?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 5.4.41
[linux/fpc-iii.git] / mm / huge_memory.c
blob0d96831b6dedb1508d1f2bc024ae2da0e52592b8
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2009 Red Hat, Inc.
4 */
5
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/sched/coredump.h>
11 #include <linux/sched/numa_balancing.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/pfn_t.h>
24 #include <linux/mman.h>
25 #include <linux/memremap.h>
26 #include <linux/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/oom.h>
34 #include <linux/numa.h>
35 #include <linux/page_owner.h>
36
37 #include <asm/tlb.h>
38 #include <asm/pgalloc.h>
39 #include "internal.h"
40
41 /*
42 * By default, transparent hugepage support is disabled in order to avoid
43 * risking an increased memory footprint for applications that are not
44 * guaranteed to benefit from it. When transparent hugepage support is
45 * enabled, it is for all mappings, and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
48 */
49 unsigned 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
60 static struct shrinker deferred_split_shrinker;
61
62 static atomic_t huge_zero_refcount;
63 struct page *huge_zero_page __read_mostly;
64
65 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
66 {
67 /* The addr is used to check if the vma size fits */
68 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
69
70 if (!transhuge_vma_suitable(vma, addr))
71 return false;
72 if (vma_is_anonymous(vma))
73 return __transparent_hugepage_enabled(vma);
74 if (vma_is_shmem(vma))
75 return shmem_huge_enabled(vma);
76
77 return false;
78 }
79
80 static struct page *get_huge_zero_page(void)
81 {
82 struct page *zero_page;
83 retry:
84 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
85 return READ_ONCE(huge_zero_page);
86
87 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
88 HPAGE_PMD_ORDER);
89 if (!zero_page) {
90 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
91 return NULL;
92 }
93 count_vm_event(THP_ZERO_PAGE_ALLOC);
94 preempt_disable();
95 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
96 preempt_enable();
97 __free_pages(zero_page, compound_order(zero_page));
98 goto retry;
99 }
100
101 /* We take additional reference here. It will be put back by shrinker */
102 atomic_set(&huge_zero_refcount, 2);
103 preempt_enable();
104 return READ_ONCE(huge_zero_page);
105 }
106
107 static void put_huge_zero_page(void)
108 {
109 /*
110 * Counter should never go to zero here. Only shrinker can put
111 * last reference.
112 */
113 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
114 }
115
116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
117 {
118 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
119 return READ_ONCE(huge_zero_page);
120
121 if (!get_huge_zero_page())
122 return NULL;
123
124 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
125 put_huge_zero_page();
126
127 return READ_ONCE(huge_zero_page);
128 }
129
130 void mm_put_huge_zero_page(struct mm_struct *mm)
131 {
132 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
133 put_huge_zero_page();
134 }
135
136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
137 struct shrink_control *sc)
138 {
139 /* we can free zero page only if last reference remains */
140 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
141 }
142
143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
144 struct shrink_control *sc)
145 {
146 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
147 struct page *zero_page = xchg(&huge_zero_page, NULL);
148 BUG_ON(zero_page == NULL);
149 __free_pages(zero_page, compound_order(zero_page));
150 return HPAGE_PMD_NR;
151 }
152
153 return 0;
154 }
155
156 static struct shrinker huge_zero_page_shrinker = {
157 .count_objects = shrink_huge_zero_page_count,
158 .scan_objects = shrink_huge_zero_page_scan,
159 .seeks = DEFAULT_SEEKS,
160 };
161
162 #ifdef CONFIG_SYSFS
163 static ssize_t enabled_show(struct kobject *kobj,
164 struct kobj_attribute *attr, char *buf)
165 {
166 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
167 return sprintf(buf, "[always] madvise never\n");
168 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
169 return sprintf(buf, "always [madvise] never\n");
170 else
171 return sprintf(buf, "always madvise [never]\n");
172 }
173
174 static ssize_t enabled_store(struct kobject *kobj,
175 struct kobj_attribute *attr,
176 const char *buf, size_t count)
177 {
178 ssize_t ret = count;
179
180 if (sysfs_streq(buf, "always")) {
181 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
182 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
183 } else if (sysfs_streq(buf, "madvise")) {
184 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
185 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
186 } else if (sysfs_streq(buf, "never")) {
187 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
188 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
189 } else
190 ret = -EINVAL;
191
192 if (ret > 0) {
193 int err = start_stop_khugepaged();
194 if (err)
195 ret = err;
196 }
197 return ret;
198 }
199 static struct kobj_attribute enabled_attr =
200 __ATTR(enabled, 0644, enabled_show, enabled_store);
201
202 ssize_t single_hugepage_flag_show(struct kobject *kobj,
203 struct kobj_attribute *attr, char *buf,
204 enum transparent_hugepage_flag flag)
205 {
206 return sprintf(buf, "%d\n",
207 !!test_bit(flag, &transparent_hugepage_flags));
208 }
209
210 ssize_t single_hugepage_flag_store(struct kobject *kobj,
211 struct kobj_attribute *attr,
212 const char *buf, size_t count,
213 enum transparent_hugepage_flag flag)
214 {
215 unsigned long value;
216 int ret;
217
218 ret = kstrtoul(buf, 10, &value);
219 if (ret < 0)
220 return ret;
221 if (value > 1)
222 return -EINVAL;
223
224 if (value)
225 set_bit(flag, &transparent_hugepage_flags);
226 else
227 clear_bit(flag, &transparent_hugepage_flags);
228
229 return count;
230 }
231
232 static ssize_t defrag_show(struct kobject *kobj,
233 struct kobj_attribute *attr, char *buf)
234 {
235 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
236 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
237 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
238 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
239 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
240 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
241 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
242 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
243 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
244 }
245
246 static ssize_t defrag_store(struct kobject *kobj,
247 struct kobj_attribute *attr,
248 const char *buf, size_t count)
249 {
250 if (sysfs_streq(buf, "always")) {
251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
255 } else if (sysfs_streq(buf, "defer+madvise")) {
256 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
259 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
260 } else if (sysfs_streq(buf, "defer")) {
261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
264 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
265 } else if (sysfs_streq(buf, "madvise")) {
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
269 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
270 } else if (sysfs_streq(buf, "never")) {
271 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
272 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
273 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
275 } else
276 return -EINVAL;
277
278 return count;
279 }
280 static struct kobj_attribute defrag_attr =
281 __ATTR(defrag, 0644, defrag_show, defrag_store);
282
283 static ssize_t use_zero_page_show(struct kobject *kobj,
284 struct kobj_attribute *attr, char *buf)
285 {
286 return single_hugepage_flag_show(kobj, attr, buf,
287 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
288 }
289 static ssize_t use_zero_page_store(struct kobject *kobj,
290 struct kobj_attribute *attr, const char *buf, size_t count)
291 {
292 return single_hugepage_flag_store(kobj, attr, buf, count,
293 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
294 }
295 static struct kobj_attribute use_zero_page_attr =
296 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
297
298 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
299 struct kobj_attribute *attr, char *buf)
300 {
301 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
302 }
303 static struct kobj_attribute hpage_pmd_size_attr =
304 __ATTR_RO(hpage_pmd_size);
305
306 #ifdef CONFIG_DEBUG_VM
307 static ssize_t debug_cow_show(struct kobject *kobj,
308 struct kobj_attribute *attr, char *buf)
309 {
310 return single_hugepage_flag_show(kobj, attr, buf,
311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312 }
313 static ssize_t debug_cow_store(struct kobject *kobj,
314 struct kobj_attribute *attr,
315 const char *buf, size_t count)
316 {
317 return single_hugepage_flag_store(kobj, attr, buf, count,
318 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
319 }
320 static struct kobj_attribute debug_cow_attr =
321 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
322 #endif /* CONFIG_DEBUG_VM */
323
324 static struct attribute *hugepage_attr[] = {
325 &enabled_attr.attr,
326 &defrag_attr.attr,
327 &use_zero_page_attr.attr,
328 &hpage_pmd_size_attr.attr,
329 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
330 &shmem_enabled_attr.attr,
331 #endif
332 #ifdef CONFIG_DEBUG_VM
333 &debug_cow_attr.attr,
334 #endif
335 NULL,
336 };
337
338 static const struct attribute_group hugepage_attr_group = {
339 .attrs = hugepage_attr,
340 };
341
342 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
343 {
344 int err;
345
346 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
347 if (unlikely(!*hugepage_kobj)) {
348 pr_err("failed to create transparent hugepage kobject\n");
349 return -ENOMEM;
350 }
351
352 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
353 if (err) {
354 pr_err("failed to register transparent hugepage group\n");
355 goto delete_obj;
356 }
357
358 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
359 if (err) {
360 pr_err("failed to register transparent hugepage group\n");
361 goto remove_hp_group;
362 }
363
364 return 0;
365
366 remove_hp_group:
367 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
368 delete_obj:
369 kobject_put(*hugepage_kobj);
370 return err;
371 }
372
373 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
374 {
375 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
376 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
377 kobject_put(hugepage_kobj);
378 }
379 #else
380 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
381 {
382 return 0;
383 }
384
385 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
386 {
387 }
388 #endif /* CONFIG_SYSFS */
389
390 static int __init hugepage_init(void)
391 {
392 int err;
393 struct kobject *hugepage_kobj;
394
395 if (!has_transparent_hugepage()) {
396 transparent_hugepage_flags = 0;
397 return -EINVAL;
398 }
399
400 /*
401 * hugepages can't be allocated by the buddy allocator
402 */
403 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
404 /*
405 * we use page->mapping and page->index in second tail page
406 * as list_head: assuming THP order >= 2
407 */
408 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
409
410 err = hugepage_init_sysfs(&hugepage_kobj);
411 if (err)
412 goto err_sysfs;
413
414 err = khugepaged_init();
415 if (err)
416 goto err_slab;
417
418 err = register_shrinker(&huge_zero_page_shrinker);
419 if (err)
420 goto err_hzp_shrinker;
421 err = register_shrinker(&deferred_split_shrinker);
422 if (err)
423 goto err_split_shrinker;
424
425 /*
426 * By default disable transparent hugepages on smaller systems,
427 * where the extra memory used could hurt more than TLB overhead
428 * is likely to save. The admin can still enable it through /sys.
429 */
430 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
431 transparent_hugepage_flags = 0;
432 return 0;
433 }
434
435 err = start_stop_khugepaged();
436 if (err)
437 goto err_khugepaged;
438
439 return 0;
440 err_khugepaged:
441 unregister_shrinker(&deferred_split_shrinker);
442 err_split_shrinker:
443 unregister_shrinker(&huge_zero_page_shrinker);
444 err_hzp_shrinker:
445 khugepaged_destroy();
446 err_slab:
447 hugepage_exit_sysfs(hugepage_kobj);
448 err_sysfs:
449 return err;
450 }
451 subsys_initcall(hugepage_init);
452
453 static int __init setup_transparent_hugepage(char *str)
454 {
455 int ret = 0;
456 if (!str)
457 goto out;
458 if (!strcmp(str, "always")) {
459 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
460 &transparent_hugepage_flags);
461 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
462 &transparent_hugepage_flags);
463 ret = 1;
464 } else if (!strcmp(str, "madvise")) {
465 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
466 &transparent_hugepage_flags);
467 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
468 &transparent_hugepage_flags);
469 ret = 1;
470 } else if (!strcmp(str, "never")) {
471 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
472 &transparent_hugepage_flags);
473 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
474 &transparent_hugepage_flags);
475 ret = 1;
476 }
477 out:
478 if (!ret)
479 pr_warn("transparent_hugepage= cannot parse, ignored\n");
480 return ret;
481 }
482 __setup("transparent_hugepage=", setup_transparent_hugepage);
483
484 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
485 {
486 if (likely(vma->vm_flags & VM_WRITE))
487 pmd = pmd_mkwrite(pmd);
488 return pmd;
489 }
490
491 #ifdef CONFIG_MEMCG
492 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
493 {
494 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
495 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
496
497 if (memcg)
498 return &memcg->deferred_split_queue;
499 else
500 return &pgdat->deferred_split_queue;
501 }
502 #else
503 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
504 {
505 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
506
507 return &pgdat->deferred_split_queue;
508 }
509 #endif
510
511 void prep_transhuge_page(struct page *page)
512 {
513 /*
514 * we use page->mapping and page->indexlru in second tail page
515 * as list_head: assuming THP order >= 2
516 */
517
518 INIT_LIST_HEAD(page_deferred_list(page));
519 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
520 }
521
522 static unsigned long __thp_get_unmapped_area(struct file *filp,
523 unsigned long addr, unsigned long len,
524 loff_t off, unsigned long flags, unsigned long size)
525 {
526 loff_t off_end = off + len;
527 loff_t off_align = round_up(off, size);
528 unsigned long len_pad, ret;
529
530 if (off_end <= off_align || (off_end - off_align) < size)
531 return 0;
532
533 len_pad = len + size;
534 if (len_pad < len || (off + len_pad) < off)
535 return 0;
536
537 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
538 off >> PAGE_SHIFT, flags);
539
540 /*
541 * The failure might be due to length padding. The caller will retry
542 * without the padding.
543 */
544 if (IS_ERR_VALUE(ret))
545 return 0;
546
547 /*
548 * Do not try to align to THP boundary if allocation at the address
549 * hint succeeds.
550 */
551 if (ret == addr)
552 return addr;
553
554 ret += (off - ret) & (size - 1);
555 return ret;
556 }
557
558 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
559 unsigned long len, unsigned long pgoff, unsigned long flags)
560 {
561 unsigned long ret;
562 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
563
564 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
565 goto out;
566
567 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
568 if (ret)
569 return ret;
570 out:
571 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
572 }
573 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
574
575 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
576 struct page *page, gfp_t gfp)
577 {
578 struct vm_area_struct *vma = vmf->vma;
579 struct mem_cgroup *memcg;
580 pgtable_t pgtable;
581 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
582 vm_fault_t ret = 0;
583
584 VM_BUG_ON_PAGE(!PageCompound(page), page);
585
586 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
587 put_page(page);
588 count_vm_event(THP_FAULT_FALLBACK);
589 return VM_FAULT_FALLBACK;
590 }
591
592 pgtable = pte_alloc_one(vma->vm_mm);
593 if (unlikely(!pgtable)) {
594 ret = VM_FAULT_OOM;
595 goto release;
596 }
597
598 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
599 /*
600 * The memory barrier inside __SetPageUptodate makes sure that
601 * clear_huge_page writes become visible before the set_pmd_at()
602 * write.
603 */
604 __SetPageUptodate(page);
605
606 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
607 if (unlikely(!pmd_none(*vmf->pmd))) {
608 goto unlock_release;
609 } else {
610 pmd_t entry;
611
612 ret = check_stable_address_space(vma->vm_mm);
613 if (ret)
614 goto unlock_release;
615
616 /* Deliver the page fault to userland */
617 if (userfaultfd_missing(vma)) {
618 vm_fault_t ret2;
619
620 spin_unlock(vmf->ptl);
621 mem_cgroup_cancel_charge(page, memcg, true);
622 put_page(page);
623 pte_free(vma->vm_mm, pgtable);
624 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
625 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
626 return ret2;
627 }
628
629 entry = mk_huge_pmd(page, vma->vm_page_prot);
630 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
631 page_add_new_anon_rmap(page, vma, haddr, true);
632 mem_cgroup_commit_charge(page, memcg, false, true);
633 lru_cache_add_active_or_unevictable(page, vma);
634 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
635 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
636 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
637 mm_inc_nr_ptes(vma->vm_mm);
638 spin_unlock(vmf->ptl);
639 count_vm_event(THP_FAULT_ALLOC);
640 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
641 }
642
643 return 0;
644 unlock_release:
645 spin_unlock(vmf->ptl);
646 release:
647 if (pgtable)
648 pte_free(vma->vm_mm, pgtable);
649 mem_cgroup_cancel_charge(page, memcg, true);
650 put_page(page);
651 return ret;
652
653 }
654
655 /*
656 * always: directly stall for all thp allocations
657 * defer: wake kswapd and fail if not immediately available
658 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
659 * fail if not immediately available
660 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
661 * available
662 * never: never stall for any thp allocation
663 */
664 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
665 {
666 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
667
668 /* Always do synchronous compaction */
669 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
670 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
671
672 /* Kick kcompactd and fail quickly */
673 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
674 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
675
676 /* Synchronous compaction if madvised, otherwise kick kcompactd */
677 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
678 return GFP_TRANSHUGE_LIGHT |
679 (vma_madvised ? __GFP_DIRECT_RECLAIM :
680 __GFP_KSWAPD_RECLAIM);
681
682 /* Only do synchronous compaction if madvised */
683 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
684 return GFP_TRANSHUGE_LIGHT |
685 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
686
687 return GFP_TRANSHUGE_LIGHT;
688 }
689
690 /* Caller must hold page table lock. */
691 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
692 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
693 struct page *zero_page)
694 {
695 pmd_t entry;
696 if (!pmd_none(*pmd))
697 return false;
698 entry = mk_pmd(zero_page, vma->vm_page_prot);
699 entry = pmd_mkhuge(entry);
700 if (pgtable)
701 pgtable_trans_huge_deposit(mm, pmd, pgtable);
702 set_pmd_at(mm, haddr, pmd, entry);
703 mm_inc_nr_ptes(mm);
704 return true;
705 }
706
707 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
708 {
709 struct vm_area_struct *vma = vmf->vma;
710 gfp_t gfp;
711 struct page *page;
712 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
713
714 if (!transhuge_vma_suitable(vma, haddr))
715 return VM_FAULT_FALLBACK;
716 if (unlikely(anon_vma_prepare(vma)))
717 return VM_FAULT_OOM;
718 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
719 return VM_FAULT_OOM;
720 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
721 !mm_forbids_zeropage(vma->vm_mm) &&
722 transparent_hugepage_use_zero_page()) {
723 pgtable_t pgtable;
724 struct page *zero_page;
725 bool set;
726 vm_fault_t ret;
727 pgtable = pte_alloc_one(vma->vm_mm);
728 if (unlikely(!pgtable))
729 return VM_FAULT_OOM;
730 zero_page = mm_get_huge_zero_page(vma->vm_mm);
731 if (unlikely(!zero_page)) {
732 pte_free(vma->vm_mm, pgtable);
733 count_vm_event(THP_FAULT_FALLBACK);
734 return VM_FAULT_FALLBACK;
735 }
736 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
737 ret = 0;
738 set = false;
739 if (pmd_none(*vmf->pmd)) {
740 ret = check_stable_address_space(vma->vm_mm);
741 if (ret) {
742 spin_unlock(vmf->ptl);
743 } else if (userfaultfd_missing(vma)) {
744 spin_unlock(vmf->ptl);
745 ret = handle_userfault(vmf, VM_UFFD_MISSING);
746 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
747 } else {
748 set_huge_zero_page(pgtable, vma->vm_mm, vma,
749 haddr, vmf->pmd, zero_page);
750 spin_unlock(vmf->ptl);
751 set = true;
752 }
753 } else
754 spin_unlock(vmf->ptl);
755 if (!set)
756 pte_free(vma->vm_mm, pgtable);
757 return ret;
758 }
759 gfp = alloc_hugepage_direct_gfpmask(vma);
760 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
761 if (unlikely(!page)) {
762 count_vm_event(THP_FAULT_FALLBACK);
763 return VM_FAULT_FALLBACK;
764 }
765 prep_transhuge_page(page);
766 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
767 }
768
769 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
770 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
771 pgtable_t pgtable)
772 {
773 struct mm_struct *mm = vma->vm_mm;
774 pmd_t entry;
775 spinlock_t *ptl;
776
777 ptl = pmd_lock(mm, pmd);
778 if (!pmd_none(*pmd)) {
779 if (write) {
780 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
781 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
782 goto out_unlock;
783 }
784 entry = pmd_mkyoung(*pmd);
785 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
786 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
787 update_mmu_cache_pmd(vma, addr, pmd);
788 }
789
790 goto out_unlock;
791 }
792
793 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
794 if (pfn_t_devmap(pfn))
795 entry = pmd_mkdevmap(entry);
796 if (write) {
797 entry = pmd_mkyoung(pmd_mkdirty(entry));
798 entry = maybe_pmd_mkwrite(entry, vma);
799 }
800
801 if (pgtable) {
802 pgtable_trans_huge_deposit(mm, pmd, pgtable);
803 mm_inc_nr_ptes(mm);
804 pgtable = NULL;
805 }
806
807 set_pmd_at(mm, addr, pmd, entry);
808 update_mmu_cache_pmd(vma, addr, pmd);
809
810 out_unlock:
811 spin_unlock(ptl);
812 if (pgtable)
813 pte_free(mm, pgtable);
814 }
815
816 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
817 {
818 unsigned long addr = vmf->address & PMD_MASK;
819 struct vm_area_struct *vma = vmf->vma;
820 pgprot_t pgprot = vma->vm_page_prot;
821 pgtable_t pgtable = NULL;
822
823 /*
824 * If we had pmd_special, we could avoid all these restrictions,
825 * but we need to be consistent with PTEs and architectures that
826 * can't support a 'special' bit.
827 */
828 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
829 !pfn_t_devmap(pfn));
830 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
831 (VM_PFNMAP|VM_MIXEDMAP));
832 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
833
834 if (addr < vma->vm_start || addr >= vma->vm_end)
835 return VM_FAULT_SIGBUS;
836
837 if (arch_needs_pgtable_deposit()) {
838 pgtable = pte_alloc_one(vma->vm_mm);
839 if (!pgtable)
840 return VM_FAULT_OOM;
841 }
842
843 track_pfn_insert(vma, &pgprot, pfn);
844
845 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
846 return VM_FAULT_NOPAGE;
847 }
848 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
849
850 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
851 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
852 {
853 if (likely(vma->vm_flags & VM_WRITE))
854 pud = pud_mkwrite(pud);
855 return pud;
856 }
857
858 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
859 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
860 {
861 struct mm_struct *mm = vma->vm_mm;
862 pud_t entry;
863 spinlock_t *ptl;
864
865 ptl = pud_lock(mm, pud);
866 if (!pud_none(*pud)) {
867 if (write) {
868 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
869 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
870 goto out_unlock;
871 }
872 entry = pud_mkyoung(*pud);
873 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
874 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
875 update_mmu_cache_pud(vma, addr, pud);
876 }
877 goto out_unlock;
878 }
879
880 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
881 if (pfn_t_devmap(pfn))
882 entry = pud_mkdevmap(entry);
883 if (write) {
884 entry = pud_mkyoung(pud_mkdirty(entry));
885 entry = maybe_pud_mkwrite(entry, vma);
886 }
887 set_pud_at(mm, addr, pud, entry);
888 update_mmu_cache_pud(vma, addr, pud);
889
890 out_unlock:
891 spin_unlock(ptl);
892 }
893
894 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
895 {
896 unsigned long addr = vmf->address & PUD_MASK;
897 struct vm_area_struct *vma = vmf->vma;
898 pgprot_t pgprot = vma->vm_page_prot;
899
900 /*
901 * If we had pud_special, we could avoid all these restrictions,
902 * but we need to be consistent with PTEs and architectures that
903 * can't support a 'special' bit.
904 */
905 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
906 !pfn_t_devmap(pfn));
907 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
908 (VM_PFNMAP|VM_MIXEDMAP));
909 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
910
911 if (addr < vma->vm_start || addr >= vma->vm_end)
912 return VM_FAULT_SIGBUS;
913
914 track_pfn_insert(vma, &pgprot, pfn);
915
916 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
917 return VM_FAULT_NOPAGE;
918 }
919 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
920 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
921
922 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
923 pmd_t *pmd, int flags)
924 {
925 pmd_t _pmd;
926
927 _pmd = pmd_mkyoung(*pmd);
928 if (flags & FOLL_WRITE)
929 _pmd = pmd_mkdirty(_pmd);
930 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
931 pmd, _pmd, flags & FOLL_WRITE))
932 update_mmu_cache_pmd(vma, addr, pmd);
933 }
934
935 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
936 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
937 {
938 unsigned long pfn = pmd_pfn(*pmd);
939 struct mm_struct *mm = vma->vm_mm;
940 struct page *page;
941
942 assert_spin_locked(pmd_lockptr(mm, pmd));
943
944 /*
945 * When we COW a devmap PMD entry, we split it into PTEs, so we should
946 * not be in this function with `flags & FOLL_COW` set.
947 */
948 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
949
950 if (flags & FOLL_WRITE && !pmd_write(*pmd))
951 return NULL;
952
953 if (pmd_present(*pmd) && pmd_devmap(*pmd))
954 /* pass */;
955 else
956 return NULL;
957
958 if (flags & FOLL_TOUCH)
959 touch_pmd(vma, addr, pmd, flags);
960
961 /*
962 * device mapped pages can only be returned if the
963 * caller will manage the page reference count.
964 */
965 if (!(flags & FOLL_GET))
966 return ERR_PTR(-EEXIST);
967
968 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
969 *pgmap = get_dev_pagemap(pfn, *pgmap);
970 if (!*pgmap)
971 return ERR_PTR(-EFAULT);
972 page = pfn_to_page(pfn);
973 get_page(page);
974
975 return page;
976 }
977
978 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
979 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
980 struct vm_area_struct *vma)
981 {
982 spinlock_t *dst_ptl, *src_ptl;
983 struct page *src_page;
984 pmd_t pmd;
985 pgtable_t pgtable = NULL;
986 int ret = -ENOMEM;
987
988 /* Skip if can be re-fill on fault */
989 if (!vma_is_anonymous(vma))
990 return 0;
991
992 pgtable = pte_alloc_one(dst_mm);
993 if (unlikely(!pgtable))
994 goto out;
995
996 dst_ptl = pmd_lock(dst_mm, dst_pmd);
997 src_ptl = pmd_lockptr(src_mm, src_pmd);
998 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
999
1000 ret = -EAGAIN;
1001 pmd = *src_pmd;
1002
1003 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1004 if (unlikely(is_swap_pmd(pmd))) {
1005 swp_entry_t entry = pmd_to_swp_entry(pmd);
1006
1007 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1008 if (is_write_migration_entry(entry)) {
1009 make_migration_entry_read(&entry);
1010 pmd = swp_entry_to_pmd(entry);
1011 if (pmd_swp_soft_dirty(*src_pmd))
1012 pmd = pmd_swp_mksoft_dirty(pmd);
1013 set_pmd_at(src_mm, addr, src_pmd, pmd);
1014 }
1015 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1016 mm_inc_nr_ptes(dst_mm);
1017 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1018 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1019 ret = 0;
1020 goto out_unlock;
1021 }
1022 #endif
1023
1024 if (unlikely(!pmd_trans_huge(pmd))) {
1025 pte_free(dst_mm, pgtable);
1026 goto out_unlock;
1027 }
1028 /*
1029 * When page table lock is held, the huge zero pmd should not be
1030 * under splitting since we don't split the page itself, only pmd to
1031 * a page table.
1032 */
1033 if (is_huge_zero_pmd(pmd)) {
1034 struct page *zero_page;
1035 /*
1036 * get_huge_zero_page() will never allocate a new page here,
1037 * since we already have a zero page to copy. It just takes a
1038 * reference.
1039 */
1040 zero_page = mm_get_huge_zero_page(dst_mm);
1041 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1042 zero_page);
1043 ret = 0;
1044 goto out_unlock;
1045 }
1046
1047 src_page = pmd_page(pmd);
1048 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1049 get_page(src_page);
1050 page_dup_rmap(src_page, true);
1051 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1052 mm_inc_nr_ptes(dst_mm);
1053 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1054
1055 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1056 pmd = pmd_mkold(pmd_wrprotect(pmd));
1057 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1058
1059 ret = 0;
1060 out_unlock:
1061 spin_unlock(src_ptl);
1062 spin_unlock(dst_ptl);
1063 out:
1064 return ret;
1065 }
1066
1067 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1068 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1069 pud_t *pud, int flags)
1070 {
1071 pud_t _pud;
1072
1073 _pud = pud_mkyoung(*pud);
1074 if (flags & FOLL_WRITE)
1075 _pud = pud_mkdirty(_pud);
1076 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1077 pud, _pud, flags & FOLL_WRITE))
1078 update_mmu_cache_pud(vma, addr, pud);
1079 }
1080
1081 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1082 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1083 {
1084 unsigned long pfn = pud_pfn(*pud);
1085 struct mm_struct *mm = vma->vm_mm;
1086 struct page *page;
1087
1088 assert_spin_locked(pud_lockptr(mm, pud));
1089
1090 if (flags & FOLL_WRITE && !pud_write(*pud))
1091 return NULL;
1092
1093 if (pud_present(*pud) && pud_devmap(*pud))
1094 /* pass */;
1095 else
1096 return NULL;
1097
1098 if (flags & FOLL_TOUCH)
1099 touch_pud(vma, addr, pud, flags);
1100
1101 /*
1102 * device mapped pages can only be returned if the
1103 * caller will manage the page reference count.
1104 */
1105 if (!(flags & FOLL_GET))
1106 return ERR_PTR(-EEXIST);
1107
1108 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1109 *pgmap = get_dev_pagemap(pfn, *pgmap);
1110 if (!*pgmap)
1111 return ERR_PTR(-EFAULT);
1112 page = pfn_to_page(pfn);
1113 get_page(page);
1114
1115 return page;
1116 }
1117
1118 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1119 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1120 struct vm_area_struct *vma)
1121 {
1122 spinlock_t *dst_ptl, *src_ptl;
1123 pud_t pud;
1124 int ret;
1125
1126 dst_ptl = pud_lock(dst_mm, dst_pud);
1127 src_ptl = pud_lockptr(src_mm, src_pud);
1128 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1129
1130 ret = -EAGAIN;
1131 pud = *src_pud;
1132 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1133 goto out_unlock;
1134
1135 /*
1136 * When page table lock is held, the huge zero pud should not be
1137 * under splitting since we don't split the page itself, only pud to
1138 * a page table.
1139 */
1140 if (is_huge_zero_pud(pud)) {
1141 /* No huge zero pud yet */
1142 }
1143
1144 pudp_set_wrprotect(src_mm, addr, src_pud);
1145 pud = pud_mkold(pud_wrprotect(pud));
1146 set_pud_at(dst_mm, addr, dst_pud, pud);
1147
1148 ret = 0;
1149 out_unlock:
1150 spin_unlock(src_ptl);
1151 spin_unlock(dst_ptl);
1152 return ret;
1153 }
1154
1155 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1156 {
1157 pud_t entry;
1158 unsigned long haddr;
1159 bool write = vmf->flags & FAULT_FLAG_WRITE;
1160
1161 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1162 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1163 goto unlock;
1164
1165 entry = pud_mkyoung(orig_pud);
1166 if (write)
1167 entry = pud_mkdirty(entry);
1168 haddr = vmf->address & HPAGE_PUD_MASK;
1169 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1170 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1171
1172 unlock:
1173 spin_unlock(vmf->ptl);
1174 }
1175 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1176
1177 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1178 {
1179 pmd_t entry;
1180 unsigned long haddr;
1181 bool write = vmf->flags & FAULT_FLAG_WRITE;
1182
1183 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1184 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1185 goto unlock;
1186
1187 entry = pmd_mkyoung(orig_pmd);
1188 if (write)
1189 entry = pmd_mkdirty(entry);
1190 haddr = vmf->address & HPAGE_PMD_MASK;
1191 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1192 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1193
1194 unlock:
1195 spin_unlock(vmf->ptl);
1196 }
1197
1198 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1199 pmd_t orig_pmd, struct page *page)
1200 {
1201 struct vm_area_struct *vma = vmf->vma;
1202 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1203 struct mem_cgroup *memcg;
1204 pgtable_t pgtable;
1205 pmd_t _pmd;
1206 int i;
1207 vm_fault_t ret = 0;
1208 struct page **pages;
1209 struct mmu_notifier_range range;
1210
1211 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1212 GFP_KERNEL);
1213 if (unlikely(!pages)) {
1214 ret |= VM_FAULT_OOM;
1215 goto out;
1216 }
1217
1218 for (i = 0; i < HPAGE_PMD_NR; i++) {
1219 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1220 vmf->address, page_to_nid(page));
1221 if (unlikely(!pages[i] ||
1222 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1223 GFP_KERNEL, &memcg, false))) {
1224 if (pages[i])
1225 put_page(pages[i]);
1226 while (--i >= 0) {
1227 memcg = (void *)page_private(pages[i]);
1228 set_page_private(pages[i], 0);
1229 mem_cgroup_cancel_charge(pages[i], memcg,
1230 false);
1231 put_page(pages[i]);
1232 }
1233 kfree(pages);
1234 ret |= VM_FAULT_OOM;
1235 goto out;
1236 }
1237 set_page_private(pages[i], (unsigned long)memcg);
1238 }
1239
1240 for (i = 0; i < HPAGE_PMD_NR; i++) {
1241 copy_user_highpage(pages[i], page + i,
1242 haddr + PAGE_SIZE * i, vma);
1243 __SetPageUptodate(pages[i]);
1244 cond_resched();
1245 }
1246
1247 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1248 haddr, haddr + HPAGE_PMD_SIZE);
1249 mmu_notifier_invalidate_range_start(&range);
1250
1251 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1252 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1253 goto out_free_pages;
1254 VM_BUG_ON_PAGE(!PageHead(page), page);
1255
1256 /*
1257 * Leave pmd empty until pte is filled note we must notify here as
1258 * concurrent CPU thread might write to new page before the call to
1259 * mmu_notifier_invalidate_range_end() happens which can lead to a
1260 * device seeing memory write in different order than CPU.
1261 *
1262 * See Documentation/vm/mmu_notifier.rst
1263 */
1264 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1265
1266 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1267 pmd_populate(vma->vm_mm, &_pmd, pgtable);
1268
1269 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1270 pte_t entry;
1271 entry = mk_pte(pages[i], vma->vm_page_prot);
1272 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1273 memcg = (void *)page_private(pages[i]);
1274 set_page_private(pages[i], 0);
1275 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1276 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1277 lru_cache_add_active_or_unevictable(pages[i], vma);
1278 vmf->pte = pte_offset_map(&_pmd, haddr);
1279 VM_BUG_ON(!pte_none(*vmf->pte));
1280 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1281 pte_unmap(vmf->pte);
1282 }
1283 kfree(pages);
1284
1285 smp_wmb(); /* make pte visible before pmd */
1286 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1287 page_remove_rmap(page, true);
1288 spin_unlock(vmf->ptl);
1289
1290 /*
1291 * No need to double call mmu_notifier->invalidate_range() callback as
1292 * the above pmdp_huge_clear_flush_notify() did already call it.
1293 */
1294 mmu_notifier_invalidate_range_only_end(&range);
1295
1296 ret |= VM_FAULT_WRITE;
1297 put_page(page);
1298
1299 out:
1300 return ret;
1301
1302 out_free_pages:
1303 spin_unlock(vmf->ptl);
1304 mmu_notifier_invalidate_range_end(&range);
1305 for (i = 0; i < HPAGE_PMD_NR; i++) {
1306 memcg = (void *)page_private(pages[i]);
1307 set_page_private(pages[i], 0);
1308 mem_cgroup_cancel_charge(pages[i], memcg, false);
1309 put_page(pages[i]);
1310 }
1311 kfree(pages);
1312 goto out;
1313 }
1314
1315 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1316 {
1317 struct vm_area_struct *vma = vmf->vma;
1318 struct page *page = NULL, *new_page;
1319 struct mem_cgroup *memcg;
1320 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1321 struct mmu_notifier_range range;
1322 gfp_t huge_gfp; /* for allocation and charge */
1323 vm_fault_t ret = 0;
1324
1325 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1326 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1327 if (is_huge_zero_pmd(orig_pmd))
1328 goto alloc;
1329 spin_lock(vmf->ptl);
1330 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1331 goto out_unlock;
1332
1333 page = pmd_page(orig_pmd);
1334 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1335 /*
1336 * We can only reuse the page if nobody else maps the huge page or it's
1337 * part.
1338 */
1339 if (!trylock_page(page)) {
1340 get_page(page);
1341 spin_unlock(vmf->ptl);
1342 lock_page(page);
1343 spin_lock(vmf->ptl);
1344 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345 unlock_page(page);
1346 put_page(page);
1347 goto out_unlock;
1348 }
1349 put_page(page);
1350 }
1351 if (reuse_swap_page(page, NULL)) {
1352 pmd_t entry;
1353 entry = pmd_mkyoung(orig_pmd);
1354 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1355 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1356 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1357 ret |= VM_FAULT_WRITE;
1358 unlock_page(page);
1359 goto out_unlock;
1360 }
1361 unlock_page(page);
1362 get_page(page);
1363 spin_unlock(vmf->ptl);
1364 alloc:
1365 if (__transparent_hugepage_enabled(vma) &&
1366 !transparent_hugepage_debug_cow()) {
1367 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1368 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1369 } else
1370 new_page = NULL;
1371
1372 if (likely(new_page)) {
1373 prep_transhuge_page(new_page);
1374 } else {
1375 if (!page) {
1376 split_huge_pmd(vma, vmf->pmd, vmf->address);
1377 ret |= VM_FAULT_FALLBACK;
1378 } else {
1379 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1380 if (ret & VM_FAULT_OOM) {
1381 split_huge_pmd(vma, vmf->pmd, vmf->address);
1382 ret |= VM_FAULT_FALLBACK;
1383 }
1384 put_page(page);
1385 }
1386 count_vm_event(THP_FAULT_FALLBACK);
1387 goto out;
1388 }
1389
1390 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1391 huge_gfp, &memcg, true))) {
1392 put_page(new_page);
1393 split_huge_pmd(vma, vmf->pmd, vmf->address);
1394 if (page)
1395 put_page(page);
1396 ret |= VM_FAULT_FALLBACK;
1397 count_vm_event(THP_FAULT_FALLBACK);
1398 goto out;
1399 }
1400
1401 count_vm_event(THP_FAULT_ALLOC);
1402 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1403
1404 if (!page)
1405 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1406 else
1407 copy_user_huge_page(new_page, page, vmf->address,
1408 vma, HPAGE_PMD_NR);
1409 __SetPageUptodate(new_page);
1410
1411 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1412 haddr, haddr + HPAGE_PMD_SIZE);
1413 mmu_notifier_invalidate_range_start(&range);
1414
1415 spin_lock(vmf->ptl);
1416 if (page)
1417 put_page(page);
1418 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1419 spin_unlock(vmf->ptl);
1420 mem_cgroup_cancel_charge(new_page, memcg, true);
1421 put_page(new_page);
1422 goto out_mn;
1423 } else {
1424 pmd_t entry;
1425 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1426 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1427 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1428 page_add_new_anon_rmap(new_page, vma, haddr, true);
1429 mem_cgroup_commit_charge(new_page, memcg, false, true);
1430 lru_cache_add_active_or_unevictable(new_page, vma);
1431 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1432 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1433 if (!page) {
1434 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1435 } else {
1436 VM_BUG_ON_PAGE(!PageHead(page), page);
1437 page_remove_rmap(page, true);
1438 put_page(page);
1439 }
1440 ret |= VM_FAULT_WRITE;
1441 }
1442 spin_unlock(vmf->ptl);
1443 out_mn:
1444 /*
1445 * No need to double call mmu_notifier->invalidate_range() callback as
1446 * the above pmdp_huge_clear_flush_notify() did already call it.
1447 */
1448 mmu_notifier_invalidate_range_only_end(&range);
1449 out:
1450 return ret;
1451 out_unlock:
1452 spin_unlock(vmf->ptl);
1453 return ret;
1454 }
1455
1456 /*
1457 * FOLL_FORCE can write to even unwritable pmd's, but only
1458 * after we've gone through a COW cycle and they are dirty.
1459 */
1460 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1461 {
1462 return pmd_write(pmd) ||
1463 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1464 }
1465
1466 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1467 unsigned long addr,
1468 pmd_t *pmd,
1469 unsigned int flags)
1470 {
1471 struct mm_struct *mm = vma->vm_mm;
1472 struct page *page = NULL;
1473
1474 assert_spin_locked(pmd_lockptr(mm, pmd));
1475
1476 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1477 goto out;
1478
1479 /* Avoid dumping huge zero page */
1480 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1481 return ERR_PTR(-EFAULT);
1482
1483 /* Full NUMA hinting faults to serialise migration in fault paths */
1484 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1485 goto out;
1486
1487 page = pmd_page(*pmd);
1488 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1489 if (flags & FOLL_TOUCH)
1490 touch_pmd(vma, addr, pmd, flags);
1491 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1492 /*
1493 * We don't mlock() pte-mapped THPs. This way we can avoid
1494 * leaking mlocked pages into non-VM_LOCKED VMAs.
1495 *
1496 * For anon THP:
1497 *
1498 * In most cases the pmd is the only mapping of the page as we
1499 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1500 * writable private mappings in populate_vma_page_range().
1501 *
1502 * The only scenario when we have the page shared here is if we
1503 * mlocking read-only mapping shared over fork(). We skip
1504 * mlocking such pages.
1505 *
1506 * For file THP:
1507 *
1508 * We can expect PageDoubleMap() to be stable under page lock:
1509 * for file pages we set it in page_add_file_rmap(), which
1510 * requires page to be locked.
1511 */
1512
1513 if (PageAnon(page) && compound_mapcount(page) != 1)
1514 goto skip_mlock;
1515 if (PageDoubleMap(page) || !page->mapping)
1516 goto skip_mlock;
1517 if (!trylock_page(page))
1518 goto skip_mlock;
1519 lru_add_drain();
1520 if (page->mapping && !PageDoubleMap(page))
1521 mlock_vma_page(page);
1522 unlock_page(page);
1523 }
1524 skip_mlock:
1525 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1526 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1527 if (flags & FOLL_GET)
1528 get_page(page);
1529
1530 out:
1531 return page;
1532 }
1533
1534 /* NUMA hinting page fault entry point for trans huge pmds */
1535 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1536 {
1537 struct vm_area_struct *vma = vmf->vma;
1538 struct anon_vma *anon_vma = NULL;
1539 struct page *page;
1540 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1541 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1542 int target_nid, last_cpupid = -1;
1543 bool page_locked;
1544 bool migrated = false;
1545 bool was_writable;
1546 int flags = 0;
1547
1548 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1549 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1550 goto out_unlock;
1551
1552 /*
1553 * If there are potential migrations, wait for completion and retry
1554 * without disrupting NUMA hinting information. Do not relock and
1555 * check_same as the page may no longer be mapped.
1556 */
1557 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1558 page = pmd_page(*vmf->pmd);
1559 if (!get_page_unless_zero(page))
1560 goto out_unlock;
1561 spin_unlock(vmf->ptl);
1562 put_and_wait_on_page_locked(page);
1563 goto out;
1564 }
1565
1566 page = pmd_page(pmd);
1567 BUG_ON(is_huge_zero_page(page));
1568 page_nid = page_to_nid(page);
1569 last_cpupid = page_cpupid_last(page);
1570 count_vm_numa_event(NUMA_HINT_FAULTS);
1571 if (page_nid == this_nid) {
1572 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1573 flags |= TNF_FAULT_LOCAL;
1574 }
1575
1576 /* See similar comment in do_numa_page for explanation */
1577 if (!pmd_savedwrite(pmd))
1578 flags |= TNF_NO_GROUP;
1579
1580 /*
1581 * Acquire the page lock to serialise THP migrations but avoid dropping
1582 * page_table_lock if at all possible
1583 */
1584 page_locked = trylock_page(page);
1585 target_nid = mpol_misplaced(page, vma, haddr);
1586 if (target_nid == NUMA_NO_NODE) {
1587 /* If the page was locked, there are no parallel migrations */
1588 if (page_locked)
1589 goto clear_pmdnuma;
1590 }
1591
1592 /* Migration could have started since the pmd_trans_migrating check */
1593 if (!page_locked) {
1594 page_nid = NUMA_NO_NODE;
1595 if (!get_page_unless_zero(page))
1596 goto out_unlock;
1597 spin_unlock(vmf->ptl);
1598 put_and_wait_on_page_locked(page);
1599 goto out;
1600 }
1601
1602 /*
1603 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1604 * to serialises splits
1605 */
1606 get_page(page);
1607 spin_unlock(vmf->ptl);
1608 anon_vma = page_lock_anon_vma_read(page);
1609
1610 /* Confirm the PMD did not change while page_table_lock was released */
1611 spin_lock(vmf->ptl);
1612 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1613 unlock_page(page);
1614 put_page(page);
1615 page_nid = NUMA_NO_NODE;
1616 goto out_unlock;
1617 }
1618
1619 /* Bail if we fail to protect against THP splits for any reason */
1620 if (unlikely(!anon_vma)) {
1621 put_page(page);
1622 page_nid = NUMA_NO_NODE;
1623 goto clear_pmdnuma;
1624 }
1625
1626 /*
1627 * Since we took the NUMA fault, we must have observed the !accessible
1628 * bit. Make sure all other CPUs agree with that, to avoid them
1629 * modifying the page we're about to migrate.
1630 *
1631 * Must be done under PTL such that we'll observe the relevant
1632 * inc_tlb_flush_pending().
1633 *
1634 * We are not sure a pending tlb flush here is for a huge page
1635 * mapping or not. Hence use the tlb range variant
1636 */
1637 if (mm_tlb_flush_pending(vma->vm_mm)) {
1638 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1639 /*
1640 * change_huge_pmd() released the pmd lock before
1641 * invalidating the secondary MMUs sharing the primary
1642 * MMU pagetables (with ->invalidate_range()). The
1643 * mmu_notifier_invalidate_range_end() (which
1644 * internally calls ->invalidate_range()) in
1645 * change_pmd_range() will run after us, so we can't
1646 * rely on it here and we need an explicit invalidate.
1647 */
1648 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1649 haddr + HPAGE_PMD_SIZE);
1650 }
1651
1652 /*
1653 * Migrate the THP to the requested node, returns with page unlocked
1654 * and access rights restored.
1655 */
1656 spin_unlock(vmf->ptl);
1657
1658 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1659 vmf->pmd, pmd, vmf->address, page, target_nid);
1660 if (migrated) {
1661 flags |= TNF_MIGRATED;
1662 page_nid = target_nid;
1663 } else
1664 flags |= TNF_MIGRATE_FAIL;
1665
1666 goto out;
1667 clear_pmdnuma:
1668 BUG_ON(!PageLocked(page));
1669 was_writable = pmd_savedwrite(pmd);
1670 pmd = pmd_modify(pmd, vma->vm_page_prot);
1671 pmd = pmd_mkyoung(pmd);
1672 if (was_writable)
1673 pmd = pmd_mkwrite(pmd);
1674 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1675 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1676 unlock_page(page);
1677 out_unlock:
1678 spin_unlock(vmf->ptl);
1679
1680 out:
1681 if (anon_vma)
1682 page_unlock_anon_vma_read(anon_vma);
1683
1684 if (page_nid != NUMA_NO_NODE)
1685 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1686 flags);
1687
1688 return 0;
1689 }
1690
1691 /*
1692 * Return true if we do MADV_FREE successfully on entire pmd page.
1693 * Otherwise, return false.
1694 */
1695 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1696 pmd_t *pmd, unsigned long addr, unsigned long next)
1697 {
1698 spinlock_t *ptl;
1699 pmd_t orig_pmd;
1700 struct page *page;
1701 struct mm_struct *mm = tlb->mm;
1702 bool ret = false;
1703
1704 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1705
1706 ptl = pmd_trans_huge_lock(pmd, vma);
1707 if (!ptl)
1708 goto out_unlocked;
1709
1710 orig_pmd = *pmd;
1711 if (is_huge_zero_pmd(orig_pmd))
1712 goto out;
1713
1714 if (unlikely(!pmd_present(orig_pmd))) {
1715 VM_BUG_ON(thp_migration_supported() &&
1716 !is_pmd_migration_entry(orig_pmd));
1717 goto out;
1718 }
1719
1720 page = pmd_page(orig_pmd);
1721 /*
1722 * If other processes are mapping this page, we couldn't discard
1723 * the page unless they all do MADV_FREE so let's skip the page.
1724 */
1725 if (page_mapcount(page) != 1)
1726 goto out;
1727
1728 if (!trylock_page(page))
1729 goto out;
1730
1731 /*
1732 * If user want to discard part-pages of THP, split it so MADV_FREE
1733 * will deactivate only them.
1734 */
1735 if (next - addr != HPAGE_PMD_SIZE) {
1736 get_page(page);
1737 spin_unlock(ptl);
1738 split_huge_page(page);
1739 unlock_page(page);
1740 put_page(page);
1741 goto out_unlocked;
1742 }
1743
1744 if (PageDirty(page))
1745 ClearPageDirty(page);
1746 unlock_page(page);
1747
1748 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1749 pmdp_invalidate(vma, addr, pmd);
1750 orig_pmd = pmd_mkold(orig_pmd);
1751 orig_pmd = pmd_mkclean(orig_pmd);
1752
1753 set_pmd_at(mm, addr, pmd, orig_pmd);
1754 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1755 }
1756
1757 mark_page_lazyfree(page);
1758 ret = true;
1759 out:
1760 spin_unlock(ptl);
1761 out_unlocked:
1762 return ret;
1763 }
1764
1765 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1766 {
1767 pgtable_t pgtable;
1768
1769 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1770 pte_free(mm, pgtable);
1771 mm_dec_nr_ptes(mm);
1772 }
1773
1774 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1775 pmd_t *pmd, unsigned long addr)
1776 {
1777 pmd_t orig_pmd;
1778 spinlock_t *ptl;
1779
1780 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1781
1782 ptl = __pmd_trans_huge_lock(pmd, vma);
1783 if (!ptl)
1784 return 0;
1785 /*
1786 * For architectures like ppc64 we look at deposited pgtable
1787 * when calling pmdp_huge_get_and_clear. So do the
1788 * pgtable_trans_huge_withdraw after finishing pmdp related
1789 * operations.
1790 */
1791 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1792 tlb->fullmm);
1793 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1794 if (vma_is_dax(vma)) {
1795 if (arch_needs_pgtable_deposit())
1796 zap_deposited_table(tlb->mm, pmd);
1797 spin_unlock(ptl);
1798 if (is_huge_zero_pmd(orig_pmd))
1799 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1800 } else if (is_huge_zero_pmd(orig_pmd)) {
1801 zap_deposited_table(tlb->mm, pmd);
1802 spin_unlock(ptl);
1803 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1804 } else {
1805 struct page *page = NULL;
1806 int flush_needed = 1;
1807
1808 if (pmd_present(orig_pmd)) {
1809 page = pmd_page(orig_pmd);
1810 page_remove_rmap(page, true);
1811 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1812 VM_BUG_ON_PAGE(!PageHead(page), page);
1813 } else if (thp_migration_supported()) {
1814 swp_entry_t entry;
1815
1816 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1817 entry = pmd_to_swp_entry(orig_pmd);
1818 page = pfn_to_page(swp_offset(entry));
1819 flush_needed = 0;
1820 } else
1821 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1822
1823 if (PageAnon(page)) {
1824 zap_deposited_table(tlb->mm, pmd);
1825 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1826 } else {
1827 if (arch_needs_pgtable_deposit())
1828 zap_deposited_table(tlb->mm, pmd);
1829 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1830 }
1831
1832 spin_unlock(ptl);
1833 if (flush_needed)
1834 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1835 }
1836 return 1;
1837 }
1838
1839 #ifndef pmd_move_must_withdraw
1840 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1841 spinlock_t *old_pmd_ptl,
1842 struct vm_area_struct *vma)
1843 {
1844 /*
1845 * With split pmd lock we also need to move preallocated
1846 * PTE page table if new_pmd is on different PMD page table.
1847 *
1848 * We also don't deposit and withdraw tables for file pages.
1849 */
1850 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1851 }
1852 #endif
1853
1854 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1855 {
1856 #ifdef CONFIG_MEM_SOFT_DIRTY
1857 if (unlikely(is_pmd_migration_entry(pmd)))
1858 pmd = pmd_swp_mksoft_dirty(pmd);
1859 else if (pmd_present(pmd))
1860 pmd = pmd_mksoft_dirty(pmd);
1861 #endif
1862 return pmd;
1863 }
1864
1865 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1866 unsigned long new_addr, unsigned long old_end,
1867 pmd_t *old_pmd, pmd_t *new_pmd)
1868 {
1869 spinlock_t *old_ptl, *new_ptl;
1870 pmd_t pmd;
1871 struct mm_struct *mm = vma->vm_mm;
1872 bool force_flush = false;
1873
1874 if ((old_addr & ~HPAGE_PMD_MASK) ||
1875 (new_addr & ~HPAGE_PMD_MASK) ||
1876 old_end - old_addr < HPAGE_PMD_SIZE)
1877 return false;
1878
1879 /*
1880 * The destination pmd shouldn't be established, free_pgtables()
1881 * should have release it.
1882 */
1883 if (WARN_ON(!pmd_none(*new_pmd))) {
1884 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1885 return false;
1886 }
1887
1888 /*
1889 * We don't have to worry about the ordering of src and dst
1890 * ptlocks because exclusive mmap_sem prevents deadlock.
1891 */
1892 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1893 if (old_ptl) {
1894 new_ptl = pmd_lockptr(mm, new_pmd);
1895 if (new_ptl != old_ptl)
1896 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1897 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1898 if (pmd_present(pmd))
1899 force_flush = true;
1900 VM_BUG_ON(!pmd_none(*new_pmd));
1901
1902 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1903 pgtable_t pgtable;
1904 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1905 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1906 }
1907 pmd = move_soft_dirty_pmd(pmd);
1908 set_pmd_at(mm, new_addr, new_pmd, pmd);
1909 if (force_flush)
1910 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1911 if (new_ptl != old_ptl)
1912 spin_unlock(new_ptl);
1913 spin_unlock(old_ptl);
1914 return true;
1915 }
1916 return false;
1917 }
1918
1919 /*
1920 * Returns
1921 * - 0 if PMD could not be locked
1922 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1923 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1924 */
1925 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1926 unsigned long addr, pgprot_t newprot, int prot_numa)
1927 {
1928 struct mm_struct *mm = vma->vm_mm;
1929 spinlock_t *ptl;
1930 pmd_t entry;
1931 bool preserve_write;
1932 int ret;
1933
1934 ptl = __pmd_trans_huge_lock(pmd, vma);
1935 if (!ptl)
1936 return 0;
1937
1938 preserve_write = prot_numa && pmd_write(*pmd);
1939 ret = 1;
1940
1941 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1942 if (is_swap_pmd(*pmd)) {
1943 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1944
1945 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1946 if (is_write_migration_entry(entry)) {
1947 pmd_t newpmd;
1948 /*
1949 * A protection check is difficult so
1950 * just be safe and disable write
1951 */
1952 make_migration_entry_read(&entry);
1953 newpmd = swp_entry_to_pmd(entry);
1954 if (pmd_swp_soft_dirty(*pmd))
1955 newpmd = pmd_swp_mksoft_dirty(newpmd);
1956 set_pmd_at(mm, addr, pmd, newpmd);
1957 }
1958 goto unlock;
1959 }
1960 #endif
1961
1962 /*
1963 * Avoid trapping faults against the zero page. The read-only
1964 * data is likely to be read-cached on the local CPU and
1965 * local/remote hits to the zero page are not interesting.
1966 */
1967 if (prot_numa && is_huge_zero_pmd(*pmd))
1968 goto unlock;
1969
1970 if (prot_numa && pmd_protnone(*pmd))
1971 goto unlock;
1972
1973 /*
1974 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1975 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1976 * which is also under down_read(mmap_sem):
1977 *
1978 * CPU0: CPU1:
1979 * change_huge_pmd(prot_numa=1)
1980 * pmdp_huge_get_and_clear_notify()
1981 * madvise_dontneed()
1982 * zap_pmd_range()
1983 * pmd_trans_huge(*pmd) == 0 (without ptl)
1984 * // skip the pmd
1985 * set_pmd_at();
1986 * // pmd is re-established
1987 *
1988 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1989 * which may break userspace.
1990 *
1991 * pmdp_invalidate() is required to make sure we don't miss
1992 * dirty/young flags set by hardware.
1993 */
1994 entry = pmdp_invalidate(vma, addr, pmd);
1995
1996 entry = pmd_modify(entry, newprot);
1997 if (preserve_write)
1998 entry = pmd_mk_savedwrite(entry);
1999 ret = HPAGE_PMD_NR;
2000 set_pmd_at(mm, addr, pmd, entry);
2001 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2002 unlock:
2003 spin_unlock(ptl);
2004 return ret;
2005 }
2006
2007 /*
2008 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2009 *
2010 * Note that if it returns page table lock pointer, this routine returns without
2011 * unlocking page table lock. So callers must unlock it.
2012 */
2013 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2014 {
2015 spinlock_t *ptl;
2016 ptl = pmd_lock(vma->vm_mm, pmd);
2017 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2018 pmd_devmap(*pmd)))
2019 return ptl;
2020 spin_unlock(ptl);
2021 return NULL;
2022 }
2023
2024 /*
2025 * Returns true if a given pud maps a thp, false otherwise.
2026 *
2027 * Note that if it returns true, this routine returns without unlocking page
2028 * table lock. So callers must unlock it.
2029 */
2030 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2031 {
2032 spinlock_t *ptl;
2033
2034 ptl = pud_lock(vma->vm_mm, pud);
2035 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2036 return ptl;
2037 spin_unlock(ptl);
2038 return NULL;
2039 }
2040
2041 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2042 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2043 pud_t *pud, unsigned long addr)
2044 {
2045 spinlock_t *ptl;
2046
2047 ptl = __pud_trans_huge_lock(pud, vma);
2048 if (!ptl)
2049 return 0;
2050 /*
2051 * For architectures like ppc64 we look at deposited pgtable
2052 * when calling pudp_huge_get_and_clear. So do the
2053 * pgtable_trans_huge_withdraw after finishing pudp related
2054 * operations.
2055 */
2056 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2057 tlb_remove_pud_tlb_entry(tlb, pud, addr);
2058 if (vma_is_dax(vma)) {
2059 spin_unlock(ptl);
2060 /* No zero page support yet */
2061 } else {
2062 /* No support for anonymous PUD pages yet */
2063 BUG();
2064 }
2065 return 1;
2066 }
2067
2068 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2069 unsigned long haddr)
2070 {
2071 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2072 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2073 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2074 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2075
2076 count_vm_event(THP_SPLIT_PUD);
2077
2078 pudp_huge_clear_flush_notify(vma, haddr, pud);
2079 }
2080
2081 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2082 unsigned long address)
2083 {
2084 spinlock_t *ptl;
2085 struct mmu_notifier_range range;
2086
2087 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2088 address & HPAGE_PUD_MASK,
2089 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2090 mmu_notifier_invalidate_range_start(&range);
2091 ptl = pud_lock(vma->vm_mm, pud);
2092 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2093 goto out;
2094 __split_huge_pud_locked(vma, pud, range.start);
2095
2096 out:
2097 spin_unlock(ptl);
2098 /*
2099 * No need to double call mmu_notifier->invalidate_range() callback as
2100 * the above pudp_huge_clear_flush_notify() did already call it.
2101 */
2102 mmu_notifier_invalidate_range_only_end(&range);
2103 }
2104 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2105
2106 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2107 unsigned long haddr, pmd_t *pmd)
2108 {
2109 struct mm_struct *mm = vma->vm_mm;
2110 pgtable_t pgtable;
2111 pmd_t _pmd;
2112 int i;
2113
2114 /*
2115 * Leave pmd empty until pte is filled note that it is fine to delay
2116 * notification until mmu_notifier_invalidate_range_end() as we are
2117 * replacing a zero pmd write protected page with a zero pte write
2118 * protected page.
2119 *
2120 * See Documentation/vm/mmu_notifier.rst
2121 */
2122 pmdp_huge_clear_flush(vma, haddr, pmd);
2123
2124 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2125 pmd_populate(mm, &_pmd, pgtable);
2126
2127 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2128 pte_t *pte, entry;
2129 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2130 entry = pte_mkspecial(entry);
2131 pte = pte_offset_map(&_pmd, haddr);
2132 VM_BUG_ON(!pte_none(*pte));
2133 set_pte_at(mm, haddr, pte, entry);
2134 pte_unmap(pte);
2135 }
2136 smp_wmb(); /* make pte visible before pmd */
2137 pmd_populate(mm, pmd, pgtable);
2138 }
2139
2140 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2141 unsigned long haddr, bool freeze)
2142 {
2143 struct mm_struct *mm = vma->vm_mm;
2144 struct page *page;
2145 pgtable_t pgtable;
2146 pmd_t old_pmd, _pmd;
2147 bool young, write, soft_dirty, pmd_migration = false;
2148 unsigned long addr;
2149 int i;
2150
2151 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2152 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2153 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2154 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2155 && !pmd_devmap(*pmd));
2156
2157 count_vm_event(THP_SPLIT_PMD);
2158
2159 if (!vma_is_anonymous(vma)) {
2160 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2161 /*
2162 * We are going to unmap this huge page. So
2163 * just go ahead and zap it
2164 */
2165 if (arch_needs_pgtable_deposit())
2166 zap_deposited_table(mm, pmd);
2167 if (vma_is_dax(vma))
2168 return;
2169 page = pmd_page(_pmd);
2170 if (!PageDirty(page) && pmd_dirty(_pmd))
2171 set_page_dirty(page);
2172 if (!PageReferenced(page) && pmd_young(_pmd))
2173 SetPageReferenced(page);
2174 page_remove_rmap(page, true);
2175 put_page(page);
2176 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2177 return;
2178 } else if (is_huge_zero_pmd(*pmd)) {
2179 /*
2180 * FIXME: Do we want to invalidate secondary mmu by calling
2181 * mmu_notifier_invalidate_range() see comments below inside
2182 * __split_huge_pmd() ?
2183 *
2184 * We are going from a zero huge page write protected to zero
2185 * small page also write protected so it does not seems useful
2186 * to invalidate secondary mmu at this time.
2187 */
2188 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2189 }
2190
2191 /*
2192 * Up to this point the pmd is present and huge and userland has the
2193 * whole access to the hugepage during the split (which happens in
2194 * place). If we overwrite the pmd with the not-huge version pointing
2195 * to the pte here (which of course we could if all CPUs were bug
2196 * free), userland could trigger a small page size TLB miss on the
2197 * small sized TLB while the hugepage TLB entry is still established in
2198 * the huge TLB. Some CPU doesn't like that.
2199 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2200 * 383 on page 93. Intel should be safe but is also warns that it's
2201 * only safe if the permission and cache attributes of the two entries
2202 * loaded in the two TLB is identical (which should be the case here).
2203 * But it is generally safer to never allow small and huge TLB entries
2204 * for the same virtual address to be loaded simultaneously. So instead
2205 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2206 * current pmd notpresent (atomically because here the pmd_trans_huge
2207 * must remain set at all times on the pmd until the split is complete
2208 * for this pmd), then we flush the SMP TLB and finally we write the
2209 * non-huge version of the pmd entry with pmd_populate.
2210 */
2211 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2212
2213 pmd_migration = is_pmd_migration_entry(old_pmd);
2214 if (unlikely(pmd_migration)) {
2215 swp_entry_t entry;
2216
2217 entry = pmd_to_swp_entry(old_pmd);
2218 page = pfn_to_page(swp_offset(entry));
2219 write = is_write_migration_entry(entry);
2220 young = false;
2221 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2222 } else {
2223 page = pmd_page(old_pmd);
2224 if (pmd_dirty(old_pmd))
2225 SetPageDirty(page);
2226 write = pmd_write(old_pmd);
2227 young = pmd_young(old_pmd);
2228 soft_dirty = pmd_soft_dirty(old_pmd);
2229 }
2230 VM_BUG_ON_PAGE(!page_count(page), page);
2231 page_ref_add(page, HPAGE_PMD_NR - 1);
2232
2233 /*
2234 * Withdraw the table only after we mark the pmd entry invalid.
2235 * This's critical for some architectures (Power).
2236 */
2237 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2238 pmd_populate(mm, &_pmd, pgtable);
2239
2240 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2241 pte_t entry, *pte;
2242 /*
2243 * Note that NUMA hinting access restrictions are not
2244 * transferred to avoid any possibility of altering
2245 * permissions across VMAs.
2246 */
2247 if (freeze || pmd_migration) {
2248 swp_entry_t swp_entry;
2249 swp_entry = make_migration_entry(page + i, write);
2250 entry = swp_entry_to_pte(swp_entry);
2251 if (soft_dirty)
2252 entry = pte_swp_mksoft_dirty(entry);
2253 } else {
2254 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2255 entry = maybe_mkwrite(entry, vma);
2256 if (!write)
2257 entry = pte_wrprotect(entry);
2258 if (!young)
2259 entry = pte_mkold(entry);
2260 if (soft_dirty)
2261 entry = pte_mksoft_dirty(entry);
2262 }
2263 pte = pte_offset_map(&_pmd, addr);
2264 BUG_ON(!pte_none(*pte));
2265 set_pte_at(mm, addr, pte, entry);
2266 atomic_inc(&page[i]._mapcount);
2267 pte_unmap(pte);
2268 }
2269
2270 /*
2271 * Set PG_double_map before dropping compound_mapcount to avoid
2272 * false-negative page_mapped().
2273 */
2274 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2275 for (i = 0; i < HPAGE_PMD_NR; i++)
2276 atomic_inc(&page[i]._mapcount);
2277 }
2278
2279 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2280 /* Last compound_mapcount is gone. */
2281 __dec_node_page_state(page, NR_ANON_THPS);
2282 if (TestClearPageDoubleMap(page)) {
2283 /* No need in mapcount reference anymore */
2284 for (i = 0; i < HPAGE_PMD_NR; i++)
2285 atomic_dec(&page[i]._mapcount);
2286 }
2287 }
2288
2289 smp_wmb(); /* make pte visible before pmd */
2290 pmd_populate(mm, pmd, pgtable);
2291
2292 if (freeze) {
2293 for (i = 0; i < HPAGE_PMD_NR; i++) {
2294 page_remove_rmap(page + i, false);
2295 put_page(page + i);
2296 }
2297 }
2298 }
2299
2300 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2301 unsigned long address, bool freeze, struct page *page)
2302 {
2303 spinlock_t *ptl;
2304 struct mmu_notifier_range range;
2305
2306 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2307 address & HPAGE_PMD_MASK,
2308 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2309 mmu_notifier_invalidate_range_start(&range);
2310 ptl = pmd_lock(vma->vm_mm, pmd);
2311
2312 /*
2313 * If caller asks to setup a migration entries, we need a page to check
2314 * pmd against. Otherwise we can end up replacing wrong page.
2315 */
2316 VM_BUG_ON(freeze && !page);
2317 if (page && page != pmd_page(*pmd))
2318 goto out;
2319
2320 if (pmd_trans_huge(*pmd)) {
2321 page = pmd_page(*pmd);
2322 if (PageMlocked(page))
2323 clear_page_mlock(page);
2324 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2325 goto out;
2326 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2327 out:
2328 spin_unlock(ptl);
2329 /*
2330 * No need to double call mmu_notifier->invalidate_range() callback.
2331 * They are 3 cases to consider inside __split_huge_pmd_locked():
2332 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2333 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2334 * fault will trigger a flush_notify before pointing to a new page
2335 * (it is fine if the secondary mmu keeps pointing to the old zero
2336 * page in the meantime)
2337 * 3) Split a huge pmd into pte pointing to the same page. No need
2338 * to invalidate secondary tlb entry they are all still valid.
2339 * any further changes to individual pte will notify. So no need
2340 * to call mmu_notifier->invalidate_range()
2341 */
2342 mmu_notifier_invalidate_range_only_end(&range);
2343 }
2344
2345 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2346 bool freeze, struct page *page)
2347 {
2348 pgd_t *pgd;
2349 p4d_t *p4d;
2350 pud_t *pud;
2351 pmd_t *pmd;
2352
2353 pgd = pgd_offset(vma->vm_mm, address);
2354 if (!pgd_present(*pgd))
2355 return;
2356
2357 p4d = p4d_offset(pgd, address);
2358 if (!p4d_present(*p4d))
2359 return;
2360
2361 pud = pud_offset(p4d, address);
2362 if (!pud_present(*pud))
2363 return;
2364
2365 pmd = pmd_offset(pud, address);
2366
2367 __split_huge_pmd(vma, pmd, address, freeze, page);
2368 }
2369
2370 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2371 unsigned long start,
2372 unsigned long end,
2373 long adjust_next)
2374 {
2375 /*
2376 * If the new start address isn't hpage aligned and it could
2377 * previously contain an hugepage: check if we need to split
2378 * an huge pmd.
2379 */
2380 if (start & ~HPAGE_PMD_MASK &&
2381 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2382 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2383 split_huge_pmd_address(vma, start, false, NULL);
2384
2385 /*
2386 * If the new end address isn't hpage aligned and it could
2387 * previously contain an hugepage: check if we need to split
2388 * an huge pmd.
2389 */
2390 if (end & ~HPAGE_PMD_MASK &&
2391 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2392 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2393 split_huge_pmd_address(vma, end, false, NULL);
2394
2395 /*
2396 * If we're also updating the vma->vm_next->vm_start, if the new
2397 * vm_next->vm_start isn't page aligned and it could previously
2398 * contain an hugepage: check if we need to split an huge pmd.
2399 */
2400 if (adjust_next > 0) {
2401 struct vm_area_struct *next = vma->vm_next;
2402 unsigned long nstart = next->vm_start;
2403 nstart += adjust_next << PAGE_SHIFT;
2404 if (nstart & ~HPAGE_PMD_MASK &&
2405 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2406 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2407 split_huge_pmd_address(next, nstart, false, NULL);
2408 }
2409 }
2410
2411 static void unmap_page(struct page *page)
2412 {
2413 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2414 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2415 bool unmap_success;
2416
2417 VM_BUG_ON_PAGE(!PageHead(page), page);
2418
2419 if (PageAnon(page))
2420 ttu_flags |= TTU_SPLIT_FREEZE;
2421
2422 unmap_success = try_to_unmap(page, ttu_flags);
2423 VM_BUG_ON_PAGE(!unmap_success, page);
2424 }
2425
2426 static void remap_page(struct page *page)
2427 {
2428 int i;
2429 if (PageTransHuge(page)) {
2430 remove_migration_ptes(page, page, true);
2431 } else {
2432 for (i = 0; i < HPAGE_PMD_NR; i++)
2433 remove_migration_ptes(page + i, page + i, true);
2434 }
2435 }
2436
2437 static void __split_huge_page_tail(struct page *head, int tail,
2438 struct lruvec *lruvec, struct list_head *list)
2439 {
2440 struct page *page_tail = head + tail;
2441
2442 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2443
2444 /*
2445 * Clone page flags before unfreezing refcount.
2446 *
2447 * After successful get_page_unless_zero() might follow flags change,
2448 * for exmaple lock_page() which set PG_waiters.
2449 */
2450 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2451 page_tail->flags |= (head->flags &
2452 ((1L << PG_referenced) |
2453 (1L << PG_swapbacked) |
2454 (1L << PG_swapcache) |
2455 (1L << PG_mlocked) |
2456 (1L << PG_uptodate) |
2457 (1L << PG_active) |
2458 (1L << PG_workingset) |
2459 (1L << PG_locked) |
2460 (1L << PG_unevictable) |
2461 (1L << PG_dirty)));
2462
2463 /* ->mapping in first tail page is compound_mapcount */
2464 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2465 page_tail);
2466 page_tail->mapping = head->mapping;
2467 page_tail->index = head->index + tail;
2468
2469 /* Page flags must be visible before we make the page non-compound. */
2470 smp_wmb();
2471
2472 /*
2473 * Clear PageTail before unfreezing page refcount.
2474 *
2475 * After successful get_page_unless_zero() might follow put_page()
2476 * which needs correct compound_head().
2477 */
2478 clear_compound_head(page_tail);
2479
2480 /* Finally unfreeze refcount. Additional reference from page cache. */
2481 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2482 PageSwapCache(head)));
2483
2484 if (page_is_young(head))
2485 set_page_young(page_tail);
2486 if (page_is_idle(head))
2487 set_page_idle(page_tail);
2488
2489 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2490
2491 /*
2492 * always add to the tail because some iterators expect new
2493 * pages to show after the currently processed elements - e.g.
2494 * migrate_pages
2495 */
2496 lru_add_page_tail(head, page_tail, lruvec, list);
2497 }
2498
2499 static void __split_huge_page(struct page *page, struct list_head *list,
2500 pgoff_t end, unsigned long flags)
2501 {
2502 struct page *head = compound_head(page);
2503 pg_data_t *pgdat = page_pgdat(head);
2504 struct lruvec *lruvec;
2505 struct address_space *swap_cache = NULL;
2506 unsigned long offset = 0;
2507 int i;
2508
2509 lruvec = mem_cgroup_page_lruvec(head, pgdat);
2510
2511 /* complete memcg works before add pages to LRU */
2512 mem_cgroup_split_huge_fixup(head);
2513
2514 if (PageAnon(head) && PageSwapCache(head)) {
2515 swp_entry_t entry = { .val = page_private(head) };
2516
2517 offset = swp_offset(entry);
2518 swap_cache = swap_address_space(entry);
2519 xa_lock(&swap_cache->i_pages);
2520 }
2521
2522 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2523 __split_huge_page_tail(head, i, lruvec, list);
2524 /* Some pages can be beyond i_size: drop them from page cache */
2525 if (head[i].index >= end) {
2526 ClearPageDirty(head + i);
2527 __delete_from_page_cache(head + i, NULL);
2528 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2529 shmem_uncharge(head->mapping->host, 1);
2530 put_page(head + i);
2531 } else if (!PageAnon(page)) {
2532 __xa_store(&head->mapping->i_pages, head[i].index,
2533 head + i, 0);
2534 } else if (swap_cache) {
2535 __xa_store(&swap_cache->i_pages, offset + i,
2536 head + i, 0);
2537 }
2538 }
2539
2540 ClearPageCompound(head);
2541
2542 split_page_owner(head, HPAGE_PMD_ORDER);
2543
2544 /* See comment in __split_huge_page_tail() */
2545 if (PageAnon(head)) {
2546 /* Additional pin to swap cache */
2547 if (PageSwapCache(head)) {
2548 page_ref_add(head, 2);
2549 xa_unlock(&swap_cache->i_pages);
2550 } else {
2551 page_ref_inc(head);
2552 }
2553 } else {
2554 /* Additional pin to page cache */
2555 page_ref_add(head, 2);
2556 xa_unlock(&head->mapping->i_pages);
2557 }
2558
2559 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2560
2561 remap_page(head);
2562
2563 for (i = 0; i < HPAGE_PMD_NR; i++) {
2564 struct page *subpage = head + i;
2565 if (subpage == page)
2566 continue;
2567 unlock_page(subpage);
2568
2569 /*
2570 * Subpages may be freed if there wasn't any mapping
2571 * like if add_to_swap() is running on a lru page that
2572 * had its mapping zapped. And freeing these pages
2573 * requires taking the lru_lock so we do the put_page
2574 * of the tail pages after the split is complete.
2575 */
2576 put_page(subpage);
2577 }
2578 }
2579
2580 int total_mapcount(struct page *page)
2581 {
2582 int i, compound, ret;
2583
2584 VM_BUG_ON_PAGE(PageTail(page), page);
2585
2586 if (likely(!PageCompound(page)))
2587 return atomic_read(&page->_mapcount) + 1;
2588
2589 compound = compound_mapcount(page);
2590 if (PageHuge(page))
2591 return compound;
2592 ret = compound;
2593 for (i = 0; i < HPAGE_PMD_NR; i++)
2594 ret += atomic_read(&page[i]._mapcount) + 1;
2595 /* File pages has compound_mapcount included in _mapcount */
2596 if (!PageAnon(page))
2597 return ret - compound * HPAGE_PMD_NR;
2598 if (PageDoubleMap(page))
2599 ret -= HPAGE_PMD_NR;
2600 return ret;
2601 }
2602
2603 /*
2604 * This calculates accurately how many mappings a transparent hugepage
2605 * has (unlike page_mapcount() which isn't fully accurate). This full
2606 * accuracy is primarily needed to know if copy-on-write faults can
2607 * reuse the page and change the mapping to read-write instead of
2608 * copying them. At the same time this returns the total_mapcount too.
2609 *
2610 * The function returns the highest mapcount any one of the subpages
2611 * has. If the return value is one, even if different processes are
2612 * mapping different subpages of the transparent hugepage, they can
2613 * all reuse it, because each process is reusing a different subpage.
2614 *
2615 * The total_mapcount is instead counting all virtual mappings of the
2616 * subpages. If the total_mapcount is equal to "one", it tells the
2617 * caller all mappings belong to the same "mm" and in turn the
2618 * anon_vma of the transparent hugepage can become the vma->anon_vma
2619 * local one as no other process may be mapping any of the subpages.
2620 *
2621 * It would be more accurate to replace page_mapcount() with
2622 * page_trans_huge_mapcount(), however we only use
2623 * page_trans_huge_mapcount() in the copy-on-write faults where we
2624 * need full accuracy to avoid breaking page pinning, because
2625 * page_trans_huge_mapcount() is slower than page_mapcount().
2626 */
2627 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2628 {
2629 int i, ret, _total_mapcount, mapcount;
2630
2631 /* hugetlbfs shouldn't call it */
2632 VM_BUG_ON_PAGE(PageHuge(page), page);
2633
2634 if (likely(!PageTransCompound(page))) {
2635 mapcount = atomic_read(&page->_mapcount) + 1;
2636 if (total_mapcount)
2637 *total_mapcount = mapcount;
2638 return mapcount;
2639 }
2640
2641 page = compound_head(page);
2642
2643 _total_mapcount = ret = 0;
2644 for (i = 0; i < HPAGE_PMD_NR; i++) {
2645 mapcount = atomic_read(&page[i]._mapcount) + 1;
2646 ret = max(ret, mapcount);
2647 _total_mapcount += mapcount;
2648 }
2649 if (PageDoubleMap(page)) {
2650 ret -= 1;
2651 _total_mapcount -= HPAGE_PMD_NR;
2652 }
2653 mapcount = compound_mapcount(page);
2654 ret += mapcount;
2655 _total_mapcount += mapcount;
2656 if (total_mapcount)
2657 *total_mapcount = _total_mapcount;
2658 return ret;
2659 }
2660
2661 /* Racy check whether the huge page can be split */
2662 bool can_split_huge_page(struct page *page, int *pextra_pins)
2663 {
2664 int extra_pins;
2665
2666 /* Additional pins from page cache */
2667 if (PageAnon(page))
2668 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2669 else
2670 extra_pins = HPAGE_PMD_NR;
2671 if (pextra_pins)
2672 *pextra_pins = extra_pins;
2673 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2674 }
2675
2676 /*
2677 * This function splits huge page into normal pages. @page can point to any
2678 * subpage of huge page to split. Split doesn't change the position of @page.
2679 *
2680 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2681 * The huge page must be locked.
2682 *
2683 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2684 *
2685 * Both head page and tail pages will inherit mapping, flags, and so on from
2686 * the hugepage.
2687 *
2688 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2689 * they are not mapped.
2690 *
2691 * Returns 0 if the hugepage is split successfully.
2692 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2693 * us.
2694 */
2695 int split_huge_page_to_list(struct page *page, struct list_head *list)
2696 {
2697 struct page *head = compound_head(page);
2698 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2699 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2700 struct anon_vma *anon_vma = NULL;
2701 struct address_space *mapping = NULL;
2702 int count, mapcount, extra_pins, ret;
2703 bool mlocked;
2704 unsigned long flags;
2705 pgoff_t end;
2706
2707 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2708 VM_BUG_ON_PAGE(!PageLocked(page), page);
2709 VM_BUG_ON_PAGE(!PageCompound(page), page);
2710
2711 if (PageWriteback(page))
2712 return -EBUSY;
2713
2714 if (PageAnon(head)) {
2715 /*
2716 * The caller does not necessarily hold an mmap_sem that would
2717 * prevent the anon_vma disappearing so we first we take a
2718 * reference to it and then lock the anon_vma for write. This
2719 * is similar to page_lock_anon_vma_read except the write lock
2720 * is taken to serialise against parallel split or collapse
2721 * operations.
2722 */
2723 anon_vma = page_get_anon_vma(head);
2724 if (!anon_vma) {
2725 ret = -EBUSY;
2726 goto out;
2727 }
2728 end = -1;
2729 mapping = NULL;
2730 anon_vma_lock_write(anon_vma);
2731 } else {
2732 mapping = head->mapping;
2733
2734 /* Truncated ? */
2735 if (!mapping) {
2736 ret = -EBUSY;
2737 goto out;
2738 }
2739
2740 anon_vma = NULL;
2741 i_mmap_lock_read(mapping);
2742
2743 /*
2744 *__split_huge_page() may need to trim off pages beyond EOF:
2745 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2746 * which cannot be nested inside the page tree lock. So note
2747 * end now: i_size itself may be changed at any moment, but
2748 * head page lock is good enough to serialize the trimming.
2749 */
2750 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2751 }
2752
2753 /*
2754 * Racy check if we can split the page, before unmap_page() will
2755 * split PMDs
2756 */
2757 if (!can_split_huge_page(head, &extra_pins)) {
2758 ret = -EBUSY;
2759 goto out_unlock;
2760 }
2761
2762 mlocked = PageMlocked(page);
2763 unmap_page(head);
2764 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2765
2766 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2767 if (mlocked)
2768 lru_add_drain();
2769
2770 /* prevent PageLRU to go away from under us, and freeze lru stats */
2771 spin_lock_irqsave(&pgdata->lru_lock, flags);
2772
2773 if (mapping) {
2774 XA_STATE(xas, &mapping->i_pages, page_index(head));
2775
2776 /*
2777 * Check if the head page is present in page cache.
2778 * We assume all tail are present too, if head is there.
2779 */
2780 xa_lock(&mapping->i_pages);
2781 if (xas_load(&xas) != head)
2782 goto fail;
2783 }
2784
2785 /* Prevent deferred_split_scan() touching ->_refcount */
2786 spin_lock(&ds_queue->split_queue_lock);
2787 count = page_count(head);
2788 mapcount = total_mapcount(head);
2789 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2790 if (!list_empty(page_deferred_list(head))) {
2791 ds_queue->split_queue_len--;
2792 list_del(page_deferred_list(head));
2793 }
2794 if (mapping) {
2795 if (PageSwapBacked(page))
2796 __dec_node_page_state(page, NR_SHMEM_THPS);
2797 else
2798 __dec_node_page_state(page, NR_FILE_THPS);
2799 }
2800
2801 spin_unlock(&ds_queue->split_queue_lock);
2802 __split_huge_page(page, list, end, flags);
2803 if (PageSwapCache(head)) {
2804 swp_entry_t entry = { .val = page_private(head) };
2805
2806 ret = split_swap_cluster(entry);
2807 } else
2808 ret = 0;
2809 } else {
2810 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2811 pr_alert("total_mapcount: %u, page_count(): %u\n",
2812 mapcount, count);
2813 if (PageTail(page))
2814 dump_page(head, NULL);
2815 dump_page(page, "total_mapcount(head) > 0");
2816 BUG();
2817 }
2818 spin_unlock(&ds_queue->split_queue_lock);
2819 fail: if (mapping)
2820 xa_unlock(&mapping->i_pages);
2821 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2822 remap_page(head);
2823 ret = -EBUSY;
2824 }
2825
2826 out_unlock:
2827 if (anon_vma) {
2828 anon_vma_unlock_write(anon_vma);
2829 put_anon_vma(anon_vma);
2830 }
2831 if (mapping)
2832 i_mmap_unlock_read(mapping);
2833 out:
2834 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2835 return ret;
2836 }
2837
2838 void free_transhuge_page(struct page *page)
2839 {
2840 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2841 unsigned long flags;
2842
2843 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2844 if (!list_empty(page_deferred_list(page))) {
2845 ds_queue->split_queue_len--;
2846 list_del(page_deferred_list(page));
2847 }
2848 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2849 free_compound_page(page);
2850 }
2851
2852 void deferred_split_huge_page(struct page *page)
2853 {
2854 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2855 #ifdef CONFIG_MEMCG
2856 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2857 #endif
2858 unsigned long flags;
2859
2860 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2861
2862 /*
2863 * The try_to_unmap() in page reclaim path might reach here too,
2864 * this may cause a race condition to corrupt deferred split queue.
2865 * And, if page reclaim is already handling the same page, it is
2866 * unnecessary to handle it again in shrinker.
2867 *
2868 * Check PageSwapCache to determine if the page is being
2869 * handled by page reclaim since THP swap would add the page into
2870 * swap cache before calling try_to_unmap().
2871 */
2872 if (PageSwapCache(page))
2873 return;
2874
2875 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2876 if (list_empty(page_deferred_list(page))) {
2877 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2878 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2879 ds_queue->split_queue_len++;
2880 #ifdef CONFIG_MEMCG
2881 if (memcg)
2882 memcg_set_shrinker_bit(memcg, page_to_nid(page),
2883 deferred_split_shrinker.id);
2884 #endif
2885 }
2886 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2887 }
2888
2889 static unsigned long deferred_split_count(struct shrinker *shrink,
2890 struct shrink_control *sc)
2891 {
2892 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2893 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2894
2895 #ifdef CONFIG_MEMCG
2896 if (sc->memcg)
2897 ds_queue = &sc->memcg->deferred_split_queue;
2898 #endif
2899 return READ_ONCE(ds_queue->split_queue_len);
2900 }
2901
2902 static unsigned long deferred_split_scan(struct shrinker *shrink,
2903 struct shrink_control *sc)
2904 {
2905 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2906 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2907 unsigned long flags;
2908 LIST_HEAD(list), *pos, *next;
2909 struct page *page;
2910 int split = 0;
2911
2912 #ifdef CONFIG_MEMCG
2913 if (sc->memcg)
2914 ds_queue = &sc->memcg->deferred_split_queue;
2915 #endif
2916
2917 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2918 /* Take pin on all head pages to avoid freeing them under us */
2919 list_for_each_safe(pos, next, &ds_queue->split_queue) {
2920 page = list_entry((void *)pos, struct page, mapping);
2921 page = compound_head(page);
2922 if (get_page_unless_zero(page)) {
2923 list_move(page_deferred_list(page), &list);
2924 } else {
2925 /* We lost race with put_compound_page() */
2926 list_del_init(page_deferred_list(page));
2927 ds_queue->split_queue_len--;
2928 }
2929 if (!--sc->nr_to_scan)
2930 break;
2931 }
2932 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2933
2934 list_for_each_safe(pos, next, &list) {
2935 page = list_entry((void *)pos, struct page, mapping);
2936 if (!trylock_page(page))
2937 goto next;
2938 /* split_huge_page() removes page from list on success */
2939 if (!split_huge_page(page))
2940 split++;
2941 unlock_page(page);
2942 next:
2943 put_page(page);
2944 }
2945
2946 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2947 list_splice_tail(&list, &ds_queue->split_queue);
2948 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2949
2950 /*
2951 * Stop shrinker if we didn't split any page, but the queue is empty.
2952 * This can happen if pages were freed under us.
2953 */
2954 if (!split && list_empty(&ds_queue->split_queue))
2955 return SHRINK_STOP;
2956 return split;
2957 }
2958
2959 static struct shrinker deferred_split_shrinker = {
2960 .count_objects = deferred_split_count,
2961 .scan_objects = deferred_split_scan,
2962 .seeks = DEFAULT_SEEKS,
2963 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2964 SHRINKER_NONSLAB,
2965 };
2966
2967 #ifdef CONFIG_DEBUG_FS
2968 static int split_huge_pages_set(void *data, u64 val)
2969 {
2970 struct zone *zone;
2971 struct page *page;
2972 unsigned long pfn, max_zone_pfn;
2973 unsigned long total = 0, split = 0;
2974
2975 if (val != 1)
2976 return -EINVAL;
2977
2978 for_each_populated_zone(zone) {
2979 max_zone_pfn = zone_end_pfn(zone);
2980 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2981 if (!pfn_valid(pfn))
2982 continue;
2983
2984 page = pfn_to_page(pfn);
2985 if (!get_page_unless_zero(page))
2986 continue;
2987
2988 if (zone != page_zone(page))
2989 goto next;
2990
2991 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2992 goto next;
2993
2994 total++;
2995 lock_page(page);
2996 if (!split_huge_page(page))
2997 split++;
2998 unlock_page(page);
2999 next:
3000 put_page(page);
3001 }
3002 }
3003
3004 pr_info("%lu of %lu THP split\n", split, total);
3005
3006 return 0;
3007 }
3008 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3009 "%llu\n");
3010
3011 static int __init split_huge_pages_debugfs(void)
3012 {
3013 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3014 &split_huge_pages_fops);
3015 return 0;
3016 }
3017 late_initcall(split_huge_pages_debugfs);
3018 #endif
3019
3020 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3021 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3022 struct page *page)
3023 {
3024 struct vm_area_struct *vma = pvmw->vma;
3025 struct mm_struct *mm = vma->vm_mm;
3026 unsigned long address = pvmw->address;
3027 pmd_t pmdval;
3028 swp_entry_t entry;
3029 pmd_t pmdswp;
3030
3031 if (!(pvmw->pmd && !pvmw->pte))
3032 return;
3033
3034 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3035 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3036 if (pmd_dirty(pmdval))
3037 set_page_dirty(page);
3038 entry = make_migration_entry(page, pmd_write(pmdval));
3039 pmdswp = swp_entry_to_pmd(entry);
3040 if (pmd_soft_dirty(pmdval))
3041 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3042 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3043 page_remove_rmap(page, true);
3044 put_page(page);
3045 }
3046
3047 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3048 {
3049 struct vm_area_struct *vma = pvmw->vma;
3050 struct mm_struct *mm = vma->vm_mm;
3051 unsigned long address = pvmw->address;
3052 unsigned long mmun_start = address & HPAGE_PMD_MASK;
3053 pmd_t pmde;
3054 swp_entry_t entry;
3055
3056 if (!(pvmw->pmd && !pvmw->pte))
3057 return;
3058
3059 entry = pmd_to_swp_entry(*pvmw->pmd);
3060 get_page(new);
3061 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3062 if (pmd_swp_soft_dirty(*pvmw->pmd))
3063 pmde = pmd_mksoft_dirty(pmde);
3064 if (is_write_migration_entry(entry))
3065 pmde = maybe_pmd_mkwrite(pmde, vma);
3066
3067 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3068 if (PageAnon(new))
3069 page_add_anon_rmap(new, vma, mmun_start, true);
3070 else
3071 page_add_file_rmap(new, true);
3072 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3073 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3074 mlock_vma_page(new);
3075 update_mmu_cache_pmd(vma, address, pvmw->pmd);
3076 }
3077 #endif
Linux with added FPC-III support