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/coredump.h>
11#include <linux/sched/numa_balancing.h>
12#include <linux/highmem.h>
13#include <linux/hugetlb.h>
14#include <linux/mmu_notifier.h>
15#include <linux/rmap.h>
16#include <linux/swap.h>
17#include <linux/shrinker.h>
18#include <linux/mm_inline.h>
19#include <linux/swapops.h>
20#include <linux/dax.h>
21#include <linux/khugepaged.h>
22#include <linux/freezer.h>
23#include <linux/pfn_t.h>
24#include <linux/mman.h>
25#include <linux/memremap.h>
26#include <linux/pagemap.h>
27#include <linux/debugfs.h>
28#include <linux/migrate.h>
29#include <linux/hashtable.h>
30#include <linux/userfaultfd_k.h>
31#include <linux/page_idle.h>
32#include <linux/shmem_fs.h>
33#include <linux/oom.h>
34#include <linux/numa.h>
35#include <linux/page_owner.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
65bool transparent_hugepage_enabled(struct vm_area_struct *vma)
66{
67 /* The addr is used to check if the vma size fits */
68 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
69
70 if (!transhuge_vma_suitable(vma, addr))
71 return false;
72 if (vma_is_anonymous(vma))
73 return __transparent_hugepage_enabled(vma);
74 if (vma_is_shmem(vma))
75 return shmem_huge_enabled(vma);
76
77 return false;
78}
79
80static struct page *get_huge_zero_page(void)
81{
82 struct page *zero_page;
83retry:
84 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
85 return READ_ONCE(huge_zero_page);
86
87 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
88 HPAGE_PMD_ORDER);
89 if (!zero_page) {
90 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
91 return NULL;
92 }
93 count_vm_event(THP_ZERO_PAGE_ALLOC);
94 preempt_disable();
95 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
96 preempt_enable();
97 __free_pages(zero_page, compound_order(zero_page));
98 goto retry;
99 }
100
101 /* We take additional reference here. It will be put back by shrinker */
102 atomic_set(&huge_zero_refcount, 2);
103 preempt_enable();
104 return READ_ONCE(huge_zero_page);
105}
106
107static void put_huge_zero_page(void)
108{
109 /*
110 * Counter should never go to zero here. Only shrinker can put
111 * last reference.
112 */
113 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
114}
115
116struct page *mm_get_huge_zero_page(struct mm_struct *mm)
117{
118 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
119 return READ_ONCE(huge_zero_page);
120
121 if (!get_huge_zero_page())
122 return NULL;
123
124 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
125 put_huge_zero_page();
126
127 return READ_ONCE(huge_zero_page);
128}
129
130void mm_put_huge_zero_page(struct mm_struct *mm)
131{
132 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
133 put_huge_zero_page();
134}
135
136static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
137 struct shrink_control *sc)
138{
139 /* we can free zero page only if last reference remains */
140 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
141}
142
143static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
144 struct shrink_control *sc)
145{
146 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
147 struct page *zero_page = xchg(&huge_zero_page, NULL);
148 BUG_ON(zero_page == NULL);
149 __free_pages(zero_page, compound_order(zero_page));
150 return HPAGE_PMD_NR;
151 }
152
153 return 0;
154}
155
156static struct shrinker huge_zero_page_shrinker = {
157 .count_objects = shrink_huge_zero_page_count,
158 .scan_objects = shrink_huge_zero_page_scan,
159 .seeks = DEFAULT_SEEKS,
160};
161
162#ifdef CONFIG_SYSFS
163static ssize_t enabled_show(struct kobject *kobj,
164 struct kobj_attribute *attr, char *buf)
165{
166 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
167 return sprintf(buf, "[always] madvise never\n");
168 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
169 return sprintf(buf, "always [madvise] never\n");
170 else
171 return sprintf(buf, "always madvise [never]\n");
172}
173
174static ssize_t enabled_store(struct kobject *kobj,
175 struct kobj_attribute *attr,
176 const char *buf, size_t count)
177{
178 ssize_t ret = count;
179
180 if (sysfs_streq(buf, "always")) {
181 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
182 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
183 } else if (sysfs_streq(buf, "madvise")) {
184 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
185 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
186 } else if (sysfs_streq(buf, "never")) {
187 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
188 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
189 } else
190 ret = -EINVAL;
191
192 if (ret > 0) {
193 int err = start_stop_khugepaged();
194 if (err)
195 ret = err;
196 }
197 return ret;
198}
199static struct kobj_attribute enabled_attr =
200 __ATTR(enabled, 0644, enabled_show, enabled_store);
201
202ssize_t single_hugepage_flag_show(struct kobject *kobj,
203 struct kobj_attribute *attr, char *buf,
204 enum transparent_hugepage_flag flag)
205{
206 return sprintf(buf, "%d\n",
207 !!test_bit(flag, &transparent_hugepage_flags));
208}
209
210ssize_t single_hugepage_flag_store(struct kobject *kobj,
211 struct kobj_attribute *attr,
212 const char *buf, size_t count,
213 enum transparent_hugepage_flag flag)
214{
215 unsigned long value;
216 int ret;
217
218 ret = kstrtoul(buf, 10, &value);
219 if (ret < 0)
220 return ret;
221 if (value > 1)
222 return -EINVAL;
223
224 if (value)
225 set_bit(flag, &transparent_hugepage_flags);
226 else
227 clear_bit(flag, &transparent_hugepage_flags);
228
229 return count;
230}
231
232static ssize_t defrag_show(struct kobject *kobj,
233 struct kobj_attribute *attr, char *buf)
234{
235 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
236 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
237 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
238 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
239 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
240 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
241 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
242 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
243 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
244}
245
246static ssize_t defrag_store(struct kobject *kobj,
247 struct kobj_attribute *attr,
248 const char *buf, size_t count)
249{
250 if (sysfs_streq(buf, "always")) {
251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
255 } else if (sysfs_streq(buf, "defer+madvise")) {
256 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
259 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
260 } else if (sysfs_streq(buf, "defer")) {
261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
264 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
265 } else if (sysfs_streq(buf, "madvise")) {
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
269 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
270 } else if (sysfs_streq(buf, "never")) {
271 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
272 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
273 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
275 } else
276 return -EINVAL;
277
278 return count;
279}
280static struct kobj_attribute defrag_attr =
281 __ATTR(defrag, 0644, defrag_show, defrag_store);
282
283static ssize_t use_zero_page_show(struct kobject *kobj,
284 struct kobj_attribute *attr, char *buf)
285{
286 return single_hugepage_flag_show(kobj, attr, buf,
287 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
288}
289static ssize_t use_zero_page_store(struct kobject *kobj,
290 struct kobj_attribute *attr, const char *buf, size_t count)
291{
292 return single_hugepage_flag_store(kobj, attr, buf, count,
293 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
294}
295static struct kobj_attribute use_zero_page_attr =
296 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
297
298static ssize_t hpage_pmd_size_show(struct kobject *kobj,
299 struct kobj_attribute *attr, char *buf)
300{
301 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
302}
303static struct kobj_attribute hpage_pmd_size_attr =
304 __ATTR_RO(hpage_pmd_size);
305
306static struct attribute *hugepage_attr[] = {
307 &enabled_attr.attr,
308 &defrag_attr.attr,
309 &use_zero_page_attr.attr,
310 &hpage_pmd_size_attr.attr,
311#ifdef CONFIG_SHMEM
312 &shmem_enabled_attr.attr,
313#endif
314 NULL,
315};
316
317static const struct attribute_group hugepage_attr_group = {
318 .attrs = hugepage_attr,
319};
320
321static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
322{
323 int err;
324
325 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
326 if (unlikely(!*hugepage_kobj)) {
327 pr_err("failed to create transparent hugepage kobject\n");
328 return -ENOMEM;
329 }
330
331 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
332 if (err) {
333 pr_err("failed to register transparent hugepage group\n");
334 goto delete_obj;
335 }
336
337 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
338 if (err) {
339 pr_err("failed to register transparent hugepage group\n");
340 goto remove_hp_group;
341 }
342
343 return 0;
344
345remove_hp_group:
346 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
347delete_obj:
348 kobject_put(*hugepage_kobj);
349 return err;
350}
351
352static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
353{
354 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
355 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
356 kobject_put(hugepage_kobj);
357}
358#else
359static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
360{
361 return 0;
362}
363
364static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
365{
366}
367#endif /* CONFIG_SYSFS */
368
369static int __init hugepage_init(void)
370{
371 int err;
372 struct kobject *hugepage_kobj;
373
374 if (!has_transparent_hugepage()) {
375 transparent_hugepage_flags = 0;
376 return -EINVAL;
377 }
378
379 /*
380 * hugepages can't be allocated by the buddy allocator
381 */
382 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
383 /*
384 * we use page->mapping and page->index in second tail page
385 * as list_head: assuming THP order >= 2
386 */
387 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
388
389 err = hugepage_init_sysfs(&hugepage_kobj);
390 if (err)
391 goto err_sysfs;
392
393 err = khugepaged_init();
394 if (err)
395 goto err_slab;
396
397 err = register_shrinker(&huge_zero_page_shrinker);
398 if (err)
399 goto err_hzp_shrinker;
400 err = register_shrinker(&deferred_split_shrinker);
401 if (err)
402 goto err_split_shrinker;
403
404 /*
405 * By default disable transparent hugepages on smaller systems,
406 * where the extra memory used could hurt more than TLB overhead
407 * is likely to save. The admin can still enable it through /sys.
408 */
409 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
410 transparent_hugepage_flags = 0;
411 return 0;
412 }
413
414 err = start_stop_khugepaged();
415 if (err)
416 goto err_khugepaged;
417
418 return 0;
419err_khugepaged:
420 unregister_shrinker(&deferred_split_shrinker);
421err_split_shrinker:
422 unregister_shrinker(&huge_zero_page_shrinker);
423err_hzp_shrinker:
424 khugepaged_destroy();
425err_slab:
426 hugepage_exit_sysfs(hugepage_kobj);
427err_sysfs:
428 return err;
429}
430subsys_initcall(hugepage_init);
431
432static int __init setup_transparent_hugepage(char *str)
433{
434 int ret = 0;
435 if (!str)
436 goto out;
437 if (!strcmp(str, "always")) {
438 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
439 &transparent_hugepage_flags);
440 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
441 &transparent_hugepage_flags);
442 ret = 1;
443 } else if (!strcmp(str, "madvise")) {
444 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
445 &transparent_hugepage_flags);
446 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
447 &transparent_hugepage_flags);
448 ret = 1;
449 } else if (!strcmp(str, "never")) {
450 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
451 &transparent_hugepage_flags);
452 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
453 &transparent_hugepage_flags);
454 ret = 1;
455 }
456out:
457 if (!ret)
458 pr_warn("transparent_hugepage= cannot parse, ignored\n");
459 return ret;
460}
461__setup("transparent_hugepage=", setup_transparent_hugepage);
462
463pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
464{
465 if (likely(vma->vm_flags & VM_WRITE))
466 pmd = pmd_mkwrite(pmd);
467 return pmd;
468}
469
470#ifdef CONFIG_MEMCG
471static inline struct deferred_split *get_deferred_split_queue(struct page *page)
472{
473 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
474 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
475
476 if (memcg)
477 return &memcg->deferred_split_queue;
478 else
479 return &pgdat->deferred_split_queue;
480}
481#else
482static inline struct deferred_split *get_deferred_split_queue(struct page *page)
483{
484 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
485
486 return &pgdat->deferred_split_queue;
487}
488#endif
489
490void prep_transhuge_page(struct page *page)
491{
492 /*
493 * we use page->mapping and page->indexlru in second tail page
494 * as list_head: assuming THP order >= 2
495 */
496
497 INIT_LIST_HEAD(page_deferred_list(page));
498 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
499}
500
501bool is_transparent_hugepage(struct page *page)
502{
503 if (!PageCompound(page))
504 return false;
505
506 page = compound_head(page);
507 return is_huge_zero_page(page) ||
508 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
509}
510EXPORT_SYMBOL_GPL(is_transparent_hugepage);
511
512static unsigned long __thp_get_unmapped_area(struct file *filp,
513 unsigned long addr, unsigned long len,
514 loff_t off, unsigned long flags, unsigned long size)
515{
516 loff_t off_end = off + len;
517 loff_t off_align = round_up(off, size);
518 unsigned long len_pad, ret;
519
520 if (off_end <= off_align || (off_end - off_align) < size)
521 return 0;
522
523 len_pad = len + size;
524 if (len_pad < len || (off + len_pad) < off)
525 return 0;
526
527 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
528 off >> PAGE_SHIFT, flags);
529
530 /*
531 * The failure might be due to length padding. The caller will retry
532 * without the padding.
533 */
534 if (IS_ERR_VALUE(ret))
535 return 0;
536
537 /*
538 * Do not try to align to THP boundary if allocation at the address
539 * hint succeeds.
540 */
541 if (ret == addr)
542 return addr;
543
544 ret += (off - ret) & (size - 1);
545 return ret;
546}
547
548unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
549 unsigned long len, unsigned long pgoff, unsigned long flags)
550{
551 unsigned long ret;
552 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
553
554 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
555 goto out;
556
557 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
558 if (ret)
559 return ret;
560out:
561 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
562}
563EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
564
565static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
566 struct page *page, gfp_t gfp)
567{
568 struct vm_area_struct *vma = vmf->vma;
569 pgtable_t pgtable;
570 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
571 vm_fault_t ret = 0;
572
573 VM_BUG_ON_PAGE(!PageCompound(page), page);
574
575 if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
576 put_page(page);
577 count_vm_event(THP_FAULT_FALLBACK);
578 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
579 return VM_FAULT_FALLBACK;
580 }
581 cgroup_throttle_swaprate(page, gfp);
582
583 pgtable = pte_alloc_one(vma->vm_mm);
584 if (unlikely(!pgtable)) {
585 ret = VM_FAULT_OOM;
586 goto release;
587 }
588
589 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
590 /*
591 * The memory barrier inside __SetPageUptodate makes sure that
592 * clear_huge_page writes become visible before the set_pmd_at()
593 * write.
594 */
595 __SetPageUptodate(page);
596
597 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
598 if (unlikely(!pmd_none(*vmf->pmd))) {
599 goto unlock_release;
600 } else {
601 pmd_t entry;
602
603 ret = check_stable_address_space(vma->vm_mm);
604 if (ret)
605 goto unlock_release;
606
607 /* Deliver the page fault to userland */
608 if (userfaultfd_missing(vma)) {
609 vm_fault_t ret2;
610
611 spin_unlock(vmf->ptl);
612 put_page(page);
613 pte_free(vma->vm_mm, pgtable);
614 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
615 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
616 return ret2;
617 }
618
619 entry = mk_huge_pmd(page, vma->vm_page_prot);
620 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
621 page_add_new_anon_rmap(page, vma, haddr, true);
622 lru_cache_add_inactive_or_unevictable(page, vma);
623 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
624 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
625 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
626 mm_inc_nr_ptes(vma->vm_mm);
627 spin_unlock(vmf->ptl);
628 count_vm_event(THP_FAULT_ALLOC);
629 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
630 }
631
632 return 0;
633unlock_release:
634 spin_unlock(vmf->ptl);
635release:
636 if (pgtable)
637 pte_free(vma->vm_mm, pgtable);
638 put_page(page);
639 return ret;
640
641}
642
643/*
644 * always: directly stall for all thp allocations
645 * defer: wake kswapd and fail if not immediately available
646 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
647 * fail if not immediately available
648 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
649 * available
650 * never: never stall for any thp allocation
651 */
652static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
653{
654 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
655
656 /* Always do synchronous compaction */
657 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
658 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
659
660 /* Kick kcompactd and fail quickly */
661 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
662 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
663
664 /* Synchronous compaction if madvised, otherwise kick kcompactd */
665 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
666 return GFP_TRANSHUGE_LIGHT |
667 (vma_madvised ? __GFP_DIRECT_RECLAIM :
668 __GFP_KSWAPD_RECLAIM);
669
670 /* Only do synchronous compaction if madvised */
671 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
672 return GFP_TRANSHUGE_LIGHT |
673 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
674
675 return GFP_TRANSHUGE_LIGHT;
676}
677
678/* Caller must hold page table lock. */
679static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
680 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
681 struct page *zero_page)
682{
683 pmd_t entry;
684 if (!pmd_none(*pmd))
685 return false;
686 entry = mk_pmd(zero_page, vma->vm_page_prot);
687 entry = pmd_mkhuge(entry);
688 if (pgtable)
689 pgtable_trans_huge_deposit(mm, pmd, pgtable);
690 set_pmd_at(mm, haddr, pmd, entry);
691 mm_inc_nr_ptes(mm);
692 return true;
693}
694
695vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
696{
697 struct vm_area_struct *vma = vmf->vma;
698 gfp_t gfp;
699 struct page *page;
700 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
701
702 if (!transhuge_vma_suitable(vma, haddr))
703 return VM_FAULT_FALLBACK;
704 if (unlikely(anon_vma_prepare(vma)))
705 return VM_FAULT_OOM;
706 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
707 return VM_FAULT_OOM;
708 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
709 !mm_forbids_zeropage(vma->vm_mm) &&
710 transparent_hugepage_use_zero_page()) {
711 pgtable_t pgtable;
712 struct page *zero_page;
713 bool set;
714 vm_fault_t ret;
715 pgtable = pte_alloc_one(vma->vm_mm);
716 if (unlikely(!pgtable))
717 return VM_FAULT_OOM;
718 zero_page = mm_get_huge_zero_page(vma->vm_mm);
719 if (unlikely(!zero_page)) {
720 pte_free(vma->vm_mm, pgtable);
721 count_vm_event(THP_FAULT_FALLBACK);
722 return VM_FAULT_FALLBACK;
723 }
724 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
725 ret = 0;
726 set = false;
727 if (pmd_none(*vmf->pmd)) {
728 ret = check_stable_address_space(vma->vm_mm);
729 if (ret) {
730 spin_unlock(vmf->ptl);
731 } else if (userfaultfd_missing(vma)) {
732 spin_unlock(vmf->ptl);
733 ret = handle_userfault(vmf, VM_UFFD_MISSING);
734 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
735 } else {
736 set_huge_zero_page(pgtable, vma->vm_mm, vma,
737 haddr, vmf->pmd, zero_page);
738 spin_unlock(vmf->ptl);
739 set = true;
740 }
741 } else
742 spin_unlock(vmf->ptl);
743 if (!set)
744 pte_free(vma->vm_mm, pgtable);
745 return ret;
746 }
747 gfp = alloc_hugepage_direct_gfpmask(vma);
748 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
749 if (unlikely(!page)) {
750 count_vm_event(THP_FAULT_FALLBACK);
751 return VM_FAULT_FALLBACK;
752 }
753 prep_transhuge_page(page);
754 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
755}
756
757static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
758 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
759 pgtable_t pgtable)
760{
761 struct mm_struct *mm = vma->vm_mm;
762 pmd_t entry;
763 spinlock_t *ptl;
764
765 ptl = pmd_lock(mm, pmd);
766 if (!pmd_none(*pmd)) {
767 if (write) {
768 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
769 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
770 goto out_unlock;
771 }
772 entry = pmd_mkyoung(*pmd);
773 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
774 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
775 update_mmu_cache_pmd(vma, addr, pmd);
776 }
777
778 goto out_unlock;
779 }
780
781 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
782 if (pfn_t_devmap(pfn))
783 entry = pmd_mkdevmap(entry);
784 if (write) {
785 entry = pmd_mkyoung(pmd_mkdirty(entry));
786 entry = maybe_pmd_mkwrite(entry, vma);
787 }
788
789 if (pgtable) {
790 pgtable_trans_huge_deposit(mm, pmd, pgtable);
791 mm_inc_nr_ptes(mm);
792 pgtable = NULL;
793 }
794
795 set_pmd_at(mm, addr, pmd, entry);
796 update_mmu_cache_pmd(vma, addr, pmd);
797
798out_unlock:
799 spin_unlock(ptl);
800 if (pgtable)
801 pte_free(mm, pgtable);
802}
803
804/**
805 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
806 * @vmf: Structure describing the fault
807 * @pfn: pfn to insert
808 * @pgprot: page protection to use
809 * @write: whether it's a write fault
810 *
811 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
812 * also consult the vmf_insert_mixed_prot() documentation when
813 * @pgprot != @vmf->vma->vm_page_prot.
814 *
815 * Return: vm_fault_t value.
816 */
817vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
818 pgprot_t pgprot, bool write)
819{
820 unsigned long addr = vmf->address & PMD_MASK;
821 struct vm_area_struct *vma = vmf->vma;
822 pgtable_t pgtable = NULL;
823
824 /*
825 * If we had pmd_special, we could avoid all these restrictions,
826 * but we need to be consistent with PTEs and architectures that
827 * can't support a 'special' bit.
828 */
829 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
830 !pfn_t_devmap(pfn));
831 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
832 (VM_PFNMAP|VM_MIXEDMAP));
833 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
834
835 if (addr < vma->vm_start || addr >= vma->vm_end)
836 return VM_FAULT_SIGBUS;
837
838 if (arch_needs_pgtable_deposit()) {
839 pgtable = pte_alloc_one(vma->vm_mm);
840 if (!pgtable)
841 return VM_FAULT_OOM;
842 }
843
844 track_pfn_insert(vma, &pgprot, pfn);
845
846 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
847 return VM_FAULT_NOPAGE;
848}
849EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
850
851#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
852static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
853{
854 if (likely(vma->vm_flags & VM_WRITE))
855 pud = pud_mkwrite(pud);
856 return pud;
857}
858
859static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
860 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
861{
862 struct mm_struct *mm = vma->vm_mm;
863 pud_t entry;
864 spinlock_t *ptl;
865
866 ptl = pud_lock(mm, pud);
867 if (!pud_none(*pud)) {
868 if (write) {
869 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
870 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
871 goto out_unlock;
872 }
873 entry = pud_mkyoung(*pud);
874 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
875 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
876 update_mmu_cache_pud(vma, addr, pud);
877 }
878 goto out_unlock;
879 }
880
881 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
882 if (pfn_t_devmap(pfn))
883 entry = pud_mkdevmap(entry);
884 if (write) {
885 entry = pud_mkyoung(pud_mkdirty(entry));
886 entry = maybe_pud_mkwrite(entry, vma);
887 }
888 set_pud_at(mm, addr, pud, entry);
889 update_mmu_cache_pud(vma, addr, pud);
890
891out_unlock:
892 spin_unlock(ptl);
893}
894
895/**
896 * vmf_insert_pfn_pud_prot - insert a pud size pfn
897 * @vmf: Structure describing the fault
898 * @pfn: pfn to insert
899 * @pgprot: page protection to use
900 * @write: whether it's a write fault
901 *
902 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
903 * also consult the vmf_insert_mixed_prot() documentation when
904 * @pgprot != @vmf->vma->vm_page_prot.
905 *
906 * Return: vm_fault_t value.
907 */
908vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
909 pgprot_t pgprot, bool write)
910{
911 unsigned long addr = vmf->address & PUD_MASK;
912 struct vm_area_struct *vma = vmf->vma;
913
914 /*
915 * If we had pud_special, we could avoid all these restrictions,
916 * but we need to be consistent with PTEs and architectures that
917 * can't support a 'special' bit.
918 */
919 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
920 !pfn_t_devmap(pfn));
921 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
922 (VM_PFNMAP|VM_MIXEDMAP));
923 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
924
925 if (addr < vma->vm_start || addr >= vma->vm_end)
926 return VM_FAULT_SIGBUS;
927
928 track_pfn_insert(vma, &pgprot, pfn);
929
930 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
931 return VM_FAULT_NOPAGE;
932}
933EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
934#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
935
936static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
937 pmd_t *pmd, int flags)
938{
939 pmd_t _pmd;
940
941 _pmd = pmd_mkyoung(*pmd);
942 if (flags & FOLL_WRITE)
943 _pmd = pmd_mkdirty(_pmd);
944 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
945 pmd, _pmd, flags & FOLL_WRITE))
946 update_mmu_cache_pmd(vma, addr, pmd);
947}
948
949struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
950 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
951{
952 unsigned long pfn = pmd_pfn(*pmd);
953 struct mm_struct *mm = vma->vm_mm;
954 struct page *page;
955
956 assert_spin_locked(pmd_lockptr(mm, pmd));
957
958 /*
959 * When we COW a devmap PMD entry, we split it into PTEs, so we should
960 * not be in this function with `flags & FOLL_COW` set.
961 */
962 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
963
964 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
965 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
966 (FOLL_PIN | FOLL_GET)))
967 return NULL;
968
969 if (flags & FOLL_WRITE && !pmd_write(*pmd))
970 return NULL;
971
972 if (pmd_present(*pmd) && pmd_devmap(*pmd))
973 /* pass */;
974 else
975 return NULL;
976
977 if (flags & FOLL_TOUCH)
978 touch_pmd(vma, addr, pmd, flags);
979
980 /*
981 * device mapped pages can only be returned if the
982 * caller will manage the page reference count.
983 */
984 if (!(flags & (FOLL_GET | FOLL_PIN)))
985 return ERR_PTR(-EEXIST);
986
987 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
988 *pgmap = get_dev_pagemap(pfn, *pgmap);
989 if (!*pgmap)
990 return ERR_PTR(-EFAULT);
991 page = pfn_to_page(pfn);
992 if (!try_grab_page(page, flags))
993 page = ERR_PTR(-ENOMEM);
994
995 return page;
996}
997
998int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
999 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1000 struct vm_area_struct *vma)
1001{
1002 spinlock_t *dst_ptl, *src_ptl;
1003 struct page *src_page;
1004 pmd_t pmd;
1005 pgtable_t pgtable = NULL;
1006 int ret = -ENOMEM;
1007
1008 /* Skip if can be re-fill on fault */
1009 if (!vma_is_anonymous(vma))
1010 return 0;
1011
1012 pgtable = pte_alloc_one(dst_mm);
1013 if (unlikely(!pgtable))
1014 goto out;
1015
1016 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1017 src_ptl = pmd_lockptr(src_mm, src_pmd);
1018 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1019
1020 ret = -EAGAIN;
1021 pmd = *src_pmd;
1022
1023 /*
1024 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1025 * does not have the VM_UFFD_WP, which means that the uffd
1026 * fork event is not enabled.
1027 */
1028 if (!(vma->vm_flags & VM_UFFD_WP))
1029 pmd = pmd_clear_uffd_wp(pmd);
1030
1031#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1032 if (unlikely(is_swap_pmd(pmd))) {
1033 swp_entry_t entry = pmd_to_swp_entry(pmd);
1034
1035 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1036 if (is_write_migration_entry(entry)) {
1037 make_migration_entry_read(&entry);
1038 pmd = swp_entry_to_pmd(entry);
1039 if (pmd_swp_soft_dirty(*src_pmd))
1040 pmd = pmd_swp_mksoft_dirty(pmd);
1041 set_pmd_at(src_mm, addr, src_pmd, pmd);
1042 }
1043 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1044 mm_inc_nr_ptes(dst_mm);
1045 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1046 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1047 ret = 0;
1048 goto out_unlock;
1049 }
1050#endif
1051
1052 if (unlikely(!pmd_trans_huge(pmd))) {
1053 pte_free(dst_mm, pgtable);
1054 goto out_unlock;
1055 }
1056 /*
1057 * When page table lock is held, the huge zero pmd should not be
1058 * under splitting since we don't split the page itself, only pmd to
1059 * a page table.
1060 */
1061 if (is_huge_zero_pmd(pmd)) {
1062 struct page *zero_page;
1063 /*
1064 * get_huge_zero_page() will never allocate a new page here,
1065 * since we already have a zero page to copy. It just takes a
1066 * reference.
1067 */
1068 zero_page = mm_get_huge_zero_page(dst_mm);
1069 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1070 zero_page);
1071 ret = 0;
1072 goto out_unlock;
1073 }
1074
1075 src_page = pmd_page(pmd);
1076 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1077
1078 /*
1079 * If this page is a potentially pinned page, split and retry the fault
1080 * with smaller page size. Normally this should not happen because the
1081 * userspace should use MADV_DONTFORK upon pinned regions. This is a
1082 * best effort that the pinned pages won't be replaced by another
1083 * random page during the coming copy-on-write.
1084 */
1085 if (unlikely(is_cow_mapping(vma->vm_flags) &&
1086 atomic_read(&src_mm->has_pinned) &&
1087 page_maybe_dma_pinned(src_page))) {
1088 pte_free(dst_mm, pgtable);
1089 spin_unlock(src_ptl);
1090 spin_unlock(dst_ptl);
1091 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1092 return -EAGAIN;
1093 }
1094
1095 get_page(src_page);
1096 page_dup_rmap(src_page, true);
1097 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1098 mm_inc_nr_ptes(dst_mm);
1099 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1100
1101 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1102 pmd = pmd_mkold(pmd_wrprotect(pmd));
1103 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1104
1105 ret = 0;
1106out_unlock:
1107 spin_unlock(src_ptl);
1108 spin_unlock(dst_ptl);
1109out:
1110 return ret;
1111}
1112
1113#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1114static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1115 pud_t *pud, int flags)
1116{
1117 pud_t _pud;
1118
1119 _pud = pud_mkyoung(*pud);
1120 if (flags & FOLL_WRITE)
1121 _pud = pud_mkdirty(_pud);
1122 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1123 pud, _pud, flags & FOLL_WRITE))
1124 update_mmu_cache_pud(vma, addr, pud);
1125}
1126
1127struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1128 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1129{
1130 unsigned long pfn = pud_pfn(*pud);
1131 struct mm_struct *mm = vma->vm_mm;
1132 struct page *page;
1133
1134 assert_spin_locked(pud_lockptr(mm, pud));
1135
1136 if (flags & FOLL_WRITE && !pud_write(*pud))
1137 return NULL;
1138
1139 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1140 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1141 (FOLL_PIN | FOLL_GET)))
1142 return NULL;
1143
1144 if (pud_present(*pud) && pud_devmap(*pud))
1145 /* pass */;
1146 else
1147 return NULL;
1148
1149 if (flags & FOLL_TOUCH)
1150 touch_pud(vma, addr, pud, flags);
1151
1152 /*
1153 * device mapped pages can only be returned if the
1154 * caller will manage the page reference count.
1155 *
1156 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1157 */
1158 if (!(flags & (FOLL_GET | FOLL_PIN)))
1159 return ERR_PTR(-EEXIST);
1160
1161 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1162 *pgmap = get_dev_pagemap(pfn, *pgmap);
1163 if (!*pgmap)
1164 return ERR_PTR(-EFAULT);
1165 page = pfn_to_page(pfn);
1166 if (!try_grab_page(page, flags))
1167 page = ERR_PTR(-ENOMEM);
1168
1169 return page;
1170}
1171
1172int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1173 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1174 struct vm_area_struct *vma)
1175{
1176 spinlock_t *dst_ptl, *src_ptl;
1177 pud_t pud;
1178 int ret;
1179
1180 dst_ptl = pud_lock(dst_mm, dst_pud);
1181 src_ptl = pud_lockptr(src_mm, src_pud);
1182 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1183
1184 ret = -EAGAIN;
1185 pud = *src_pud;
1186 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1187 goto out_unlock;
1188
1189 /*
1190 * When page table lock is held, the huge zero pud should not be
1191 * under splitting since we don't split the page itself, only pud to
1192 * a page table.
1193 */
1194 if (is_huge_zero_pud(pud)) {
1195 /* No huge zero pud yet */
1196 }
1197
1198 /* Please refer to comments in copy_huge_pmd() */
1199 if (unlikely(is_cow_mapping(vma->vm_flags) &&
1200 atomic_read(&src_mm->has_pinned) &&
1201 page_maybe_dma_pinned(pud_page(pud)))) {
1202 spin_unlock(src_ptl);
1203 spin_unlock(dst_ptl);
1204 __split_huge_pud(vma, src_pud, addr);
1205 return -EAGAIN;
1206 }
1207
1208 pudp_set_wrprotect(src_mm, addr, src_pud);
1209 pud = pud_mkold(pud_wrprotect(pud));
1210 set_pud_at(dst_mm, addr, dst_pud, pud);
1211
1212 ret = 0;
1213out_unlock:
1214 spin_unlock(src_ptl);
1215 spin_unlock(dst_ptl);
1216 return ret;
1217}
1218
1219void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1220{
1221 pud_t entry;
1222 unsigned long haddr;
1223 bool write = vmf->flags & FAULT_FLAG_WRITE;
1224
1225 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1226 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1227 goto unlock;
1228
1229 entry = pud_mkyoung(orig_pud);
1230 if (write)
1231 entry = pud_mkdirty(entry);
1232 haddr = vmf->address & HPAGE_PUD_MASK;
1233 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1234 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1235
1236unlock:
1237 spin_unlock(vmf->ptl);
1238}
1239#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1240
1241void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1242{
1243 pmd_t entry;
1244 unsigned long haddr;
1245 bool write = vmf->flags & FAULT_FLAG_WRITE;
1246
1247 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1248 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1249 goto unlock;
1250
1251 entry = pmd_mkyoung(orig_pmd);
1252 if (write)
1253 entry = pmd_mkdirty(entry);
1254 haddr = vmf->address & HPAGE_PMD_MASK;
1255 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1256 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1257
1258unlock:
1259 spin_unlock(vmf->ptl);
1260}
1261
1262vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1263{
1264 struct vm_area_struct *vma = vmf->vma;
1265 struct page *page;
1266 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1267
1268 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1269 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1270
1271 if (is_huge_zero_pmd(orig_pmd))
1272 goto fallback;
1273
1274 spin_lock(vmf->ptl);
1275
1276 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1277 spin_unlock(vmf->ptl);
1278 return 0;
1279 }
1280
1281 page = pmd_page(orig_pmd);
1282 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1283
1284 /* Lock page for reuse_swap_page() */
1285 if (!trylock_page(page)) {
1286 get_page(page);
1287 spin_unlock(vmf->ptl);
1288 lock_page(page);
1289 spin_lock(vmf->ptl);
1290 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1291 spin_unlock(vmf->ptl);
1292 unlock_page(page);
1293 put_page(page);
1294 return 0;
1295 }
1296 put_page(page);
1297 }
1298
1299 /*
1300 * We can only reuse the page if nobody else maps the huge page or it's
1301 * part.
1302 */
1303 if (reuse_swap_page(page, NULL)) {
1304 pmd_t entry;
1305 entry = pmd_mkyoung(orig_pmd);
1306 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1307 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1308 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1309 unlock_page(page);
1310 spin_unlock(vmf->ptl);
1311 return VM_FAULT_WRITE;
1312 }
1313
1314 unlock_page(page);
1315 spin_unlock(vmf->ptl);
1316fallback:
1317 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1318 return VM_FAULT_FALLBACK;
1319}
1320
1321/*
1322 * FOLL_FORCE can write to even unwritable pmd's, but only
1323 * after we've gone through a COW cycle and they are dirty.
1324 */
1325static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1326{
1327 return pmd_write(pmd) ||
1328 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1329}
1330
1331struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1332 unsigned long addr,
1333 pmd_t *pmd,
1334 unsigned int flags)
1335{
1336 struct mm_struct *mm = vma->vm_mm;
1337 struct page *page = NULL;
1338
1339 assert_spin_locked(pmd_lockptr(mm, pmd));
1340
1341 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1342 goto out;
1343
1344 /* Avoid dumping huge zero page */
1345 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1346 return ERR_PTR(-EFAULT);
1347
1348 /* Full NUMA hinting faults to serialise migration in fault paths */
1349 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1350 goto out;
1351
1352 page = pmd_page(*pmd);
1353 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1354
1355 if (!try_grab_page(page, flags))
1356 return ERR_PTR(-ENOMEM);
1357
1358 if (flags & FOLL_TOUCH)
1359 touch_pmd(vma, addr, pmd, flags);
1360
1361 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1362 /*
1363 * We don't mlock() pte-mapped THPs. This way we can avoid
1364 * leaking mlocked pages into non-VM_LOCKED VMAs.
1365 *
1366 * For anon THP:
1367 *
1368 * In most cases the pmd is the only mapping of the page as we
1369 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1370 * writable private mappings in populate_vma_page_range().
1371 *
1372 * The only scenario when we have the page shared here is if we
1373 * mlocking read-only mapping shared over fork(). We skip
1374 * mlocking such pages.
1375 *
1376 * For file THP:
1377 *
1378 * We can expect PageDoubleMap() to be stable under page lock:
1379 * for file pages we set it in page_add_file_rmap(), which
1380 * requires page to be locked.
1381 */
1382
1383 if (PageAnon(page) && compound_mapcount(page) != 1)
1384 goto skip_mlock;
1385 if (PageDoubleMap(page) || !page->mapping)
1386 goto skip_mlock;
1387 if (!trylock_page(page))
1388 goto skip_mlock;
1389 if (page->mapping && !PageDoubleMap(page))
1390 mlock_vma_page(page);
1391 unlock_page(page);
1392 }
1393skip_mlock:
1394 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1395 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1396
1397out:
1398 return page;
1399}
1400
1401/* NUMA hinting page fault entry point for trans huge pmds */
1402vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1403{
1404 struct vm_area_struct *vma = vmf->vma;
1405 struct anon_vma *anon_vma = NULL;
1406 struct page *page;
1407 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1408 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1409 int target_nid, last_cpupid = -1;
1410 bool page_locked;
1411 bool migrated = false;
1412 bool was_writable;
1413 int flags = 0;
1414
1415 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1416 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1417 goto out_unlock;
1418
1419 /*
1420 * If there are potential migrations, wait for completion and retry
1421 * without disrupting NUMA hinting information. Do not relock and
1422 * check_same as the page may no longer be mapped.
1423 */
1424 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1425 page = pmd_page(*vmf->pmd);
1426 if (!get_page_unless_zero(page))
1427 goto out_unlock;
1428 spin_unlock(vmf->ptl);
1429 put_and_wait_on_page_locked(page);
1430 goto out;
1431 }
1432
1433 page = pmd_page(pmd);
1434 BUG_ON(is_huge_zero_page(page));
1435 page_nid = page_to_nid(page);
1436 last_cpupid = page_cpupid_last(page);
1437 count_vm_numa_event(NUMA_HINT_FAULTS);
1438 if (page_nid == this_nid) {
1439 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1440 flags |= TNF_FAULT_LOCAL;
1441 }
1442
1443 /* See similar comment in do_numa_page for explanation */
1444 if (!pmd_savedwrite(pmd))
1445 flags |= TNF_NO_GROUP;
1446
1447 /*
1448 * Acquire the page lock to serialise THP migrations but avoid dropping
1449 * page_table_lock if at all possible
1450 */
1451 page_locked = trylock_page(page);
1452 target_nid = mpol_misplaced(page, vma, haddr);
1453 if (target_nid == NUMA_NO_NODE) {
1454 /* If the page was locked, there are no parallel migrations */
1455 if (page_locked)
1456 goto clear_pmdnuma;
1457 }
1458
1459 /* Migration could have started since the pmd_trans_migrating check */
1460 if (!page_locked) {
1461 page_nid = NUMA_NO_NODE;
1462 if (!get_page_unless_zero(page))
1463 goto out_unlock;
1464 spin_unlock(vmf->ptl);
1465 put_and_wait_on_page_locked(page);
1466 goto out;
1467 }
1468
1469 /*
1470 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1471 * to serialises splits
1472 */
1473 get_page(page);
1474 spin_unlock(vmf->ptl);
1475 anon_vma = page_lock_anon_vma_read(page);
1476
1477 /* Confirm the PMD did not change while page_table_lock was released */
1478 spin_lock(vmf->ptl);
1479 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1480 unlock_page(page);
1481 put_page(page);
1482 page_nid = NUMA_NO_NODE;
1483 goto out_unlock;
1484 }
1485
1486 /* Bail if we fail to protect against THP splits for any reason */
1487 if (unlikely(!anon_vma)) {
1488 put_page(page);
1489 page_nid = NUMA_NO_NODE;
1490 goto clear_pmdnuma;
1491 }
1492
1493 /*
1494 * Since we took the NUMA fault, we must have observed the !accessible
1495 * bit. Make sure all other CPUs agree with that, to avoid them
1496 * modifying the page we're about to migrate.
1497 *
1498 * Must be done under PTL such that we'll observe the relevant
1499 * inc_tlb_flush_pending().
1500 *
1501 * We are not sure a pending tlb flush here is for a huge page
1502 * mapping or not. Hence use the tlb range variant
1503 */
1504 if (mm_tlb_flush_pending(vma->vm_mm)) {
1505 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1506 /*
1507 * change_huge_pmd() released the pmd lock before
1508 * invalidating the secondary MMUs sharing the primary
1509 * MMU pagetables (with ->invalidate_range()). The
1510 * mmu_notifier_invalidate_range_end() (which
1511 * internally calls ->invalidate_range()) in
1512 * change_pmd_range() will run after us, so we can't
1513 * rely on it here and we need an explicit invalidate.
1514 */
1515 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1516 haddr + HPAGE_PMD_SIZE);
1517 }
1518
1519 /*
1520 * Migrate the THP to the requested node, returns with page unlocked
1521 * and access rights restored.
1522 */
1523 spin_unlock(vmf->ptl);
1524
1525 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1526 vmf->pmd, pmd, vmf->address, page, target_nid);
1527 if (migrated) {
1528 flags |= TNF_MIGRATED;
1529 page_nid = target_nid;
1530 } else
1531 flags |= TNF_MIGRATE_FAIL;
1532
1533 goto out;
1534clear_pmdnuma:
1535 BUG_ON(!PageLocked(page));
1536 was_writable = pmd_savedwrite(pmd);
1537 pmd = pmd_modify(pmd, vma->vm_page_prot);
1538 pmd = pmd_mkyoung(pmd);
1539 if (was_writable)
1540 pmd = pmd_mkwrite(pmd);
1541 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1542 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1543 unlock_page(page);
1544out_unlock:
1545 spin_unlock(vmf->ptl);
1546
1547out:
1548 if (anon_vma)
1549 page_unlock_anon_vma_read(anon_vma);
1550
1551 if (page_nid != NUMA_NO_NODE)
1552 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1553 flags);
1554
1555 return 0;
1556}
1557
1558/*
1559 * Return true if we do MADV_FREE successfully on entire pmd page.
1560 * Otherwise, return false.
1561 */
1562bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1563 pmd_t *pmd, unsigned long addr, unsigned long next)
1564{
1565 spinlock_t *ptl;
1566 pmd_t orig_pmd;
1567 struct page *page;
1568 struct mm_struct *mm = tlb->mm;
1569 bool ret = false;
1570
1571 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1572
1573 ptl = pmd_trans_huge_lock(pmd, vma);
1574 if (!ptl)
1575 goto out_unlocked;
1576
1577 orig_pmd = *pmd;
1578 if (is_huge_zero_pmd(orig_pmd))
1579 goto out;
1580
1581 if (unlikely(!pmd_present(orig_pmd))) {
1582 VM_BUG_ON(thp_migration_supported() &&
1583 !is_pmd_migration_entry(orig_pmd));
1584 goto out;
1585 }
1586
1587 page = pmd_page(orig_pmd);
1588 /*
1589 * If other processes are mapping this page, we couldn't discard
1590 * the page unless they all do MADV_FREE so let's skip the page.
1591 */
1592 if (page_mapcount(page) != 1)
1593 goto out;
1594
1595 if (!trylock_page(page))
1596 goto out;
1597
1598 /*
1599 * If user want to discard part-pages of THP, split it so MADV_FREE
1600 * will deactivate only them.
1601 */
1602 if (next - addr != HPAGE_PMD_SIZE) {
1603 get_page(page);
1604 spin_unlock(ptl);
1605 split_huge_page(page);
1606 unlock_page(page);
1607 put_page(page);
1608 goto out_unlocked;
1609 }
1610
1611 if (PageDirty(page))
1612 ClearPageDirty(page);
1613 unlock_page(page);
1614
1615 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1616 pmdp_invalidate(vma, addr, pmd);
1617 orig_pmd = pmd_mkold(orig_pmd);
1618 orig_pmd = pmd_mkclean(orig_pmd);
1619
1620 set_pmd_at(mm, addr, pmd, orig_pmd);
1621 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1622 }
1623
1624 mark_page_lazyfree(page);
1625 ret = true;
1626out:
1627 spin_unlock(ptl);
1628out_unlocked:
1629 return ret;
1630}
1631
1632static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1633{
1634 pgtable_t pgtable;
1635
1636 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1637 pte_free(mm, pgtable);
1638 mm_dec_nr_ptes(mm);
1639}
1640
1641int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1642 pmd_t *pmd, unsigned long addr)
1643{
1644 pmd_t orig_pmd;
1645 spinlock_t *ptl;
1646
1647 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1648
1649 ptl = __pmd_trans_huge_lock(pmd, vma);
1650 if (!ptl)
1651 return 0;
1652 /*
1653 * For architectures like ppc64 we look at deposited pgtable
1654 * when calling pmdp_huge_get_and_clear. So do the
1655 * pgtable_trans_huge_withdraw after finishing pmdp related
1656 * operations.
1657 */
1658 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1659 tlb->fullmm);
1660 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1661 if (vma_is_special_huge(vma)) {
1662 if (arch_needs_pgtable_deposit())
1663 zap_deposited_table(tlb->mm, pmd);
1664 spin_unlock(ptl);
1665 if (is_huge_zero_pmd(orig_pmd))
1666 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1667 } else if (is_huge_zero_pmd(orig_pmd)) {
1668 zap_deposited_table(tlb->mm, pmd);
1669 spin_unlock(ptl);
1670 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1671 } else {
1672 struct page *page = NULL;
1673 int flush_needed = 1;
1674
1675 if (pmd_present(orig_pmd)) {
1676 page = pmd_page(orig_pmd);
1677 page_remove_rmap(page, true);
1678 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1679 VM_BUG_ON_PAGE(!PageHead(page), page);
1680 } else if (thp_migration_supported()) {
1681 swp_entry_t entry;
1682
1683 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1684 entry = pmd_to_swp_entry(orig_pmd);
1685 page = pfn_to_page(swp_offset(entry));
1686 flush_needed = 0;
1687 } else
1688 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1689
1690 if (PageAnon(page)) {
1691 zap_deposited_table(tlb->mm, pmd);
1692 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1693 } else {
1694 if (arch_needs_pgtable_deposit())
1695 zap_deposited_table(tlb->mm, pmd);
1696 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1697 }
1698
1699 spin_unlock(ptl);
1700 if (flush_needed)
1701 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1702 }
1703 return 1;
1704}
1705
1706#ifndef pmd_move_must_withdraw
1707static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1708 spinlock_t *old_pmd_ptl,
1709 struct vm_area_struct *vma)
1710{
1711 /*
1712 * With split pmd lock we also need to move preallocated
1713 * PTE page table if new_pmd is on different PMD page table.
1714 *
1715 * We also don't deposit and withdraw tables for file pages.
1716 */
1717 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1718}
1719#endif
1720
1721static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1722{
1723#ifdef CONFIG_MEM_SOFT_DIRTY
1724 if (unlikely(is_pmd_migration_entry(pmd)))
1725 pmd = pmd_swp_mksoft_dirty(pmd);
1726 else if (pmd_present(pmd))
1727 pmd = pmd_mksoft_dirty(pmd);
1728#endif
1729 return pmd;
1730}
1731
1732bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1733 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1734{
1735 spinlock_t *old_ptl, *new_ptl;
1736 pmd_t pmd;
1737 struct mm_struct *mm = vma->vm_mm;
1738 bool force_flush = false;
1739
1740 /*
1741 * The destination pmd shouldn't be established, free_pgtables()
1742 * should have release it.
1743 */
1744 if (WARN_ON(!pmd_none(*new_pmd))) {
1745 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1746 return false;
1747 }
1748
1749 /*
1750 * We don't have to worry about the ordering of src and dst
1751 * ptlocks because exclusive mmap_lock prevents deadlock.
1752 */
1753 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1754 if (old_ptl) {
1755 new_ptl = pmd_lockptr(mm, new_pmd);
1756 if (new_ptl != old_ptl)
1757 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1758 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1759 if (pmd_present(pmd))
1760 force_flush = true;
1761 VM_BUG_ON(!pmd_none(*new_pmd));
1762
1763 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1764 pgtable_t pgtable;
1765 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1766 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1767 }
1768 pmd = move_soft_dirty_pmd(pmd);
1769 set_pmd_at(mm, new_addr, new_pmd, pmd);
1770 if (force_flush)
1771 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1772 if (new_ptl != old_ptl)
1773 spin_unlock(new_ptl);
1774 spin_unlock(old_ptl);
1775 return true;
1776 }
1777 return false;
1778}
1779
1780/*
1781 * Returns
1782 * - 0 if PMD could not be locked
1783 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1784 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1785 */
1786int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1787 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1788{
1789 struct mm_struct *mm = vma->vm_mm;
1790 spinlock_t *ptl;
1791 pmd_t entry;
1792 bool preserve_write;
1793 int ret;
1794 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1795 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1796 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1797
1798 ptl = __pmd_trans_huge_lock(pmd, vma);
1799 if (!ptl)
1800 return 0;
1801
1802 preserve_write = prot_numa && pmd_write(*pmd);
1803 ret = 1;
1804
1805#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1806 if (is_swap_pmd(*pmd)) {
1807 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1808
1809 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1810 if (is_write_migration_entry(entry)) {
1811 pmd_t newpmd;
1812 /*
1813 * A protection check is difficult so
1814 * just be safe and disable write
1815 */
1816 make_migration_entry_read(&entry);
1817 newpmd = swp_entry_to_pmd(entry);
1818 if (pmd_swp_soft_dirty(*pmd))
1819 newpmd = pmd_swp_mksoft_dirty(newpmd);
1820 set_pmd_at(mm, addr, pmd, newpmd);
1821 }
1822 goto unlock;
1823 }
1824#endif
1825
1826 /*
1827 * Avoid trapping faults against the zero page. The read-only
1828 * data is likely to be read-cached on the local CPU and
1829 * local/remote hits to the zero page are not interesting.
1830 */
1831 if (prot_numa && is_huge_zero_pmd(*pmd))
1832 goto unlock;
1833
1834 if (prot_numa && pmd_protnone(*pmd))
1835 goto unlock;
1836
1837 /*
1838 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1839 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1840 * which is also under mmap_read_lock(mm):
1841 *
1842 * CPU0: CPU1:
1843 * change_huge_pmd(prot_numa=1)
1844 * pmdp_huge_get_and_clear_notify()
1845 * madvise_dontneed()
1846 * zap_pmd_range()
1847 * pmd_trans_huge(*pmd) == 0 (without ptl)
1848 * // skip the pmd
1849 * set_pmd_at();
1850 * // pmd is re-established
1851 *
1852 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1853 * which may break userspace.
1854 *
1855 * pmdp_invalidate() is required to make sure we don't miss
1856 * dirty/young flags set by hardware.
1857 */
1858 entry = pmdp_invalidate(vma, addr, pmd);
1859
1860 entry = pmd_modify(entry, newprot);
1861 if (preserve_write)
1862 entry = pmd_mk_savedwrite(entry);
1863 if (uffd_wp) {
1864 entry = pmd_wrprotect(entry);
1865 entry = pmd_mkuffd_wp(entry);
1866 } else if (uffd_wp_resolve) {
1867 /*
1868 * Leave the write bit to be handled by PF interrupt
1869 * handler, then things like COW could be properly
1870 * handled.
1871 */
1872 entry = pmd_clear_uffd_wp(entry);
1873 }
1874 ret = HPAGE_PMD_NR;
1875 set_pmd_at(mm, addr, pmd, entry);
1876 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1877unlock:
1878 spin_unlock(ptl);
1879 return ret;
1880}
1881
1882/*
1883 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1884 *
1885 * Note that if it returns page table lock pointer, this routine returns without
1886 * unlocking page table lock. So callers must unlock it.
1887 */
1888spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1889{
1890 spinlock_t *ptl;
1891 ptl = pmd_lock(vma->vm_mm, pmd);
1892 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1893 pmd_devmap(*pmd)))
1894 return ptl;
1895 spin_unlock(ptl);
1896 return NULL;
1897}
1898
1899/*
1900 * Returns true if a given pud maps a thp, false otherwise.
1901 *
1902 * Note that if it returns true, this routine returns without unlocking page
1903 * table lock. So callers must unlock it.
1904 */
1905spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1906{
1907 spinlock_t *ptl;
1908
1909 ptl = pud_lock(vma->vm_mm, pud);
1910 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1911 return ptl;
1912 spin_unlock(ptl);
1913 return NULL;
1914}
1915
1916#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1917int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1918 pud_t *pud, unsigned long addr)
1919{
1920 spinlock_t *ptl;
1921
1922 ptl = __pud_trans_huge_lock(pud, vma);
1923 if (!ptl)
1924 return 0;
1925 /*
1926 * For architectures like ppc64 we look at deposited pgtable
1927 * when calling pudp_huge_get_and_clear. So do the
1928 * pgtable_trans_huge_withdraw after finishing pudp related
1929 * operations.
1930 */
1931 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1932 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1933 if (vma_is_special_huge(vma)) {
1934 spin_unlock(ptl);
1935 /* No zero page support yet */
1936 } else {
1937 /* No support for anonymous PUD pages yet */
1938 BUG();
1939 }
1940 return 1;
1941}
1942
1943static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1944 unsigned long haddr)
1945{
1946 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1947 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1948 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1949 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1950
1951 count_vm_event(THP_SPLIT_PUD);
1952
1953 pudp_huge_clear_flush_notify(vma, haddr, pud);
1954}
1955
1956void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1957 unsigned long address)
1958{
1959 spinlock_t *ptl;
1960 struct mmu_notifier_range range;
1961
1962 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1963 address & HPAGE_PUD_MASK,
1964 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1965 mmu_notifier_invalidate_range_start(&range);
1966 ptl = pud_lock(vma->vm_mm, pud);
1967 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1968 goto out;
1969 __split_huge_pud_locked(vma, pud, range.start);
1970
1971out:
1972 spin_unlock(ptl);
1973 /*
1974 * No need to double call mmu_notifier->invalidate_range() callback as
1975 * the above pudp_huge_clear_flush_notify() did already call it.
1976 */
1977 mmu_notifier_invalidate_range_only_end(&range);
1978}
1979#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1980
1981static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1982 unsigned long haddr, pmd_t *pmd)
1983{
1984 struct mm_struct *mm = vma->vm_mm;
1985 pgtable_t pgtable;
1986 pmd_t _pmd;
1987 int i;
1988
1989 /*
1990 * Leave pmd empty until pte is filled note that it is fine to delay
1991 * notification until mmu_notifier_invalidate_range_end() as we are
1992 * replacing a zero pmd write protected page with a zero pte write
1993 * protected page.
1994 *
1995 * See Documentation/vm/mmu_notifier.rst
1996 */
1997 pmdp_huge_clear_flush(vma, haddr, pmd);
1998
1999 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2000 pmd_populate(mm, &_pmd, pgtable);
2001
2002 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2003 pte_t *pte, entry;
2004 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2005 entry = pte_mkspecial(entry);
2006 pte = pte_offset_map(&_pmd, haddr);
2007 VM_BUG_ON(!pte_none(*pte));
2008 set_pte_at(mm, haddr, pte, entry);
2009 pte_unmap(pte);
2010 }
2011 smp_wmb(); /* make pte visible before pmd */
2012 pmd_populate(mm, pmd, pgtable);
2013}
2014
2015static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2016 unsigned long haddr, bool freeze)
2017{
2018 struct mm_struct *mm = vma->vm_mm;
2019 struct page *page;
2020 pgtable_t pgtable;
2021 pmd_t old_pmd, _pmd;
2022 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2023 unsigned long addr;
2024 int i;
2025
2026 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2027 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2028 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2029 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2030 && !pmd_devmap(*pmd));
2031
2032 count_vm_event(THP_SPLIT_PMD);
2033
2034 if (!vma_is_anonymous(vma)) {
2035 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2036 /*
2037 * We are going to unmap this huge page. So
2038 * just go ahead and zap it
2039 */
2040 if (arch_needs_pgtable_deposit())
2041 zap_deposited_table(mm, pmd);
2042 if (vma_is_special_huge(vma))
2043 return;
2044 page = pmd_page(_pmd);
2045 if (!PageDirty(page) && pmd_dirty(_pmd))
2046 set_page_dirty(page);
2047 if (!PageReferenced(page) && pmd_young(_pmd))
2048 SetPageReferenced(page);
2049 page_remove_rmap(page, true);
2050 put_page(page);
2051 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2052 return;
2053 } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) {
2054 /*
2055 * FIXME: Do we want to invalidate secondary mmu by calling
2056 * mmu_notifier_invalidate_range() see comments below inside
2057 * __split_huge_pmd() ?
2058 *
2059 * We are going from a zero huge page write protected to zero
2060 * small page also write protected so it does not seems useful
2061 * to invalidate secondary mmu at this time.
2062 */
2063 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2064 }
2065
2066 /*
2067 * Up to this point the pmd is present and huge and userland has the
2068 * whole access to the hugepage during the split (which happens in
2069 * place). If we overwrite the pmd with the not-huge version pointing
2070 * to the pte here (which of course we could if all CPUs were bug
2071 * free), userland could trigger a small page size TLB miss on the
2072 * small sized TLB while the hugepage TLB entry is still established in
2073 * the huge TLB. Some CPU doesn't like that.
2074 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2075 * 383 on page 105. Intel should be safe but is also warns that it's
2076 * only safe if the permission and cache attributes of the two entries
2077 * loaded in the two TLB is identical (which should be the case here).
2078 * But it is generally safer to never allow small and huge TLB entries
2079 * for the same virtual address to be loaded simultaneously. So instead
2080 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2081 * current pmd notpresent (atomically because here the pmd_trans_huge
2082 * must remain set at all times on the pmd until the split is complete
2083 * for this pmd), then we flush the SMP TLB and finally we write the
2084 * non-huge version of the pmd entry with pmd_populate.
2085 */
2086 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2087
2088 pmd_migration = is_pmd_migration_entry(old_pmd);
2089 if (unlikely(pmd_migration)) {
2090 swp_entry_t entry;
2091
2092 entry = pmd_to_swp_entry(old_pmd);
2093 page = pfn_to_page(swp_offset(entry));
2094 write = is_write_migration_entry(entry);
2095 young = false;
2096 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2097 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2098 } else {
2099 page = pmd_page(old_pmd);
2100 if (pmd_dirty(old_pmd))
2101 SetPageDirty(page);
2102 write = pmd_write(old_pmd);
2103 young = pmd_young(old_pmd);
2104 soft_dirty = pmd_soft_dirty(old_pmd);
2105 uffd_wp = pmd_uffd_wp(old_pmd);
2106 }
2107 VM_BUG_ON_PAGE(!page_count(page), page);
2108 page_ref_add(page, HPAGE_PMD_NR - 1);
2109
2110 /*
2111 * Withdraw the table only after we mark the pmd entry invalid.
2112 * This's critical for some architectures (Power).
2113 */
2114 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2115 pmd_populate(mm, &_pmd, pgtable);
2116
2117 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2118 pte_t entry, *pte;
2119 /*
2120 * Note that NUMA hinting access restrictions are not
2121 * transferred to avoid any possibility of altering
2122 * permissions across VMAs.
2123 */
2124 if (freeze || pmd_migration) {
2125 swp_entry_t swp_entry;
2126 swp_entry = make_migration_entry(page + i, write);
2127 entry = swp_entry_to_pte(swp_entry);
2128 if (soft_dirty)
2129 entry = pte_swp_mksoft_dirty(entry);
2130 if (uffd_wp)
2131 entry = pte_swp_mkuffd_wp(entry);
2132 } else {
2133 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2134 entry = maybe_mkwrite(entry, vma);
2135 if (!write)
2136 entry = pte_wrprotect(entry);
2137 if (!young)
2138 entry = pte_mkold(entry);
2139 if (soft_dirty)
2140 entry = pte_mksoft_dirty(entry);
2141 if (uffd_wp)
2142 entry = pte_mkuffd_wp(entry);
2143 }
2144 pte = pte_offset_map(&_pmd, addr);
2145 BUG_ON(!pte_none(*pte));
2146 set_pte_at(mm, addr, pte, entry);
2147 if (!pmd_migration)
2148 atomic_inc(&page[i]._mapcount);
2149 pte_unmap(pte);
2150 }
2151
2152 if (!pmd_migration) {
2153 /*
2154 * Set PG_double_map before dropping compound_mapcount to avoid
2155 * false-negative page_mapped().
2156 */
2157 if (compound_mapcount(page) > 1 &&
2158 !TestSetPageDoubleMap(page)) {
2159 for (i = 0; i < HPAGE_PMD_NR; i++)
2160 atomic_inc(&page[i]._mapcount);
2161 }
2162
2163 lock_page_memcg(page);
2164 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2165 /* Last compound_mapcount is gone. */
2166 __dec_lruvec_page_state(page, NR_ANON_THPS);
2167 if (TestClearPageDoubleMap(page)) {
2168 /* No need in mapcount reference anymore */
2169 for (i = 0; i < HPAGE_PMD_NR; i++)
2170 atomic_dec(&page[i]._mapcount);
2171 }
2172 }
2173 unlock_page_memcg(page);
2174 }
2175
2176 smp_wmb(); /* make pte visible before pmd */
2177 pmd_populate(mm, pmd, pgtable);
2178
2179 if (freeze) {
2180 for (i = 0; i < HPAGE_PMD_NR; i++) {
2181 page_remove_rmap(page + i, false);
2182 put_page(page + i);
2183 }
2184 }
2185}
2186
2187void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2188 unsigned long address, bool freeze, struct page *page)
2189{
2190 spinlock_t *ptl;
2191 struct mmu_notifier_range range;
2192 bool was_locked = false;
2193 pmd_t _pmd;
2194
2195 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2196 address & HPAGE_PMD_MASK,
2197 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2198 mmu_notifier_invalidate_range_start(&range);
2199 ptl = pmd_lock(vma->vm_mm, pmd);
2200
2201 /*
2202 * If caller asks to setup a migration entries, we need a page to check
2203 * pmd against. Otherwise we can end up replacing wrong page.
2204 */
2205 VM_BUG_ON(freeze && !page);
2206 if (page) {
2207 VM_WARN_ON_ONCE(!PageLocked(page));
2208 was_locked = true;
2209 if (page != pmd_page(*pmd))
2210 goto out;
2211 }
2212
2213repeat:
2214 if (pmd_trans_huge(*pmd)) {
2215 if (!page) {
2216 page = pmd_page(*pmd);
2217 if (unlikely(!trylock_page(page))) {
2218 get_page(page);
2219 _pmd = *pmd;
2220 spin_unlock(ptl);
2221 lock_page(page);
2222 spin_lock(ptl);
2223 if (unlikely(!pmd_same(*pmd, _pmd))) {
2224 unlock_page(page);
2225 put_page(page);
2226 page = NULL;
2227 goto repeat;
2228 }
2229 put_page(page);
2230 }
2231 }
2232 if (PageMlocked(page))
2233 clear_page_mlock(page);
2234 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2235 goto out;
2236 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2237out:
2238 spin_unlock(ptl);
2239 if (!was_locked && page)
2240 unlock_page(page);
2241 /*
2242 * No need to double call mmu_notifier->invalidate_range() callback.
2243 * They are 3 cases to consider inside __split_huge_pmd_locked():
2244 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2245 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2246 * fault will trigger a flush_notify before pointing to a new page
2247 * (it is fine if the secondary mmu keeps pointing to the old zero
2248 * page in the meantime)
2249 * 3) Split a huge pmd into pte pointing to the same page. No need
2250 * to invalidate secondary tlb entry they are all still valid.
2251 * any further changes to individual pte will notify. So no need
2252 * to call mmu_notifier->invalidate_range()
2253 */
2254 mmu_notifier_invalidate_range_only_end(&range);
2255}
2256
2257void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2258 bool freeze, struct page *page)
2259{
2260 pgd_t *pgd;
2261 p4d_t *p4d;
2262 pud_t *pud;
2263 pmd_t *pmd;
2264
2265 pgd = pgd_offset(vma->vm_mm, address);
2266 if (!pgd_present(*pgd))
2267 return;
2268
2269 p4d = p4d_offset(pgd, address);
2270 if (!p4d_present(*p4d))
2271 return;
2272
2273 pud = pud_offset(p4d, address);
2274 if (!pud_present(*pud))
2275 return;
2276
2277 pmd = pmd_offset(pud, address);
2278
2279 __split_huge_pmd(vma, pmd, address, freeze, page);
2280}
2281
2282void vma_adjust_trans_huge(struct vm_area_struct *vma,
2283 unsigned long start,
2284 unsigned long end,
2285 long adjust_next)
2286{
2287 /*
2288 * If the new start address isn't hpage aligned and it could
2289 * previously contain an hugepage: check if we need to split
2290 * an huge pmd.
2291 */
2292 if (start & ~HPAGE_PMD_MASK &&
2293 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2294 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2295 split_huge_pmd_address(vma, start, false, NULL);
2296
2297 /*
2298 * If the new end address isn't hpage aligned and it could
2299 * previously contain an hugepage: check if we need to split
2300 * an huge pmd.
2301 */
2302 if (end & ~HPAGE_PMD_MASK &&
2303 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2304 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2305 split_huge_pmd_address(vma, end, false, NULL);
2306
2307 /*
2308 * If we're also updating the vma->vm_next->vm_start, if the new
2309 * vm_next->vm_start isn't page aligned and it could previously
2310 * contain an hugepage: check if we need to split an huge pmd.
2311 */
2312 if (adjust_next > 0) {
2313 struct vm_area_struct *next = vma->vm_next;
2314 unsigned long nstart = next->vm_start;
2315 nstart += adjust_next << PAGE_SHIFT;
2316 if (nstart & ~HPAGE_PMD_MASK &&
2317 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2318 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2319 split_huge_pmd_address(next, nstart, false, NULL);
2320 }
2321}
2322
2323static void unmap_page(struct page *page)
2324{
2325 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2326 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2327 bool unmap_success;
2328
2329 VM_BUG_ON_PAGE(!PageHead(page), page);
2330
2331 if (PageAnon(page))
2332 ttu_flags |= TTU_SPLIT_FREEZE;
2333
2334 unmap_success = try_to_unmap(page, ttu_flags);
2335 VM_BUG_ON_PAGE(!unmap_success, page);
2336}
2337
2338static void remap_page(struct page *page)
2339{
2340 int i;
2341 if (PageTransHuge(page)) {
2342 remove_migration_ptes(page, page, true);
2343 } else {
2344 for (i = 0; i < HPAGE_PMD_NR; i++)
2345 remove_migration_ptes(page + i, page + i, true);
2346 }
2347}
2348
2349static void __split_huge_page_tail(struct page *head, int tail,
2350 struct lruvec *lruvec, struct list_head *list)
2351{
2352 struct page *page_tail = head + tail;
2353
2354 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2355
2356 /*
2357 * Clone page flags before unfreezing refcount.
2358 *
2359 * After successful get_page_unless_zero() might follow flags change,
2360 * for exmaple lock_page() which set PG_waiters.
2361 */
2362 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2363 page_tail->flags |= (head->flags &
2364 ((1L << PG_referenced) |
2365 (1L << PG_swapbacked) |
2366 (1L << PG_swapcache) |
2367 (1L << PG_mlocked) |
2368 (1L << PG_uptodate) |
2369 (1L << PG_active) |
2370 (1L << PG_workingset) |
2371 (1L << PG_locked) |
2372 (1L << PG_unevictable) |
2373 (1L << PG_dirty)));
2374
2375 /* ->mapping in first tail page is compound_mapcount */
2376 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2377 page_tail);
2378 page_tail->mapping = head->mapping;
2379 page_tail->index = head->index + tail;
2380
2381 /* Page flags must be visible before we make the page non-compound. */
2382 smp_wmb();
2383
2384 /*
2385 * Clear PageTail before unfreezing page refcount.
2386 *
2387 * After successful get_page_unless_zero() might follow put_page()
2388 * which needs correct compound_head().
2389 */
2390 clear_compound_head(page_tail);
2391
2392 /* Finally unfreeze refcount. Additional reference from page cache. */
2393 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2394 PageSwapCache(head)));
2395
2396 if (page_is_young(head))
2397 set_page_young(page_tail);
2398 if (page_is_idle(head))
2399 set_page_idle(page_tail);
2400
2401 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2402
2403 /*
2404 * always add to the tail because some iterators expect new
2405 * pages to show after the currently processed elements - e.g.
2406 * migrate_pages
2407 */
2408 lru_add_page_tail(head, page_tail, lruvec, list);
2409}
2410
2411static void __split_huge_page(struct page *page, struct list_head *list,
2412 pgoff_t end, unsigned long flags)
2413{
2414 struct page *head = compound_head(page);
2415 pg_data_t *pgdat = page_pgdat(head);
2416 struct lruvec *lruvec;
2417 struct address_space *swap_cache = NULL;
2418 unsigned long offset = 0;
2419 int i;
2420
2421 lruvec = mem_cgroup_page_lruvec(head, pgdat);
2422
2423 /* complete memcg works before add pages to LRU */
2424 mem_cgroup_split_huge_fixup(head);
2425
2426 if (PageAnon(head) && PageSwapCache(head)) {
2427 swp_entry_t entry = { .val = page_private(head) };
2428
2429 offset = swp_offset(entry);
2430 swap_cache = swap_address_space(entry);
2431 xa_lock(&swap_cache->i_pages);
2432 }
2433
2434 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2435 __split_huge_page_tail(head, i, lruvec, list);
2436 /* Some pages can be beyond i_size: drop them from page cache */
2437 if (head[i].index >= end) {
2438 ClearPageDirty(head + i);
2439 __delete_from_page_cache(head + i, NULL);
2440 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2441 shmem_uncharge(head->mapping->host, 1);
2442 put_page(head + i);
2443 } else if (!PageAnon(page)) {
2444 __xa_store(&head->mapping->i_pages, head[i].index,
2445 head + i, 0);
2446 } else if (swap_cache) {
2447 __xa_store(&swap_cache->i_pages, offset + i,
2448 head + i, 0);
2449 }
2450 }
2451
2452 ClearPageCompound(head);
2453
2454 split_page_owner(head, HPAGE_PMD_ORDER);
2455
2456 /* See comment in __split_huge_page_tail() */
2457 if (PageAnon(head)) {
2458 /* Additional pin to swap cache */
2459 if (PageSwapCache(head)) {
2460 page_ref_add(head, 2);
2461 xa_unlock(&swap_cache->i_pages);
2462 } else {
2463 page_ref_inc(head);
2464 }
2465 } else {
2466 /* Additional pin to page cache */
2467 page_ref_add(head, 2);
2468 xa_unlock(&head->mapping->i_pages);
2469 }
2470
2471 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2472
2473 remap_page(head);
2474
2475 for (i = 0; i < HPAGE_PMD_NR; i++) {
2476 struct page *subpage = head + i;
2477 if (subpage == page)
2478 continue;
2479 unlock_page(subpage);
2480
2481 /*
2482 * Subpages may be freed if there wasn't any mapping
2483 * like if add_to_swap() is running on a lru page that
2484 * had its mapping zapped. And freeing these pages
2485 * requires taking the lru_lock so we do the put_page
2486 * of the tail pages after the split is complete.
2487 */
2488 put_page(subpage);
2489 }
2490}
2491
2492int total_mapcount(struct page *page)
2493{
2494 int i, compound, ret;
2495
2496 VM_BUG_ON_PAGE(PageTail(page), page);
2497
2498 if (likely(!PageCompound(page)))
2499 return atomic_read(&page->_mapcount) + 1;
2500
2501 compound = compound_mapcount(page);
2502 if (PageHuge(page))
2503 return compound;
2504 ret = compound;
2505 for (i = 0; i < HPAGE_PMD_NR; i++)
2506 ret += atomic_read(&page[i]._mapcount) + 1;
2507 /* File pages has compound_mapcount included in _mapcount */
2508 if (!PageAnon(page))
2509 return ret - compound * HPAGE_PMD_NR;
2510 if (PageDoubleMap(page))
2511 ret -= HPAGE_PMD_NR;
2512 return ret;
2513}
2514
2515/*
2516 * This calculates accurately how many mappings a transparent hugepage
2517 * has (unlike page_mapcount() which isn't fully accurate). This full
2518 * accuracy is primarily needed to know if copy-on-write faults can
2519 * reuse the page and change the mapping to read-write instead of
2520 * copying them. At the same time this returns the total_mapcount too.
2521 *
2522 * The function returns the highest mapcount any one of the subpages
2523 * has. If the return value is one, even if different processes are
2524 * mapping different subpages of the transparent hugepage, they can
2525 * all reuse it, because each process is reusing a different subpage.
2526 *
2527 * The total_mapcount is instead counting all virtual mappings of the
2528 * subpages. If the total_mapcount is equal to "one", it tells the
2529 * caller all mappings belong to the same "mm" and in turn the
2530 * anon_vma of the transparent hugepage can become the vma->anon_vma
2531 * local one as no other process may be mapping any of the subpages.
2532 *
2533 * It would be more accurate to replace page_mapcount() with
2534 * page_trans_huge_mapcount(), however we only use
2535 * page_trans_huge_mapcount() in the copy-on-write faults where we
2536 * need full accuracy to avoid breaking page pinning, because
2537 * page_trans_huge_mapcount() is slower than page_mapcount().
2538 */
2539int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2540{
2541 int i, ret, _total_mapcount, mapcount;
2542
2543 /* hugetlbfs shouldn't call it */
2544 VM_BUG_ON_PAGE(PageHuge(page), page);
2545
2546 if (likely(!PageTransCompound(page))) {
2547 mapcount = atomic_read(&page->_mapcount) + 1;
2548 if (total_mapcount)
2549 *total_mapcount = mapcount;
2550 return mapcount;
2551 }
2552
2553 page = compound_head(page);
2554
2555 _total_mapcount = ret = 0;
2556 for (i = 0; i < HPAGE_PMD_NR; i++) {
2557 mapcount = atomic_read(&page[i]._mapcount) + 1;
2558 ret = max(ret, mapcount);
2559 _total_mapcount += mapcount;
2560 }
2561 if (PageDoubleMap(page)) {
2562 ret -= 1;
2563 _total_mapcount -= HPAGE_PMD_NR;
2564 }
2565 mapcount = compound_mapcount(page);
2566 ret += mapcount;
2567 _total_mapcount += mapcount;
2568 if (total_mapcount)
2569 *total_mapcount = _total_mapcount;
2570 return ret;
2571}
2572
2573/* Racy check whether the huge page can be split */
2574bool can_split_huge_page(struct page *page, int *pextra_pins)
2575{
2576 int extra_pins;
2577
2578 /* Additional pins from page cache */
2579 if (PageAnon(page))
2580 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2581 else
2582 extra_pins = HPAGE_PMD_NR;
2583 if (pextra_pins)
2584 *pextra_pins = extra_pins;
2585 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2586}
2587
2588/*
2589 * This function splits huge page into normal pages. @page can point to any
2590 * subpage of huge page to split. Split doesn't change the position of @page.
2591 *
2592 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2593 * The huge page must be locked.
2594 *
2595 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2596 *
2597 * Both head page and tail pages will inherit mapping, flags, and so on from
2598 * the hugepage.
2599 *
2600 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2601 * they are not mapped.
2602 *
2603 * Returns 0 if the hugepage is split successfully.
2604 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2605 * us.
2606 */
2607int split_huge_page_to_list(struct page *page, struct list_head *list)
2608{
2609 struct page *head = compound_head(page);
2610 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2611 struct deferred_split *ds_queue = get_deferred_split_queue(head);
2612 struct anon_vma *anon_vma = NULL;
2613 struct address_space *mapping = NULL;
2614 int count, mapcount, extra_pins, ret;
2615 unsigned long flags;
2616 pgoff_t end;
2617
2618 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2619 VM_BUG_ON_PAGE(!PageLocked(head), head);
2620 VM_BUG_ON_PAGE(!PageCompound(head), head);
2621
2622 if (PageWriteback(head))
2623 return -EBUSY;
2624
2625 if (PageAnon(head)) {
2626 /*
2627 * The caller does not necessarily hold an mmap_lock that would
2628 * prevent the anon_vma disappearing so we first we take a
2629 * reference to it and then lock the anon_vma for write. This
2630 * is similar to page_lock_anon_vma_read except the write lock
2631 * is taken to serialise against parallel split or collapse
2632 * operations.
2633 */
2634 anon_vma = page_get_anon_vma(head);
2635 if (!anon_vma) {
2636 ret = -EBUSY;
2637 goto out;
2638 }
2639 end = -1;
2640 mapping = NULL;
2641 anon_vma_lock_write(anon_vma);
2642 } else {
2643 mapping = head->mapping;
2644
2645 /* Truncated ? */
2646 if (!mapping) {
2647 ret = -EBUSY;
2648 goto out;
2649 }
2650
2651 anon_vma = NULL;
2652 i_mmap_lock_read(mapping);
2653
2654 /*
2655 *__split_huge_page() may need to trim off pages beyond EOF:
2656 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2657 * which cannot be nested inside the page tree lock. So note
2658 * end now: i_size itself may be changed at any moment, but
2659 * head page lock is good enough to serialize the trimming.
2660 */
2661 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2662 }
2663
2664 /*
2665 * Racy check if we can split the page, before unmap_page() will
2666 * split PMDs
2667 */
2668 if (!can_split_huge_page(head, &extra_pins)) {
2669 ret = -EBUSY;
2670 goto out_unlock;
2671 }
2672
2673 unmap_page(head);
2674 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2675
2676 /* prevent PageLRU to go away from under us, and freeze lru stats */
2677 spin_lock_irqsave(&pgdata->lru_lock, flags);
2678
2679 if (mapping) {
2680 XA_STATE(xas, &mapping->i_pages, page_index(head));
2681
2682 /*
2683 * Check if the head page is present in page cache.
2684 * We assume all tail are present too, if head is there.
2685 */
2686 xa_lock(&mapping->i_pages);
2687 if (xas_load(&xas) != head)
2688 goto fail;
2689 }
2690
2691 /* Prevent deferred_split_scan() touching ->_refcount */
2692 spin_lock(&ds_queue->split_queue_lock);
2693 count = page_count(head);
2694 mapcount = total_mapcount(head);
2695 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2696 if (!list_empty(page_deferred_list(head))) {
2697 ds_queue->split_queue_len--;
2698 list_del(page_deferred_list(head));
2699 }
2700 spin_unlock(&ds_queue->split_queue_lock);
2701 if (mapping) {
2702 if (PageSwapBacked(head))
2703 __dec_node_page_state(head, NR_SHMEM_THPS);
2704 else
2705 __dec_node_page_state(head, NR_FILE_THPS);
2706 }
2707
2708 __split_huge_page(page, list, end, flags);
2709 if (PageSwapCache(head)) {
2710 swp_entry_t entry = { .val = page_private(head) };
2711
2712 ret = split_swap_cluster(entry);
2713 } else
2714 ret = 0;
2715 } else {
2716 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2717 pr_alert("total_mapcount: %u, page_count(): %u\n",
2718 mapcount, count);
2719 if (PageTail(page))
2720 dump_page(head, NULL);
2721 dump_page(page, "total_mapcount(head) > 0");
2722 BUG();
2723 }
2724 spin_unlock(&ds_queue->split_queue_lock);
2725fail: if (mapping)
2726 xa_unlock(&mapping->i_pages);
2727 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2728 remap_page(head);
2729 ret = -EBUSY;
2730 }
2731
2732out_unlock:
2733 if (anon_vma) {
2734 anon_vma_unlock_write(anon_vma);
2735 put_anon_vma(anon_vma);
2736 }
2737 if (mapping)
2738 i_mmap_unlock_read(mapping);
2739out:
2740 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2741 return ret;
2742}
2743
2744void free_transhuge_page(struct page *page)
2745{
2746 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2747 unsigned long flags;
2748
2749 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2750 if (!list_empty(page_deferred_list(page))) {
2751 ds_queue->split_queue_len--;
2752 list_del(page_deferred_list(page));
2753 }
2754 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2755 free_compound_page(page);
2756}
2757
2758void deferred_split_huge_page(struct page *page)
2759{
2760 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2761#ifdef CONFIG_MEMCG
2762 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2763#endif
2764 unsigned long flags;
2765
2766 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2767
2768 /*
2769 * The try_to_unmap() in page reclaim path might reach here too,
2770 * this may cause a race condition to corrupt deferred split queue.
2771 * And, if page reclaim is already handling the same page, it is
2772 * unnecessary to handle it again in shrinker.
2773 *
2774 * Check PageSwapCache to determine if the page is being
2775 * handled by page reclaim since THP swap would add the page into
2776 * swap cache before calling try_to_unmap().
2777 */
2778 if (PageSwapCache(page))
2779 return;
2780
2781 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2782 if (list_empty(page_deferred_list(page))) {
2783 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2784 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2785 ds_queue->split_queue_len++;
2786#ifdef CONFIG_MEMCG
2787 if (memcg)
2788 memcg_set_shrinker_bit(memcg, page_to_nid(page),
2789 deferred_split_shrinker.id);
2790#endif
2791 }
2792 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2793}
2794
2795static unsigned long deferred_split_count(struct shrinker *shrink,
2796 struct shrink_control *sc)
2797{
2798 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2799 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2800
2801#ifdef CONFIG_MEMCG
2802 if (sc->memcg)
2803 ds_queue = &sc->memcg->deferred_split_queue;
2804#endif
2805 return READ_ONCE(ds_queue->split_queue_len);
2806}
2807
2808static unsigned long deferred_split_scan(struct shrinker *shrink,
2809 struct shrink_control *sc)
2810{
2811 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2812 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2813 unsigned long flags;
2814 LIST_HEAD(list), *pos, *next;
2815 struct page *page;
2816 int split = 0;
2817
2818#ifdef CONFIG_MEMCG
2819 if (sc->memcg)
2820 ds_queue = &sc->memcg->deferred_split_queue;
2821#endif
2822
2823 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2824 /* Take pin on all head pages to avoid freeing them under us */
2825 list_for_each_safe(pos, next, &ds_queue->split_queue) {
2826 page = list_entry((void *)pos, struct page, mapping);
2827 page = compound_head(page);
2828 if (get_page_unless_zero(page)) {
2829 list_move(page_deferred_list(page), &list);
2830 } else {
2831 /* We lost race with put_compound_page() */
2832 list_del_init(page_deferred_list(page));
2833 ds_queue->split_queue_len--;
2834 }
2835 if (!--sc->nr_to_scan)
2836 break;
2837 }
2838 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2839
2840 list_for_each_safe(pos, next, &list) {
2841 page = list_entry((void *)pos, struct page, mapping);
2842 if (!trylock_page(page))
2843 goto next;
2844 /* split_huge_page() removes page from list on success */
2845 if (!split_huge_page(page))
2846 split++;
2847 unlock_page(page);
2848next:
2849 put_page(page);
2850 }
2851
2852 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2853 list_splice_tail(&list, &ds_queue->split_queue);
2854 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2855
2856 /*
2857 * Stop shrinker if we didn't split any page, but the queue is empty.
2858 * This can happen if pages were freed under us.
2859 */
2860 if (!split && list_empty(&ds_queue->split_queue))
2861 return SHRINK_STOP;
2862 return split;
2863}
2864
2865static struct shrinker deferred_split_shrinker = {
2866 .count_objects = deferred_split_count,
2867 .scan_objects = deferred_split_scan,
2868 .seeks = DEFAULT_SEEKS,
2869 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2870 SHRINKER_NONSLAB,
2871};
2872
2873#ifdef CONFIG_DEBUG_FS
2874static int split_huge_pages_set(void *data, u64 val)
2875{
2876 struct zone *zone;
2877 struct page *page;
2878 unsigned long pfn, max_zone_pfn;
2879 unsigned long total = 0, split = 0;
2880
2881 if (val != 1)
2882 return -EINVAL;
2883
2884 for_each_populated_zone(zone) {
2885 max_zone_pfn = zone_end_pfn(zone);
2886 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2887 if (!pfn_valid(pfn))
2888 continue;
2889
2890 page = pfn_to_page(pfn);
2891 if (!get_page_unless_zero(page))
2892 continue;
2893
2894 if (zone != page_zone(page))
2895 goto next;
2896
2897 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2898 goto next;
2899
2900 total++;
2901 lock_page(page);
2902 if (!split_huge_page(page))
2903 split++;
2904 unlock_page(page);
2905next:
2906 put_page(page);
2907 }
2908 }
2909
2910 pr_info("%lu of %lu THP split\n", split, total);
2911
2912 return 0;
2913}
2914DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2915 "%llu\n");
2916
2917static int __init split_huge_pages_debugfs(void)
2918{
2919 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2920 &split_huge_pages_fops);
2921 return 0;
2922}
2923late_initcall(split_huge_pages_debugfs);
2924#endif
2925
2926#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2927void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2928 struct page *page)
2929{
2930 struct vm_area_struct *vma = pvmw->vma;
2931 struct mm_struct *mm = vma->vm_mm;
2932 unsigned long address = pvmw->address;
2933 pmd_t pmdval;
2934 swp_entry_t entry;
2935 pmd_t pmdswp;
2936
2937 if (!(pvmw->pmd && !pvmw->pte))
2938 return;
2939
2940 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2941 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2942 if (pmd_dirty(pmdval))
2943 set_page_dirty(page);
2944 entry = make_migration_entry(page, pmd_write(pmdval));
2945 pmdswp = swp_entry_to_pmd(entry);
2946 if (pmd_soft_dirty(pmdval))
2947 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2948 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2949 page_remove_rmap(page, true);
2950 put_page(page);
2951}
2952
2953void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2954{
2955 struct vm_area_struct *vma = pvmw->vma;
2956 struct mm_struct *mm = vma->vm_mm;
2957 unsigned long address = pvmw->address;
2958 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2959 pmd_t pmde;
2960 swp_entry_t entry;
2961
2962 if (!(pvmw->pmd && !pvmw->pte))
2963 return;
2964
2965 entry = pmd_to_swp_entry(*pvmw->pmd);
2966 get_page(new);
2967 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2968 if (pmd_swp_soft_dirty(*pvmw->pmd))
2969 pmde = pmd_mksoft_dirty(pmde);
2970 if (is_write_migration_entry(entry))
2971 pmde = maybe_pmd_mkwrite(pmde, vma);
2972
2973 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2974 if (PageAnon(new))
2975 page_add_anon_rmap(new, vma, mmun_start, true);
2976 else
2977 page_add_file_rmap(new, true);
2978 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2979 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2980 mlock_vma_page(new);
2981 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2982}
2983#endif