
3181 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3182 list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
3183 ds_queue->split_queue_len++;
3184#ifdef CONFIG_MEMCG
3185 if (memcg)
3186 set_shrinker_bit(memcg, folio_nid(folio),
3187 deferred_split_shrinker->id);
3188#endif
3189 }
3190 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3191}
3192
3193static unsigned long deferred_split_count(struct shrinker *shrink,
3194 struct shrink_control *sc)
3195{
3196 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3197 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
3198
3199#ifdef CONFIG_MEMCG
3200 if (sc->memcg)
3201 ds_queue = &sc->memcg->deferred_split_queue;
3202#endif
3203 return READ_ONCE(ds_queue->split_queue_len);
3204}
3205
3206static unsigned long deferred_split_scan(struct shrinker *shrink,
3207 struct shrink_control *sc)
3208{
3209 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3210 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
3211 unsigned long flags;
3212 LIST_HEAD(list);
3213 struct folio *folio, *next;
3214 int split = 0;
3215
3216#ifdef CONFIG_MEMCG
3217 if (sc->memcg)
3218 ds_queue = &sc->memcg->deferred_split_queue;
3219#endif
3220
3221 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3222 /* Take pin on all head pages to avoid freeing them under us */
3223 list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
3224 _deferred_list) {
3225 if (folio_try_get(folio)) {
3226 list_move(&folio->_deferred_list, &list);
3227 } else {
3228 /* We lost race with folio_put() */
3229 list_del_init(&folio->_deferred_list);
3230 ds_queue->split_queue_len--;
3231 }
3232 if (!--sc->nr_to_scan)
3233 break;
3234 }
3235 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3236
3237 list_for_each_entry_safe(folio, next, &list, _deferred_list) {
3238 if (!folio_trylock(folio))
3239 goto next;
3240 /* split_huge_page() removes page from list on success */
3241 if (!split_folio(folio))
3242 split++;
3243 folio_unlock(folio);
3244next:
3245 folio_put(folio);
3246 }
3247
3248 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3249 list_splice_tail(&list, &ds_queue->split_queue);
3250 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3251
3252 /*
3253 * Stop shrinker if we didn't split any page, but the queue is empty.
3254 * This can happen if pages were freed under us.
3255 */
3256 if (!split && list_empty(&ds_queue->split_queue))
3257 return SHRINK_STOP;
3258 return split;
3259}
3260
3261#ifdef CONFIG_DEBUG_FS
3262static void split_huge_pages_all(void)
3263{
3264 struct zone *zone;
3265 struct page *page;
3266 struct folio *folio;
3267 unsigned long pfn, max_zone_pfn;
3268 unsigned long total = 0, split = 0;
3269
3270 pr_debug("Split all THPs\n");
3271 for_each_zone(zone) {
3272 if (!managed_zone(zone))
3273 continue;
3274 max_zone_pfn = zone_end_pfn(zone);
3275 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3276 int nr_pages;
3277
3278 page = pfn_to_online_page(pfn);
3279 if (!page || PageTail(page))
3280 continue;
3281 folio = page_folio(page);
3282 if (!folio_try_get(folio))
3283 continue;
3284
3285 if (unlikely(page_folio(page) != folio))
3286 goto next;
3287
3288 if (zone != folio_zone(folio))
3289 goto next;
3290
3291 if (!folio_test_large(folio)
3292 || folio_test_hugetlb(folio)
3293 || !folio_test_lru(folio))
3294 goto next;
3295
3296 total++;
3297 folio_lock(folio);
3298 nr_pages = folio_nr_pages(folio);
3299 if (!split_folio(folio))
3300 split++;
3301 pfn += nr_pages - 1;
3302 folio_unlock(folio);
3303next:
3304 folio_put(folio);
3305 cond_resched();
3306 }
3307 }
3308
3309 pr_debug("%lu of %lu THP split\n", split, total);
3310}
3311
3312static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
3313{
3314 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
3315 is_vm_hugetlb_page(vma);
3316}
3317
3318static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
3319 unsigned long vaddr_end)
3320{
3321 int ret = 0;
3322 struct task_struct *task;
3323 struct mm_struct *mm;
3324 unsigned long total = 0, split = 0;
3325 unsigned long addr;
3326
3327 vaddr_start &= PAGE_MASK;
3328 vaddr_end &= PAGE_MASK;
3329
3330 /* Find the task_struct from pid */
3331 rcu_read_lock();
3332 task = find_task_by_vpid(pid);
3333 if (!task) {
3334 rcu_read_unlock();
3335 ret = -ESRCH;
3336 goto out;
3337 }
3338 get_task_struct(task);
3339 rcu_read_unlock();
3340
3341 /* Find the mm_struct */
3342 mm = get_task_mm(task);
3343 put_task_struct(task);
3344
3345 if (!mm) {
3346 ret = -EINVAL;
3347 goto out;
3348 }
3349
3350 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3351 pid, vaddr_start, vaddr_end);
3352
3353 mmap_read_lock(mm);
3354 /*
3355 * always increase addr by PAGE_SIZE, since we could have a PTE page
3356 * table filled with PTE-mapped THPs, each of which is distinct.
3357 */
3358 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3359 struct vm_area_struct *vma = vma_lookup(mm, addr);
3360 struct page *page;
3361 struct folio *folio;
3362
3363 if (!vma)
3364 break;
3365
3366 /* skip special VMA and hugetlb VMA */
3367 if (vma_not_suitable_for_thp_split(vma)) {
3368 addr = vma->vm_end;
3369 continue;
3370 }
3371
3372 /* FOLL_DUMP to ignore special (like zero) pages */
3373 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3374
3375 if (IS_ERR_OR_NULL(page))
3376 continue;
3377
3378 folio = page_folio(page);
3379 if (!is_transparent_hugepage(folio))
3380 goto next;
3381
3382 total++;
3383 if (!can_split_folio(folio, NULL))
3384 goto next;
3385
3386 if (!folio_trylock(folio))
3387 goto next;
3388
3389 if (!split_folio(folio))
3390 split++;
3391
3392 folio_unlock(folio);
3393next:
3394 folio_put(folio);
3395 cond_resched();
3396 }
3397 mmap_read_unlock(mm);
3398 mmput(mm);
3399
3400 pr_debug("%lu of %lu THP split\n", split, total);
3401
3402out:
3403 return ret;
3404}
3405
3406static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3407 pgoff_t off_end)
3408{
3409 struct filename *file;
3410 struct file *candidate;
3411 struct address_space *mapping;
3412 int ret = -EINVAL;
3413 pgoff_t index;
3414 int nr_pages = 1;
3415 unsigned long total = 0, split = 0;
3416
3417 file = getname_kernel(file_path);
3418 if (IS_ERR(file))
3419 return ret;
3420
3421 candidate = file_open_name(file, O_RDONLY, 0);
3422 if (IS_ERR(candidate))
3423 goto out;
3424
3425 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3426 file_path, off_start, off_end);
3427
3428 mapping = candidate->f_mapping;
3429
3430 for (index = off_start; index < off_end; index += nr_pages) {
3431 struct folio *folio = filemap_get_folio(mapping, index);
3432
3433 nr_pages = 1;
3434 if (IS_ERR(folio))
3435 continue;
3436
3437 if (!folio_test_large(folio))
3438 goto next;
3439
3440 total++;
3441 nr_pages = folio_nr_pages(folio);
3442
3443 if (!folio_trylock(folio))
3444 goto next;
3445
3446 if (!split_folio(folio))
3447 split++;
3448
3449 folio_unlock(folio);
3450next:
3451 folio_put(folio);
3452 cond_resched();
3453 }
3454
3455 filp_close(candidate, NULL);
3456 ret = 0;
3457
3458 pr_debug("%lu of %lu file-backed THP split\n", split, total);
3459out:
3460 putname(file);
3461 return ret;
3462}
3463
3464#define MAX_INPUT_BUF_SZ 255
3465
3466static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3467 size_t count, loff_t *ppops)
3468{
3469 static DEFINE_MUTEX(split_debug_mutex);
3470 ssize_t ret;
3471 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3472 char input_buf[MAX_INPUT_BUF_SZ];
3473 int pid;
3474 unsigned long vaddr_start, vaddr_end;
3475
3476 ret = mutex_lock_interruptible(&split_debug_mutex);
3477 if (ret)
3478 return ret;
3479
3480 ret = -EFAULT;
3481
3482 memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3483 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3484 goto out;
3485
3486 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3487
3488 if (input_buf[0] == '/') {
3489 char *tok;
3490 char *buf = input_buf;
3491 char file_path[MAX_INPUT_BUF_SZ];
3492 pgoff_t off_start = 0, off_end = 0;
3493 size_t input_len = strlen(input_buf);
3494
3495 tok = strsep(&buf, ",");
3496 if (tok) {
3497 strcpy(file_path, tok);
3498 } else {
3499 ret = -EINVAL;
3500 goto out;
3501 }
3502
3503 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3504 if (ret != 2) {
3505 ret = -EINVAL;
3506 goto out;
3507 }
3508 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3509 if (!ret)
3510 ret = input_len;
3511
3512 goto out;
3513 }
3514
3515 ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3516 if (ret == 1 && pid == 1) {
3517 split_huge_pages_all();
3518 ret = strlen(input_buf);
3519 goto out;
3520 } else if (ret != 3) {
3521 ret = -EINVAL;
3522 goto out;
3523 }
3524
3525 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3526 if (!ret)
3527 ret = strlen(input_buf);
3528out:
3529 mutex_unlock(&split_debug_mutex);
3530 return ret;
3531
3532}
3533
3534static const struct file_operations split_huge_pages_fops = {
3535 .owner = THIS_MODULE,
3536 .write = split_huge_pages_write,
3537 .llseek = no_llseek,
3538};
3539
3540static int __init split_huge_pages_debugfs(void)
3541{
3542 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3543 &split_huge_pages_fops);
3544 return 0;
3545}
3546late_initcall(split_huge_pages_debugfs);
3547#endif
3548
3549#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3550int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3551 struct page *page)
3552{
3553 struct folio *folio = page_folio(page);
3554 struct vm_area_struct *vma = pvmw->vma;
3555 struct mm_struct *mm = vma->vm_mm;
3556 unsigned long address = pvmw->address;
3557 bool anon_exclusive;
3558 pmd_t pmdval;
3559 swp_entry_t entry;
3560 pmd_t pmdswp;
3561
3562 if (!(pvmw->pmd && !pvmw->pte))
3563 return 0;
3564
3565 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3566 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3567
3568 /* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */
3569 anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page);
3570 if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) {
3571 set_pmd_at(mm, address, pvmw->pmd, pmdval);
3572 return -EBUSY;
3573 }
3574
3575 if (pmd_dirty(pmdval))
3576 folio_mark_dirty(folio);
3577 if (pmd_write(pmdval))
3578 entry = make_writable_migration_entry(page_to_pfn(page));
3579 else if (anon_exclusive)
3580 entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3581 else
3582 entry = make_readable_migration_entry(page_to_pfn(page));
3583 if (pmd_young(pmdval))
3584 entry = make_migration_entry_young(entry);
3585 if (pmd_dirty(pmdval))
3586 entry = make_migration_entry_dirty(entry);
3587 pmdswp = swp_entry_to_pmd(entry);
3588 if (pmd_soft_dirty(pmdval))
3589 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3590 if (pmd_uffd_wp(pmdval))
3591 pmdswp = pmd_swp_mkuffd_wp(pmdswp);
3592 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3593 folio_remove_rmap_pmd(folio, page, vma);
3594 folio_put(folio);
3595 trace_set_migration_pmd(address, pmd_val(pmdswp));
3596
3597 return 0;
3598}
3599
3600void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3601{
3602 struct folio *folio = page_folio(new);
3603 struct vm_area_struct *vma = pvmw->vma;
3604 struct mm_struct *mm = vma->vm_mm;
3605 unsigned long address = pvmw->address;
3606 unsigned long haddr = address & HPAGE_PMD_MASK;
3607 pmd_t pmde;
3608 swp_entry_t entry;
3609
3610 if (!(pvmw->pmd && !pvmw->pte))
3611 return;
3612
3613 entry = pmd_to_swp_entry(*pvmw->pmd);
3614 folio_get(folio);
3615 pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3616 if (pmd_swp_soft_dirty(*pvmw->pmd))
3617 pmde = pmd_mksoft_dirty(pmde);
3618 if (is_writable_migration_entry(entry))
3619 pmde = pmd_mkwrite(pmde, vma);
3620 if (pmd_swp_uffd_wp(*pvmw->pmd))
3621 pmde = pmd_mkuffd_wp(pmde);
3622 if (!is_migration_entry_young(entry))
3623 pmde = pmd_mkold(pmde);
3624 /* NOTE: this may contain setting soft-dirty on some archs */
3625 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
3626 pmde = pmd_mkdirty(pmde);
3627
3628 if (folio_test_anon(folio)) {
3629 rmap_t rmap_flags = RMAP_NONE;
3630
3631 if (!is_readable_migration_entry(entry))
3632 rmap_flags |= RMAP_EXCLUSIVE;
3633
3634 folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags);
3635 } else {
3636 folio_add_file_rmap_pmd(folio, new, vma);
3637 }
3638 VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new));
3639 set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3640
3641 /* No need to invalidate - it was non-present before */
3642 update_mmu_cache_pmd(vma, address, pvmw->pmd);
3643 trace_remove_migration_pmd(address, pmd_val(pmde));
3644}
3645#endif