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/*
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#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/highmem.h>
11#include <linux/hugetlb.h>
12#include <linux/mmu_notifier.h>
13#include <linux/rmap.h>
14#include <linux/swap.h>
15#include <linux/mm_inline.h>
16#include <linux/kthread.h>
17#include <linux/khugepaged.h>
18#include <linux/freezer.h>
19#include <linux/mman.h>
20#include <asm/tlb.h>
21#include <asm/pgalloc.h>
22#include "internal.h"
23
24/*
25 * By default transparent hugepage support is enabled for all mappings
26 * and khugepaged scans all mappings. Defrag is only invoked by
27 * khugepaged hugepage allocations and by page faults inside
28 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
29 * allocations.
30 */
31unsigned long transparent_hugepage_flags __read_mostly =
32#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
33 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
34#endif
35#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
37#endif
38 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
40
41/* default scan 8*512 pte (or vmas) every 30 second */
42static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43static unsigned int khugepaged_pages_collapsed;
44static unsigned int khugepaged_full_scans;
45static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46/* during fragmentation poll the hugepage allocator once every minute */
47static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48static struct task_struct *khugepaged_thread __read_mostly;
49static DEFINE_MUTEX(khugepaged_mutex);
50static DEFINE_SPINLOCK(khugepaged_mm_lock);
51static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
52/*
53 * default collapse hugepages if there is at least one pte mapped like
54 * it would have happened if the vma was large enough during page
55 * fault.
56 */
57static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
58
59static int khugepaged(void *none);
60static int mm_slots_hash_init(void);
61static int khugepaged_slab_init(void);
62static void khugepaged_slab_free(void);
63
64#define MM_SLOTS_HASH_HEADS 1024
65static struct hlist_head *mm_slots_hash __read_mostly;
66static struct kmem_cache *mm_slot_cache __read_mostly;
67
68/**
69 * struct mm_slot - hash lookup from mm to mm_slot
70 * @hash: hash collision list
71 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72 * @mm: the mm that this information is valid for
73 */
74struct mm_slot {
75 struct hlist_node hash;
76 struct list_head mm_node;
77 struct mm_struct *mm;
78};
79
80/**
81 * struct khugepaged_scan - cursor for scanning
82 * @mm_head: the head of the mm list to scan
83 * @mm_slot: the current mm_slot we are scanning
84 * @address: the next address inside that to be scanned
85 *
86 * There is only the one khugepaged_scan instance of this cursor structure.
87 */
88struct khugepaged_scan {
89 struct list_head mm_head;
90 struct mm_slot *mm_slot;
91 unsigned long address;
92};
93static struct khugepaged_scan khugepaged_scan = {
94 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
95};
96
97
98static int set_recommended_min_free_kbytes(void)
99{
100 struct zone *zone;
101 int nr_zones = 0;
102 unsigned long recommended_min;
103 extern int min_free_kbytes;
104
105 if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
106 &transparent_hugepage_flags) &&
107 !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
108 &transparent_hugepage_flags))
109 return 0;
110
111 for_each_populated_zone(zone)
112 nr_zones++;
113
114 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
115 recommended_min = pageblock_nr_pages * nr_zones * 2;
116
117 /*
118 * Make sure that on average at least two pageblocks are almost free
119 * of another type, one for a migratetype to fall back to and a
120 * second to avoid subsequent fallbacks of other types There are 3
121 * MIGRATE_TYPES we care about.
122 */
123 recommended_min += pageblock_nr_pages * nr_zones *
124 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
125
126 /* don't ever allow to reserve more than 5% of the lowmem */
127 recommended_min = min(recommended_min,
128 (unsigned long) nr_free_buffer_pages() / 20);
129 recommended_min <<= (PAGE_SHIFT-10);
130
131 if (recommended_min > min_free_kbytes)
132 min_free_kbytes = recommended_min;
133 setup_per_zone_wmarks();
134 return 0;
135}
136late_initcall(set_recommended_min_free_kbytes);
137
138static int start_khugepaged(void)
139{
140 int err = 0;
141 if (khugepaged_enabled()) {
142 int wakeup;
143 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
144 err = -ENOMEM;
145 goto out;
146 }
147 mutex_lock(&khugepaged_mutex);
148 if (!khugepaged_thread)
149 khugepaged_thread = kthread_run(khugepaged, NULL,
150 "khugepaged");
151 if (unlikely(IS_ERR(khugepaged_thread))) {
152 printk(KERN_ERR
153 "khugepaged: kthread_run(khugepaged) failed\n");
154 err = PTR_ERR(khugepaged_thread);
155 khugepaged_thread = NULL;
156 }
157 wakeup = !list_empty(&khugepaged_scan.mm_head);
158 mutex_unlock(&khugepaged_mutex);
159 if (wakeup)
160 wake_up_interruptible(&khugepaged_wait);
161
162 set_recommended_min_free_kbytes();
163 } else
164 /* wakeup to exit */
165 wake_up_interruptible(&khugepaged_wait);
166out:
167 return err;
168}
169
170#ifdef CONFIG_SYSFS
171
172static ssize_t double_flag_show(struct kobject *kobj,
173 struct kobj_attribute *attr, char *buf,
174 enum transparent_hugepage_flag enabled,
175 enum transparent_hugepage_flag req_madv)
176{
177 if (test_bit(enabled, &transparent_hugepage_flags)) {
178 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
179 return sprintf(buf, "[always] madvise never\n");
180 } else if (test_bit(req_madv, &transparent_hugepage_flags))
181 return sprintf(buf, "always [madvise] never\n");
182 else
183 return sprintf(buf, "always madvise [never]\n");
184}
185static ssize_t double_flag_store(struct kobject *kobj,
186 struct kobj_attribute *attr,
187 const char *buf, size_t count,
188 enum transparent_hugepage_flag enabled,
189 enum transparent_hugepage_flag req_madv)
190{
191 if (!memcmp("always", buf,
192 min(sizeof("always")-1, count))) {
193 set_bit(enabled, &transparent_hugepage_flags);
194 clear_bit(req_madv, &transparent_hugepage_flags);
195 } else if (!memcmp("madvise", buf,
196 min(sizeof("madvise")-1, count))) {
197 clear_bit(enabled, &transparent_hugepage_flags);
198 set_bit(req_madv, &transparent_hugepage_flags);
199 } else if (!memcmp("never", buf,
200 min(sizeof("never")-1, count))) {
201 clear_bit(enabled, &transparent_hugepage_flags);
202 clear_bit(req_madv, &transparent_hugepage_flags);
203 } else
204 return -EINVAL;
205
206 return count;
207}
208
209static ssize_t enabled_show(struct kobject *kobj,
210 struct kobj_attribute *attr, char *buf)
211{
212 return double_flag_show(kobj, attr, buf,
213 TRANSPARENT_HUGEPAGE_FLAG,
214 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
215}
216static ssize_t enabled_store(struct kobject *kobj,
217 struct kobj_attribute *attr,
218 const char *buf, size_t count)
219{
220 ssize_t ret;
221
222 ret = double_flag_store(kobj, attr, buf, count,
223 TRANSPARENT_HUGEPAGE_FLAG,
224 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
225
226 if (ret > 0) {
227 int err = start_khugepaged();
228 if (err)
229 ret = err;
230 }
231
232 if (ret > 0 &&
233 (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
234 &transparent_hugepage_flags) ||
235 test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
236 &transparent_hugepage_flags)))
237 set_recommended_min_free_kbytes();
238
239 return ret;
240}
241static struct kobj_attribute enabled_attr =
242 __ATTR(enabled, 0644, enabled_show, enabled_store);
243
244static ssize_t single_flag_show(struct kobject *kobj,
245 struct kobj_attribute *attr, char *buf,
246 enum transparent_hugepage_flag flag)
247{
248 return sprintf(buf, "%d\n",
249 !!test_bit(flag, &transparent_hugepage_flags));
250}
251
252static ssize_t single_flag_store(struct kobject *kobj,
253 struct kobj_attribute *attr,
254 const char *buf, size_t count,
255 enum transparent_hugepage_flag flag)
256{
257 unsigned long value;
258 int ret;
259
260 ret = kstrtoul(buf, 10, &value);
261 if (ret < 0)
262 return ret;
263 if (value > 1)
264 return -EINVAL;
265
266 if (value)
267 set_bit(flag, &transparent_hugepage_flags);
268 else
269 clear_bit(flag, &transparent_hugepage_flags);
270
271 return count;
272}
273
274/*
275 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
276 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
277 * memory just to allocate one more hugepage.
278 */
279static ssize_t defrag_show(struct kobject *kobj,
280 struct kobj_attribute *attr, char *buf)
281{
282 return double_flag_show(kobj, attr, buf,
283 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
284 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
285}
286static ssize_t defrag_store(struct kobject *kobj,
287 struct kobj_attribute *attr,
288 const char *buf, size_t count)
289{
290 return double_flag_store(kobj, attr, buf, count,
291 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
292 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
293}
294static struct kobj_attribute defrag_attr =
295 __ATTR(defrag, 0644, defrag_show, defrag_store);
296
297#ifdef CONFIG_DEBUG_VM
298static ssize_t debug_cow_show(struct kobject *kobj,
299 struct kobj_attribute *attr, char *buf)
300{
301 return single_flag_show(kobj, attr, buf,
302 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303}
304static ssize_t debug_cow_store(struct kobject *kobj,
305 struct kobj_attribute *attr,
306 const char *buf, size_t count)
307{
308 return single_flag_store(kobj, attr, buf, count,
309 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310}
311static struct kobj_attribute debug_cow_attr =
312 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
313#endif /* CONFIG_DEBUG_VM */
314
315static struct attribute *hugepage_attr[] = {
316 &enabled_attr.attr,
317 &defrag_attr.attr,
318#ifdef CONFIG_DEBUG_VM
319 &debug_cow_attr.attr,
320#endif
321 NULL,
322};
323
324static struct attribute_group hugepage_attr_group = {
325 .attrs = hugepage_attr,
326};
327
328static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
329 struct kobj_attribute *attr,
330 char *buf)
331{
332 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
333}
334
335static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
336 struct kobj_attribute *attr,
337 const char *buf, size_t count)
338{
339 unsigned long msecs;
340 int err;
341
342 err = strict_strtoul(buf, 10, &msecs);
343 if (err || msecs > UINT_MAX)
344 return -EINVAL;
345
346 khugepaged_scan_sleep_millisecs = msecs;
347 wake_up_interruptible(&khugepaged_wait);
348
349 return count;
350}
351static struct kobj_attribute scan_sleep_millisecs_attr =
352 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
353 scan_sleep_millisecs_store);
354
355static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
356 struct kobj_attribute *attr,
357 char *buf)
358{
359 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
360}
361
362static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
363 struct kobj_attribute *attr,
364 const char *buf, size_t count)
365{
366 unsigned long msecs;
367 int err;
368
369 err = strict_strtoul(buf, 10, &msecs);
370 if (err || msecs > UINT_MAX)
371 return -EINVAL;
372
373 khugepaged_alloc_sleep_millisecs = msecs;
374 wake_up_interruptible(&khugepaged_wait);
375
376 return count;
377}
378static struct kobj_attribute alloc_sleep_millisecs_attr =
379 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
380 alloc_sleep_millisecs_store);
381
382static ssize_t pages_to_scan_show(struct kobject *kobj,
383 struct kobj_attribute *attr,
384 char *buf)
385{
386 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
387}
388static ssize_t pages_to_scan_store(struct kobject *kobj,
389 struct kobj_attribute *attr,
390 const char *buf, size_t count)
391{
392 int err;
393 unsigned long pages;
394
395 err = strict_strtoul(buf, 10, &pages);
396 if (err || !pages || pages > UINT_MAX)
397 return -EINVAL;
398
399 khugepaged_pages_to_scan = pages;
400
401 return count;
402}
403static struct kobj_attribute pages_to_scan_attr =
404 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
405 pages_to_scan_store);
406
407static ssize_t pages_collapsed_show(struct kobject *kobj,
408 struct kobj_attribute *attr,
409 char *buf)
410{
411 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
412}
413static struct kobj_attribute pages_collapsed_attr =
414 __ATTR_RO(pages_collapsed);
415
416static ssize_t full_scans_show(struct kobject *kobj,
417 struct kobj_attribute *attr,
418 char *buf)
419{
420 return sprintf(buf, "%u\n", khugepaged_full_scans);
421}
422static struct kobj_attribute full_scans_attr =
423 __ATTR_RO(full_scans);
424
425static ssize_t khugepaged_defrag_show(struct kobject *kobj,
426 struct kobj_attribute *attr, char *buf)
427{
428 return single_flag_show(kobj, attr, buf,
429 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
430}
431static ssize_t khugepaged_defrag_store(struct kobject *kobj,
432 struct kobj_attribute *attr,
433 const char *buf, size_t count)
434{
435 return single_flag_store(kobj, attr, buf, count,
436 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
437}
438static struct kobj_attribute khugepaged_defrag_attr =
439 __ATTR(defrag, 0644, khugepaged_defrag_show,
440 khugepaged_defrag_store);
441
442/*
443 * max_ptes_none controls if khugepaged should collapse hugepages over
444 * any unmapped ptes in turn potentially increasing the memory
445 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
446 * reduce the available free memory in the system as it
447 * runs. Increasing max_ptes_none will instead potentially reduce the
448 * free memory in the system during the khugepaged scan.
449 */
450static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
451 struct kobj_attribute *attr,
452 char *buf)
453{
454 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
455}
456static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
457 struct kobj_attribute *attr,
458 const char *buf, size_t count)
459{
460 int err;
461 unsigned long max_ptes_none;
462
463 err = strict_strtoul(buf, 10, &max_ptes_none);
464 if (err || max_ptes_none > HPAGE_PMD_NR-1)
465 return -EINVAL;
466
467 khugepaged_max_ptes_none = max_ptes_none;
468
469 return count;
470}
471static struct kobj_attribute khugepaged_max_ptes_none_attr =
472 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
473 khugepaged_max_ptes_none_store);
474
475static struct attribute *khugepaged_attr[] = {
476 &khugepaged_defrag_attr.attr,
477 &khugepaged_max_ptes_none_attr.attr,
478 &pages_to_scan_attr.attr,
479 &pages_collapsed_attr.attr,
480 &full_scans_attr.attr,
481 &scan_sleep_millisecs_attr.attr,
482 &alloc_sleep_millisecs_attr.attr,
483 NULL,
484};
485
486static struct attribute_group khugepaged_attr_group = {
487 .attrs = khugepaged_attr,
488 .name = "khugepaged",
489};
490
491static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
492{
493 int err;
494
495 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
496 if (unlikely(!*hugepage_kobj)) {
497 printk(KERN_ERR "hugepage: failed kobject create\n");
498 return -ENOMEM;
499 }
500
501 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
502 if (err) {
503 printk(KERN_ERR "hugepage: failed register hugeage group\n");
504 goto delete_obj;
505 }
506
507 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
508 if (err) {
509 printk(KERN_ERR "hugepage: failed register hugeage group\n");
510 goto remove_hp_group;
511 }
512
513 return 0;
514
515remove_hp_group:
516 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
517delete_obj:
518 kobject_put(*hugepage_kobj);
519 return err;
520}
521
522static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
523{
524 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
525 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
526 kobject_put(hugepage_kobj);
527}
528#else
529static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
530{
531 return 0;
532}
533
534static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
535{
536}
537#endif /* CONFIG_SYSFS */
538
539static int __init hugepage_init(void)
540{
541 int err;
542 struct kobject *hugepage_kobj;
543
544 if (!has_transparent_hugepage()) {
545 transparent_hugepage_flags = 0;
546 return -EINVAL;
547 }
548
549 err = hugepage_init_sysfs(&hugepage_kobj);
550 if (err)
551 return err;
552
553 err = khugepaged_slab_init();
554 if (err)
555 goto out;
556
557 err = mm_slots_hash_init();
558 if (err) {
559 khugepaged_slab_free();
560 goto out;
561 }
562
563 /*
564 * By default disable transparent hugepages on smaller systems,
565 * where the extra memory used could hurt more than TLB overhead
566 * is likely to save. The admin can still enable it through /sys.
567 */
568 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
569 transparent_hugepage_flags = 0;
570
571 start_khugepaged();
572
573 set_recommended_min_free_kbytes();
574
575 return 0;
576out:
577 hugepage_exit_sysfs(hugepage_kobj);
578 return err;
579}
580module_init(hugepage_init)
581
582static int __init setup_transparent_hugepage(char *str)
583{
584 int ret = 0;
585 if (!str)
586 goto out;
587 if (!strcmp(str, "always")) {
588 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
589 &transparent_hugepage_flags);
590 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
591 &transparent_hugepage_flags);
592 ret = 1;
593 } else if (!strcmp(str, "madvise")) {
594 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
595 &transparent_hugepage_flags);
596 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
597 &transparent_hugepage_flags);
598 ret = 1;
599 } else if (!strcmp(str, "never")) {
600 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
601 &transparent_hugepage_flags);
602 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
603 &transparent_hugepage_flags);
604 ret = 1;
605 }
606out:
607 if (!ret)
608 printk(KERN_WARNING
609 "transparent_hugepage= cannot parse, ignored\n");
610 return ret;
611}
612__setup("transparent_hugepage=", setup_transparent_hugepage);
613
614static void prepare_pmd_huge_pte(pgtable_t pgtable,
615 struct mm_struct *mm)
616{
617 assert_spin_locked(&mm->page_table_lock);
618
619 /* FIFO */
620 if (!mm->pmd_huge_pte)
621 INIT_LIST_HEAD(&pgtable->lru);
622 else
623 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
624 mm->pmd_huge_pte = pgtable;
625}
626
627static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
628{
629 if (likely(vma->vm_flags & VM_WRITE))
630 pmd = pmd_mkwrite(pmd);
631 return pmd;
632}
633
634static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
635 struct vm_area_struct *vma,
636 unsigned long haddr, pmd_t *pmd,
637 struct page *page)
638{
639 pgtable_t pgtable;
640
641 VM_BUG_ON(!PageCompound(page));
642 pgtable = pte_alloc_one(mm, haddr);
643 if (unlikely(!pgtable))
644 return VM_FAULT_OOM;
645
646 clear_huge_page(page, haddr, HPAGE_PMD_NR);
647 __SetPageUptodate(page);
648
649 spin_lock(&mm->page_table_lock);
650 if (unlikely(!pmd_none(*pmd))) {
651 spin_unlock(&mm->page_table_lock);
652 mem_cgroup_uncharge_page(page);
653 put_page(page);
654 pte_free(mm, pgtable);
655 } else {
656 pmd_t entry;
657 entry = mk_pmd(page, vma->vm_page_prot);
658 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
659 entry = pmd_mkhuge(entry);
660 /*
661 * The spinlocking to take the lru_lock inside
662 * page_add_new_anon_rmap() acts as a full memory
663 * barrier to be sure clear_huge_page writes become
664 * visible after the set_pmd_at() write.
665 */
666 page_add_new_anon_rmap(page, vma, haddr);
667 set_pmd_at(mm, haddr, pmd, entry);
668 prepare_pmd_huge_pte(pgtable, mm);
669 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
670 mm->nr_ptes++;
671 spin_unlock(&mm->page_table_lock);
672 }
673
674 return 0;
675}
676
677static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
678{
679 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
680}
681
682static inline struct page *alloc_hugepage_vma(int defrag,
683 struct vm_area_struct *vma,
684 unsigned long haddr, int nd,
685 gfp_t extra_gfp)
686{
687 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
688 HPAGE_PMD_ORDER, vma, haddr, nd);
689}
690
691#ifndef CONFIG_NUMA
692static inline struct page *alloc_hugepage(int defrag)
693{
694 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
695 HPAGE_PMD_ORDER);
696}
697#endif
698
699int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
700 unsigned long address, pmd_t *pmd,
701 unsigned int flags)
702{
703 struct page *page;
704 unsigned long haddr = address & HPAGE_PMD_MASK;
705 pte_t *pte;
706
707 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
708 if (unlikely(anon_vma_prepare(vma)))
709 return VM_FAULT_OOM;
710 if (unlikely(khugepaged_enter(vma)))
711 return VM_FAULT_OOM;
712 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
713 vma, haddr, numa_node_id(), 0);
714 if (unlikely(!page)) {
715 count_vm_event(THP_FAULT_FALLBACK);
716 goto out;
717 }
718 count_vm_event(THP_FAULT_ALLOC);
719 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
720 put_page(page);
721 goto out;
722 }
723 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
724 page))) {
725 mem_cgroup_uncharge_page(page);
726 put_page(page);
727 goto out;
728 }
729
730 return 0;
731 }
732out:
733 /*
734 * Use __pte_alloc instead of pte_alloc_map, because we can't
735 * run pte_offset_map on the pmd, if an huge pmd could
736 * materialize from under us from a different thread.
737 */
738 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
739 return VM_FAULT_OOM;
740 /* if an huge pmd materialized from under us just retry later */
741 if (unlikely(pmd_trans_huge(*pmd)))
742 return 0;
743 /*
744 * A regular pmd is established and it can't morph into a huge pmd
745 * from under us anymore at this point because we hold the mmap_sem
746 * read mode and khugepaged takes it in write mode. So now it's
747 * safe to run pte_offset_map().
748 */
749 pte = pte_offset_map(pmd, address);
750 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
751}
752
753int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
754 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
755 struct vm_area_struct *vma)
756{
757 struct page *src_page;
758 pmd_t pmd;
759 pgtable_t pgtable;
760 int ret;
761
762 ret = -ENOMEM;
763 pgtable = pte_alloc_one(dst_mm, addr);
764 if (unlikely(!pgtable))
765 goto out;
766
767 spin_lock(&dst_mm->page_table_lock);
768 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
769
770 ret = -EAGAIN;
771 pmd = *src_pmd;
772 if (unlikely(!pmd_trans_huge(pmd))) {
773 pte_free(dst_mm, pgtable);
774 goto out_unlock;
775 }
776 if (unlikely(pmd_trans_splitting(pmd))) {
777 /* split huge page running from under us */
778 spin_unlock(&src_mm->page_table_lock);
779 spin_unlock(&dst_mm->page_table_lock);
780 pte_free(dst_mm, pgtable);
781
782 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
783 goto out;
784 }
785 src_page = pmd_page(pmd);
786 VM_BUG_ON(!PageHead(src_page));
787 get_page(src_page);
788 page_dup_rmap(src_page);
789 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
790
791 pmdp_set_wrprotect(src_mm, addr, src_pmd);
792 pmd = pmd_mkold(pmd_wrprotect(pmd));
793 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
794 prepare_pmd_huge_pte(pgtable, dst_mm);
795 dst_mm->nr_ptes++;
796
797 ret = 0;
798out_unlock:
799 spin_unlock(&src_mm->page_table_lock);
800 spin_unlock(&dst_mm->page_table_lock);
801out:
802 return ret;
803}
804
805/* no "address" argument so destroys page coloring of some arch */
806pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
807{
808 pgtable_t pgtable;
809
810 assert_spin_locked(&mm->page_table_lock);
811
812 /* FIFO */
813 pgtable = mm->pmd_huge_pte;
814 if (list_empty(&pgtable->lru))
815 mm->pmd_huge_pte = NULL;
816 else {
817 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
818 struct page, lru);
819 list_del(&pgtable->lru);
820 }
821 return pgtable;
822}
823
824static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
825 struct vm_area_struct *vma,
826 unsigned long address,
827 pmd_t *pmd, pmd_t orig_pmd,
828 struct page *page,
829 unsigned long haddr)
830{
831 pgtable_t pgtable;
832 pmd_t _pmd;
833 int ret = 0, i;
834 struct page **pages;
835
836 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
837 GFP_KERNEL);
838 if (unlikely(!pages)) {
839 ret |= VM_FAULT_OOM;
840 goto out;
841 }
842
843 for (i = 0; i < HPAGE_PMD_NR; i++) {
844 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
845 __GFP_OTHER_NODE,
846 vma, address, page_to_nid(page));
847 if (unlikely(!pages[i] ||
848 mem_cgroup_newpage_charge(pages[i], mm,
849 GFP_KERNEL))) {
850 if (pages[i])
851 put_page(pages[i]);
852 mem_cgroup_uncharge_start();
853 while (--i >= 0) {
854 mem_cgroup_uncharge_page(pages[i]);
855 put_page(pages[i]);
856 }
857 mem_cgroup_uncharge_end();
858 kfree(pages);
859 ret |= VM_FAULT_OOM;
860 goto out;
861 }
862 }
863
864 for (i = 0; i < HPAGE_PMD_NR; i++) {
865 copy_user_highpage(pages[i], page + i,
866 haddr + PAGE_SIZE * i, vma);
867 __SetPageUptodate(pages[i]);
868 cond_resched();
869 }
870
871 spin_lock(&mm->page_table_lock);
872 if (unlikely(!pmd_same(*pmd, orig_pmd)))
873 goto out_free_pages;
874 VM_BUG_ON(!PageHead(page));
875
876 pmdp_clear_flush_notify(vma, haddr, pmd);
877 /* leave pmd empty until pte is filled */
878
879 pgtable = get_pmd_huge_pte(mm);
880 pmd_populate(mm, &_pmd, pgtable);
881
882 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
883 pte_t *pte, entry;
884 entry = mk_pte(pages[i], vma->vm_page_prot);
885 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
886 page_add_new_anon_rmap(pages[i], vma, haddr);
887 pte = pte_offset_map(&_pmd, haddr);
888 VM_BUG_ON(!pte_none(*pte));
889 set_pte_at(mm, haddr, pte, entry);
890 pte_unmap(pte);
891 }
892 kfree(pages);
893
894 smp_wmb(); /* make pte visible before pmd */
895 pmd_populate(mm, pmd, pgtable);
896 page_remove_rmap(page);
897 spin_unlock(&mm->page_table_lock);
898
899 ret |= VM_FAULT_WRITE;
900 put_page(page);
901
902out:
903 return ret;
904
905out_free_pages:
906 spin_unlock(&mm->page_table_lock);
907 mem_cgroup_uncharge_start();
908 for (i = 0; i < HPAGE_PMD_NR; i++) {
909 mem_cgroup_uncharge_page(pages[i]);
910 put_page(pages[i]);
911 }
912 mem_cgroup_uncharge_end();
913 kfree(pages);
914 goto out;
915}
916
917int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
918 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
919{
920 int ret = 0;
921 struct page *page, *new_page;
922 unsigned long haddr;
923
924 VM_BUG_ON(!vma->anon_vma);
925 spin_lock(&mm->page_table_lock);
926 if (unlikely(!pmd_same(*pmd, orig_pmd)))
927 goto out_unlock;
928
929 page = pmd_page(orig_pmd);
930 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
931 haddr = address & HPAGE_PMD_MASK;
932 if (page_mapcount(page) == 1) {
933 pmd_t entry;
934 entry = pmd_mkyoung(orig_pmd);
935 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
936 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
937 update_mmu_cache(vma, address, entry);
938 ret |= VM_FAULT_WRITE;
939 goto out_unlock;
940 }
941 get_page(page);
942 spin_unlock(&mm->page_table_lock);
943
944 if (transparent_hugepage_enabled(vma) &&
945 !transparent_hugepage_debug_cow())
946 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
947 vma, haddr, numa_node_id(), 0);
948 else
949 new_page = NULL;
950
951 if (unlikely(!new_page)) {
952 count_vm_event(THP_FAULT_FALLBACK);
953 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
954 pmd, orig_pmd, page, haddr);
955 if (ret & VM_FAULT_OOM)
956 split_huge_page(page);
957 put_page(page);
958 goto out;
959 }
960 count_vm_event(THP_FAULT_ALLOC);
961
962 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
963 put_page(new_page);
964 split_huge_page(page);
965 put_page(page);
966 ret |= VM_FAULT_OOM;
967 goto out;
968 }
969
970 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
971 __SetPageUptodate(new_page);
972
973 spin_lock(&mm->page_table_lock);
974 put_page(page);
975 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
976 spin_unlock(&mm->page_table_lock);
977 mem_cgroup_uncharge_page(new_page);
978 put_page(new_page);
979 goto out;
980 } else {
981 pmd_t entry;
982 VM_BUG_ON(!PageHead(page));
983 entry = mk_pmd(new_page, vma->vm_page_prot);
984 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
985 entry = pmd_mkhuge(entry);
986 pmdp_clear_flush_notify(vma, haddr, pmd);
987 page_add_new_anon_rmap(new_page, vma, haddr);
988 set_pmd_at(mm, haddr, pmd, entry);
989 update_mmu_cache(vma, address, entry);
990 page_remove_rmap(page);
991 put_page(page);
992 ret |= VM_FAULT_WRITE;
993 }
994out_unlock:
995 spin_unlock(&mm->page_table_lock);
996out:
997 return ret;
998}
999
1000struct page *follow_trans_huge_pmd(struct mm_struct *mm,
1001 unsigned long addr,
1002 pmd_t *pmd,
1003 unsigned int flags)
1004{
1005 struct page *page = NULL;
1006
1007 assert_spin_locked(&mm->page_table_lock);
1008
1009 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1010 goto out;
1011
1012 page = pmd_page(*pmd);
1013 VM_BUG_ON(!PageHead(page));
1014 if (flags & FOLL_TOUCH) {
1015 pmd_t _pmd;
1016 /*
1017 * We should set the dirty bit only for FOLL_WRITE but
1018 * for now the dirty bit in the pmd is meaningless.
1019 * And if the dirty bit will become meaningful and
1020 * we'll only set it with FOLL_WRITE, an atomic
1021 * set_bit will be required on the pmd to set the
1022 * young bit, instead of the current set_pmd_at.
1023 */
1024 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1025 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1026 }
1027 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1028 VM_BUG_ON(!PageCompound(page));
1029 if (flags & FOLL_GET)
1030 get_page_foll(page);
1031
1032out:
1033 return page;
1034}
1035
1036int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1037 pmd_t *pmd, unsigned long addr)
1038{
1039 int ret = 0;
1040
1041 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1042 struct page *page;
1043 pgtable_t pgtable;
1044 pgtable = get_pmd_huge_pte(tlb->mm);
1045 page = pmd_page(*pmd);
1046 pmd_clear(pmd);
1047 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1048 page_remove_rmap(page);
1049 VM_BUG_ON(page_mapcount(page) < 0);
1050 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1051 VM_BUG_ON(!PageHead(page));
1052 tlb->mm->nr_ptes--;
1053 spin_unlock(&tlb->mm->page_table_lock);
1054 tlb_remove_page(tlb, page);
1055 pte_free(tlb->mm, pgtable);
1056 ret = 1;
1057 }
1058 return ret;
1059}
1060
1061int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1062 unsigned long addr, unsigned long end,
1063 unsigned char *vec)
1064{
1065 int ret = 0;
1066
1067 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1068 /*
1069 * All logical pages in the range are present
1070 * if backed by a huge page.
1071 */
1072 spin_unlock(&vma->vm_mm->page_table_lock);
1073 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1074 ret = 1;
1075 }
1076
1077 return ret;
1078}
1079
1080int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1081 unsigned long old_addr,
1082 unsigned long new_addr, unsigned long old_end,
1083 pmd_t *old_pmd, pmd_t *new_pmd)
1084{
1085 int ret = 0;
1086 pmd_t pmd;
1087
1088 struct mm_struct *mm = vma->vm_mm;
1089
1090 if ((old_addr & ~HPAGE_PMD_MASK) ||
1091 (new_addr & ~HPAGE_PMD_MASK) ||
1092 old_end - old_addr < HPAGE_PMD_SIZE ||
1093 (new_vma->vm_flags & VM_NOHUGEPAGE))
1094 goto out;
1095
1096 /*
1097 * The destination pmd shouldn't be established, free_pgtables()
1098 * should have release it.
1099 */
1100 if (WARN_ON(!pmd_none(*new_pmd))) {
1101 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1102 goto out;
1103 }
1104
1105 ret = __pmd_trans_huge_lock(old_pmd, vma);
1106 if (ret == 1) {
1107 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1108 VM_BUG_ON(!pmd_none(*new_pmd));
1109 set_pmd_at(mm, new_addr, new_pmd, pmd);
1110 spin_unlock(&mm->page_table_lock);
1111 }
1112out:
1113 return ret;
1114}
1115
1116int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1117 unsigned long addr, pgprot_t newprot)
1118{
1119 struct mm_struct *mm = vma->vm_mm;
1120 int ret = 0;
1121
1122 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1123 pmd_t entry;
1124 entry = pmdp_get_and_clear(mm, addr, pmd);
1125 entry = pmd_modify(entry, newprot);
1126 set_pmd_at(mm, addr, pmd, entry);
1127 spin_unlock(&vma->vm_mm->page_table_lock);
1128 ret = 1;
1129 }
1130
1131 return ret;
1132}
1133
1134/*
1135 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1136 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1137 *
1138 * Note that if it returns 1, this routine returns without unlocking page
1139 * table locks. So callers must unlock them.
1140 */
1141int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1142{
1143 spin_lock(&vma->vm_mm->page_table_lock);
1144 if (likely(pmd_trans_huge(*pmd))) {
1145 if (unlikely(pmd_trans_splitting(*pmd))) {
1146 spin_unlock(&vma->vm_mm->page_table_lock);
1147 wait_split_huge_page(vma->anon_vma, pmd);
1148 return -1;
1149 } else {
1150 /* Thp mapped by 'pmd' is stable, so we can
1151 * handle it as it is. */
1152 return 1;
1153 }
1154 }
1155 spin_unlock(&vma->vm_mm->page_table_lock);
1156 return 0;
1157}
1158
1159pmd_t *page_check_address_pmd(struct page *page,
1160 struct mm_struct *mm,
1161 unsigned long address,
1162 enum page_check_address_pmd_flag flag)
1163{
1164 pgd_t *pgd;
1165 pud_t *pud;
1166 pmd_t *pmd, *ret = NULL;
1167
1168 if (address & ~HPAGE_PMD_MASK)
1169 goto out;
1170
1171 pgd = pgd_offset(mm, address);
1172 if (!pgd_present(*pgd))
1173 goto out;
1174
1175 pud = pud_offset(pgd, address);
1176 if (!pud_present(*pud))
1177 goto out;
1178
1179 pmd = pmd_offset(pud, address);
1180 if (pmd_none(*pmd))
1181 goto out;
1182 if (pmd_page(*pmd) != page)
1183 goto out;
1184 /*
1185 * split_vma() may create temporary aliased mappings. There is
1186 * no risk as long as all huge pmd are found and have their
1187 * splitting bit set before __split_huge_page_refcount
1188 * runs. Finding the same huge pmd more than once during the
1189 * same rmap walk is not a problem.
1190 */
1191 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1192 pmd_trans_splitting(*pmd))
1193 goto out;
1194 if (pmd_trans_huge(*pmd)) {
1195 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1196 !pmd_trans_splitting(*pmd));
1197 ret = pmd;
1198 }
1199out:
1200 return ret;
1201}
1202
1203static int __split_huge_page_splitting(struct page *page,
1204 struct vm_area_struct *vma,
1205 unsigned long address)
1206{
1207 struct mm_struct *mm = vma->vm_mm;
1208 pmd_t *pmd;
1209 int ret = 0;
1210
1211 spin_lock(&mm->page_table_lock);
1212 pmd = page_check_address_pmd(page, mm, address,
1213 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1214 if (pmd) {
1215 /*
1216 * We can't temporarily set the pmd to null in order
1217 * to split it, the pmd must remain marked huge at all
1218 * times or the VM won't take the pmd_trans_huge paths
1219 * and it won't wait on the anon_vma->root->mutex to
1220 * serialize against split_huge_page*.
1221 */
1222 pmdp_splitting_flush_notify(vma, address, pmd);
1223 ret = 1;
1224 }
1225 spin_unlock(&mm->page_table_lock);
1226
1227 return ret;
1228}
1229
1230static void __split_huge_page_refcount(struct page *page)
1231{
1232 int i;
1233 struct zone *zone = page_zone(page);
1234 struct lruvec *lruvec;
1235 int tail_count = 0;
1236
1237 /* prevent PageLRU to go away from under us, and freeze lru stats */
1238 spin_lock_irq(&zone->lru_lock);
1239 lruvec = mem_cgroup_page_lruvec(page, zone);
1240
1241 compound_lock(page);
1242 /* complete memcg works before add pages to LRU */
1243 mem_cgroup_split_huge_fixup(page);
1244
1245 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1246 struct page *page_tail = page + i;
1247
1248 /* tail_page->_mapcount cannot change */
1249 BUG_ON(page_mapcount(page_tail) < 0);
1250 tail_count += page_mapcount(page_tail);
1251 /* check for overflow */
1252 BUG_ON(tail_count < 0);
1253 BUG_ON(atomic_read(&page_tail->_count) != 0);
1254 /*
1255 * tail_page->_count is zero and not changing from
1256 * under us. But get_page_unless_zero() may be running
1257 * from under us on the tail_page. If we used
1258 * atomic_set() below instead of atomic_add(), we
1259 * would then run atomic_set() concurrently with
1260 * get_page_unless_zero(), and atomic_set() is
1261 * implemented in C not using locked ops. spin_unlock
1262 * on x86 sometime uses locked ops because of PPro
1263 * errata 66, 92, so unless somebody can guarantee
1264 * atomic_set() here would be safe on all archs (and
1265 * not only on x86), it's safer to use atomic_add().
1266 */
1267 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1268 &page_tail->_count);
1269
1270 /* after clearing PageTail the gup refcount can be released */
1271 smp_mb();
1272
1273 /*
1274 * retain hwpoison flag of the poisoned tail page:
1275 * fix for the unsuitable process killed on Guest Machine(KVM)
1276 * by the memory-failure.
1277 */
1278 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1279 page_tail->flags |= (page->flags &
1280 ((1L << PG_referenced) |
1281 (1L << PG_swapbacked) |
1282 (1L << PG_mlocked) |
1283 (1L << PG_uptodate)));
1284 page_tail->flags |= (1L << PG_dirty);
1285
1286 /* clear PageTail before overwriting first_page */
1287 smp_wmb();
1288
1289 /*
1290 * __split_huge_page_splitting() already set the
1291 * splitting bit in all pmd that could map this
1292 * hugepage, that will ensure no CPU can alter the
1293 * mapcount on the head page. The mapcount is only
1294 * accounted in the head page and it has to be
1295 * transferred to all tail pages in the below code. So
1296 * for this code to be safe, the split the mapcount
1297 * can't change. But that doesn't mean userland can't
1298 * keep changing and reading the page contents while
1299 * we transfer the mapcount, so the pmd splitting
1300 * status is achieved setting a reserved bit in the
1301 * pmd, not by clearing the present bit.
1302 */
1303 page_tail->_mapcount = page->_mapcount;
1304
1305 BUG_ON(page_tail->mapping);
1306 page_tail->mapping = page->mapping;
1307
1308 page_tail->index = page->index + i;
1309
1310 BUG_ON(!PageAnon(page_tail));
1311 BUG_ON(!PageUptodate(page_tail));
1312 BUG_ON(!PageDirty(page_tail));
1313 BUG_ON(!PageSwapBacked(page_tail));
1314
1315 lru_add_page_tail(page, page_tail, lruvec);
1316 }
1317 atomic_sub(tail_count, &page->_count);
1318 BUG_ON(atomic_read(&page->_count) <= 0);
1319
1320 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1321 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1322
1323 ClearPageCompound(page);
1324 compound_unlock(page);
1325 spin_unlock_irq(&zone->lru_lock);
1326
1327 for (i = 1; i < HPAGE_PMD_NR; i++) {
1328 struct page *page_tail = page + i;
1329 BUG_ON(page_count(page_tail) <= 0);
1330 /*
1331 * Tail pages may be freed if there wasn't any mapping
1332 * like if add_to_swap() is running on a lru page that
1333 * had its mapping zapped. And freeing these pages
1334 * requires taking the lru_lock so we do the put_page
1335 * of the tail pages after the split is complete.
1336 */
1337 put_page(page_tail);
1338 }
1339
1340 /*
1341 * Only the head page (now become a regular page) is required
1342 * to be pinned by the caller.
1343 */
1344 BUG_ON(page_count(page) <= 0);
1345}
1346
1347static int __split_huge_page_map(struct page *page,
1348 struct vm_area_struct *vma,
1349 unsigned long address)
1350{
1351 struct mm_struct *mm = vma->vm_mm;
1352 pmd_t *pmd, _pmd;
1353 int ret = 0, i;
1354 pgtable_t pgtable;
1355 unsigned long haddr;
1356
1357 spin_lock(&mm->page_table_lock);
1358 pmd = page_check_address_pmd(page, mm, address,
1359 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1360 if (pmd) {
1361 pgtable = get_pmd_huge_pte(mm);
1362 pmd_populate(mm, &_pmd, pgtable);
1363
1364 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1365 i++, haddr += PAGE_SIZE) {
1366 pte_t *pte, entry;
1367 BUG_ON(PageCompound(page+i));
1368 entry = mk_pte(page + i, vma->vm_page_prot);
1369 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1370 if (!pmd_write(*pmd))
1371 entry = pte_wrprotect(entry);
1372 else
1373 BUG_ON(page_mapcount(page) != 1);
1374 if (!pmd_young(*pmd))
1375 entry = pte_mkold(entry);
1376 pte = pte_offset_map(&_pmd, haddr);
1377 BUG_ON(!pte_none(*pte));
1378 set_pte_at(mm, haddr, pte, entry);
1379 pte_unmap(pte);
1380 }
1381
1382 smp_wmb(); /* make pte visible before pmd */
1383 /*
1384 * Up to this point the pmd is present and huge and
1385 * userland has the whole access to the hugepage
1386 * during the split (which happens in place). If we
1387 * overwrite the pmd with the not-huge version
1388 * pointing to the pte here (which of course we could
1389 * if all CPUs were bug free), userland could trigger
1390 * a small page size TLB miss on the small sized TLB
1391 * while the hugepage TLB entry is still established
1392 * in the huge TLB. Some CPU doesn't like that. See
1393 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1394 * Erratum 383 on page 93. Intel should be safe but is
1395 * also warns that it's only safe if the permission
1396 * and cache attributes of the two entries loaded in
1397 * the two TLB is identical (which should be the case
1398 * here). But it is generally safer to never allow
1399 * small and huge TLB entries for the same virtual
1400 * address to be loaded simultaneously. So instead of
1401 * doing "pmd_populate(); flush_tlb_range();" we first
1402 * mark the current pmd notpresent (atomically because
1403 * here the pmd_trans_huge and pmd_trans_splitting
1404 * must remain set at all times on the pmd until the
1405 * split is complete for this pmd), then we flush the
1406 * SMP TLB and finally we write the non-huge version
1407 * of the pmd entry with pmd_populate.
1408 */
1409 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1410 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1411 pmd_populate(mm, pmd, pgtable);
1412 ret = 1;
1413 }
1414 spin_unlock(&mm->page_table_lock);
1415
1416 return ret;
1417}
1418
1419/* must be called with anon_vma->root->mutex hold */
1420static void __split_huge_page(struct page *page,
1421 struct anon_vma *anon_vma)
1422{
1423 int mapcount, mapcount2;
1424 struct anon_vma_chain *avc;
1425
1426 BUG_ON(!PageHead(page));
1427 BUG_ON(PageTail(page));
1428
1429 mapcount = 0;
1430 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1431 struct vm_area_struct *vma = avc->vma;
1432 unsigned long addr = vma_address(page, vma);
1433 BUG_ON(is_vma_temporary_stack(vma));
1434 if (addr == -EFAULT)
1435 continue;
1436 mapcount += __split_huge_page_splitting(page, vma, addr);
1437 }
1438 /*
1439 * It is critical that new vmas are added to the tail of the
1440 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1441 * and establishes a child pmd before
1442 * __split_huge_page_splitting() freezes the parent pmd (so if
1443 * we fail to prevent copy_huge_pmd() from running until the
1444 * whole __split_huge_page() is complete), we will still see
1445 * the newly established pmd of the child later during the
1446 * walk, to be able to set it as pmd_trans_splitting too.
1447 */
1448 if (mapcount != page_mapcount(page))
1449 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1450 mapcount, page_mapcount(page));
1451 BUG_ON(mapcount != page_mapcount(page));
1452
1453 __split_huge_page_refcount(page);
1454
1455 mapcount2 = 0;
1456 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1457 struct vm_area_struct *vma = avc->vma;
1458 unsigned long addr = vma_address(page, vma);
1459 BUG_ON(is_vma_temporary_stack(vma));
1460 if (addr == -EFAULT)
1461 continue;
1462 mapcount2 += __split_huge_page_map(page, vma, addr);
1463 }
1464 if (mapcount != mapcount2)
1465 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1466 mapcount, mapcount2, page_mapcount(page));
1467 BUG_ON(mapcount != mapcount2);
1468}
1469
1470int split_huge_page(struct page *page)
1471{
1472 struct anon_vma *anon_vma;
1473 int ret = 1;
1474
1475 BUG_ON(!PageAnon(page));
1476 anon_vma = page_lock_anon_vma(page);
1477 if (!anon_vma)
1478 goto out;
1479 ret = 0;
1480 if (!PageCompound(page))
1481 goto out_unlock;
1482
1483 BUG_ON(!PageSwapBacked(page));
1484 __split_huge_page(page, anon_vma);
1485 count_vm_event(THP_SPLIT);
1486
1487 BUG_ON(PageCompound(page));
1488out_unlock:
1489 page_unlock_anon_vma(anon_vma);
1490out:
1491 return ret;
1492}
1493
1494#define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1495 VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1496
1497int hugepage_madvise(struct vm_area_struct *vma,
1498 unsigned long *vm_flags, int advice)
1499{
1500 switch (advice) {
1501 case MADV_HUGEPAGE:
1502 /*
1503 * Be somewhat over-protective like KSM for now!
1504 */
1505 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1506 return -EINVAL;
1507 *vm_flags &= ~VM_NOHUGEPAGE;
1508 *vm_flags |= VM_HUGEPAGE;
1509 /*
1510 * If the vma become good for khugepaged to scan,
1511 * register it here without waiting a page fault that
1512 * may not happen any time soon.
1513 */
1514 if (unlikely(khugepaged_enter_vma_merge(vma)))
1515 return -ENOMEM;
1516 break;
1517 case MADV_NOHUGEPAGE:
1518 /*
1519 * Be somewhat over-protective like KSM for now!
1520 */
1521 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1522 return -EINVAL;
1523 *vm_flags &= ~VM_HUGEPAGE;
1524 *vm_flags |= VM_NOHUGEPAGE;
1525 /*
1526 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1527 * this vma even if we leave the mm registered in khugepaged if
1528 * it got registered before VM_NOHUGEPAGE was set.
1529 */
1530 break;
1531 }
1532
1533 return 0;
1534}
1535
1536static int __init khugepaged_slab_init(void)
1537{
1538 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1539 sizeof(struct mm_slot),
1540 __alignof__(struct mm_slot), 0, NULL);
1541 if (!mm_slot_cache)
1542 return -ENOMEM;
1543
1544 return 0;
1545}
1546
1547static void __init khugepaged_slab_free(void)
1548{
1549 kmem_cache_destroy(mm_slot_cache);
1550 mm_slot_cache = NULL;
1551}
1552
1553static inline struct mm_slot *alloc_mm_slot(void)
1554{
1555 if (!mm_slot_cache) /* initialization failed */
1556 return NULL;
1557 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1558}
1559
1560static inline void free_mm_slot(struct mm_slot *mm_slot)
1561{
1562 kmem_cache_free(mm_slot_cache, mm_slot);
1563}
1564
1565static int __init mm_slots_hash_init(void)
1566{
1567 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1568 GFP_KERNEL);
1569 if (!mm_slots_hash)
1570 return -ENOMEM;
1571 return 0;
1572}
1573
1574#if 0
1575static void __init mm_slots_hash_free(void)
1576{
1577 kfree(mm_slots_hash);
1578 mm_slots_hash = NULL;
1579}
1580#endif
1581
1582static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1583{
1584 struct mm_slot *mm_slot;
1585 struct hlist_head *bucket;
1586 struct hlist_node *node;
1587
1588 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1589 % MM_SLOTS_HASH_HEADS];
1590 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1591 if (mm == mm_slot->mm)
1592 return mm_slot;
1593 }
1594 return NULL;
1595}
1596
1597static void insert_to_mm_slots_hash(struct mm_struct *mm,
1598 struct mm_slot *mm_slot)
1599{
1600 struct hlist_head *bucket;
1601
1602 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1603 % MM_SLOTS_HASH_HEADS];
1604 mm_slot->mm = mm;
1605 hlist_add_head(&mm_slot->hash, bucket);
1606}
1607
1608static inline int khugepaged_test_exit(struct mm_struct *mm)
1609{
1610 return atomic_read(&mm->mm_users) == 0;
1611}
1612
1613int __khugepaged_enter(struct mm_struct *mm)
1614{
1615 struct mm_slot *mm_slot;
1616 int wakeup;
1617
1618 mm_slot = alloc_mm_slot();
1619 if (!mm_slot)
1620 return -ENOMEM;
1621
1622 /* __khugepaged_exit() must not run from under us */
1623 VM_BUG_ON(khugepaged_test_exit(mm));
1624 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1625 free_mm_slot(mm_slot);
1626 return 0;
1627 }
1628
1629 spin_lock(&khugepaged_mm_lock);
1630 insert_to_mm_slots_hash(mm, mm_slot);
1631 /*
1632 * Insert just behind the scanning cursor, to let the area settle
1633 * down a little.
1634 */
1635 wakeup = list_empty(&khugepaged_scan.mm_head);
1636 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1637 spin_unlock(&khugepaged_mm_lock);
1638
1639 atomic_inc(&mm->mm_count);
1640 if (wakeup)
1641 wake_up_interruptible(&khugepaged_wait);
1642
1643 return 0;
1644}
1645
1646int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1647{
1648 unsigned long hstart, hend;
1649 if (!vma->anon_vma)
1650 /*
1651 * Not yet faulted in so we will register later in the
1652 * page fault if needed.
1653 */
1654 return 0;
1655 if (vma->vm_ops)
1656 /* khugepaged not yet working on file or special mappings */
1657 return 0;
1658 /*
1659 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1660 * true too, verify it here.
1661 */
1662 VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1663 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1664 hend = vma->vm_end & HPAGE_PMD_MASK;
1665 if (hstart < hend)
1666 return khugepaged_enter(vma);
1667 return 0;
1668}
1669
1670void __khugepaged_exit(struct mm_struct *mm)
1671{
1672 struct mm_slot *mm_slot;
1673 int free = 0;
1674
1675 spin_lock(&khugepaged_mm_lock);
1676 mm_slot = get_mm_slot(mm);
1677 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1678 hlist_del(&mm_slot->hash);
1679 list_del(&mm_slot->mm_node);
1680 free = 1;
1681 }
1682 spin_unlock(&khugepaged_mm_lock);
1683
1684 if (free) {
1685 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1686 free_mm_slot(mm_slot);
1687 mmdrop(mm);
1688 } else if (mm_slot) {
1689 /*
1690 * This is required to serialize against
1691 * khugepaged_test_exit() (which is guaranteed to run
1692 * under mmap sem read mode). Stop here (after we
1693 * return all pagetables will be destroyed) until
1694 * khugepaged has finished working on the pagetables
1695 * under the mmap_sem.
1696 */
1697 down_write(&mm->mmap_sem);
1698 up_write(&mm->mmap_sem);
1699 }
1700}
1701
1702static void release_pte_page(struct page *page)
1703{
1704 /* 0 stands for page_is_file_cache(page) == false */
1705 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1706 unlock_page(page);
1707 putback_lru_page(page);
1708}
1709
1710static void release_pte_pages(pte_t *pte, pte_t *_pte)
1711{
1712 while (--_pte >= pte) {
1713 pte_t pteval = *_pte;
1714 if (!pte_none(pteval))
1715 release_pte_page(pte_page(pteval));
1716 }
1717}
1718
1719static void release_all_pte_pages(pte_t *pte)
1720{
1721 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1722}
1723
1724static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1725 unsigned long address,
1726 pte_t *pte)
1727{
1728 struct page *page;
1729 pte_t *_pte;
1730 int referenced = 0, isolated = 0, none = 0;
1731 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1732 _pte++, address += PAGE_SIZE) {
1733 pte_t pteval = *_pte;
1734 if (pte_none(pteval)) {
1735 if (++none <= khugepaged_max_ptes_none)
1736 continue;
1737 else {
1738 release_pte_pages(pte, _pte);
1739 goto out;
1740 }
1741 }
1742 if (!pte_present(pteval) || !pte_write(pteval)) {
1743 release_pte_pages(pte, _pte);
1744 goto out;
1745 }
1746 page = vm_normal_page(vma, address, pteval);
1747 if (unlikely(!page)) {
1748 release_pte_pages(pte, _pte);
1749 goto out;
1750 }
1751 VM_BUG_ON(PageCompound(page));
1752 BUG_ON(!PageAnon(page));
1753 VM_BUG_ON(!PageSwapBacked(page));
1754
1755 /* cannot use mapcount: can't collapse if there's a gup pin */
1756 if (page_count(page) != 1) {
1757 release_pte_pages(pte, _pte);
1758 goto out;
1759 }
1760 /*
1761 * We can do it before isolate_lru_page because the
1762 * page can't be freed from under us. NOTE: PG_lock
1763 * is needed to serialize against split_huge_page
1764 * when invoked from the VM.
1765 */
1766 if (!trylock_page(page)) {
1767 release_pte_pages(pte, _pte);
1768 goto out;
1769 }
1770 /*
1771 * Isolate the page to avoid collapsing an hugepage
1772 * currently in use by the VM.
1773 */
1774 if (isolate_lru_page(page)) {
1775 unlock_page(page);
1776 release_pte_pages(pte, _pte);
1777 goto out;
1778 }
1779 /* 0 stands for page_is_file_cache(page) == false */
1780 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1781 VM_BUG_ON(!PageLocked(page));
1782 VM_BUG_ON(PageLRU(page));
1783
1784 /* If there is no mapped pte young don't collapse the page */
1785 if (pte_young(pteval) || PageReferenced(page) ||
1786 mmu_notifier_test_young(vma->vm_mm, address))
1787 referenced = 1;
1788 }
1789 if (unlikely(!referenced))
1790 release_all_pte_pages(pte);
1791 else
1792 isolated = 1;
1793out:
1794 return isolated;
1795}
1796
1797static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1798 struct vm_area_struct *vma,
1799 unsigned long address,
1800 spinlock_t *ptl)
1801{
1802 pte_t *_pte;
1803 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1804 pte_t pteval = *_pte;
1805 struct page *src_page;
1806
1807 if (pte_none(pteval)) {
1808 clear_user_highpage(page, address);
1809 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1810 } else {
1811 src_page = pte_page(pteval);
1812 copy_user_highpage(page, src_page, address, vma);
1813 VM_BUG_ON(page_mapcount(src_page) != 1);
1814 VM_BUG_ON(page_count(src_page) != 2);
1815 release_pte_page(src_page);
1816 /*
1817 * ptl mostly unnecessary, but preempt has to
1818 * be disabled to update the per-cpu stats
1819 * inside page_remove_rmap().
1820 */
1821 spin_lock(ptl);
1822 /*
1823 * paravirt calls inside pte_clear here are
1824 * superfluous.
1825 */
1826 pte_clear(vma->vm_mm, address, _pte);
1827 page_remove_rmap(src_page);
1828 spin_unlock(ptl);
1829 free_page_and_swap_cache(src_page);
1830 }
1831
1832 address += PAGE_SIZE;
1833 page++;
1834 }
1835}
1836
1837static void collapse_huge_page(struct mm_struct *mm,
1838 unsigned long address,
1839 struct page **hpage,
1840 struct vm_area_struct *vma,
1841 int node)
1842{
1843 pgd_t *pgd;
1844 pud_t *pud;
1845 pmd_t *pmd, _pmd;
1846 pte_t *pte;
1847 pgtable_t pgtable;
1848 struct page *new_page;
1849 spinlock_t *ptl;
1850 int isolated;
1851 unsigned long hstart, hend;
1852
1853 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1854#ifndef CONFIG_NUMA
1855 up_read(&mm->mmap_sem);
1856 VM_BUG_ON(!*hpage);
1857 new_page = *hpage;
1858#else
1859 VM_BUG_ON(*hpage);
1860 /*
1861 * Allocate the page while the vma is still valid and under
1862 * the mmap_sem read mode so there is no memory allocation
1863 * later when we take the mmap_sem in write mode. This is more
1864 * friendly behavior (OTOH it may actually hide bugs) to
1865 * filesystems in userland with daemons allocating memory in
1866 * the userland I/O paths. Allocating memory with the
1867 * mmap_sem in read mode is good idea also to allow greater
1868 * scalability.
1869 */
1870 new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1871 node, __GFP_OTHER_NODE);
1872
1873 /*
1874 * After allocating the hugepage, release the mmap_sem read lock in
1875 * preparation for taking it in write mode.
1876 */
1877 up_read(&mm->mmap_sem);
1878 if (unlikely(!new_page)) {
1879 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1880 *hpage = ERR_PTR(-ENOMEM);
1881 return;
1882 }
1883#endif
1884
1885 count_vm_event(THP_COLLAPSE_ALLOC);
1886 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1887#ifdef CONFIG_NUMA
1888 put_page(new_page);
1889#endif
1890 return;
1891 }
1892
1893 /*
1894 * Prevent all access to pagetables with the exception of
1895 * gup_fast later hanlded by the ptep_clear_flush and the VM
1896 * handled by the anon_vma lock + PG_lock.
1897 */
1898 down_write(&mm->mmap_sem);
1899 if (unlikely(khugepaged_test_exit(mm)))
1900 goto out;
1901
1902 vma = find_vma(mm, address);
1903 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1904 hend = vma->vm_end & HPAGE_PMD_MASK;
1905 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1906 goto out;
1907
1908 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1909 (vma->vm_flags & VM_NOHUGEPAGE))
1910 goto out;
1911
1912 if (!vma->anon_vma || vma->vm_ops)
1913 goto out;
1914 if (is_vma_temporary_stack(vma))
1915 goto out;
1916 /*
1917 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1918 * true too, verify it here.
1919 */
1920 VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1921
1922 pgd = pgd_offset(mm, address);
1923 if (!pgd_present(*pgd))
1924 goto out;
1925
1926 pud = pud_offset(pgd, address);
1927 if (!pud_present(*pud))
1928 goto out;
1929
1930 pmd = pmd_offset(pud, address);
1931 /* pmd can't go away or become huge under us */
1932 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1933 goto out;
1934
1935 anon_vma_lock(vma->anon_vma);
1936
1937 pte = pte_offset_map(pmd, address);
1938 ptl = pte_lockptr(mm, pmd);
1939
1940 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1941 /*
1942 * After this gup_fast can't run anymore. This also removes
1943 * any huge TLB entry from the CPU so we won't allow
1944 * huge and small TLB entries for the same virtual address
1945 * to avoid the risk of CPU bugs in that area.
1946 */
1947 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1948 spin_unlock(&mm->page_table_lock);
1949
1950 spin_lock(ptl);
1951 isolated = __collapse_huge_page_isolate(vma, address, pte);
1952 spin_unlock(ptl);
1953
1954 if (unlikely(!isolated)) {
1955 pte_unmap(pte);
1956 spin_lock(&mm->page_table_lock);
1957 BUG_ON(!pmd_none(*pmd));
1958 set_pmd_at(mm, address, pmd, _pmd);
1959 spin_unlock(&mm->page_table_lock);
1960 anon_vma_unlock(vma->anon_vma);
1961 goto out;
1962 }
1963
1964 /*
1965 * All pages are isolated and locked so anon_vma rmap
1966 * can't run anymore.
1967 */
1968 anon_vma_unlock(vma->anon_vma);
1969
1970 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1971 pte_unmap(pte);
1972 __SetPageUptodate(new_page);
1973 pgtable = pmd_pgtable(_pmd);
1974 VM_BUG_ON(page_count(pgtable) != 1);
1975 VM_BUG_ON(page_mapcount(pgtable) != 0);
1976
1977 _pmd = mk_pmd(new_page, vma->vm_page_prot);
1978 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1979 _pmd = pmd_mkhuge(_pmd);
1980
1981 /*
1982 * spin_lock() below is not the equivalent of smp_wmb(), so
1983 * this is needed to avoid the copy_huge_page writes to become
1984 * visible after the set_pmd_at() write.
1985 */
1986 smp_wmb();
1987
1988 spin_lock(&mm->page_table_lock);
1989 BUG_ON(!pmd_none(*pmd));
1990 page_add_new_anon_rmap(new_page, vma, address);
1991 set_pmd_at(mm, address, pmd, _pmd);
1992 update_mmu_cache(vma, address, _pmd);
1993 prepare_pmd_huge_pte(pgtable, mm);
1994 spin_unlock(&mm->page_table_lock);
1995
1996#ifndef CONFIG_NUMA
1997 *hpage = NULL;
1998#endif
1999 khugepaged_pages_collapsed++;
2000out_up_write:
2001 up_write(&mm->mmap_sem);
2002 return;
2003
2004out:
2005 mem_cgroup_uncharge_page(new_page);
2006#ifdef CONFIG_NUMA
2007 put_page(new_page);
2008#endif
2009 goto out_up_write;
2010}
2011
2012static int khugepaged_scan_pmd(struct mm_struct *mm,
2013 struct vm_area_struct *vma,
2014 unsigned long address,
2015 struct page **hpage)
2016{
2017 pgd_t *pgd;
2018 pud_t *pud;
2019 pmd_t *pmd;
2020 pte_t *pte, *_pte;
2021 int ret = 0, referenced = 0, none = 0;
2022 struct page *page;
2023 unsigned long _address;
2024 spinlock_t *ptl;
2025 int node = -1;
2026
2027 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2028
2029 pgd = pgd_offset(mm, address);
2030 if (!pgd_present(*pgd))
2031 goto out;
2032
2033 pud = pud_offset(pgd, address);
2034 if (!pud_present(*pud))
2035 goto out;
2036
2037 pmd = pmd_offset(pud, address);
2038 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2039 goto out;
2040
2041 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2042 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2043 _pte++, _address += PAGE_SIZE) {
2044 pte_t pteval = *_pte;
2045 if (pte_none(pteval)) {
2046 if (++none <= khugepaged_max_ptes_none)
2047 continue;
2048 else
2049 goto out_unmap;
2050 }
2051 if (!pte_present(pteval) || !pte_write(pteval))
2052 goto out_unmap;
2053 page = vm_normal_page(vma, _address, pteval);
2054 if (unlikely(!page))
2055 goto out_unmap;
2056 /*
2057 * Chose the node of the first page. This could
2058 * be more sophisticated and look at more pages,
2059 * but isn't for now.
2060 */
2061 if (node == -1)
2062 node = page_to_nid(page);
2063 VM_BUG_ON(PageCompound(page));
2064 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2065 goto out_unmap;
2066 /* cannot use mapcount: can't collapse if there's a gup pin */
2067 if (page_count(page) != 1)
2068 goto out_unmap;
2069 if (pte_young(pteval) || PageReferenced(page) ||
2070 mmu_notifier_test_young(vma->vm_mm, address))
2071 referenced = 1;
2072 }
2073 if (referenced)
2074 ret = 1;
2075out_unmap:
2076 pte_unmap_unlock(pte, ptl);
2077 if (ret)
2078 /* collapse_huge_page will return with the mmap_sem released */
2079 collapse_huge_page(mm, address, hpage, vma, node);
2080out:
2081 return ret;
2082}
2083
2084static void collect_mm_slot(struct mm_slot *mm_slot)
2085{
2086 struct mm_struct *mm = mm_slot->mm;
2087
2088 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2089
2090 if (khugepaged_test_exit(mm)) {
2091 /* free mm_slot */
2092 hlist_del(&mm_slot->hash);
2093 list_del(&mm_slot->mm_node);
2094
2095 /*
2096 * Not strictly needed because the mm exited already.
2097 *
2098 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2099 */
2100
2101 /* khugepaged_mm_lock actually not necessary for the below */
2102 free_mm_slot(mm_slot);
2103 mmdrop(mm);
2104 }
2105}
2106
2107static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2108 struct page **hpage)
2109 __releases(&khugepaged_mm_lock)
2110 __acquires(&khugepaged_mm_lock)
2111{
2112 struct mm_slot *mm_slot;
2113 struct mm_struct *mm;
2114 struct vm_area_struct *vma;
2115 int progress = 0;
2116
2117 VM_BUG_ON(!pages);
2118 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2119
2120 if (khugepaged_scan.mm_slot)
2121 mm_slot = khugepaged_scan.mm_slot;
2122 else {
2123 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2124 struct mm_slot, mm_node);
2125 khugepaged_scan.address = 0;
2126 khugepaged_scan.mm_slot = mm_slot;
2127 }
2128 spin_unlock(&khugepaged_mm_lock);
2129
2130 mm = mm_slot->mm;
2131 down_read(&mm->mmap_sem);
2132 if (unlikely(khugepaged_test_exit(mm)))
2133 vma = NULL;
2134 else
2135 vma = find_vma(mm, khugepaged_scan.address);
2136
2137 progress++;
2138 for (; vma; vma = vma->vm_next) {
2139 unsigned long hstart, hend;
2140
2141 cond_resched();
2142 if (unlikely(khugepaged_test_exit(mm))) {
2143 progress++;
2144 break;
2145 }
2146
2147 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2148 !khugepaged_always()) ||
2149 (vma->vm_flags & VM_NOHUGEPAGE)) {
2150 skip:
2151 progress++;
2152 continue;
2153 }
2154 if (!vma->anon_vma || vma->vm_ops)
2155 goto skip;
2156 if (is_vma_temporary_stack(vma))
2157 goto skip;
2158 /*
2159 * If is_pfn_mapping() is true is_learn_pfn_mapping()
2160 * must be true too, verify it here.
2161 */
2162 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2163 vma->vm_flags & VM_NO_THP);
2164
2165 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2166 hend = vma->vm_end & HPAGE_PMD_MASK;
2167 if (hstart >= hend)
2168 goto skip;
2169 if (khugepaged_scan.address > hend)
2170 goto skip;
2171 if (khugepaged_scan.address < hstart)
2172 khugepaged_scan.address = hstart;
2173 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2174
2175 while (khugepaged_scan.address < hend) {
2176 int ret;
2177 cond_resched();
2178 if (unlikely(khugepaged_test_exit(mm)))
2179 goto breakouterloop;
2180
2181 VM_BUG_ON(khugepaged_scan.address < hstart ||
2182 khugepaged_scan.address + HPAGE_PMD_SIZE >
2183 hend);
2184 ret = khugepaged_scan_pmd(mm, vma,
2185 khugepaged_scan.address,
2186 hpage);
2187 /* move to next address */
2188 khugepaged_scan.address += HPAGE_PMD_SIZE;
2189 progress += HPAGE_PMD_NR;
2190 if (ret)
2191 /* we released mmap_sem so break loop */
2192 goto breakouterloop_mmap_sem;
2193 if (progress >= pages)
2194 goto breakouterloop;
2195 }
2196 }
2197breakouterloop:
2198 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2199breakouterloop_mmap_sem:
2200
2201 spin_lock(&khugepaged_mm_lock);
2202 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2203 /*
2204 * Release the current mm_slot if this mm is about to die, or
2205 * if we scanned all vmas of this mm.
2206 */
2207 if (khugepaged_test_exit(mm) || !vma) {
2208 /*
2209 * Make sure that if mm_users is reaching zero while
2210 * khugepaged runs here, khugepaged_exit will find
2211 * mm_slot not pointing to the exiting mm.
2212 */
2213 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2214 khugepaged_scan.mm_slot = list_entry(
2215 mm_slot->mm_node.next,
2216 struct mm_slot, mm_node);
2217 khugepaged_scan.address = 0;
2218 } else {
2219 khugepaged_scan.mm_slot = NULL;
2220 khugepaged_full_scans++;
2221 }
2222
2223 collect_mm_slot(mm_slot);
2224 }
2225
2226 return progress;
2227}
2228
2229static int khugepaged_has_work(void)
2230{
2231 return !list_empty(&khugepaged_scan.mm_head) &&
2232 khugepaged_enabled();
2233}
2234
2235static int khugepaged_wait_event(void)
2236{
2237 return !list_empty(&khugepaged_scan.mm_head) ||
2238 !khugepaged_enabled();
2239}
2240
2241static void khugepaged_do_scan(struct page **hpage)
2242{
2243 unsigned int progress = 0, pass_through_head = 0;
2244 unsigned int pages = khugepaged_pages_to_scan;
2245
2246 barrier(); /* write khugepaged_pages_to_scan to local stack */
2247
2248 while (progress < pages) {
2249 cond_resched();
2250
2251#ifndef CONFIG_NUMA
2252 if (!*hpage) {
2253 *hpage = alloc_hugepage(khugepaged_defrag());
2254 if (unlikely(!*hpage)) {
2255 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2256 break;
2257 }
2258 count_vm_event(THP_COLLAPSE_ALLOC);
2259 }
2260#else
2261 if (IS_ERR(*hpage))
2262 break;
2263#endif
2264
2265 if (unlikely(kthread_should_stop() || freezing(current)))
2266 break;
2267
2268 spin_lock(&khugepaged_mm_lock);
2269 if (!khugepaged_scan.mm_slot)
2270 pass_through_head++;
2271 if (khugepaged_has_work() &&
2272 pass_through_head < 2)
2273 progress += khugepaged_scan_mm_slot(pages - progress,
2274 hpage);
2275 else
2276 progress = pages;
2277 spin_unlock(&khugepaged_mm_lock);
2278 }
2279}
2280
2281static void khugepaged_alloc_sleep(void)
2282{
2283 wait_event_freezable_timeout(khugepaged_wait, false,
2284 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2285}
2286
2287#ifndef CONFIG_NUMA
2288static struct page *khugepaged_alloc_hugepage(void)
2289{
2290 struct page *hpage;
2291
2292 do {
2293 hpage = alloc_hugepage(khugepaged_defrag());
2294 if (!hpage) {
2295 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2296 khugepaged_alloc_sleep();
2297 } else
2298 count_vm_event(THP_COLLAPSE_ALLOC);
2299 } while (unlikely(!hpage) &&
2300 likely(khugepaged_enabled()));
2301 return hpage;
2302}
2303#endif
2304
2305static void khugepaged_loop(void)
2306{
2307 struct page *hpage;
2308
2309#ifdef CONFIG_NUMA
2310 hpage = NULL;
2311#endif
2312 while (likely(khugepaged_enabled())) {
2313#ifndef CONFIG_NUMA
2314 hpage = khugepaged_alloc_hugepage();
2315 if (unlikely(!hpage))
2316 break;
2317#else
2318 if (IS_ERR(hpage)) {
2319 khugepaged_alloc_sleep();
2320 hpage = NULL;
2321 }
2322#endif
2323
2324 khugepaged_do_scan(&hpage);
2325#ifndef CONFIG_NUMA
2326 if (hpage)
2327 put_page(hpage);
2328#endif
2329 try_to_freeze();
2330 if (unlikely(kthread_should_stop()))
2331 break;
2332 if (khugepaged_has_work()) {
2333 if (!khugepaged_scan_sleep_millisecs)
2334 continue;
2335 wait_event_freezable_timeout(khugepaged_wait, false,
2336 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2337 } else if (khugepaged_enabled())
2338 wait_event_freezable(khugepaged_wait,
2339 khugepaged_wait_event());
2340 }
2341}
2342
2343static int khugepaged(void *none)
2344{
2345 struct mm_slot *mm_slot;
2346
2347 set_freezable();
2348 set_user_nice(current, 19);
2349
2350 /* serialize with start_khugepaged() */
2351 mutex_lock(&khugepaged_mutex);
2352
2353 for (;;) {
2354 mutex_unlock(&khugepaged_mutex);
2355 VM_BUG_ON(khugepaged_thread != current);
2356 khugepaged_loop();
2357 VM_BUG_ON(khugepaged_thread != current);
2358
2359 mutex_lock(&khugepaged_mutex);
2360 if (!khugepaged_enabled())
2361 break;
2362 if (unlikely(kthread_should_stop()))
2363 break;
2364 }
2365
2366 spin_lock(&khugepaged_mm_lock);
2367 mm_slot = khugepaged_scan.mm_slot;
2368 khugepaged_scan.mm_slot = NULL;
2369 if (mm_slot)
2370 collect_mm_slot(mm_slot);
2371 spin_unlock(&khugepaged_mm_lock);
2372
2373 khugepaged_thread = NULL;
2374 mutex_unlock(&khugepaged_mutex);
2375
2376 return 0;
2377}
2378
2379void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2380{
2381 struct page *page;
2382
2383 spin_lock(&mm->page_table_lock);
2384 if (unlikely(!pmd_trans_huge(*pmd))) {
2385 spin_unlock(&mm->page_table_lock);
2386 return;
2387 }
2388 page = pmd_page(*pmd);
2389 VM_BUG_ON(!page_count(page));
2390 get_page(page);
2391 spin_unlock(&mm->page_table_lock);
2392
2393 split_huge_page(page);
2394
2395 put_page(page);
2396 BUG_ON(pmd_trans_huge(*pmd));
2397}
2398
2399static void split_huge_page_address(struct mm_struct *mm,
2400 unsigned long address)
2401{
2402 pgd_t *pgd;
2403 pud_t *pud;
2404 pmd_t *pmd;
2405
2406 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2407
2408 pgd = pgd_offset(mm, address);
2409 if (!pgd_present(*pgd))
2410 return;
2411
2412 pud = pud_offset(pgd, address);
2413 if (!pud_present(*pud))
2414 return;
2415
2416 pmd = pmd_offset(pud, address);
2417 if (!pmd_present(*pmd))
2418 return;
2419 /*
2420 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2421 * materialize from under us.
2422 */
2423 split_huge_page_pmd(mm, pmd);
2424}
2425
2426void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2427 unsigned long start,
2428 unsigned long end,
2429 long adjust_next)
2430{
2431 /*
2432 * If the new start address isn't hpage aligned and it could
2433 * previously contain an hugepage: check if we need to split
2434 * an huge pmd.
2435 */
2436 if (start & ~HPAGE_PMD_MASK &&
2437 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2438 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2439 split_huge_page_address(vma->vm_mm, start);
2440
2441 /*
2442 * If the new end address isn't hpage aligned and it could
2443 * previously contain an hugepage: check if we need to split
2444 * an huge pmd.
2445 */
2446 if (end & ~HPAGE_PMD_MASK &&
2447 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2448 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2449 split_huge_page_address(vma->vm_mm, end);
2450
2451 /*
2452 * If we're also updating the vma->vm_next->vm_start, if the new
2453 * vm_next->vm_start isn't page aligned and it could previously
2454 * contain an hugepage: check if we need to split an huge pmd.
2455 */
2456 if (adjust_next > 0) {
2457 struct vm_area_struct *next = vma->vm_next;
2458 unsigned long nstart = next->vm_start;
2459 nstart += adjust_next << PAGE_SHIFT;
2460 if (nstart & ~HPAGE_PMD_MASK &&
2461 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2462 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2463 split_huge_page_address(next->vm_mm, nstart);
2464 }
2465}