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