brintos / linux-shallow public Read only

1// SPDX-License-Identifier: GPL-2.0-only2/*3 * Generic hugetlb support.4 * (C) Nadia Yvette Chambers, April 20045 */6#include <linux/list.h>7#include <linux/init.h>8#include <linux/mm.h>9#include <linux/seq_file.h>10#include <linux/sysctl.h>11#include <linux/highmem.h>12#include <linux/mmu_notifier.h>13#include <linux/nodemask.h>14#include <linux/pagemap.h>15#include <linux/mempolicy.h>16#include <linux/compiler.h>17#include <linux/cpuset.h>18#include <linux/mutex.h>19#include <linux/memblock.h>20#include <linux/sysfs.h>21#include <linux/slab.h>22#include <linux/sched/mm.h>23#include <linux/mmdebug.h>24#include <linux/sched/signal.h>25#include <linux/rmap.h>26#include <linux/string_helpers.h>27#include <linux/swap.h>28#include <linux/swapops.h>29#include <linux/jhash.h>30#include <linux/numa.h>31#include <linux/llist.h>32#include <linux/cma.h>33#include <linux/migrate.h>34#include <linux/nospec.h>35#include <linux/delayacct.h>36#include <linux/memory.h>37#include <linux/mm_inline.h>38#include <linux/padata.h>39 40#include <asm/page.h>41#include <asm/pgalloc.h>42#include <asm/tlb.h>43 44#include <linux/io.h>45#include <linux/hugetlb.h>46#include <linux/hugetlb_cgroup.h>47#include <linux/node.h>48#include <linux/page_owner.h>49#include "internal.h"50#include "hugetlb_vmemmap.h"51 52int hugetlb_max_hstate __read_mostly;53unsigned int default_hstate_idx;54struct hstate hstates[HUGE_MAX_HSTATE];55 56#ifdef CONFIG_CMA57static struct cma *hugetlb_cma[MAX_NUMNODES];58static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;59#endif60static unsigned long hugetlb_cma_size __initdata;61 62__initdata struct list_head huge_boot_pages[MAX_NUMNODES];63 64/* for command line parsing */65static struct hstate * __initdata parsed_hstate;66static unsigned long __initdata default_hstate_max_huge_pages;67static bool __initdata parsed_valid_hugepagesz = true;68static bool __initdata parsed_default_hugepagesz;69static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;70 71/*72 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,73 * free_huge_pages, and surplus_huge_pages.74 */75__cacheline_aligned_in_smp DEFINE_SPINLOCK(hugetlb_lock);76 77/*78 * Serializes faults on the same logical page.  This is used to79 * prevent spurious OOMs when the hugepage pool is fully utilized.80 */81static int num_fault_mutexes __ro_after_init;82struct mutex *hugetlb_fault_mutex_table __ro_after_init;83 84/* Forward declaration */85static int hugetlb_acct_memory(struct hstate *h, long delta);86static void hugetlb_vma_lock_free(struct vm_area_struct *vma);87static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);88static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);89static void hugetlb_unshare_pmds(struct vm_area_struct *vma,90		unsigned long start, unsigned long end);91static struct resv_map *vma_resv_map(struct vm_area_struct *vma);92 93static void hugetlb_free_folio(struct folio *folio)94{95#ifdef CONFIG_CMA96	int nid = folio_nid(folio);97 98	if (cma_free_folio(hugetlb_cma[nid], folio))99		return;100#endif101	folio_put(folio);102}103 104static inline bool subpool_is_free(struct hugepage_subpool *spool)105{106	if (spool->count)107		return false;108	if (spool->max_hpages != -1)109		return spool->used_hpages == 0;110	if (spool->min_hpages != -1)111		return spool->rsv_hpages == spool->min_hpages;112 113	return true;114}115 116static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,117						unsigned long irq_flags)118{119	spin_unlock_irqrestore(&spool->lock, irq_flags);120 121	/* If no pages are used, and no other handles to the subpool122	 * remain, give up any reservations based on minimum size and123	 * free the subpool */124	if (subpool_is_free(spool)) {125		if (spool->min_hpages != -1)126			hugetlb_acct_memory(spool->hstate,127						-spool->min_hpages);128		kfree(spool);129	}130}131 132struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,133						long min_hpages)134{135	struct hugepage_subpool *spool;136 137	spool = kzalloc(sizeof(*spool), GFP_KERNEL);138	if (!spool)139		return NULL;140 141	spin_lock_init(&spool->lock);142	spool->count = 1;143	spool->max_hpages = max_hpages;144	spool->hstate = h;145	spool->min_hpages = min_hpages;146 147	if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {148		kfree(spool);149		return NULL;150	}151	spool->rsv_hpages = min_hpages;152 153	return spool;154}155 156void hugepage_put_subpool(struct hugepage_subpool *spool)157{158	unsigned long flags;159 160	spin_lock_irqsave(&spool->lock, flags);161	BUG_ON(!spool->count);162	spool->count--;163	unlock_or_release_subpool(spool, flags);164}165 166/*167 * Subpool accounting for allocating and reserving pages.168 * Return -ENOMEM if there are not enough resources to satisfy the169 * request.  Otherwise, return the number of pages by which the170 * global pools must be adjusted (upward).  The returned value may171 * only be different than the passed value (delta) in the case where172 * a subpool minimum size must be maintained.173 */174static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,175				      long delta)176{177	long ret = delta;178 179	if (!spool)180		return ret;181 182	spin_lock_irq(&spool->lock);183 184	if (spool->max_hpages != -1) {		/* maximum size accounting */185		if ((spool->used_hpages + delta) <= spool->max_hpages)186			spool->used_hpages += delta;187		else {188			ret = -ENOMEM;189			goto unlock_ret;190		}191	}192 193	/* minimum size accounting */194	if (spool->min_hpages != -1 && spool->rsv_hpages) {195		if (delta > spool->rsv_hpages) {196			/*197			 * Asking for more reserves than those already taken on198			 * behalf of subpool.  Return difference.199			 */200			ret = delta - spool->rsv_hpages;201			spool->rsv_hpages = 0;202		} else {203			ret = 0;	/* reserves already accounted for */204			spool->rsv_hpages -= delta;205		}206	}207 208unlock_ret:209	spin_unlock_irq(&spool->lock);210	return ret;211}212 213/*214 * Subpool accounting for freeing and unreserving pages.215 * Return the number of global page reservations that must be dropped.216 * The return value may only be different than the passed value (delta)217 * in the case where a subpool minimum size must be maintained.218 */219static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,220				       long delta)221{222	long ret = delta;223	unsigned long flags;224 225	if (!spool)226		return delta;227 228	spin_lock_irqsave(&spool->lock, flags);229 230	if (spool->max_hpages != -1)		/* maximum size accounting */231		spool->used_hpages -= delta;232 233	 /* minimum size accounting */234	if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {235		if (spool->rsv_hpages + delta <= spool->min_hpages)236			ret = 0;237		else238			ret = spool->rsv_hpages + delta - spool->min_hpages;239 240		spool->rsv_hpages += delta;241		if (spool->rsv_hpages > spool->min_hpages)242			spool->rsv_hpages = spool->min_hpages;243	}244 245	/*246	 * If hugetlbfs_put_super couldn't free spool due to an outstanding247	 * quota reference, free it now.248	 */249	unlock_or_release_subpool(spool, flags);250 251	return ret;252}253 254static inline struct hugepage_subpool *subpool_inode(struct inode *inode)255{256	return HUGETLBFS_SB(inode->i_sb)->spool;257}258 259static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)260{261	return subpool_inode(file_inode(vma->vm_file));262}263 264/*265 * hugetlb vma_lock helper routines266 */267void hugetlb_vma_lock_read(struct vm_area_struct *vma)268{269	if (__vma_shareable_lock(vma)) {270		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;271 272		down_read(&vma_lock->rw_sema);273	} else if (__vma_private_lock(vma)) {274		struct resv_map *resv_map = vma_resv_map(vma);275 276		down_read(&resv_map->rw_sema);277	}278}279 280void hugetlb_vma_unlock_read(struct vm_area_struct *vma)281{282	if (__vma_shareable_lock(vma)) {283		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;284 285		up_read(&vma_lock->rw_sema);286	} else if (__vma_private_lock(vma)) {287		struct resv_map *resv_map = vma_resv_map(vma);288 289		up_read(&resv_map->rw_sema);290	}291}292 293void hugetlb_vma_lock_write(struct vm_area_struct *vma)294{295	if (__vma_shareable_lock(vma)) {296		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;297 298		down_write(&vma_lock->rw_sema);299	} else if (__vma_private_lock(vma)) {300		struct resv_map *resv_map = vma_resv_map(vma);301 302		down_write(&resv_map->rw_sema);303	}304}305 306void hugetlb_vma_unlock_write(struct vm_area_struct *vma)307{308	if (__vma_shareable_lock(vma)) {309		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;310 311		up_write(&vma_lock->rw_sema);312	} else if (__vma_private_lock(vma)) {313		struct resv_map *resv_map = vma_resv_map(vma);314 315		up_write(&resv_map->rw_sema);316	}317}318 319int hugetlb_vma_trylock_write(struct vm_area_struct *vma)320{321 322	if (__vma_shareable_lock(vma)) {323		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;324 325		return down_write_trylock(&vma_lock->rw_sema);326	} else if (__vma_private_lock(vma)) {327		struct resv_map *resv_map = vma_resv_map(vma);328 329		return down_write_trylock(&resv_map->rw_sema);330	}331 332	return 1;333}334 335void hugetlb_vma_assert_locked(struct vm_area_struct *vma)336{337	if (__vma_shareable_lock(vma)) {338		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;339 340		lockdep_assert_held(&vma_lock->rw_sema);341	} else if (__vma_private_lock(vma)) {342		struct resv_map *resv_map = vma_resv_map(vma);343 344		lockdep_assert_held(&resv_map->rw_sema);345	}346}347 348void hugetlb_vma_lock_release(struct kref *kref)349{350	struct hugetlb_vma_lock *vma_lock = container_of(kref,351			struct hugetlb_vma_lock, refs);352 353	kfree(vma_lock);354}355 356static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)357{358	struct vm_area_struct *vma = vma_lock->vma;359 360	/*361	 * vma_lock structure may or not be released as a result of put,362	 * it certainly will no longer be attached to vma so clear pointer.363	 * Semaphore synchronizes access to vma_lock->vma field.364	 */365	vma_lock->vma = NULL;366	vma->vm_private_data = NULL;367	up_write(&vma_lock->rw_sema);368	kref_put(&vma_lock->refs, hugetlb_vma_lock_release);369}370 371static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)372{373	if (__vma_shareable_lock(vma)) {374		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;375 376		__hugetlb_vma_unlock_write_put(vma_lock);377	} else if (__vma_private_lock(vma)) {378		struct resv_map *resv_map = vma_resv_map(vma);379 380		/* no free for anon vmas, but still need to unlock */381		up_write(&resv_map->rw_sema);382	}383}384 385static void hugetlb_vma_lock_free(struct vm_area_struct *vma)386{387	/*388	 * Only present in sharable vmas.389	 */390	if (!vma || !__vma_shareable_lock(vma))391		return;392 393	if (vma->vm_private_data) {394		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;395 396		down_write(&vma_lock->rw_sema);397		__hugetlb_vma_unlock_write_put(vma_lock);398	}399}400 401static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)402{403	struct hugetlb_vma_lock *vma_lock;404 405	/* Only establish in (flags) sharable vmas */406	if (!vma || !(vma->vm_flags & VM_MAYSHARE))407		return;408 409	/* Should never get here with non-NULL vm_private_data */410	if (vma->vm_private_data)411		return;412 413	vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);414	if (!vma_lock) {415		/*416		 * If we can not allocate structure, then vma can not417		 * participate in pmd sharing.  This is only a possible418		 * performance enhancement and memory saving issue.419		 * However, the lock is also used to synchronize page420		 * faults with truncation.  If the lock is not present,421		 * unlikely races could leave pages in a file past i_size422		 * until the file is removed.  Warn in the unlikely case of423		 * allocation failure.424		 */425		pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");426		return;427	}428 429	kref_init(&vma_lock->refs);430	init_rwsem(&vma_lock->rw_sema);431	vma_lock->vma = vma;432	vma->vm_private_data = vma_lock;433}434 435/* Helper that removes a struct file_region from the resv_map cache and returns436 * it for use.437 */438static struct file_region *439get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)440{441	struct file_region *nrg;442 443	VM_BUG_ON(resv->region_cache_count <= 0);444 445	resv->region_cache_count--;446	nrg = list_first_entry(&resv->region_cache, struct file_region, link);447	list_del(&nrg->link);448 449	nrg->from = from;450	nrg->to = to;451 452	return nrg;453}454 455static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,456					      struct file_region *rg)457{458#ifdef CONFIG_CGROUP_HUGETLB459	nrg->reservation_counter = rg->reservation_counter;460	nrg->css = rg->css;461	if (rg->css)462		css_get(rg->css);463#endif464}465 466/* Helper that records hugetlb_cgroup uncharge info. */467static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,468						struct hstate *h,469						struct resv_map *resv,470						struct file_region *nrg)471{472#ifdef CONFIG_CGROUP_HUGETLB473	if (h_cg) {474		nrg->reservation_counter =475			&h_cg->rsvd_hugepage[hstate_index(h)];476		nrg->css = &h_cg->css;477		/*478		 * The caller will hold exactly one h_cg->css reference for the479		 * whole contiguous reservation region. But this area might be480		 * scattered when there are already some file_regions reside in481		 * it. As a result, many file_regions may share only one css482		 * reference. In order to ensure that one file_region must hold483		 * exactly one h_cg->css reference, we should do css_get for484		 * each file_region and leave the reference held by caller485		 * untouched.486		 */487		css_get(&h_cg->css);488		if (!resv->pages_per_hpage)489			resv->pages_per_hpage = pages_per_huge_page(h);490		/* pages_per_hpage should be the same for all entries in491		 * a resv_map.492		 */493		VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));494	} else {495		nrg->reservation_counter = NULL;496		nrg->css = NULL;497	}498#endif499}500 501static void put_uncharge_info(struct file_region *rg)502{503#ifdef CONFIG_CGROUP_HUGETLB504	if (rg->css)505		css_put(rg->css);506#endif507}508 509static bool has_same_uncharge_info(struct file_region *rg,510				   struct file_region *org)511{512#ifdef CONFIG_CGROUP_HUGETLB513	return rg->reservation_counter == org->reservation_counter &&514	       rg->css == org->css;515 516#else517	return true;518#endif519}520 521static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)522{523	struct file_region *nrg, *prg;524 525	prg = list_prev_entry(rg, link);526	if (&prg->link != &resv->regions && prg->to == rg->from &&527	    has_same_uncharge_info(prg, rg)) {528		prg->to = rg->to;529 530		list_del(&rg->link);531		put_uncharge_info(rg);532		kfree(rg);533 534		rg = prg;535	}536 537	nrg = list_next_entry(rg, link);538	if (&nrg->link != &resv->regions && nrg->from == rg->to &&539	    has_same_uncharge_info(nrg, rg)) {540		nrg->from = rg->from;541 542		list_del(&rg->link);543		put_uncharge_info(rg);544		kfree(rg);545	}546}547 548static inline long549hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,550		     long to, struct hstate *h, struct hugetlb_cgroup *cg,551		     long *regions_needed)552{553	struct file_region *nrg;554 555	if (!regions_needed) {556		nrg = get_file_region_entry_from_cache(map, from, to);557		record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);558		list_add(&nrg->link, rg);559		coalesce_file_region(map, nrg);560	} else561		*regions_needed += 1;562 563	return to - from;564}565 566/*567 * Must be called with resv->lock held.568 *569 * Calling this with regions_needed != NULL will count the number of pages570 * to be added but will not modify the linked list. And regions_needed will571 * indicate the number of file_regions needed in the cache to carry out to add572 * the regions for this range.573 */574static long add_reservation_in_range(struct resv_map *resv, long f, long t,575				     struct hugetlb_cgroup *h_cg,576				     struct hstate *h, long *regions_needed)577{578	long add = 0;579	struct list_head *head = &resv->regions;580	long last_accounted_offset = f;581	struct file_region *iter, *trg = NULL;582	struct list_head *rg = NULL;583 584	if (regions_needed)585		*regions_needed = 0;586 587	/* In this loop, we essentially handle an entry for the range588	 * [last_accounted_offset, iter->from), at every iteration, with some589	 * bounds checking.590	 */591	list_for_each_entry_safe(iter, trg, head, link) {592		/* Skip irrelevant regions that start before our range. */593		if (iter->from < f) {594			/* If this region ends after the last accounted offset,595			 * then we need to update last_accounted_offset.596			 */597			if (iter->to > last_accounted_offset)598				last_accounted_offset = iter->to;599			continue;600		}601 602		/* When we find a region that starts beyond our range, we've603		 * finished.604		 */605		if (iter->from >= t) {606			rg = iter->link.prev;607			break;608		}609 610		/* Add an entry for last_accounted_offset -> iter->from, and611		 * update last_accounted_offset.612		 */613		if (iter->from > last_accounted_offset)614			add += hugetlb_resv_map_add(resv, iter->link.prev,615						    last_accounted_offset,616						    iter->from, h, h_cg,617						    regions_needed);618 619		last_accounted_offset = iter->to;620	}621 622	/* Handle the case where our range extends beyond623	 * last_accounted_offset.624	 */625	if (!rg)626		rg = head->prev;627	if (last_accounted_offset < t)628		add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,629					    t, h, h_cg, regions_needed);630 631	return add;632}633 634/* Must be called with resv->lock acquired. Will drop lock to allocate entries.635 */636static int allocate_file_region_entries(struct resv_map *resv,637					int regions_needed)638	__must_hold(&resv->lock)639{640	LIST_HEAD(allocated_regions);641	int to_allocate = 0, i = 0;642	struct file_region *trg = NULL, *rg = NULL;643 644	VM_BUG_ON(regions_needed < 0);645 646	/*647	 * Check for sufficient descriptors in the cache to accommodate648	 * the number of in progress add operations plus regions_needed.649	 *650	 * This is a while loop because when we drop the lock, some other call651	 * to region_add or region_del may have consumed some region_entries,652	 * so we keep looping here until we finally have enough entries for653	 * (adds_in_progress + regions_needed).654	 */655	while (resv->region_cache_count <656	       (resv->adds_in_progress + regions_needed)) {657		to_allocate = resv->adds_in_progress + regions_needed -658			      resv->region_cache_count;659 660		/* At this point, we should have enough entries in the cache661		 * for all the existing adds_in_progress. We should only be662		 * needing to allocate for regions_needed.663		 */664		VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);665 666		spin_unlock(&resv->lock);667		for (i = 0; i < to_allocate; i++) {668			trg = kmalloc(sizeof(*trg), GFP_KERNEL);669			if (!trg)670				goto out_of_memory;671			list_add(&trg->link, &allocated_regions);672		}673 674		spin_lock(&resv->lock);675 676		list_splice(&allocated_regions, &resv->region_cache);677		resv->region_cache_count += to_allocate;678	}679 680	return 0;681 682out_of_memory:683	list_for_each_entry_safe(rg, trg, &allocated_regions, link) {684		list_del(&rg->link);685		kfree(rg);686	}687	return -ENOMEM;688}689 690/*691 * Add the huge page range represented by [f, t) to the reserve692 * map.  Regions will be taken from the cache to fill in this range.693 * Sufficient regions should exist in the cache due to the previous694 * call to region_chg with the same range, but in some cases the cache will not695 * have sufficient entries due to races with other code doing region_add or696 * region_del.  The extra needed entries will be allocated.697 *698 * regions_needed is the out value provided by a previous call to region_chg.699 *700 * Return the number of new huge pages added to the map.  This number is greater701 * than or equal to zero.  If file_region entries needed to be allocated for702 * this operation and we were not able to allocate, it returns -ENOMEM.703 * region_add of regions of length 1 never allocate file_regions and cannot704 * fail; region_chg will always allocate at least 1 entry and a region_add for705 * 1 page will only require at most 1 entry.706 */707static long region_add(struct resv_map *resv, long f, long t,708		       long in_regions_needed, struct hstate *h,709		       struct hugetlb_cgroup *h_cg)710{711	long add = 0, actual_regions_needed = 0;712 713	spin_lock(&resv->lock);714retry:715 716	/* Count how many regions are actually needed to execute this add. */717	add_reservation_in_range(resv, f, t, NULL, NULL,718				 &actual_regions_needed);719 720	/*721	 * Check for sufficient descriptors in the cache to accommodate722	 * this add operation. Note that actual_regions_needed may be greater723	 * than in_regions_needed, as the resv_map may have been modified since724	 * the region_chg call. In this case, we need to make sure that we725	 * allocate extra entries, such that we have enough for all the726	 * existing adds_in_progress, plus the excess needed for this727	 * operation.728	 */729	if (actual_regions_needed > in_regions_needed &&730	    resv->region_cache_count <731		    resv->adds_in_progress +732			    (actual_regions_needed - in_regions_needed)) {733		/* region_add operation of range 1 should never need to734		 * allocate file_region entries.735		 */736		VM_BUG_ON(t - f <= 1);737 738		if (allocate_file_region_entries(739			    resv, actual_regions_needed - in_regions_needed)) {740			return -ENOMEM;741		}742 743		goto retry;744	}745 746	add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);747 748	resv->adds_in_progress -= in_regions_needed;749 750	spin_unlock(&resv->lock);751	return add;752}753 754/*755 * Examine the existing reserve map and determine how many756 * huge pages in the specified range [f, t) are NOT currently757 * represented.  This routine is called before a subsequent758 * call to region_add that will actually modify the reserve759 * map to add the specified range [f, t).  region_chg does760 * not change the number of huge pages represented by the761 * map.  A number of new file_region structures is added to the cache as a762 * placeholder, for the subsequent region_add call to use. At least 1763 * file_region structure is added.764 *765 * out_regions_needed is the number of regions added to the766 * resv->adds_in_progress.  This value needs to be provided to a follow up call767 * to region_add or region_abort for proper accounting.768 *769 * Returns the number of huge pages that need to be added to the existing770 * reservation map for the range [f, t).  This number is greater or equal to771 * zero.  -ENOMEM is returned if a new file_region structure or cache entry772 * is needed and can not be allocated.773 */774static long region_chg(struct resv_map *resv, long f, long t,775		       long *out_regions_needed)776{777	long chg = 0;778 779	spin_lock(&resv->lock);780 781	/* Count how many hugepages in this range are NOT represented. */782	chg = add_reservation_in_range(resv, f, t, NULL, NULL,783				       out_regions_needed);784 785	if (*out_regions_needed == 0)786		*out_regions_needed = 1;787 788	if (allocate_file_region_entries(resv, *out_regions_needed))789		return -ENOMEM;790 791	resv->adds_in_progress += *out_regions_needed;792 793	spin_unlock(&resv->lock);794	return chg;795}796 797/*798 * Abort the in progress add operation.  The adds_in_progress field799 * of the resv_map keeps track of the operations in progress between800 * calls to region_chg and region_add.  Operations are sometimes801 * aborted after the call to region_chg.  In such cases, region_abort802 * is called to decrement the adds_in_progress counter. regions_needed803 * is the value returned by the region_chg call, it is used to decrement804 * the adds_in_progress counter.805 *806 * NOTE: The range arguments [f, t) are not needed or used in this807 * routine.  They are kept to make reading the calling code easier as808 * arguments will match the associated region_chg call.809 */810static void region_abort(struct resv_map *resv, long f, long t,811			 long regions_needed)812{813	spin_lock(&resv->lock);814	VM_BUG_ON(!resv->region_cache_count);815	resv->adds_in_progress -= regions_needed;816	spin_unlock(&resv->lock);817}818 819/*820 * Delete the specified range [f, t) from the reserve map.  If the821 * t parameter is LONG_MAX, this indicates that ALL regions after f822 * should be deleted.  Locate the regions which intersect [f, t)823 * and either trim, delete or split the existing regions.824 *825 * Returns the number of huge pages deleted from the reserve map.826 * In the normal case, the return value is zero or more.  In the827 * case where a region must be split, a new region descriptor must828 * be allocated.  If the allocation fails, -ENOMEM will be returned.829 * NOTE: If the parameter t == LONG_MAX, then we will never split830 * a region and possibly return -ENOMEM.  Callers specifying831 * t == LONG_MAX do not need to check for -ENOMEM error.832 */833static long region_del(struct resv_map *resv, long f, long t)834{835	struct list_head *head = &resv->regions;836	struct file_region *rg, *trg;837	struct file_region *nrg = NULL;838	long del = 0;839 840retry:841	spin_lock(&resv->lock);842	list_for_each_entry_safe(rg, trg, head, link) {843		/*844		 * Skip regions before the range to be deleted.  file_region845		 * ranges are normally of the form [from, to).  However, there846		 * may be a "placeholder" entry in the map which is of the form847		 * (from, to) with from == to.  Check for placeholder entries848		 * at the beginning of the range to be deleted.849		 */850		if (rg->to <= f && (rg->to != rg->from || rg->to != f))851			continue;852 853		if (rg->from >= t)854			break;855 856		if (f > rg->from && t < rg->to) { /* Must split region */857			/*858			 * Check for an entry in the cache before dropping859			 * lock and attempting allocation.860			 */861			if (!nrg &&862			    resv->region_cache_count > resv->adds_in_progress) {863				nrg = list_first_entry(&resv->region_cache,864							struct file_region,865							link);866				list_del(&nrg->link);867				resv->region_cache_count--;868			}869 870			if (!nrg) {871				spin_unlock(&resv->lock);872				nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);873				if (!nrg)874					return -ENOMEM;875				goto retry;876			}877 878			del += t - f;879			hugetlb_cgroup_uncharge_file_region(880				resv, rg, t - f, false);881 882			/* New entry for end of split region */883			nrg->from = t;884			nrg->to = rg->to;885 886			copy_hugetlb_cgroup_uncharge_info(nrg, rg);887 888			INIT_LIST_HEAD(&nrg->link);889 890			/* Original entry is trimmed */891			rg->to = f;892 893			list_add(&nrg->link, &rg->link);894			nrg = NULL;895			break;896		}897 898		if (f <= rg->from && t >= rg->to) { /* Remove entire region */899			del += rg->to - rg->from;900			hugetlb_cgroup_uncharge_file_region(resv, rg,901							    rg->to - rg->from, true);902			list_del(&rg->link);903			kfree(rg);904			continue;905		}906 907		if (f <= rg->from) {	/* Trim beginning of region */908			hugetlb_cgroup_uncharge_file_region(resv, rg,909							    t - rg->from, false);910 911			del += t - rg->from;912			rg->from = t;913		} else {		/* Trim end of region */914			hugetlb_cgroup_uncharge_file_region(resv, rg,915							    rg->to - f, false);916 917			del += rg->to - f;918			rg->to = f;919		}920	}921 922	spin_unlock(&resv->lock);923	kfree(nrg);924	return del;925}926 927/*928 * A rare out of memory error was encountered which prevented removal of929 * the reserve map region for a page.  The huge page itself was free'ed930 * and removed from the page cache.  This routine will adjust the subpool931 * usage count, and the global reserve count if needed.  By incrementing932 * these counts, the reserve map entry which could not be deleted will933 * appear as a "reserved" entry instead of simply dangling with incorrect934 * counts.935 */936void hugetlb_fix_reserve_counts(struct inode *inode)937{938	struct hugepage_subpool *spool = subpool_inode(inode);939	long rsv_adjust;940	bool reserved = false;941 942	rsv_adjust = hugepage_subpool_get_pages(spool, 1);943	if (rsv_adjust > 0) {944		struct hstate *h = hstate_inode(inode);945 946		if (!hugetlb_acct_memory(h, 1))947			reserved = true;948	} else if (!rsv_adjust) {949		reserved = true;950	}951 952	if (!reserved)953		pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");954}955 956/*957 * Count and return the number of huge pages in the reserve map958 * that intersect with the range [f, t).959 */960static long region_count(struct resv_map *resv, long f, long t)961{962	struct list_head *head = &resv->regions;963	struct file_region *rg;964	long chg = 0;965 966	spin_lock(&resv->lock);967	/* Locate each segment we overlap with, and count that overlap. */968	list_for_each_entry(rg, head, link) {969		long seg_from;970		long seg_to;971 972		if (rg->to <= f)973			continue;974		if (rg->from >= t)975			break;976 977		seg_from = max(rg->from, f);978		seg_to = min(rg->to, t);979 980		chg += seg_to - seg_from;981	}982	spin_unlock(&resv->lock);983 984	return chg;985}986 987/*988 * Convert the address within this vma to the page offset within989 * the mapping, huge page units here.990 */991static pgoff_t vma_hugecache_offset(struct hstate *h,992			struct vm_area_struct *vma, unsigned long address)993{994	return ((address - vma->vm_start) >> huge_page_shift(h)) +995			(vma->vm_pgoff >> huge_page_order(h));996}997 998/**999 * vma_kernel_pagesize - Page size granularity for this VMA.1000 * @vma: The user mapping.1001 *1002 * Folios in this VMA will be aligned to, and at least the size of the1003 * number of bytes returned by this function.1004 *1005 * Return: The default size of the folios allocated when backing a VMA.1006 */1007unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)1008{1009	if (vma->vm_ops && vma->vm_ops->pagesize)1010		return vma->vm_ops->pagesize(vma);1011	return PAGE_SIZE;1012}1013EXPORT_SYMBOL_GPL(vma_kernel_pagesize);1014 1015/*1016 * Return the page size being used by the MMU to back a VMA. In the majority1017 * of cases, the page size used by the kernel matches the MMU size. On1018 * architectures where it differs, an architecture-specific 'strong'1019 * version of this symbol is required.1020 */1021__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)1022{1023	return vma_kernel_pagesize(vma);1024}1025 1026/*1027 * Flags for MAP_PRIVATE reservations.  These are stored in the bottom1028 * bits of the reservation map pointer, which are always clear due to1029 * alignment.1030 */1031#define HPAGE_RESV_OWNER    (1UL << 0)1032#define HPAGE_RESV_UNMAPPED (1UL << 1)1033#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)1034 1035/*1036 * These helpers are used to track how many pages are reserved for1037 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()1038 * is guaranteed to have their future faults succeed.1039 *1040 * With the exception of hugetlb_dup_vma_private() which is called at fork(),1041 * the reserve counters are updated with the hugetlb_lock held. It is safe1042 * to reset the VMA at fork() time as it is not in use yet and there is no1043 * chance of the global counters getting corrupted as a result of the values.1044 *1045 * The private mapping reservation is represented in a subtly different1046 * manner to a shared mapping.  A shared mapping has a region map associated1047 * with the underlying file, this region map represents the backing file1048 * pages which have ever had a reservation assigned which this persists even1049 * after the page is instantiated.  A private mapping has a region map1050 * associated with the original mmap which is attached to all VMAs which1051 * reference it, this region map represents those offsets which have consumed1052 * reservation ie. where pages have been instantiated.1053 */1054static unsigned long get_vma_private_data(struct vm_area_struct *vma)1055{1056	return (unsigned long)vma->vm_private_data;1057}1058 1059static void set_vma_private_data(struct vm_area_struct *vma,1060							unsigned long value)1061{1062	vma->vm_private_data = (void *)value;1063}1064 1065static void1066resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,1067					  struct hugetlb_cgroup *h_cg,1068					  struct hstate *h)1069{1070#ifdef CONFIG_CGROUP_HUGETLB1071	if (!h_cg || !h) {1072		resv_map->reservation_counter = NULL;1073		resv_map->pages_per_hpage = 0;1074		resv_map->css = NULL;1075	} else {1076		resv_map->reservation_counter =1077			&h_cg->rsvd_hugepage[hstate_index(h)];1078		resv_map->pages_per_hpage = pages_per_huge_page(h);1079		resv_map->css = &h_cg->css;1080	}1081#endif1082}1083 1084struct resv_map *resv_map_alloc(void)1085{1086	struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);1087	struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);1088 1089	if (!resv_map || !rg) {1090		kfree(resv_map);1091		kfree(rg);1092		return NULL;1093	}1094 1095	kref_init(&resv_map->refs);1096	spin_lock_init(&resv_map->lock);1097	INIT_LIST_HEAD(&resv_map->regions);1098	init_rwsem(&resv_map->rw_sema);1099 1100	resv_map->adds_in_progress = 0;1101	/*1102	 * Initialize these to 0. On shared mappings, 0's here indicate these1103	 * fields don't do cgroup accounting. On private mappings, these will be1104	 * re-initialized to the proper values, to indicate that hugetlb cgroup1105	 * reservations are to be un-charged from here.1106	 */1107	resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);1108 1109	INIT_LIST_HEAD(&resv_map->region_cache);1110	list_add(&rg->link, &resv_map->region_cache);1111	resv_map->region_cache_count = 1;1112 1113	return resv_map;1114}1115 1116void resv_map_release(struct kref *ref)1117{1118	struct resv_map *resv_map = container_of(ref, struct resv_map, refs);1119	struct list_head *head = &resv_map->region_cache;1120	struct file_region *rg, *trg;1121 1122	/* Clear out any active regions before we release the map. */1123	region_del(resv_map, 0, LONG_MAX);1124 1125	/* ... and any entries left in the cache */1126	list_for_each_entry_safe(rg, trg, head, link) {1127		list_del(&rg->link);1128		kfree(rg);1129	}1130 1131	VM_BUG_ON(resv_map->adds_in_progress);1132 1133	kfree(resv_map);1134}1135 1136static inline struct resv_map *inode_resv_map(struct inode *inode)1137{1138	/*1139	 * At inode evict time, i_mapping may not point to the original1140	 * address space within the inode.  This original address space1141	 * contains the pointer to the resv_map.  So, always use the1142	 * address space embedded within the inode.1143	 * The VERY common case is inode->mapping == &inode->i_data but,1144	 * this may not be true for device special inodes.1145	 */1146	return (struct resv_map *)(&inode->i_data)->i_private_data;1147}1148 1149static struct resv_map *vma_resv_map(struct vm_area_struct *vma)1150{1151	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);1152	if (vma->vm_flags & VM_MAYSHARE) {1153		struct address_space *mapping = vma->vm_file->f_mapping;1154		struct inode *inode = mapping->host;1155 1156		return inode_resv_map(inode);1157 1158	} else {1159		return (struct resv_map *)(get_vma_private_data(vma) &1160							~HPAGE_RESV_MASK);1161	}1162}1163 1164static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)1165{1166	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);1167	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);1168 1169	set_vma_private_data(vma, (unsigned long)map);1170}1171 1172static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)1173{1174	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);1175	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);1176 1177	set_vma_private_data(vma, get_vma_private_data(vma) | flags);1178}1179 1180static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)1181{1182	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);1183 1184	return (get_vma_private_data(vma) & flag) != 0;1185}1186 1187bool __vma_private_lock(struct vm_area_struct *vma)1188{1189	return !(vma->vm_flags & VM_MAYSHARE) &&1190		get_vma_private_data(vma) & ~HPAGE_RESV_MASK &&1191		is_vma_resv_set(vma, HPAGE_RESV_OWNER);1192}1193 1194void hugetlb_dup_vma_private(struct vm_area_struct *vma)1195{1196	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);1197	/*1198	 * Clear vm_private_data1199	 * - For shared mappings this is a per-vma semaphore that may be1200	 *   allocated in a subsequent call to hugetlb_vm_op_open.1201	 *   Before clearing, make sure pointer is not associated with vma1202	 *   as this will leak the structure.  This is the case when called1203	 *   via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already1204	 *   been called to allocate a new structure.1205	 * - For MAP_PRIVATE mappings, this is the reserve map which does1206	 *   not apply to children.  Faults generated by the children are1207	 *   not guaranteed to succeed, even if read-only.1208	 */1209	if (vma->vm_flags & VM_MAYSHARE) {1210		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;1211 1212		if (vma_lock && vma_lock->vma != vma)1213			vma->vm_private_data = NULL;1214	} else1215		vma->vm_private_data = NULL;1216}1217 1218/*1219 * Reset and decrement one ref on hugepage private reservation.1220 * Called with mm->mmap_lock writer semaphore held.1221 * This function should be only used by move_vma() and operate on1222 * same sized vma. It should never come here with last ref on the1223 * reservation.1224 */1225void clear_vma_resv_huge_pages(struct vm_area_struct *vma)1226{1227	/*1228	 * Clear the old hugetlb private page reservation.1229	 * It has already been transferred to new_vma.1230	 *1231	 * During a mremap() operation of a hugetlb vma we call move_vma()1232	 * which copies vma into new_vma and unmaps vma. After the copy1233	 * operation both new_vma and vma share a reference to the resv_map1234	 * struct, and at that point vma is about to be unmapped. We don't1235	 * want to return the reservation to the pool at unmap of vma because1236	 * the reservation still lives on in new_vma, so simply decrement the1237	 * ref here and remove the resv_map reference from this vma.1238	 */1239	struct resv_map *reservations = vma_resv_map(vma);1240 1241	if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {1242		resv_map_put_hugetlb_cgroup_uncharge_info(reservations);1243		kref_put(&reservations->refs, resv_map_release);1244	}1245 1246	hugetlb_dup_vma_private(vma);1247}1248 1249/* Returns true if the VMA has associated reserve pages */1250static bool vma_has_reserves(struct vm_area_struct *vma, long chg)1251{1252	if (vma->vm_flags & VM_NORESERVE) {1253		/*1254		 * This address is already reserved by other process(chg == 0),1255		 * so, we should decrement reserved count. Without decrementing,1256		 * reserve count remains after releasing inode, because this1257		 * allocated page will go into page cache and is regarded as1258		 * coming from reserved pool in releasing step.  Currently, we1259		 * don't have any other solution to deal with this situation1260		 * properly, so add work-around here.1261		 */1262		if (vma->vm_flags & VM_MAYSHARE && chg == 0)1263			return true;1264		else1265			return false;1266	}1267 1268	/* Shared mappings always use reserves */1269	if (vma->vm_flags & VM_MAYSHARE) {1270		/*1271		 * We know VM_NORESERVE is not set.  Therefore, there SHOULD1272		 * be a region map for all pages.  The only situation where1273		 * there is no region map is if a hole was punched via1274		 * fallocate.  In this case, there really are no reserves to1275		 * use.  This situation is indicated if chg != 0.1276		 */1277		if (chg)1278			return false;1279		else1280			return true;1281	}1282 1283	/*1284	 * Only the process that called mmap() has reserves for1285	 * private mappings.1286	 */1287	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {1288		/*1289		 * Like the shared case above, a hole punch or truncate1290		 * could have been performed on the private mapping.1291		 * Examine the value of chg to determine if reserves1292		 * actually exist or were previously consumed.1293		 * Very Subtle - The value of chg comes from a previous1294		 * call to vma_needs_reserves().  The reserve map for1295		 * private mappings has different (opposite) semantics1296		 * than that of shared mappings.  vma_needs_reserves()1297		 * has already taken this difference in semantics into1298		 * account.  Therefore, the meaning of chg is the same1299		 * as in the shared case above.  Code could easily be1300		 * combined, but keeping it separate draws attention to1301		 * subtle differences.1302		 */1303		if (chg)1304			return false;1305		else1306			return true;1307	}1308 1309	return false;1310}1311 1312static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)1313{1314	int nid = folio_nid(folio);1315 1316	lockdep_assert_held(&hugetlb_lock);1317	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);1318 1319	list_move(&folio->lru, &h->hugepage_freelists[nid]);1320	h->free_huge_pages++;1321	h->free_huge_pages_node[nid]++;1322	folio_set_hugetlb_freed(folio);1323}1324 1325static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,1326								int nid)1327{1328	struct folio *folio;1329	bool pin = !!(current->flags & PF_MEMALLOC_PIN);1330 1331	lockdep_assert_held(&hugetlb_lock);1332	list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {1333		if (pin && !folio_is_longterm_pinnable(folio))1334			continue;1335 1336		if (folio_test_hwpoison(folio))1337			continue;1338 1339		list_move(&folio->lru, &h->hugepage_activelist);1340		folio_ref_unfreeze(folio, 1);1341		folio_clear_hugetlb_freed(folio);1342		h->free_huge_pages--;1343		h->free_huge_pages_node[nid]--;1344		return folio;1345	}1346 1347	return NULL;1348}1349 1350static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,1351							int nid, nodemask_t *nmask)1352{1353	unsigned int cpuset_mems_cookie;1354	struct zonelist *zonelist;1355	struct zone *zone;1356	struct zoneref *z;1357	int node = NUMA_NO_NODE;1358 1359	/* 'nid' should not be NUMA_NO_NODE. Try to catch any misuse of it and rectifiy. */1360	if (nid == NUMA_NO_NODE)1361		nid = numa_node_id();1362 1363	zonelist = node_zonelist(nid, gfp_mask);1364 1365retry_cpuset:1366	cpuset_mems_cookie = read_mems_allowed_begin();1367	for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {1368		struct folio *folio;1369 1370		if (!cpuset_zone_allowed(zone, gfp_mask))1371			continue;1372		/*1373		 * no need to ask again on the same node. Pool is node rather than1374		 * zone aware1375		 */1376		if (zone_to_nid(zone) == node)1377			continue;1378		node = zone_to_nid(zone);1379 1380		folio = dequeue_hugetlb_folio_node_exact(h, node);1381		if (folio)1382			return folio;1383	}1384	if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))1385		goto retry_cpuset;1386 1387	return NULL;1388}1389 1390static unsigned long available_huge_pages(struct hstate *h)1391{1392	return h->free_huge_pages - h->resv_huge_pages;1393}1394 1395static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,1396				struct vm_area_struct *vma,1397				unsigned long address, int avoid_reserve,1398				long chg)1399{1400	struct folio *folio = NULL;1401	struct mempolicy *mpol;1402	gfp_t gfp_mask;1403	nodemask_t *nodemask;1404	int nid;1405 1406	/*1407	 * A child process with MAP_PRIVATE mappings created by their parent1408	 * have no page reserves. This check ensures that reservations are1409	 * not "stolen". The child may still get SIGKILLed1410	 */1411	if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))1412		goto err;1413 1414	/* If reserves cannot be used, ensure enough pages are in the pool */1415	if (avoid_reserve && !available_huge_pages(h))1416		goto err;1417 1418	gfp_mask = htlb_alloc_mask(h);1419	nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);1420 1421	if (mpol_is_preferred_many(mpol)) {1422		folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,1423							nid, nodemask);1424 1425		/* Fallback to all nodes if page==NULL */1426		nodemask = NULL;1427	}1428 1429	if (!folio)1430		folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,1431							nid, nodemask);1432 1433	if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {1434		folio_set_hugetlb_restore_reserve(folio);1435		h->resv_huge_pages--;1436	}1437 1438	mpol_cond_put(mpol);1439	return folio;1440 1441err:1442	return NULL;1443}1444 1445/*1446 * common helper functions for hstate_next_node_to_{alloc|free}.1447 * We may have allocated or freed a huge page based on a different1448 * nodes_allowed previously, so h->next_node_to_{alloc|free} might1449 * be outside of *nodes_allowed.  Ensure that we use an allowed1450 * node for alloc or free.1451 */1452static int next_node_allowed(int nid, nodemask_t *nodes_allowed)1453{1454	nid = next_node_in(nid, *nodes_allowed);1455	VM_BUG_ON(nid >= MAX_NUMNODES);1456 1457	return nid;1458}1459 1460static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)1461{1462	if (!node_isset(nid, *nodes_allowed))1463		nid = next_node_allowed(nid, nodes_allowed);1464	return nid;1465}1466 1467/*1468 * returns the previously saved node ["this node"] from which to1469 * allocate a persistent huge page for the pool and advance the1470 * next node from which to allocate, handling wrap at end of node1471 * mask.1472 */1473static int hstate_next_node_to_alloc(int *next_node,1474					nodemask_t *nodes_allowed)1475{1476	int nid;1477 1478	VM_BUG_ON(!nodes_allowed);1479 1480	nid = get_valid_node_allowed(*next_node, nodes_allowed);1481	*next_node = next_node_allowed(nid, nodes_allowed);1482 1483	return nid;1484}1485 1486/*1487 * helper for remove_pool_hugetlb_folio() - return the previously saved1488 * node ["this node"] from which to free a huge page.  Advance the1489 * next node id whether or not we find a free huge page to free so1490 * that the next attempt to free addresses the next node.1491 */1492static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)1493{1494	int nid;1495 1496	VM_BUG_ON(!nodes_allowed);1497 1498	nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);1499	h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);1500 1501	return nid;1502}1503 1504#define for_each_node_mask_to_alloc(next_node, nr_nodes, node, mask)		\1505	for (nr_nodes = nodes_weight(*mask);				\1506		nr_nodes > 0 &&						\1507		((node = hstate_next_node_to_alloc(next_node, mask)) || 1);	\1508		nr_nodes--)1509 1510#define for_each_node_mask_to_free(hs, nr_nodes, node, mask)		\1511	for (nr_nodes = nodes_weight(*mask);				\1512		nr_nodes > 0 &&						\1513		((node = hstate_next_node_to_free(hs, mask)) || 1);	\1514		nr_nodes--)1515 1516#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE1517#ifdef CONFIG_CONTIG_ALLOC1518static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,1519		int nid, nodemask_t *nodemask)1520{1521	struct folio *folio;1522	int order = huge_page_order(h);1523	bool retried = false;1524 1525	if (nid == NUMA_NO_NODE)1526		nid = numa_mem_id();1527retry:1528	folio = NULL;1529#ifdef CONFIG_CMA1530	{1531		int node;1532 1533		if (hugetlb_cma[nid])1534			folio = cma_alloc_folio(hugetlb_cma[nid], order, gfp_mask);1535 1536		if (!folio && !(gfp_mask & __GFP_THISNODE)) {1537			for_each_node_mask(node, *nodemask) {1538				if (node == nid || !hugetlb_cma[node])1539					continue;1540 1541				folio = cma_alloc_folio(hugetlb_cma[node], order, gfp_mask);1542				if (folio)1543					break;1544			}1545		}1546	}1547#endif1548	if (!folio) {1549		folio = folio_alloc_gigantic(order, gfp_mask, nid, nodemask);1550		if (!folio)1551			return NULL;1552	}1553 1554	if (folio_ref_freeze(folio, 1))1555		return folio;1556 1557	pr_warn("HugeTLB: unexpected refcount on PFN %lu\n", folio_pfn(folio));1558	hugetlb_free_folio(folio);1559	if (!retried) {1560		retried = true;1561		goto retry;1562	}1563	return NULL;1564}1565 1566#else /* !CONFIG_CONTIG_ALLOC */1567static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,1568					int nid, nodemask_t *nodemask)1569{1570	return NULL;1571}1572#endif /* CONFIG_CONTIG_ALLOC */1573 1574#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */1575static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,1576					int nid, nodemask_t *nodemask)1577{1578	return NULL;1579}1580#endif1581 1582/*1583 * Remove hugetlb folio from lists.1584 * If vmemmap exists for the folio, clear the hugetlb flag so that the1585 * folio appears as just a compound page.  Otherwise, wait until after1586 * allocating vmemmap to clear the flag.1587 *1588 * Must be called with hugetlb lock held.1589 */1590static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,1591							bool adjust_surplus)1592{1593	int nid = folio_nid(folio);1594 1595	VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);1596	VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);1597 1598	lockdep_assert_held(&hugetlb_lock);1599	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())1600		return;1601 1602	list_del(&folio->lru);1603 1604	if (folio_test_hugetlb_freed(folio)) {1605		folio_clear_hugetlb_freed(folio);1606		h->free_huge_pages--;1607		h->free_huge_pages_node[nid]--;1608	}1609	if (adjust_surplus) {1610		h->surplus_huge_pages--;1611		h->surplus_huge_pages_node[nid]--;1612	}1613 1614	/*1615	 * We can only clear the hugetlb flag after allocating vmemmap1616	 * pages.  Otherwise, someone (memory error handling) may try to write1617	 * to tail struct pages.1618	 */1619	if (!folio_test_hugetlb_vmemmap_optimized(folio))1620		__folio_clear_hugetlb(folio);1621 1622	h->nr_huge_pages--;1623	h->nr_huge_pages_node[nid]--;1624}1625 1626static void add_hugetlb_folio(struct hstate *h, struct folio *folio,1627			     bool adjust_surplus)1628{1629	int nid = folio_nid(folio);1630 1631	VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);1632 1633	lockdep_assert_held(&hugetlb_lock);1634 1635	INIT_LIST_HEAD(&folio->lru);1636	h->nr_huge_pages++;1637	h->nr_huge_pages_node[nid]++;1638 1639	if (adjust_surplus) {1640		h->surplus_huge_pages++;1641		h->surplus_huge_pages_node[nid]++;1642	}1643 1644	__folio_set_hugetlb(folio);1645	folio_change_private(folio, NULL);1646	/*1647	 * We have to set hugetlb_vmemmap_optimized again as above1648	 * folio_change_private(folio, NULL) cleared it.1649	 */1650	folio_set_hugetlb_vmemmap_optimized(folio);1651 1652	arch_clear_hugetlb_flags(folio);1653	enqueue_hugetlb_folio(h, folio);1654}1655 1656static void __update_and_free_hugetlb_folio(struct hstate *h,1657						struct folio *folio)1658{1659	bool clear_flag = folio_test_hugetlb_vmemmap_optimized(folio);1660 1661	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())1662		return;1663 1664	/*1665	 * If we don't know which subpages are hwpoisoned, we can't free1666	 * the hugepage, so it's leaked intentionally.1667	 */1668	if (folio_test_hugetlb_raw_hwp_unreliable(folio))1669		return;1670 1671	/*1672	 * If folio is not vmemmap optimized (!clear_flag), then the folio1673	 * is no longer identified as a hugetlb page.  hugetlb_vmemmap_restore_folio1674	 * can only be passed hugetlb pages and will BUG otherwise.1675	 */1676	if (clear_flag && hugetlb_vmemmap_restore_folio(h, folio)) {1677		spin_lock_irq(&hugetlb_lock);1678		/*1679		 * If we cannot allocate vmemmap pages, just refuse to free the1680		 * page and put the page back on the hugetlb free list and treat1681		 * as a surplus page.1682		 */1683		add_hugetlb_folio(h, folio, true);1684		spin_unlock_irq(&hugetlb_lock);1685		return;1686	}1687 1688	/*1689	 * If vmemmap pages were allocated above, then we need to clear the1690	 * hugetlb flag under the hugetlb lock.1691	 */1692	if (folio_test_hugetlb(folio)) {1693		spin_lock_irq(&hugetlb_lock);1694		__folio_clear_hugetlb(folio);1695		spin_unlock_irq(&hugetlb_lock);1696	}1697 1698	/*1699	 * Move PageHWPoison flag from head page to the raw error pages,1700	 * which makes any healthy subpages reusable.1701	 */1702	if (unlikely(folio_test_hwpoison(folio)))1703		folio_clear_hugetlb_hwpoison(folio);1704 1705	folio_ref_unfreeze(folio, 1);1706 1707	INIT_LIST_HEAD(&folio->_deferred_list);1708	hugetlb_free_folio(folio);1709}1710 1711/*1712 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot1713 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the1714 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate1715 * the vmemmap pages.1716 *1717 * free_hpage_workfn() locklessly retrieves the linked list of pages to be1718 * freed and frees them one-by-one. As the page->mapping pointer is going1719 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node1720 * structure of a lockless linked list of huge pages to be freed.1721 */1722static LLIST_HEAD(hpage_freelist);1723 1724static void free_hpage_workfn(struct work_struct *work)1725{1726	struct llist_node *node;1727 1728	node = llist_del_all(&hpage_freelist);1729 1730	while (node) {1731		struct folio *folio;1732		struct hstate *h;1733 1734		folio = container_of((struct address_space **)node,1735				     struct folio, mapping);1736		node = node->next;1737		folio->mapping = NULL;1738		/*1739		 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in1740		 * folio_hstate() is going to trigger because a previous call to1741		 * remove_hugetlb_folio() will clear the hugetlb bit, so do1742		 * not use folio_hstate() directly.1743		 */1744		h = size_to_hstate(folio_size(folio));1745 1746		__update_and_free_hugetlb_folio(h, folio);1747 1748		cond_resched();1749	}1750}1751static DECLARE_WORK(free_hpage_work, free_hpage_workfn);1752 1753static inline void flush_free_hpage_work(struct hstate *h)1754{1755	if (hugetlb_vmemmap_optimizable(h))1756		flush_work(&free_hpage_work);1757}1758 1759static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,1760				 bool atomic)1761{1762	if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {1763		__update_and_free_hugetlb_folio(h, folio);1764		return;1765	}1766 1767	/*1768	 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.1769	 *1770	 * Only call schedule_work() if hpage_freelist is previously1771	 * empty. Otherwise, schedule_work() had been called but the workfn1772	 * hasn't retrieved the list yet.1773	 */1774	if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))1775		schedule_work(&free_hpage_work);1776}1777 1778static void bulk_vmemmap_restore_error(struct hstate *h,1779					struct list_head *folio_list,1780					struct list_head *non_hvo_folios)1781{1782	struct folio *folio, *t_folio;1783 1784	if (!list_empty(non_hvo_folios)) {1785		/*1786		 * Free any restored hugetlb pages so that restore of the1787		 * entire list can be retried.1788		 * The idea is that in the common case of ENOMEM errors freeing1789		 * hugetlb pages with vmemmap we will free up memory so that we1790		 * can allocate vmemmap for more hugetlb pages.1791		 */1792		list_for_each_entry_safe(folio, t_folio, non_hvo_folios, lru) {1793			list_del(&folio->lru);1794			spin_lock_irq(&hugetlb_lock);1795			__folio_clear_hugetlb(folio);1796			spin_unlock_irq(&hugetlb_lock);1797			update_and_free_hugetlb_folio(h, folio, false);1798			cond_resched();1799		}1800	} else {1801		/*1802		 * In the case where there are no folios which can be1803		 * immediately freed, we loop through the list trying to restore1804		 * vmemmap individually in the hope that someone elsewhere may1805		 * have done something to cause success (such as freeing some1806		 * memory).  If unable to restore a hugetlb page, the hugetlb1807		 * page is made a surplus page and removed from the list.1808		 * If are able to restore vmemmap and free one hugetlb page, we1809		 * quit processing the list to retry the bulk operation.1810		 */1811		list_for_each_entry_safe(folio, t_folio, folio_list, lru)1812			if (hugetlb_vmemmap_restore_folio(h, folio)) {1813				list_del(&folio->lru);1814				spin_lock_irq(&hugetlb_lock);1815				add_hugetlb_folio(h, folio, true);1816				spin_unlock_irq(&hugetlb_lock);1817			} else {1818				list_del(&folio->lru);1819				spin_lock_irq(&hugetlb_lock);1820				__folio_clear_hugetlb(folio);1821				spin_unlock_irq(&hugetlb_lock);1822				update_and_free_hugetlb_folio(h, folio, false);1823				cond_resched();1824				break;1825			}1826	}1827}1828 1829static void update_and_free_pages_bulk(struct hstate *h,1830						struct list_head *folio_list)1831{1832	long ret;1833	struct folio *folio, *t_folio;1834	LIST_HEAD(non_hvo_folios);1835 1836	/*1837	 * First allocate required vmemmmap (if necessary) for all folios.1838	 * Carefully handle errors and free up any available hugetlb pages1839	 * in an effort to make forward progress.1840	 */1841retry:1842	ret = hugetlb_vmemmap_restore_folios(h, folio_list, &non_hvo_folios);1843	if (ret < 0) {1844		bulk_vmemmap_restore_error(h, folio_list, &non_hvo_folios);1845		goto retry;1846	}1847 1848	/*1849	 * At this point, list should be empty, ret should be >= 0 and there1850	 * should only be pages on the non_hvo_folios list.1851	 * Do note that the non_hvo_folios list could be empty.1852	 * Without HVO enabled, ret will be 0 and there is no need to call1853	 * __folio_clear_hugetlb as this was done previously.1854	 */1855	VM_WARN_ON(!list_empty(folio_list));1856	VM_WARN_ON(ret < 0);1857	if (!list_empty(&non_hvo_folios) && ret) {1858		spin_lock_irq(&hugetlb_lock);1859		list_for_each_entry(folio, &non_hvo_folios, lru)1860			__folio_clear_hugetlb(folio);1861		spin_unlock_irq(&hugetlb_lock);1862	}1863 1864	list_for_each_entry_safe(folio, t_folio, &non_hvo_folios, lru) {1865		update_and_free_hugetlb_folio(h, folio, false);1866		cond_resched();1867	}1868}1869 1870struct hstate *size_to_hstate(unsigned long size)1871{1872	struct hstate *h;1873 1874	for_each_hstate(h) {1875		if (huge_page_size(h) == size)1876			return h;1877	}1878	return NULL;1879}1880 1881void free_huge_folio(struct folio *folio)1882{1883	/*1884	 * Can't pass hstate in here because it is called from the1885	 * generic mm code.1886	 */1887	struct hstate *h = folio_hstate(folio);1888	int nid = folio_nid(folio);1889	struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);1890	bool restore_reserve;1891	unsigned long flags;1892 1893	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);1894	VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);1895 1896	hugetlb_set_folio_subpool(folio, NULL);1897	if (folio_test_anon(folio))1898		__ClearPageAnonExclusive(&folio->page);1899	folio->mapping = NULL;1900	restore_reserve = folio_test_hugetlb_restore_reserve(folio);1901	folio_clear_hugetlb_restore_reserve(folio);1902 1903	/*1904	 * If HPageRestoreReserve was set on page, page allocation consumed a1905	 * reservation.  If the page was associated with a subpool, there1906	 * would have been a page reserved in the subpool before allocation1907	 * via hugepage_subpool_get_pages().  Since we are 'restoring' the1908	 * reservation, do not call hugepage_subpool_put_pages() as this will1909	 * remove the reserved page from the subpool.1910	 */1911	if (!restore_reserve) {1912		/*1913		 * A return code of zero implies that the subpool will be1914		 * under its minimum size if the reservation is not restored1915		 * after page is free.  Therefore, force restore_reserve1916		 * operation.1917		 */1918		if (hugepage_subpool_put_pages(spool, 1) == 0)1919			restore_reserve = true;1920	}1921 1922	spin_lock_irqsave(&hugetlb_lock, flags);1923	folio_clear_hugetlb_migratable(folio);1924	hugetlb_cgroup_uncharge_folio(hstate_index(h),1925				     pages_per_huge_page(h), folio);1926	hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),1927					  pages_per_huge_page(h), folio);1928	mem_cgroup_uncharge(folio);1929	if (restore_reserve)1930		h->resv_huge_pages++;1931 1932	if (folio_test_hugetlb_temporary(folio)) {1933		remove_hugetlb_folio(h, folio, false);1934		spin_unlock_irqrestore(&hugetlb_lock, flags);1935		update_and_free_hugetlb_folio(h, folio, true);1936	} else if (h->surplus_huge_pages_node[nid]) {1937		/* remove the page from active list */1938		remove_hugetlb_folio(h, folio, true);1939		spin_unlock_irqrestore(&hugetlb_lock, flags);1940		update_and_free_hugetlb_folio(h, folio, true);1941	} else {1942		arch_clear_hugetlb_flags(folio);1943		enqueue_hugetlb_folio(h, folio);1944		spin_unlock_irqrestore(&hugetlb_lock, flags);1945	}1946}1947 1948/*1949 * Must be called with the hugetlb lock held1950 */1951static void __prep_account_new_huge_page(struct hstate *h, int nid)1952{1953	lockdep_assert_held(&hugetlb_lock);1954	h->nr_huge_pages++;1955	h->nr_huge_pages_node[nid]++;1956}1957 1958static void init_new_hugetlb_folio(struct hstate *h, struct folio *folio)1959{1960	__folio_set_hugetlb(folio);1961	INIT_LIST_HEAD(&folio->lru);1962	hugetlb_set_folio_subpool(folio, NULL);1963	set_hugetlb_cgroup(folio, NULL);1964	set_hugetlb_cgroup_rsvd(folio, NULL);1965}1966 1967static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)1968{1969	init_new_hugetlb_folio(h, folio);1970	hugetlb_vmemmap_optimize_folio(h, folio);1971}1972 1973static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)1974{1975	__prep_new_hugetlb_folio(h, folio);1976	spin_lock_irq(&hugetlb_lock);1977	__prep_account_new_huge_page(h, nid);1978	spin_unlock_irq(&hugetlb_lock);1979}1980 1981/*1982 * Find and lock address space (mapping) in write mode.1983 *1984 * Upon entry, the folio is locked which means that folio_mapping() is1985 * stable.  Due to locking order, we can only trylock_write.  If we can1986 * not get the lock, simply return NULL to caller.1987 */1988struct address_space *hugetlb_folio_mapping_lock_write(struct folio *folio)1989{1990	struct address_space *mapping = folio_mapping(folio);1991 1992	if (!mapping)1993		return mapping;1994 1995	if (i_mmap_trylock_write(mapping))1996		return mapping;1997 1998	return NULL;1999}2000 2001static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,2002		gfp_t gfp_mask, int nid, nodemask_t *nmask,2003		nodemask_t *node_alloc_noretry)2004{2005	int order = huge_page_order(h);2006	struct folio *folio;2007	bool alloc_try_hard = true;2008	bool retry = true;2009 2010	/*2011	 * By default we always try hard to allocate the folio with2012	 * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating folios in2013	 * a loop (to adjust global huge page counts) and previous allocation2014	 * failed, do not continue to try hard on the same node.  Use the2015	 * node_alloc_noretry bitmap to manage this state information.2016	 */2017	if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))2018		alloc_try_hard = false;2019	if (alloc_try_hard)2020		gfp_mask |= __GFP_RETRY_MAYFAIL;2021	if (nid == NUMA_NO_NODE)2022		nid = numa_mem_id();2023retry:2024	folio = __folio_alloc(gfp_mask, order, nid, nmask);2025	/* Ensure hugetlb folio won't have large_rmappable flag set. */2026	if (folio)2027		folio_clear_large_rmappable(folio);2028 2029	if (folio && !folio_ref_freeze(folio, 1)) {2030		folio_put(folio);2031		if (retry) {	/* retry once */2032			retry = false;2033			goto retry;2034		}2035		/* WOW!  twice in a row. */2036		pr_warn("HugeTLB unexpected inflated folio ref count\n");2037		folio = NULL;2038	}2039 2040	/*2041	 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a2042	 * folio this indicates an overall state change.  Clear bit so2043	 * that we resume normal 'try hard' allocations.2044	 */2045	if (node_alloc_noretry && folio && !alloc_try_hard)2046		node_clear(nid, *node_alloc_noretry);2047 2048	/*2049	 * If we tried hard to get a folio but failed, set bit so that2050	 * subsequent attempts will not try as hard until there is an2051	 * overall state change.2052	 */2053	if (node_alloc_noretry && !folio && alloc_try_hard)2054		node_set(nid, *node_alloc_noretry);2055 2056	if (!folio) {2057		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);2058		return NULL;2059	}2060 2061	__count_vm_event(HTLB_BUDDY_PGALLOC);2062	return folio;2063}2064 2065static struct folio *only_alloc_fresh_hugetlb_folio(struct hstate *h,2066		gfp_t gfp_mask, int nid, nodemask_t *nmask,2067		nodemask_t *node_alloc_noretry)2068{2069	struct folio *folio;2070 2071	if (hstate_is_gigantic(h))2072		folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);2073	else2074		folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, nmask, node_alloc_noretry);2075	if (folio)2076		init_new_hugetlb_folio(h, folio);2077	return folio;2078}2079 2080/*2081 * Common helper to allocate a fresh hugetlb page. All specific allocators2082 * should use this function to get new hugetlb pages2083 *2084 * Note that returned page is 'frozen':  ref count of head page and all tail2085 * pages is zero.2086 */2087static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,2088		gfp_t gfp_mask, int nid, nodemask_t *nmask)2089{2090	struct folio *folio;2091 2092	if (hstate_is_gigantic(h))2093		folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);2094	else2095		folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);2096	if (!folio)2097		return NULL;2098 2099	prep_new_hugetlb_folio(h, folio, folio_nid(folio));2100	return folio;2101}2102 2103static void prep_and_add_allocated_folios(struct hstate *h,2104					struct list_head *folio_list)2105{2106	unsigned long flags;2107	struct folio *folio, *tmp_f;2108 2109	/* Send list for bulk vmemmap optimization processing */2110	hugetlb_vmemmap_optimize_folios(h, folio_list);2111 2112	/* Add all new pool pages to free lists in one lock cycle */2113	spin_lock_irqsave(&hugetlb_lock, flags);2114	list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {2115		__prep_account_new_huge_page(h, folio_nid(folio));2116		enqueue_hugetlb_folio(h, folio);2117	}2118	spin_unlock_irqrestore(&hugetlb_lock, flags);2119}2120 2121/*2122 * Allocates a fresh hugetlb page in a node interleaved manner.  The page2123 * will later be added to the appropriate hugetlb pool.2124 */2125static struct folio *alloc_pool_huge_folio(struct hstate *h,2126					nodemask_t *nodes_allowed,2127					nodemask_t *node_alloc_noretry,2128					int *next_node)2129{2130	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;2131	int nr_nodes, node;2132 2133	for_each_node_mask_to_alloc(next_node, nr_nodes, node, nodes_allowed) {2134		struct folio *folio;2135 2136		folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, node,2137					nodes_allowed, node_alloc_noretry);2138		if (folio)2139			return folio;2140	}2141 2142	return NULL;2143}2144 2145/*2146 * Remove huge page from pool from next node to free.  Attempt to keep2147 * persistent huge pages more or less balanced over allowed nodes.2148 * This routine only 'removes' the hugetlb page.  The caller must make2149 * an additional call to free the page to low level allocators.2150 * Called with hugetlb_lock locked.2151 */2152static struct folio *remove_pool_hugetlb_folio(struct hstate *h,2153		nodemask_t *nodes_allowed, bool acct_surplus)2154{2155	int nr_nodes, node;2156	struct folio *folio = NULL;2157 2158	lockdep_assert_held(&hugetlb_lock);2159	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {2160		/*2161		 * If we're returning unused surplus pages, only examine2162		 * nodes with surplus pages.2163		 */2164		if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&2165		    !list_empty(&h->hugepage_freelists[node])) {2166			folio = list_entry(h->hugepage_freelists[node].next,2167					  struct folio, lru);2168			remove_hugetlb_folio(h, folio, acct_surplus);2169			break;2170		}2171	}2172 2173	return folio;2174}2175 2176/*2177 * Dissolve a given free hugetlb folio into free buddy pages. This function2178 * does nothing for in-use hugetlb folios and non-hugetlb folios.2179 * This function returns values like below:2180 *2181 *  -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages2182 *           when the system is under memory pressure and the feature of2183 *           freeing unused vmemmap pages associated with each hugetlb page2184 *           is enabled.2185 *  -EBUSY:  failed to dissolved free hugepages or the hugepage is in-use2186 *           (allocated or reserved.)2187 *       0:  successfully dissolved free hugepages or the page is not a2188 *           hugepage (considered as already dissolved)2189 */2190int dissolve_free_hugetlb_folio(struct folio *folio)2191{2192	int rc = -EBUSY;2193 2194retry:2195	/* Not to disrupt normal path by vainly holding hugetlb_lock */2196	if (!folio_test_hugetlb(folio))2197		return 0;2198 2199	spin_lock_irq(&hugetlb_lock);2200	if (!folio_test_hugetlb(folio)) {2201		rc = 0;2202		goto out;2203	}2204 2205	if (!folio_ref_count(folio)) {2206		struct hstate *h = folio_hstate(folio);2207		if (!available_huge_pages(h))2208			goto out;2209 2210		/*2211		 * We should make sure that the page is already on the free list2212		 * when it is dissolved.2213		 */2214		if (unlikely(!folio_test_hugetlb_freed(folio))) {2215			spin_unlock_irq(&hugetlb_lock);2216			cond_resched();2217 2218			/*2219			 * Theoretically, we should return -EBUSY when we2220			 * encounter this race. In fact, we have a chance2221			 * to successfully dissolve the page if we do a2222			 * retry. Because the race window is quite small.2223			 * If we seize this opportunity, it is an optimization2224			 * for increasing the success rate of dissolving page.2225			 */2226			goto retry;2227		}2228 2229		remove_hugetlb_folio(h, folio, false);2230		h->max_huge_pages--;2231		spin_unlock_irq(&hugetlb_lock);2232 2233		/*2234		 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap2235		 * before freeing the page.  update_and_free_hugtlb_folio will fail to2236		 * free the page if it can not allocate required vmemmap.  We2237		 * need to adjust max_huge_pages if the page is not freed.2238		 * Attempt to allocate vmemmmap here so that we can take2239		 * appropriate action on failure.2240		 *2241		 * The folio_test_hugetlb check here is because2242		 * remove_hugetlb_folio will clear hugetlb folio flag for2243		 * non-vmemmap optimized hugetlb folios.2244		 */2245		if (folio_test_hugetlb(folio)) {2246			rc = hugetlb_vmemmap_restore_folio(h, folio);2247			if (rc) {2248				spin_lock_irq(&hugetlb_lock);2249				add_hugetlb_folio(h, folio, false);2250				h->max_huge_pages++;2251				goto out;2252			}2253		} else2254			rc = 0;2255 2256		update_and_free_hugetlb_folio(h, folio, false);2257		return rc;2258	}2259out:2260	spin_unlock_irq(&hugetlb_lock);2261	return rc;2262}2263 2264/*2265 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to2266 * make specified memory blocks removable from the system.2267 * Note that this will dissolve a free gigantic hugepage completely, if any2268 * part of it lies within the given range.2269 * Also note that if dissolve_free_hugetlb_folio() returns with an error, all2270 * free hugetlb folios that were dissolved before that error are lost.2271 */2272int dissolve_free_hugetlb_folios(unsigned long start_pfn, unsigned long end_pfn)2273{2274	unsigned long pfn;2275	struct folio *folio;2276	int rc = 0;2277	unsigned int order;2278	struct hstate *h;2279 2280	if (!hugepages_supported())2281		return rc;2282 2283	order = huge_page_order(&default_hstate);2284	for_each_hstate(h)2285		order = min(order, huge_page_order(h));2286 2287	for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {2288		folio = pfn_folio(pfn);2289		rc = dissolve_free_hugetlb_folio(folio);2290		if (rc)2291			break;2292	}2293 2294	return rc;2295}2296 2297/*2298 * Allocates a fresh surplus page from the page allocator.2299 */2300static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,2301				gfp_t gfp_mask,	int nid, nodemask_t *nmask)2302{2303	struct folio *folio = NULL;2304 2305	if (hstate_is_gigantic(h))2306		return NULL;2307 2308	spin_lock_irq(&hugetlb_lock);2309	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)2310		goto out_unlock;2311	spin_unlock_irq(&hugetlb_lock);2312 2313	folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask);2314	if (!folio)2315		return NULL;2316 2317	spin_lock_irq(&hugetlb_lock);2318	/*2319	 * We could have raced with the pool size change.2320	 * Double check that and simply deallocate the new page2321	 * if we would end up overcommiting the surpluses. Abuse2322	 * temporary page to workaround the nasty free_huge_folio2323	 * codeflow2324	 */2325	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {2326		folio_set_hugetlb_temporary(folio);2327		spin_unlock_irq(&hugetlb_lock);2328		free_huge_folio(folio);2329		return NULL;2330	}2331 2332	h->surplus_huge_pages++;2333	h->surplus_huge_pages_node[folio_nid(folio)]++;2334 2335out_unlock:2336	spin_unlock_irq(&hugetlb_lock);2337 2338	return folio;2339}2340 2341static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,2342				     int nid, nodemask_t *nmask)2343{2344	struct folio *folio;2345 2346	if (hstate_is_gigantic(h))2347		return NULL;2348 2349	folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask);2350	if (!folio)2351		return NULL;2352 2353	/* fresh huge pages are frozen */2354	folio_ref_unfreeze(folio, 1);2355	/*2356	 * We do not account these pages as surplus because they are only2357	 * temporary and will be released properly on the last reference2358	 */2359	folio_set_hugetlb_temporary(folio);2360 2361	return folio;2362}2363 2364/*2365 * Use the VMA's mpolicy to allocate a huge page from the buddy.2366 */2367static2368struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,2369		struct vm_area_struct *vma, unsigned long addr)2370{2371	struct folio *folio = NULL;2372	struct mempolicy *mpol;2373	gfp_t gfp_mask = htlb_alloc_mask(h);2374	int nid;2375	nodemask_t *nodemask;2376 2377	nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);2378	if (mpol_is_preferred_many(mpol)) {2379		gfp_t gfp = gfp_mask & ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);2380 2381		folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);2382 2383		/* Fallback to all nodes if page==NULL */2384		nodemask = NULL;2385	}2386 2387	if (!folio)2388		folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);2389	mpol_cond_put(mpol);2390	return folio;2391}2392 2393struct folio *alloc_hugetlb_folio_reserve(struct hstate *h, int preferred_nid,2394		nodemask_t *nmask, gfp_t gfp_mask)2395{2396	struct folio *folio;2397 2398	spin_lock_irq(&hugetlb_lock);2399	folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, preferred_nid,2400					       nmask);2401	if (folio) {2402		VM_BUG_ON(!h->resv_huge_pages);2403		h->resv_huge_pages--;2404	}2405 2406	spin_unlock_irq(&hugetlb_lock);2407	return folio;2408}2409 2410/* folio migration callback function */2411struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,2412		nodemask_t *nmask, gfp_t gfp_mask, bool allow_alloc_fallback)2413{2414	spin_lock_irq(&hugetlb_lock);2415	if (available_huge_pages(h)) {2416		struct folio *folio;2417 2418		folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,2419						preferred_nid, nmask);2420		if (folio) {2421			spin_unlock_irq(&hugetlb_lock);2422			return folio;2423		}2424	}2425	spin_unlock_irq(&hugetlb_lock);2426 2427	/* We cannot fallback to other nodes, as we could break the per-node pool. */2428	if (!allow_alloc_fallback)2429		gfp_mask |= __GFP_THISNODE;2430 2431	return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);2432}2433 2434static nodemask_t *policy_mbind_nodemask(gfp_t gfp)2435{2436#ifdef CONFIG_NUMA2437	struct mempolicy *mpol = get_task_policy(current);2438 2439	/*2440	 * Only enforce MPOL_BIND policy which overlaps with cpuset policy2441	 * (from policy_nodemask) specifically for hugetlb case2442	 */2443	if (mpol->mode == MPOL_BIND &&2444		(apply_policy_zone(mpol, gfp_zone(gfp)) &&2445		 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))2446		return &mpol->nodes;2447#endif2448	return NULL;2449}2450 2451/*2452 * Increase the hugetlb pool such that it can accommodate a reservation2453 * of size 'delta'.2454 */2455static int gather_surplus_pages(struct hstate *h, long delta)2456	__must_hold(&hugetlb_lock)2457{2458	LIST_HEAD(surplus_list);2459	struct folio *folio, *tmp;2460	int ret;2461	long i;2462	long needed, allocated;2463	bool alloc_ok = true;2464	int node;2465	nodemask_t *mbind_nodemask = policy_mbind_nodemask(htlb_alloc_mask(h));2466 2467	lockdep_assert_held(&hugetlb_lock);2468	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;2469	if (needed <= 0) {2470		h->resv_huge_pages += delta;2471		return 0;2472	}2473 2474	allocated = 0;2475 2476	ret = -ENOMEM;2477retry:2478	spin_unlock_irq(&hugetlb_lock);2479	for (i = 0; i < needed; i++) {2480		folio = NULL;2481		for_each_node_mask(node, cpuset_current_mems_allowed) {2482			if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) {2483				folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),2484						node, NULL);2485				if (folio)2486					break;2487			}2488		}2489		if (!folio) {2490			alloc_ok = false;2491			break;2492		}2493		list_add(&folio->lru, &surplus_list);2494		cond_resched();2495	}2496	allocated += i;2497 2498	/*2499	 * After retaking hugetlb_lock, we need to recalculate 'needed'2500	 * because either resv_huge_pages or free_huge_pages may have changed.2501	 */2502	spin_lock_irq(&hugetlb_lock);2503	needed = (h->resv_huge_pages + delta) -2504			(h->free_huge_pages + allocated);2505	if (needed > 0) {2506		if (alloc_ok)2507			goto retry;2508		/*2509		 * We were not able to allocate enough pages to2510		 * satisfy the entire reservation so we free what2511		 * we've allocated so far.2512		 */2513		goto free;2514	}2515	/*2516	 * The surplus_list now contains _at_least_ the number of extra pages2517	 * needed to accommodate the reservation.  Add the appropriate number2518	 * of pages to the hugetlb pool and free the extras back to the buddy2519	 * allocator.  Commit the entire reservation here to prevent another2520	 * process from stealing the pages as they are added to the pool but2521	 * before they are reserved.2522	 */2523	needed += allocated;2524	h->resv_huge_pages += delta;2525	ret = 0;2526 2527	/* Free the needed pages to the hugetlb pool */2528	list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {2529		if ((--needed) < 0)2530			break;2531		/* Add the page to the hugetlb allocator */2532		enqueue_hugetlb_folio(h, folio);2533	}2534free:2535	spin_unlock_irq(&hugetlb_lock);2536 2537	/*2538	 * Free unnecessary surplus pages to the buddy allocator.2539	 * Pages have no ref count, call free_huge_folio directly.2540	 */2541	list_for_each_entry_safe(folio, tmp, &surplus_list, lru)2542		free_huge_folio(folio);2543	spin_lock_irq(&hugetlb_lock);2544 2545	return ret;2546}2547 2548/*2549 * This routine has two main purposes:2550 * 1) Decrement the reservation count (resv_huge_pages) by the value passed2551 *    in unused_resv_pages.  This corresponds to the prior adjustments made2552 *    to the associated reservation map.2553 * 2) Free any unused surplus pages that may have been allocated to satisfy2554 *    the reservation.  As many as unused_resv_pages may be freed.2555 */2556static void return_unused_surplus_pages(struct hstate *h,2557					unsigned long unused_resv_pages)2558{2559	unsigned long nr_pages;2560	LIST_HEAD(page_list);2561 2562	lockdep_assert_held(&hugetlb_lock);2563	/* Uncommit the reservation */2564	h->resv_huge_pages -= unused_resv_pages;2565 2566	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())2567		goto out;2568 2569	/*2570	 * Part (or even all) of the reservation could have been backed2571	 * by pre-allocated pages. Only free surplus pages.2572	 */2573	nr_pages = min(unused_resv_pages, h->surplus_huge_pages);2574 2575	/*2576	 * We want to release as many surplus pages as possible, spread2577	 * evenly across all nodes with memory. Iterate across these nodes2578	 * until we can no longer free unreserved surplus pages. This occurs2579	 * when the nodes with surplus pages have no free pages.2580	 * remove_pool_hugetlb_folio() will balance the freed pages across the2581	 * on-line nodes with memory and will handle the hstate accounting.2582	 */2583	while (nr_pages--) {2584		struct folio *folio;2585 2586		folio = remove_pool_hugetlb_folio(h, &node_states[N_MEMORY], 1);2587		if (!folio)2588			goto out;2589 2590		list_add(&folio->lru, &page_list);2591	}2592 2593out:2594	spin_unlock_irq(&hugetlb_lock);2595	update_and_free_pages_bulk(h, &page_list);2596	spin_lock_irq(&hugetlb_lock);2597}2598 2599 2600/*2601 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation2602 * are used by the huge page allocation routines to manage reservations.2603 *2604 * vma_needs_reservation is called to determine if the huge page at addr2605 * within the vma has an associated reservation.  If a reservation is2606 * needed, the value 1 is returned.  The caller is then responsible for2607 * managing the global reservation and subpool usage counts.  After2608 * the huge page has been allocated, vma_commit_reservation is called2609 * to add the page to the reservation map.  If the page allocation fails,2610 * the reservation must be ended instead of committed.  vma_end_reservation2611 * is called in such cases.2612 *2613 * In the normal case, vma_commit_reservation returns the same value2614 * as the preceding vma_needs_reservation call.  The only time this2615 * is not the case is if a reserve map was changed between calls.  It2616 * is the responsibility of the caller to notice the difference and2617 * take appropriate action.2618 *2619 * vma_add_reservation is used in error paths where a reservation must2620 * be restored when a newly allocated huge page must be freed.  It is2621 * to be called after calling vma_needs_reservation to determine if a2622 * reservation exists.2623 *2624 * vma_del_reservation is used in error paths where an entry in the reserve2625 * map was created during huge page allocation and must be removed.  It is to2626 * be called after calling vma_needs_reservation to determine if a reservation2627 * exists.2628 */2629enum vma_resv_mode {2630	VMA_NEEDS_RESV,2631	VMA_COMMIT_RESV,2632	VMA_END_RESV,2633	VMA_ADD_RESV,2634	VMA_DEL_RESV,2635};2636static long __vma_reservation_common(struct hstate *h,2637				struct vm_area_struct *vma, unsigned long addr,2638				enum vma_resv_mode mode)2639{2640	struct resv_map *resv;2641	pgoff_t idx;2642	long ret;2643	long dummy_out_regions_needed;2644 2645	resv = vma_resv_map(vma);2646	if (!resv)2647		return 1;2648 2649	idx = vma_hugecache_offset(h, vma, addr);2650	switch (mode) {2651	case VMA_NEEDS_RESV:2652		ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);2653		/* We assume that vma_reservation_* routines always operate on2654		 * 1 page, and that adding to resv map a 1 page entry can only2655		 * ever require 1 region.2656		 */2657		VM_BUG_ON(dummy_out_regions_needed != 1);2658		break;2659	case VMA_COMMIT_RESV:2660		ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);2661		/* region_add calls of range 1 should never fail. */2662		VM_BUG_ON(ret < 0);2663		break;2664	case VMA_END_RESV:2665		region_abort(resv, idx, idx + 1, 1);2666		ret = 0;2667		break;2668	case VMA_ADD_RESV:2669		if (vma->vm_flags & VM_MAYSHARE) {2670			ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);2671			/* region_add calls of range 1 should never fail. */2672			VM_BUG_ON(ret < 0);2673		} else {2674			region_abort(resv, idx, idx + 1, 1);2675			ret = region_del(resv, idx, idx + 1);2676		}2677		break;2678	case VMA_DEL_RESV:2679		if (vma->vm_flags & VM_MAYSHARE) {2680			region_abort(resv, idx, idx + 1, 1);2681			ret = region_del(resv, idx, idx + 1);2682		} else {2683			ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);2684			/* region_add calls of range 1 should never fail. */2685			VM_BUG_ON(ret < 0);2686		}2687		break;2688	default:2689		BUG();2690	}2691 2692	if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)2693		return ret;2694	/*2695	 * We know private mapping must have HPAGE_RESV_OWNER set.2696	 *2697	 * In most cases, reserves always exist for private mappings.2698	 * However, a file associated with mapping could have been2699	 * hole punched or truncated after reserves were consumed.2700	 * As subsequent fault on such a range will not use reserves.2701	 * Subtle - The reserve map for private mappings has the2702	 * opposite meaning than that of shared mappings.  If NO2703	 * entry is in the reserve map, it means a reservation exists.2704	 * If an entry exists in the reserve map, it means the2705	 * reservation has already been consumed.  As a result, the2706	 * return value of this routine is the opposite of the2707	 * value returned from reserve map manipulation routines above.2708	 */2709	if (ret > 0)2710		return 0;2711	if (ret == 0)2712		return 1;2713	return ret;2714}2715 2716static long vma_needs_reservation(struct hstate *h,2717			struct vm_area_struct *vma, unsigned long addr)2718{2719	return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);2720}2721 2722static long vma_commit_reservation(struct hstate *h,2723			struct vm_area_struct *vma, unsigned long addr)2724{2725	return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);2726}2727 2728static void vma_end_reservation(struct hstate *h,2729			struct vm_area_struct *vma, unsigned long addr)2730{2731	(void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);2732}2733 2734static long vma_add_reservation(struct hstate *h,2735			struct vm_area_struct *vma, unsigned long addr)2736{2737	return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);2738}2739 2740static long vma_del_reservation(struct hstate *h,2741			struct vm_area_struct *vma, unsigned long addr)2742{2743	return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);2744}2745 2746/*2747 * This routine is called to restore reservation information on error paths.2748 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),2749 * and the hugetlb mutex should remain held when calling this routine.2750 *2751 * It handles two specific cases:2752 * 1) A reservation was in place and the folio consumed the reservation.2753 *    hugetlb_restore_reserve is set in the folio.2754 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is2755 *    not set.  However, alloc_hugetlb_folio always updates the reserve map.2756 *2757 * In case 1, free_huge_folio later in the error path will increment the2758 * global reserve count.  But, free_huge_folio does not have enough context2759 * to adjust the reservation map.  This case deals primarily with private2760 * mappings.  Adjust the reserve map here to be consistent with global2761 * reserve count adjustments to be made by free_huge_folio.  Make sure the2762 * reserve map indicates there is a reservation present.2763 *2764 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.2765 */2766void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,2767			unsigned long address, struct folio *folio)2768{2769	long rc = vma_needs_reservation(h, vma, address);2770 2771	if (folio_test_hugetlb_restore_reserve(folio)) {2772		if (unlikely(rc < 0))2773			/*2774			 * Rare out of memory condition in reserve map2775			 * manipulation.  Clear hugetlb_restore_reserve so2776			 * that global reserve count will not be incremented2777			 * by free_huge_folio.  This will make it appear2778			 * as though the reservation for this folio was2779			 * consumed.  This may prevent the task from2780			 * faulting in the folio at a later time.  This2781			 * is better than inconsistent global huge page2782			 * accounting of reserve counts.2783			 */2784			folio_clear_hugetlb_restore_reserve(folio);2785		else if (rc)2786			(void)vma_add_reservation(h, vma, address);2787		else2788			vma_end_reservation(h, vma, address);2789	} else {2790		if (!rc) {2791			/*2792			 * This indicates there is an entry in the reserve map2793			 * not added by alloc_hugetlb_folio.  We know it was added2794			 * before the alloc_hugetlb_folio call, otherwise2795			 * hugetlb_restore_reserve would be set on the folio.2796			 * Remove the entry so that a subsequent allocation2797			 * does not consume a reservation.2798			 */2799			rc = vma_del_reservation(h, vma, address);2800			if (rc < 0)2801				/*2802				 * VERY rare out of memory condition.  Since2803				 * we can not delete the entry, set2804				 * hugetlb_restore_reserve so that the reserve2805				 * count will be incremented when the folio2806				 * is freed.  This reserve will be consumed2807				 * on a subsequent allocation.2808				 */2809				folio_set_hugetlb_restore_reserve(folio);2810		} else if (rc < 0) {2811			/*2812			 * Rare out of memory condition from2813			 * vma_needs_reservation call.  Memory allocation is2814			 * only attempted if a new entry is needed.  Therefore,2815			 * this implies there is not an entry in the2816			 * reserve map.2817			 *2818			 * For shared mappings, no entry in the map indicates2819			 * no reservation.  We are done.2820			 */2821			if (!(vma->vm_flags & VM_MAYSHARE))2822				/*2823				 * For private mappings, no entry indicates2824				 * a reservation is present.  Since we can2825				 * not add an entry, set hugetlb_restore_reserve2826				 * on the folio so reserve count will be2827				 * incremented when freed.  This reserve will2828				 * be consumed on a subsequent allocation.2829				 */2830				folio_set_hugetlb_restore_reserve(folio);2831		} else2832			/*2833			 * No reservation present, do nothing2834			 */2835			 vma_end_reservation(h, vma, address);2836	}2837}2838 2839/*2840 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve2841 * the old one2842 * @h: struct hstate old page belongs to2843 * @old_folio: Old folio to dissolve2844 * @list: List to isolate the page in case we need to2845 * Returns 0 on success, otherwise negated error.2846 */2847static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,2848			struct folio *old_folio, struct list_head *list)2849{2850	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;2851	int nid = folio_nid(old_folio);2852	struct folio *new_folio = NULL;2853	int ret = 0;2854 2855retry:2856	spin_lock_irq(&hugetlb_lock);2857	if (!folio_test_hugetlb(old_folio)) {2858		/*2859		 * Freed from under us. Drop new_folio too.2860		 */2861		goto free_new;2862	} else if (folio_ref_count(old_folio)) {2863		bool isolated;2864 2865		/*2866		 * Someone has grabbed the folio, try to isolate it here.2867		 * Fail with -EBUSY if not possible.2868		 */2869		spin_unlock_irq(&hugetlb_lock);2870		isolated = isolate_hugetlb(old_folio, list);2871		ret = isolated ? 0 : -EBUSY;2872		spin_lock_irq(&hugetlb_lock);2873		goto free_new;2874	} else if (!folio_test_hugetlb_freed(old_folio)) {2875		/*2876		 * Folio's refcount is 0 but it has not been enqueued in the2877		 * freelist yet. Race window is small, so we can succeed here if2878		 * we retry.2879		 */2880		spin_unlock_irq(&hugetlb_lock);2881		cond_resched();2882		goto retry;2883	} else {2884		if (!new_folio) {2885			spin_unlock_irq(&hugetlb_lock);2886			new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid,2887							      NULL, NULL);2888			if (!new_folio)2889				return -ENOMEM;2890			__prep_new_hugetlb_folio(h, new_folio);2891			goto retry;2892		}2893 2894		/*2895		 * Ok, old_folio is still a genuine free hugepage. Remove it from2896		 * the freelist and decrease the counters. These will be2897		 * incremented again when calling __prep_account_new_huge_page()2898		 * and enqueue_hugetlb_folio() for new_folio. The counters will2899		 * remain stable since this happens under the lock.2900		 */2901		remove_hugetlb_folio(h, old_folio, false);2902 2903		/*2904		 * Ref count on new_folio is already zero as it was dropped2905		 * earlier.  It can be directly added to the pool free list.2906		 */2907		__prep_account_new_huge_page(h, nid);2908		enqueue_hugetlb_folio(h, new_folio);2909 2910		/*2911		 * Folio has been replaced, we can safely free the old one.2912		 */2913		spin_unlock_irq(&hugetlb_lock);2914		update_and_free_hugetlb_folio(h, old_folio, false);2915	}2916 2917	return ret;2918 2919free_new:2920	spin_unlock_irq(&hugetlb_lock);2921	if (new_folio)2922		update_and_free_hugetlb_folio(h, new_folio, false);2923 2924	return ret;2925}2926 2927int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)2928{2929	struct hstate *h;2930	struct folio *folio = page_folio(page);2931	int ret = -EBUSY;2932 2933	/*2934	 * The page might have been dissolved from under our feet, so make sure2935	 * to carefully check the state under the lock.2936	 * Return success when racing as if we dissolved the page ourselves.2937	 */2938	spin_lock_irq(&hugetlb_lock);2939	if (folio_test_hugetlb(folio)) {2940		h = folio_hstate(folio);2941	} else {2942		spin_unlock_irq(&hugetlb_lock);2943		return 0;2944	}2945	spin_unlock_irq(&hugetlb_lock);2946 2947	/*2948	 * Fence off gigantic pages as there is a cyclic dependency between2949	 * alloc_contig_range and them. Return -ENOMEM as this has the effect2950	 * of bailing out right away without further retrying.2951	 */2952	if (hstate_is_gigantic(h))2953		return -ENOMEM;2954 2955	if (folio_ref_count(folio) && isolate_hugetlb(folio, list))2956		ret = 0;2957	else if (!folio_ref_count(folio))2958		ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);2959 2960	return ret;2961}2962 2963struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,2964				    unsigned long addr, int avoid_reserve)2965{2966	struct hugepage_subpool *spool = subpool_vma(vma);2967	struct hstate *h = hstate_vma(vma);2968	struct folio *folio;2969	long map_chg, map_commit, nr_pages = pages_per_huge_page(h);2970	long gbl_chg;2971	int memcg_charge_ret, ret, idx;2972	struct hugetlb_cgroup *h_cg = NULL;2973	struct mem_cgroup *memcg;2974	bool deferred_reserve;2975	gfp_t gfp = htlb_alloc_mask(h) | __GFP_RETRY_MAYFAIL;2976 2977	memcg = get_mem_cgroup_from_current();2978	memcg_charge_ret = mem_cgroup_hugetlb_try_charge(memcg, gfp, nr_pages);2979	if (memcg_charge_ret == -ENOMEM) {2980		mem_cgroup_put(memcg);2981		return ERR_PTR(-ENOMEM);2982	}2983 2984	idx = hstate_index(h);2985	/*2986	 * Examine the region/reserve map to determine if the process2987	 * has a reservation for the page to be allocated.  A return2988	 * code of zero indicates a reservation exists (no change).2989	 */2990	map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);2991	if (map_chg < 0) {2992		if (!memcg_charge_ret)2993			mem_cgroup_cancel_charge(memcg, nr_pages);2994		mem_cgroup_put(memcg);2995		return ERR_PTR(-ENOMEM);2996	}2997 2998	/*2999	 * Processes that did not create the mapping will have no3000	 * reserves as indicated by the region/reserve map. Check3001	 * that the allocation will not exceed the subpool limit.3002	 * Allocations for MAP_NORESERVE mappings also need to be3003	 * checked against any subpool limit.3004	 */3005	if (map_chg || avoid_reserve) {3006		gbl_chg = hugepage_subpool_get_pages(spool, 1);3007		if (gbl_chg < 0)3008			goto out_end_reservation;3009 3010		/*3011		 * Even though there was no reservation in the region/reserve3012		 * map, there could be reservations associated with the3013		 * subpool that can be used.  This would be indicated if the3014		 * return value of hugepage_subpool_get_pages() is zero.3015		 * However, if avoid_reserve is specified we still avoid even3016		 * the subpool reservations.3017		 */3018		if (avoid_reserve)3019			gbl_chg = 1;3020	}3021 3022	/* If this allocation is not consuming a reservation, charge it now.3023	 */3024	deferred_reserve = map_chg || avoid_reserve;3025	if (deferred_reserve) {3026		ret = hugetlb_cgroup_charge_cgroup_rsvd(3027			idx, pages_per_huge_page(h), &h_cg);3028		if (ret)3029			goto out_subpool_put;3030	}3031 3032	ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);3033	if (ret)3034		goto out_uncharge_cgroup_reservation;3035 3036	spin_lock_irq(&hugetlb_lock);3037	/*3038	 * glb_chg is passed to indicate whether or not a page must be taken3039	 * from the global free pool (global change).  gbl_chg == 0 indicates3040	 * a reservation exists for the allocation.3041	 */3042	folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);3043	if (!folio) {3044		spin_unlock_irq(&hugetlb_lock);3045		folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);3046		if (!folio)3047			goto out_uncharge_cgroup;3048		spin_lock_irq(&hugetlb_lock);3049		if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {3050			folio_set_hugetlb_restore_reserve(folio);3051			h->resv_huge_pages--;3052		}3053		list_add(&folio->lru, &h->hugepage_activelist);3054		folio_ref_unfreeze(folio, 1);3055		/* Fall through */3056	}3057 3058	hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);3059	/* If allocation is not consuming a reservation, also store the3060	 * hugetlb_cgroup pointer on the page.3061	 */3062	if (deferred_reserve) {3063		hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),3064						  h_cg, folio);3065	}3066 3067	spin_unlock_irq(&hugetlb_lock);3068 3069	hugetlb_set_folio_subpool(folio, spool);3070 3071	map_commit = vma_commit_reservation(h, vma, addr);3072	if (unlikely(map_chg > map_commit)) {3073		/*3074		 * The page was added to the reservation map between3075		 * vma_needs_reservation and vma_commit_reservation.3076		 * This indicates a race with hugetlb_reserve_pages.3077		 * Adjust for the subpool count incremented above AND3078		 * in hugetlb_reserve_pages for the same page.  Also,3079		 * the reservation count added in hugetlb_reserve_pages3080		 * no longer applies.3081		 */3082		long rsv_adjust;3083 3084		rsv_adjust = hugepage_subpool_put_pages(spool, 1);3085		hugetlb_acct_memory(h, -rsv_adjust);3086		if (deferred_reserve) {3087			spin_lock_irq(&hugetlb_lock);3088			hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),3089					pages_per_huge_page(h), folio);3090			spin_unlock_irq(&hugetlb_lock);3091		}3092	}3093 3094	if (!memcg_charge_ret)3095		mem_cgroup_commit_charge(folio, memcg);3096	mem_cgroup_put(memcg);3097 3098	return folio;3099 3100out_uncharge_cgroup:3101	hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);3102out_uncharge_cgroup_reservation:3103	if (deferred_reserve)3104		hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),3105						    h_cg);3106out_subpool_put:3107	if (map_chg || avoid_reserve)3108		hugepage_subpool_put_pages(spool, 1);3109out_end_reservation:3110	vma_end_reservation(h, vma, addr);3111	if (!memcg_charge_ret)3112		mem_cgroup_cancel_charge(memcg, nr_pages);3113	mem_cgroup_put(memcg);3114	return ERR_PTR(-ENOSPC);3115}3116 3117int alloc_bootmem_huge_page(struct hstate *h, int nid)3118	__attribute__ ((weak, alias("__alloc_bootmem_huge_page")));3119int __alloc_bootmem_huge_page(struct hstate *h, int nid)3120{3121	struct huge_bootmem_page *m = NULL; /* initialize for clang */3122	int nr_nodes, node = nid;3123 3124	/* do node specific alloc */3125	if (nid != NUMA_NO_NODE) {3126		m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),3127				0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);3128		if (!m)3129			return 0;3130		goto found;3131	}3132	/* allocate from next node when distributing huge pages */3133	for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, &node_states[N_MEMORY]) {3134		m = memblock_alloc_try_nid_raw(3135				huge_page_size(h), huge_page_size(h),3136				0, MEMBLOCK_ALLOC_ACCESSIBLE, node);3137		/*3138		 * Use the beginning of the huge page to store the3139		 * huge_bootmem_page struct (until gather_bootmem3140		 * puts them into the mem_map).3141		 */3142		if (!m)3143			return 0;3144		goto found;3145	}3146 3147found:3148 3149	/*3150	 * Only initialize the head struct page in memmap_init_reserved_pages,3151	 * rest of the struct pages will be initialized by the HugeTLB3152	 * subsystem itself.3153	 * The head struct page is used to get folio information by the HugeTLB3154	 * subsystem like zone id and node id.3155	 */3156	memblock_reserved_mark_noinit(virt_to_phys((void *)m + PAGE_SIZE),3157		huge_page_size(h) - PAGE_SIZE);3158	/* Put them into a private list first because mem_map is not up yet */3159	INIT_LIST_HEAD(&m->list);3160	list_add(&m->list, &huge_boot_pages[node]);3161	m->hstate = h;3162	return 1;3163}3164 3165/* Initialize [start_page:end_page_number] tail struct pages of a hugepage */3166static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,3167					unsigned long start_page_number,3168					unsigned long end_page_number)3169{3170	enum zone_type zone = zone_idx(folio_zone(folio));3171	int nid = folio_nid(folio);3172	unsigned long head_pfn = folio_pfn(folio);3173	unsigned long pfn, end_pfn = head_pfn + end_page_number;3174	int ret;3175 3176	for (pfn = head_pfn + start_page_number; pfn < end_pfn; pfn++) {3177		struct page *page = pfn_to_page(pfn);3178 3179		__ClearPageReserved(folio_page(folio, pfn - head_pfn));3180		__init_single_page(page, pfn, zone, nid);3181		prep_compound_tail((struct page *)folio, pfn - head_pfn);3182		ret = page_ref_freeze(page, 1);3183		VM_BUG_ON(!ret);3184	}3185}3186 3187static void __init hugetlb_folio_init_vmemmap(struct folio *folio,3188					      struct hstate *h,3189					      unsigned long nr_pages)3190{3191	int ret;3192 3193	/* Prepare folio head */3194	__folio_clear_reserved(folio);3195	__folio_set_head(folio);3196	ret = folio_ref_freeze(folio, 1);3197	VM_BUG_ON(!ret);3198	/* Initialize the necessary tail struct pages */3199	hugetlb_folio_init_tail_vmemmap(folio, 1, nr_pages);3200	prep_compound_head((struct page *)folio, huge_page_order(h));3201}3202 3203static void __init prep_and_add_bootmem_folios(struct hstate *h,3204					struct list_head *folio_list)3205{3206	unsigned long flags;3207	struct folio *folio, *tmp_f;3208 3209	/* Send list for bulk vmemmap optimization processing */3210	hugetlb_vmemmap_optimize_folios(h, folio_list);3211 3212	list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {3213		if (!folio_test_hugetlb_vmemmap_optimized(folio)) {3214			/*3215			 * If HVO fails, initialize all tail struct pages3216			 * We do not worry about potential long lock hold3217			 * time as this is early in boot and there should3218			 * be no contention.3219			 */3220			hugetlb_folio_init_tail_vmemmap(folio,3221					HUGETLB_VMEMMAP_RESERVE_PAGES,3222					pages_per_huge_page(h));3223		}3224		/* Subdivide locks to achieve better parallel performance */3225		spin_lock_irqsave(&hugetlb_lock, flags);3226		__prep_account_new_huge_page(h, folio_nid(folio));3227		enqueue_hugetlb_folio(h, folio);3228		spin_unlock_irqrestore(&hugetlb_lock, flags);3229	}3230}3231 3232/*3233 * Put bootmem huge pages into the standard lists after mem_map is up.3234 * Note: This only applies to gigantic (order > MAX_PAGE_ORDER) pages.3235 */3236static void __init gather_bootmem_prealloc_node(unsigned long nid)3237{3238	LIST_HEAD(folio_list);3239	struct huge_bootmem_page *m;3240	struct hstate *h = NULL, *prev_h = NULL;3241 3242	list_for_each_entry(m, &huge_boot_pages[nid], list) {3243		struct page *page = virt_to_page(m);3244		struct folio *folio = (void *)page;3245 3246		h = m->hstate;3247		/*3248		 * It is possible to have multiple huge page sizes (hstates)3249		 * in this list.  If so, process each size separately.3250		 */3251		if (h != prev_h && prev_h != NULL)3252			prep_and_add_bootmem_folios(prev_h, &folio_list);3253		prev_h = h;3254 3255		VM_BUG_ON(!hstate_is_gigantic(h));3256		WARN_ON(folio_ref_count(folio) != 1);3257 3258		hugetlb_folio_init_vmemmap(folio, h,3259					   HUGETLB_VMEMMAP_RESERVE_PAGES);3260		init_new_hugetlb_folio(h, folio);3261		list_add(&folio->lru, &folio_list);3262 3263		/*3264		 * We need to restore the 'stolen' pages to totalram_pages3265		 * in order to fix confusing memory reports from free(1) and3266		 * other side-effects, like CommitLimit going negative.3267		 */3268		adjust_managed_page_count(page, pages_per_huge_page(h));3269		cond_resched();3270	}3271 3272	prep_and_add_bootmem_folios(h, &folio_list);3273}3274 3275static void __init gather_bootmem_prealloc_parallel(unsigned long start,3276						    unsigned long end, void *arg)3277{3278	int nid;3279 3280	for (nid = start; nid < end; nid++)3281		gather_bootmem_prealloc_node(nid);3282}3283 3284static void __init gather_bootmem_prealloc(void)3285{3286	struct padata_mt_job job = {3287		.thread_fn	= gather_bootmem_prealloc_parallel,3288		.fn_arg		= NULL,3289		.start		= 0,3290		.size		= num_node_state(N_MEMORY),3291		.align		= 1,3292		.min_chunk	= 1,3293		.max_threads	= num_node_state(N_MEMORY),3294		.numa_aware	= true,3295	};3296 3297	padata_do_multithreaded(&job);3298}3299 3300static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)3301{3302	unsigned long i;3303	char buf[32];3304 3305	for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {3306		if (hstate_is_gigantic(h)) {3307			if (!alloc_bootmem_huge_page(h, nid))3308				break;3309		} else {3310			struct folio *folio;3311			gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;3312 3313			folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,3314					&node_states[N_MEMORY]);3315			if (!folio)3316				break;3317			free_huge_folio(folio); /* free it into the hugepage allocator */3318		}3319		cond_resched();3320	}3321	if (i == h->max_huge_pages_node[nid])3322		return;3323 3324	string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);3325	pr_warn("HugeTLB: allocating %u of page size %s failed node%d.  Only allocated %lu hugepages.\n",3326		h->max_huge_pages_node[nid], buf, nid, i);3327	h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);3328	h->max_huge_pages_node[nid] = i;3329}3330 3331static bool __init hugetlb_hstate_alloc_pages_specific_nodes(struct hstate *h)3332{3333	int i;3334	bool node_specific_alloc = false;3335 3336	for_each_online_node(i) {3337		if (h->max_huge_pages_node[i] > 0) {3338			hugetlb_hstate_alloc_pages_onenode(h, i);3339			node_specific_alloc = true;3340		}3341	}3342 3343	return node_specific_alloc;3344}3345 3346static void __init hugetlb_hstate_alloc_pages_errcheck(unsigned long allocated, struct hstate *h)3347{3348	if (allocated < h->max_huge_pages) {3349		char buf[32];3350 3351		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);3352		pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",3353			h->max_huge_pages, buf, allocated);3354		h->max_huge_pages = allocated;3355	}3356}3357 3358static void __init hugetlb_pages_alloc_boot_node(unsigned long start, unsigned long end, void *arg)3359{3360	struct hstate *h = (struct hstate *)arg;3361	int i, num = end - start;3362	nodemask_t node_alloc_noretry;3363	LIST_HEAD(folio_list);3364	int next_node = first_online_node;3365 3366	/* Bit mask controlling how hard we retry per-node allocations.*/3367	nodes_clear(node_alloc_noretry);3368 3369	for (i = 0; i < num; ++i) {3370		struct folio *folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],3371						&node_alloc_noretry, &next_node);3372		if (!folio)3373			break;3374 3375		list_move(&folio->lru, &folio_list);3376		cond_resched();3377	}3378 3379	prep_and_add_allocated_folios(h, &folio_list);3380}3381 3382static unsigned long __init hugetlb_gigantic_pages_alloc_boot(struct hstate *h)3383{3384	unsigned long i;3385 3386	for (i = 0; i < h->max_huge_pages; ++i) {3387		if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))3388			break;3389		cond_resched();3390	}3391 3392	return i;3393}3394 3395static unsigned long __init hugetlb_pages_alloc_boot(struct hstate *h)3396{3397	struct padata_mt_job job = {3398		.fn_arg		= h,3399		.align		= 1,3400		.numa_aware	= true3401	};3402 3403	job.thread_fn	= hugetlb_pages_alloc_boot_node;3404	job.start	= 0;3405	job.size	= h->max_huge_pages;3406 3407	/*3408	 * job.max_threads is twice the num_node_state(N_MEMORY),3409	 *3410	 * Tests below indicate that a multiplier of 2 significantly improves3411	 * performance, and although larger values also provide improvements,3412	 * the gains are marginal.3413	 *3414	 * Therefore, choosing 2 as the multiplier strikes a good balance between3415	 * enhancing parallel processing capabilities and maintaining efficient3416	 * resource management.3417	 *3418	 * +------------+-------+-------+-------+-------+-------+3419	 * | multiplier |   1   |   2   |   3   |   4   |   5   |3420	 * +------------+-------+-------+-------+-------+-------+3421	 * | 256G 2node | 358ms | 215ms | 157ms | 134ms | 126ms |3422	 * | 2T   4node | 979ms | 679ms | 543ms | 489ms | 481ms |3423	 * | 50G  2node | 71ms  | 44ms  | 37ms  | 30ms  | 31ms  |3424	 * +------------+-------+-------+-------+-------+-------+3425	 */3426	job.max_threads	= num_node_state(N_MEMORY) * 2;3427	job.min_chunk	= h->max_huge_pages / num_node_state(N_MEMORY) / 2;3428	padata_do_multithreaded(&job);3429 3430	return h->nr_huge_pages;3431}3432 3433/*3434 * NOTE: this routine is called in different contexts for gigantic and3435 * non-gigantic pages.3436 * - For gigantic pages, this is called early in the boot process and3437 *   pages are allocated from memblock allocated or something similar.3438 *   Gigantic pages are actually added to pools later with the routine3439 *   gather_bootmem_prealloc.3440 * - For non-gigantic pages, this is called later in the boot process after3441 *   all of mm is up and functional.  Pages are allocated from buddy and3442 *   then added to hugetlb pools.3443 */3444static void __init hugetlb_hstate_alloc_pages(struct hstate *h)3445{3446	unsigned long allocated;3447	static bool initialized __initdata;3448 3449	/* skip gigantic hugepages allocation if hugetlb_cma enabled */3450	if (hstate_is_gigantic(h) && hugetlb_cma_size) {3451		pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");3452		return;3453	}3454 3455	/* hugetlb_hstate_alloc_pages will be called many times, initialize huge_boot_pages once */3456	if (!initialized) {3457		int i = 0;3458 3459		for (i = 0; i < MAX_NUMNODES; i++)3460			INIT_LIST_HEAD(&huge_boot_pages[i]);3461		initialized = true;3462	}3463 3464	/* do node specific alloc */3465	if (hugetlb_hstate_alloc_pages_specific_nodes(h))3466		return;3467 3468	/* below will do all node balanced alloc */3469	if (hstate_is_gigantic(h))3470		allocated = hugetlb_gigantic_pages_alloc_boot(h);3471	else3472		allocated = hugetlb_pages_alloc_boot(h);3473 3474	hugetlb_hstate_alloc_pages_errcheck(allocated, h);3475}3476 3477static void __init hugetlb_init_hstates(void)3478{3479	struct hstate *h, *h2;3480 3481	for_each_hstate(h) {3482		/* oversize hugepages were init'ed in early boot */3483		if (!hstate_is_gigantic(h))3484			hugetlb_hstate_alloc_pages(h);3485 3486		/*3487		 * Set demote order for each hstate.  Note that3488		 * h->demote_order is initially 0.3489		 * - We can not demote gigantic pages if runtime freeing3490		 *   is not supported, so skip this.3491		 * - If CMA allocation is possible, we can not demote3492		 *   HUGETLB_PAGE_ORDER or smaller size pages.3493		 */3494		if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())3495			continue;3496		if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)3497			continue;3498		for_each_hstate(h2) {3499			if (h2 == h)3500				continue;3501			if (h2->order < h->order &&3502			    h2->order > h->demote_order)3503				h->demote_order = h2->order;3504		}3505	}3506}3507 3508static void __init report_hugepages(void)3509{3510	struct hstate *h;3511 3512	for_each_hstate(h) {3513		char buf[32];3514 3515		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);3516		pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",3517			buf, h->free_huge_pages);3518		pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",3519			hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);3520	}3521}3522 3523#ifdef CONFIG_HIGHMEM3524static void try_to_free_low(struct hstate *h, unsigned long count,3525						nodemask_t *nodes_allowed)3526{3527	int i;3528	LIST_HEAD(page_list);3529 3530	lockdep_assert_held(&hugetlb_lock);3531	if (hstate_is_gigantic(h))3532		return;3533 3534	/*3535	 * Collect pages to be freed on a list, and free after dropping lock3536	 */3537	for_each_node_mask(i, *nodes_allowed) {3538		struct folio *folio, *next;3539		struct list_head *freel = &h->hugepage_freelists[i];3540		list_for_each_entry_safe(folio, next, freel, lru) {3541			if (count >= h->nr_huge_pages)3542				goto out;3543			if (folio_test_highmem(folio))3544				continue;3545			remove_hugetlb_folio(h, folio, false);3546			list_add(&folio->lru, &page_list);3547		}3548	}3549 3550out:3551	spin_unlock_irq(&hugetlb_lock);3552	update_and_free_pages_bulk(h, &page_list);3553	spin_lock_irq(&hugetlb_lock);3554}3555#else3556static inline void try_to_free_low(struct hstate *h, unsigned long count,3557						nodemask_t *nodes_allowed)3558{3559}3560#endif3561 3562/*3563 * Increment or decrement surplus_huge_pages.  Keep node-specific counters3564 * balanced by operating on them in a round-robin fashion.3565 * Returns 1 if an adjustment was made.3566 */3567static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,3568				int delta)3569{3570	int nr_nodes, node;3571 3572	lockdep_assert_held(&hugetlb_lock);3573	VM_BUG_ON(delta != -1 && delta != 1);3574 3575	if (delta < 0) {3576		for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, nodes_allowed) {3577			if (h->surplus_huge_pages_node[node])3578				goto found;3579		}3580	} else {3581		for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {3582			if (h->surplus_huge_pages_node[node] <3583					h->nr_huge_pages_node[node])3584				goto found;3585		}3586	}3587	return 0;3588 3589found:3590	h->surplus_huge_pages += delta;3591	h->surplus_huge_pages_node[node] += delta;3592	return 1;3593}3594 3595#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)3596static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,3597			      nodemask_t *nodes_allowed)3598{3599	unsigned long min_count;3600	unsigned long allocated;3601	struct folio *folio;3602	LIST_HEAD(page_list);3603	NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);3604 3605	/*3606	 * Bit mask controlling how hard we retry per-node allocations.3607	 * If we can not allocate the bit mask, do not attempt to allocate3608	 * the requested huge pages.3609	 */3610	if (node_alloc_noretry)3611		nodes_clear(*node_alloc_noretry);3612	else3613		return -ENOMEM;3614 3615	/*3616	 * resize_lock mutex prevents concurrent adjustments to number of3617	 * pages in hstate via the proc/sysfs interfaces.3618	 */3619	mutex_lock(&h->resize_lock);3620	flush_free_hpage_work(h);3621	spin_lock_irq(&hugetlb_lock);3622 3623	/*3624	 * Check for a node specific request.3625	 * Changing node specific huge page count may require a corresponding3626	 * change to the global count.  In any case, the passed node mask3627	 * (nodes_allowed) will restrict alloc/free to the specified node.3628	 */3629	if (nid != NUMA_NO_NODE) {3630		unsigned long old_count = count;3631 3632		count += persistent_huge_pages(h) -3633			 (h->nr_huge_pages_node[nid] -3634			  h->surplus_huge_pages_node[nid]);3635		/*3636		 * User may have specified a large count value which caused the3637		 * above calculation to overflow.  In this case, they wanted3638		 * to allocate as many huge pages as possible.  Set count to3639		 * largest possible value to align with their intention.3640		 */3641		if (count < old_count)3642			count = ULONG_MAX;3643	}3644 3645	/*3646	 * Gigantic pages runtime allocation depend on the capability for large3647	 * page range allocation.3648	 * If the system does not provide this feature, return an error when3649	 * the user tries to allocate gigantic pages but let the user free the3650	 * boottime allocated gigantic pages.3651	 */3652	if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {3653		if (count > persistent_huge_pages(h)) {3654			spin_unlock_irq(&hugetlb_lock);3655			mutex_unlock(&h->resize_lock);3656			NODEMASK_FREE(node_alloc_noretry);3657			return -EINVAL;3658		}3659		/* Fall through to decrease pool */3660	}3661 3662	/*3663	 * Increase the pool size3664	 * First take pages out of surplus state.  Then make up the3665	 * remaining difference by allocating fresh huge pages.3666	 *3667	 * We might race with alloc_surplus_hugetlb_folio() here and be unable3668	 * to convert a surplus huge page to a normal huge page. That is3669	 * not critical, though, it just means the overall size of the3670	 * pool might be one hugepage larger than it needs to be, but3671	 * within all the constraints specified by the sysctls.3672	 */3673	while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {3674		if (!adjust_pool_surplus(h, nodes_allowed, -1))3675			break;3676	}3677 3678	allocated = 0;3679	while (count > (persistent_huge_pages(h) + allocated)) {3680		/*3681		 * If this allocation races such that we no longer need the3682		 * page, free_huge_folio will handle it by freeing the page3683		 * and reducing the surplus.3684		 */3685		spin_unlock_irq(&hugetlb_lock);3686 3687		/* yield cpu to avoid soft lockup */3688		cond_resched();3689 3690		folio = alloc_pool_huge_folio(h, nodes_allowed,3691						node_alloc_noretry,3692						&h->next_nid_to_alloc);3693		if (!folio) {3694			prep_and_add_allocated_folios(h, &page_list);3695			spin_lock_irq(&hugetlb_lock);3696			goto out;3697		}3698 3699		list_add(&folio->lru, &page_list);3700		allocated++;3701 3702		/* Bail for signals. Probably ctrl-c from user */3703		if (signal_pending(current)) {3704			prep_and_add_allocated_folios(h, &page_list);3705			spin_lock_irq(&hugetlb_lock);3706			goto out;3707		}3708 3709		spin_lock_irq(&hugetlb_lock);3710	}3711 3712	/* Add allocated pages to the pool */3713	if (!list_empty(&page_list)) {3714		spin_unlock_irq(&hugetlb_lock);3715		prep_and_add_allocated_folios(h, &page_list);3716		spin_lock_irq(&hugetlb_lock);3717	}3718 3719	/*3720	 * Decrease the pool size3721	 * First return free pages to the buddy allocator (being careful3722	 * to keep enough around to satisfy reservations).  Then place3723	 * pages into surplus state as needed so the pool will shrink3724	 * to the desired size as pages become free.3725	 *3726	 * By placing pages into the surplus state independent of the3727	 * overcommit value, we are allowing the surplus pool size to3728	 * exceed overcommit. There are few sane options here. Since3729	 * alloc_surplus_hugetlb_folio() is checking the global counter,3730	 * though, we'll note that we're not allowed to exceed surplus3731	 * and won't grow the pool anywhere else. Not until one of the3732	 * sysctls are changed, or the surplus pages go out of use.3733	 */3734	min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;3735	min_count = max(count, min_count);3736	try_to_free_low(h, min_count, nodes_allowed);3737 3738	/*3739	 * Collect pages to be removed on list without dropping lock3740	 */3741	while (min_count < persistent_huge_pages(h)) {3742		folio = remove_pool_hugetlb_folio(h, nodes_allowed, 0);3743		if (!folio)3744			break;3745 3746		list_add(&folio->lru, &page_list);3747	}3748	/* free the pages after dropping lock */3749	spin_unlock_irq(&hugetlb_lock);3750	update_and_free_pages_bulk(h, &page_list);3751	flush_free_hpage_work(h);3752	spin_lock_irq(&hugetlb_lock);3753 3754	while (count < persistent_huge_pages(h)) {3755		if (!adjust_pool_surplus(h, nodes_allowed, 1))3756			break;3757	}3758out:3759	h->max_huge_pages = persistent_huge_pages(h);3760	spin_unlock_irq(&hugetlb_lock);3761	mutex_unlock(&h->resize_lock);3762 3763	NODEMASK_FREE(node_alloc_noretry);3764 3765	return 0;3766}3767 3768static long demote_free_hugetlb_folios(struct hstate *src, struct hstate *dst,3769				       struct list_head *src_list)3770{3771	long rc;3772	struct folio *folio, *next;3773	LIST_HEAD(dst_list);3774	LIST_HEAD(ret_list);3775 3776	rc = hugetlb_vmemmap_restore_folios(src, src_list, &ret_list);3777	list_splice_init(&ret_list, src_list);3778 3779	/*3780	 * Taking target hstate mutex synchronizes with set_max_huge_pages.3781	 * Without the mutex, pages added to target hstate could be marked3782	 * as surplus.3783	 *3784	 * Note that we already hold src->resize_lock.  To prevent deadlock,3785	 * use the convention of always taking larger size hstate mutex first.3786	 */3787	mutex_lock(&dst->resize_lock);3788 3789	list_for_each_entry_safe(folio, next, src_list, lru) {3790		int i;3791 3792		if (folio_test_hugetlb_vmemmap_optimized(folio))3793			continue;3794 3795		list_del(&folio->lru);3796 3797		split_page_owner(&folio->page, huge_page_order(src), huge_page_order(dst));3798		pgalloc_tag_split(folio, huge_page_order(src), huge_page_order(dst));3799 3800		for (i = 0; i < pages_per_huge_page(src); i += pages_per_huge_page(dst)) {3801			struct page *page = folio_page(folio, i);3802 3803			page->mapping = NULL;3804			clear_compound_head(page);3805			prep_compound_page(page, dst->order);3806 3807			init_new_hugetlb_folio(dst, page_folio(page));3808			list_add(&page->lru, &dst_list);3809		}3810	}3811 3812	prep_and_add_allocated_folios(dst, &dst_list);3813 3814	mutex_unlock(&dst->resize_lock);3815 3816	return rc;3817}3818 3819static long demote_pool_huge_page(struct hstate *src, nodemask_t *nodes_allowed,3820				  unsigned long nr_to_demote)3821	__must_hold(&hugetlb_lock)3822{3823	int nr_nodes, node;3824	struct hstate *dst;3825	long rc = 0;3826	long nr_demoted = 0;3827 3828	lockdep_assert_held(&hugetlb_lock);3829 3830	/* We should never get here if no demote order */3831	if (!src->demote_order) {3832		pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");3833		return -EINVAL;		/* internal error */3834	}3835	dst = size_to_hstate(PAGE_SIZE << src->demote_order);3836 3837	for_each_node_mask_to_free(src, nr_nodes, node, nodes_allowed) {3838		LIST_HEAD(list);3839		struct folio *folio, *next;3840 3841		list_for_each_entry_safe(folio, next, &src->hugepage_freelists[node], lru) {3842			if (folio_test_hwpoison(folio))3843				continue;3844 3845			remove_hugetlb_folio(src, folio, false);3846			list_add(&folio->lru, &list);3847 3848			if (++nr_demoted == nr_to_demote)3849				break;3850		}3851 3852		spin_unlock_irq(&hugetlb_lock);3853 3854		rc = demote_free_hugetlb_folios(src, dst, &list);3855 3856		spin_lock_irq(&hugetlb_lock);3857 3858		list_for_each_entry_safe(folio, next, &list, lru) {3859			list_del(&folio->lru);3860			add_hugetlb_folio(src, folio, false);3861 3862			nr_demoted--;3863		}3864 3865		if (rc < 0 || nr_demoted == nr_to_demote)3866			break;3867	}3868 3869	/*3870	 * Not absolutely necessary, but for consistency update max_huge_pages3871	 * based on pool changes for the demoted page.3872	 */3873	src->max_huge_pages -= nr_demoted;3874	dst->max_huge_pages += nr_demoted << (huge_page_order(src) - huge_page_order(dst));3875 3876	if (rc < 0)3877		return rc;3878 3879	if (nr_demoted)3880		return nr_demoted;3881	/*3882	 * Only way to get here is if all pages on free lists are poisoned.3883	 * Return -EBUSY so that caller will not retry.3884	 */3885	return -EBUSY;3886}3887 3888#define HSTATE_ATTR_RO(_name) \3889	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)3890 3891#define HSTATE_ATTR_WO(_name) \3892	static struct kobj_attribute _name##_attr = __ATTR_WO(_name)3893 3894#define HSTATE_ATTR(_name) \3895	static struct kobj_attribute _name##_attr = __ATTR_RW(_name)3896 3897static struct kobject *hugepages_kobj;3898static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];3899 3900static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);3901 3902static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)3903{3904	int i;3905 3906	for (i = 0; i < HUGE_MAX_HSTATE; i++)3907		if (hstate_kobjs[i] == kobj) {3908			if (nidp)3909				*nidp = NUMA_NO_NODE;3910			return &hstates[i];3911		}3912 3913	return kobj_to_node_hstate(kobj, nidp);3914}3915 3916static ssize_t nr_hugepages_show_common(struct kobject *kobj,3917					struct kobj_attribute *attr, char *buf)3918{3919	struct hstate *h;3920	unsigned long nr_huge_pages;3921	int nid;3922 3923	h = kobj_to_hstate(kobj, &nid);3924	if (nid == NUMA_NO_NODE)3925		nr_huge_pages = h->nr_huge_pages;3926	else3927		nr_huge_pages = h->nr_huge_pages_node[nid];3928 3929	return sysfs_emit(buf, "%lu\n", nr_huge_pages);3930}3931 3932static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,3933					   struct hstate *h, int nid,3934					   unsigned long count, size_t len)3935{3936	int err;3937	nodemask_t nodes_allowed, *n_mask;3938 3939	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())3940		return -EINVAL;3941 3942	if (nid == NUMA_NO_NODE) {3943		/*3944		 * global hstate attribute3945		 */3946		if (!(obey_mempolicy &&3947				init_nodemask_of_mempolicy(&nodes_allowed)))3948			n_mask = &node_states[N_MEMORY];3949		else3950			n_mask = &nodes_allowed;3951	} else {3952		/*3953		 * Node specific request.  count adjustment happens in3954		 * set_max_huge_pages() after acquiring hugetlb_lock.3955		 */3956		init_nodemask_of_node(&nodes_allowed, nid);3957		n_mask = &nodes_allowed;3958	}3959 3960	err = set_max_huge_pages(h, count, nid, n_mask);3961 3962	return err ? err : len;3963}3964 3965static ssize_t nr_hugepages_store_common(bool obey_mempolicy,3966					 struct kobject *kobj, const char *buf,3967					 size_t len)3968{3969	struct hstate *h;3970	unsigned long count;3971	int nid;3972	int err;3973 3974	err = kstrtoul(buf, 10, &count);3975	if (err)3976		return err;3977 3978	h = kobj_to_hstate(kobj, &nid);3979	return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);3980}3981 3982static ssize_t nr_hugepages_show(struct kobject *kobj,3983				       struct kobj_attribute *attr, char *buf)3984{3985	return nr_hugepages_show_common(kobj, attr, buf);3986}3987 3988static ssize_t nr_hugepages_store(struct kobject *kobj,3989	       struct kobj_attribute *attr, const char *buf, size_t len)3990{3991	return nr_hugepages_store_common(false, kobj, buf, len);3992}3993HSTATE_ATTR(nr_hugepages);3994 3995#ifdef CONFIG_NUMA3996 3997/*3998 * hstate attribute for optionally mempolicy-based constraint on persistent3999 * huge page alloc/free.4000 */4001static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,4002					   struct kobj_attribute *attr,4003					   char *buf)4004{4005	return nr_hugepages_show_common(kobj, attr, buf);4006}4007 4008static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,4009	       struct kobj_attribute *attr, const char *buf, size_t len)4010{4011	return nr_hugepages_store_common(true, kobj, buf, len);4012}4013HSTATE_ATTR(nr_hugepages_mempolicy);4014#endif4015 4016 4017static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,4018					struct kobj_attribute *attr, char *buf)4019{4020	struct hstate *h = kobj_to_hstate(kobj, NULL);4021	return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);4022}4023 4024static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,4025		struct kobj_attribute *attr, const char *buf, size_t count)4026{4027	int err;4028	unsigned long input;4029	struct hstate *h = kobj_to_hstate(kobj, NULL);4030 4031	if (hstate_is_gigantic(h))4032		return -EINVAL;4033 4034	err = kstrtoul(buf, 10, &input);4035	if (err)4036		return err;4037 4038	spin_lock_irq(&hugetlb_lock);4039	h->nr_overcommit_huge_pages = input;4040	spin_unlock_irq(&hugetlb_lock);4041 4042	return count;4043}4044HSTATE_ATTR(nr_overcommit_hugepages);4045 4046static ssize_t free_hugepages_show(struct kobject *kobj,4047					struct kobj_attribute *attr, char *buf)4048{4049	struct hstate *h;4050	unsigned long free_huge_pages;4051	int nid;4052 4053	h = kobj_to_hstate(kobj, &nid);4054	if (nid == NUMA_NO_NODE)4055		free_huge_pages = h->free_huge_pages;4056	else4057		free_huge_pages = h->free_huge_pages_node[nid];4058 4059	return sysfs_emit(buf, "%lu\n", free_huge_pages);4060}4061HSTATE_ATTR_RO(free_hugepages);4062 4063static ssize_t resv_hugepages_show(struct kobject *kobj,4064					struct kobj_attribute *attr, char *buf)4065{4066	struct hstate *h = kobj_to_hstate(kobj, NULL);4067	return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);4068}4069HSTATE_ATTR_RO(resv_hugepages);4070 4071static ssize_t surplus_hugepages_show(struct kobject *kobj,4072					struct kobj_attribute *attr, char *buf)4073{4074	struct hstate *h;4075	unsigned long surplus_huge_pages;4076	int nid;4077 4078	h = kobj_to_hstate(kobj, &nid);4079	if (nid == NUMA_NO_NODE)4080		surplus_huge_pages = h->surplus_huge_pages;4081	else4082		surplus_huge_pages = h->surplus_huge_pages_node[nid];4083 4084	return sysfs_emit(buf, "%lu\n", surplus_huge_pages);4085}4086HSTATE_ATTR_RO(surplus_hugepages);4087 4088static ssize_t demote_store(struct kobject *kobj,4089	       struct kobj_attribute *attr, const char *buf, size_t len)4090{4091	unsigned long nr_demote;4092	unsigned long nr_available;4093	nodemask_t nodes_allowed, *n_mask;4094	struct hstate *h;4095	int err;4096	int nid;4097 4098	err = kstrtoul(buf, 10, &nr_demote);4099	if (err)4100		return err;4101	h = kobj_to_hstate(kobj, &nid);4102 4103	if (nid != NUMA_NO_NODE) {4104		init_nodemask_of_node(&nodes_allowed, nid);4105		n_mask = &nodes_allowed;4106	} else {4107		n_mask = &node_states[N_MEMORY];4108	}4109 4110	/* Synchronize with other sysfs operations modifying huge pages */4111	mutex_lock(&h->resize_lock);4112	spin_lock_irq(&hugetlb_lock);4113 4114	while (nr_demote) {4115		long rc;4116 4117		/*4118		 * Check for available pages to demote each time thorough the4119		 * loop as demote_pool_huge_page will drop hugetlb_lock.4120		 */4121		if (nid != NUMA_NO_NODE)4122			nr_available = h->free_huge_pages_node[nid];4123		else4124			nr_available = h->free_huge_pages;4125		nr_available -= h->resv_huge_pages;4126		if (!nr_available)4127			break;4128 4129		rc = demote_pool_huge_page(h, n_mask, nr_demote);4130		if (rc < 0) {4131			err = rc;4132			break;4133		}4134 4135		nr_demote -= rc;4136	}4137 4138	spin_unlock_irq(&hugetlb_lock);4139	mutex_unlock(&h->resize_lock);4140 4141	if (err)4142		return err;4143	return len;4144}4145HSTATE_ATTR_WO(demote);4146 4147static ssize_t demote_size_show(struct kobject *kobj,4148					struct kobj_attribute *attr, char *buf)4149{4150	struct hstate *h = kobj_to_hstate(kobj, NULL);4151	unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;4152 4153	return sysfs_emit(buf, "%lukB\n", demote_size);4154}4155 4156static ssize_t demote_size_store(struct kobject *kobj,4157					struct kobj_attribute *attr,4158					const char *buf, size_t count)4159{4160	struct hstate *h, *demote_hstate;4161	unsigned long demote_size;4162	unsigned int demote_order;4163 4164	demote_size = (unsigned long)memparse(buf, NULL);4165 4166	demote_hstate = size_to_hstate(demote_size);4167	if (!demote_hstate)4168		return -EINVAL;4169	demote_order = demote_hstate->order;4170	if (demote_order < HUGETLB_PAGE_ORDER)4171		return -EINVAL;4172 4173	/* demote order must be smaller than hstate order */4174	h = kobj_to_hstate(kobj, NULL);4175	if (demote_order >= h->order)4176		return -EINVAL;4177 4178	/* resize_lock synchronizes access to demote size and writes */4179	mutex_lock(&h->resize_lock);4180	h->demote_order = demote_order;4181	mutex_unlock(&h->resize_lock);4182 4183	return count;4184}4185HSTATE_ATTR(demote_size);4186 4187static struct attribute *hstate_attrs[] = {4188	&nr_hugepages_attr.attr,4189	&nr_overcommit_hugepages_attr.attr,4190	&free_hugepages_attr.attr,4191	&resv_hugepages_attr.attr,4192	&surplus_hugepages_attr.attr,4193#ifdef CONFIG_NUMA4194	&nr_hugepages_mempolicy_attr.attr,4195#endif4196	NULL,4197};4198 4199static const struct attribute_group hstate_attr_group = {4200	.attrs = hstate_attrs,4201};4202 4203static struct attribute *hstate_demote_attrs[] = {4204	&demote_size_attr.attr,4205	&demote_attr.attr,4206	NULL,4207};4208 4209static const struct attribute_group hstate_demote_attr_group = {4210	.attrs = hstate_demote_attrs,4211};4212 4213static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,4214				    struct kobject **hstate_kobjs,4215				    const struct attribute_group *hstate_attr_group)4216{4217	int retval;4218	int hi = hstate_index(h);4219 4220	hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);4221	if (!hstate_kobjs[hi])4222		return -ENOMEM;4223 4224	retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);4225	if (retval) {4226		kobject_put(hstate_kobjs[hi]);4227		hstate_kobjs[hi] = NULL;4228		return retval;4229	}4230 4231	if (h->demote_order) {4232		retval = sysfs_create_group(hstate_kobjs[hi],4233					    &hstate_demote_attr_group);4234		if (retval) {4235			pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);4236			sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);4237			kobject_put(hstate_kobjs[hi]);4238			hstate_kobjs[hi] = NULL;4239			return retval;4240		}4241	}4242 4243	return 0;4244}4245 4246#ifdef CONFIG_NUMA4247static bool hugetlb_sysfs_initialized __ro_after_init;4248 4249/*4250 * node_hstate/s - associate per node hstate attributes, via their kobjects,4251 * with node devices in node_devices[] using a parallel array.  The array4252 * index of a node device or _hstate == node id.4253 * This is here to avoid any static dependency of the node device driver, in4254 * the base kernel, on the hugetlb module.4255 */4256struct node_hstate {4257	struct kobject		*hugepages_kobj;4258	struct kobject		*hstate_kobjs[HUGE_MAX_HSTATE];4259};4260static struct node_hstate node_hstates[MAX_NUMNODES];4261 4262/*4263 * A subset of global hstate attributes for node devices4264 */4265static struct attribute *per_node_hstate_attrs[] = {4266	&nr_hugepages_attr.attr,4267	&free_hugepages_attr.attr,4268	&surplus_hugepages_attr.attr,4269	NULL,4270};4271 4272static const struct attribute_group per_node_hstate_attr_group = {4273	.attrs = per_node_hstate_attrs,4274};4275 4276/*4277 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.4278 * Returns node id via non-NULL nidp.4279 */4280static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)4281{4282	int nid;4283 4284	for (nid = 0; nid < nr_node_ids; nid++) {4285		struct node_hstate *nhs = &node_hstates[nid];4286		int i;4287		for (i = 0; i < HUGE_MAX_HSTATE; i++)4288			if (nhs->hstate_kobjs[i] == kobj) {4289				if (nidp)4290					*nidp = nid;4291				return &hstates[i];4292			}4293	}4294 4295	BUG();4296	return NULL;4297}4298 4299/*4300 * Unregister hstate attributes from a single node device.4301 * No-op if no hstate attributes attached.4302 */4303void hugetlb_unregister_node(struct node *node)4304{4305	struct hstate *h;4306	struct node_hstate *nhs = &node_hstates[node->dev.id];4307 4308	if (!nhs->hugepages_kobj)4309		return;		/* no hstate attributes */4310 4311	for_each_hstate(h) {4312		int idx = hstate_index(h);4313		struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];4314 4315		if (!hstate_kobj)4316			continue;4317		if (h->demote_order)4318			sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);4319		sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);4320		kobject_put(hstate_kobj);4321		nhs->hstate_kobjs[idx] = NULL;4322	}4323 4324	kobject_put(nhs->hugepages_kobj);4325	nhs->hugepages_kobj = NULL;4326}4327 4328 4329/*4330 * Register hstate attributes for a single node device.4331 * No-op if attributes already registered.4332 */4333void hugetlb_register_node(struct node *node)4334{4335	struct hstate *h;4336	struct node_hstate *nhs = &node_hstates[node->dev.id];4337	int err;4338 4339	if (!hugetlb_sysfs_initialized)4340		return;4341 4342	if (nhs->hugepages_kobj)4343		return;		/* already allocated */4344 4345	nhs->hugepages_kobj = kobject_create_and_add("hugepages",4346							&node->dev.kobj);4347	if (!nhs->hugepages_kobj)4348		return;4349 4350	for_each_hstate(h) {4351		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,4352						nhs->hstate_kobjs,4353						&per_node_hstate_attr_group);4354		if (err) {4355			pr_err("HugeTLB: Unable to add hstate %s for node %d\n",4356				h->name, node->dev.id);4357			hugetlb_unregister_node(node);4358			break;4359		}4360	}4361}4362 4363/*4364 * hugetlb init time:  register hstate attributes for all registered node4365 * devices of nodes that have memory.  All on-line nodes should have4366 * registered their associated device by this time.4367 */4368static void __init hugetlb_register_all_nodes(void)4369{4370	int nid;4371 4372	for_each_online_node(nid)4373		hugetlb_register_node(node_devices[nid]);4374}4375#else	/* !CONFIG_NUMA */4376 4377static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)4378{4379	BUG();4380	if (nidp)4381		*nidp = -1;4382	return NULL;4383}4384 4385static void hugetlb_register_all_nodes(void) { }4386 4387#endif4388 4389#ifdef CONFIG_CMA4390static void __init hugetlb_cma_check(void);4391#else4392static inline __init void hugetlb_cma_check(void)4393{4394}4395#endif4396 4397static void __init hugetlb_sysfs_init(void)4398{4399	struct hstate *h;4400	int err;4401 4402	hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);4403	if (!hugepages_kobj)4404		return;4405 4406	for_each_hstate(h) {4407		err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,4408					 hstate_kobjs, &hstate_attr_group);4409		if (err)4410			pr_err("HugeTLB: Unable to add hstate %s", h->name);4411	}4412 4413#ifdef CONFIG_NUMA4414	hugetlb_sysfs_initialized = true;4415#endif4416	hugetlb_register_all_nodes();4417}4418 4419#ifdef CONFIG_SYSCTL4420static void hugetlb_sysctl_init(void);4421#else4422static inline void hugetlb_sysctl_init(void) { }4423#endif4424 4425static int __init hugetlb_init(void)4426{4427	int i;4428 4429	BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <4430			__NR_HPAGEFLAGS);4431 4432	if (!hugepages_supported()) {4433		if (hugetlb_max_hstate || default_hstate_max_huge_pages)4434			pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");4435		return 0;4436	}4437 4438	/*4439	 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some4440	 * architectures depend on setup being done here.4441	 */4442	hugetlb_add_hstate(HUGETLB_PAGE_ORDER);4443	if (!parsed_default_hugepagesz) {4444		/*4445		 * If we did not parse a default huge page size, set4446		 * default_hstate_idx to HPAGE_SIZE hstate. And, if the4447		 * number of huge pages for this default size was implicitly4448		 * specified, set that here as well.4449		 * Note that the implicit setting will overwrite an explicit4450		 * setting.  A warning will be printed in this case.4451		 */4452		default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));4453		if (default_hstate_max_huge_pages) {4454			if (default_hstate.max_huge_pages) {4455				char buf[32];4456 4457				string_get_size(huge_page_size(&default_hstate),4458					1, STRING_UNITS_2, buf, 32);4459				pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",4460					default_hstate.max_huge_pages, buf);4461				pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",4462					default_hstate_max_huge_pages);4463			}4464			default_hstate.max_huge_pages =4465				default_hstate_max_huge_pages;4466 4467			for_each_online_node(i)4468				default_hstate.max_huge_pages_node[i] =4469					default_hugepages_in_node[i];4470		}4471	}4472 4473	hugetlb_cma_check();4474	hugetlb_init_hstates();4475	gather_bootmem_prealloc();4476	report_hugepages();4477 4478	hugetlb_sysfs_init();4479	hugetlb_cgroup_file_init();4480	hugetlb_sysctl_init();4481 4482#ifdef CONFIG_SMP4483	num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());4484#else4485	num_fault_mutexes = 1;4486#endif4487	hugetlb_fault_mutex_table =4488		kmalloc_array(num_fault_mutexes, sizeof(struct mutex),4489			      GFP_KERNEL);4490	BUG_ON(!hugetlb_fault_mutex_table);4491 4492	for (i = 0; i < num_fault_mutexes; i++)4493		mutex_init(&hugetlb_fault_mutex_table[i]);4494	return 0;4495}4496subsys_initcall(hugetlb_init);4497 4498/* Overwritten by architectures with more huge page sizes */4499bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)4500{4501	return size == HPAGE_SIZE;4502}4503 4504void __init hugetlb_add_hstate(unsigned int order)4505{4506	struct hstate *h;4507	unsigned long i;4508 4509	if (size_to_hstate(PAGE_SIZE << order)) {4510		return;4511	}4512	BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);4513	BUG_ON(order < order_base_2(__NR_USED_SUBPAGE));4514	h = &hstates[hugetlb_max_hstate++];4515	__mutex_init(&h->resize_lock, "resize mutex", &h->resize_key);4516	h->order = order;4517	h->mask = ~(huge_page_size(h) - 1);4518	for (i = 0; i < MAX_NUMNODES; ++i)4519		INIT_LIST_HEAD(&h->hugepage_freelists[i]);4520	INIT_LIST_HEAD(&h->hugepage_activelist);4521	h->next_nid_to_alloc = first_memory_node;4522	h->next_nid_to_free = first_memory_node;4523	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",4524					huge_page_size(h)/SZ_1K);4525 4526	parsed_hstate = h;4527}4528 4529bool __init __weak hugetlb_node_alloc_supported(void)4530{4531	return true;4532}4533 4534static void __init hugepages_clear_pages_in_node(void)4535{4536	if (!hugetlb_max_hstate) {4537		default_hstate_max_huge_pages = 0;4538		memset(default_hugepages_in_node, 0,4539			sizeof(default_hugepages_in_node));4540	} else {4541		parsed_hstate->max_huge_pages = 0;4542		memset(parsed_hstate->max_huge_pages_node, 0,4543			sizeof(parsed_hstate->max_huge_pages_node));4544	}4545}4546 4547/*4548 * hugepages command line processing4549 * hugepages normally follows a valid hugepagsz or default_hugepagsz4550 * specification.  If not, ignore the hugepages value.  hugepages can also4551 * be the first huge page command line  option in which case it implicitly4552 * specifies the number of huge pages for the default size.4553 */4554static int __init hugepages_setup(char *s)4555{4556	unsigned long *mhp;4557	static unsigned long *last_mhp;4558	int node = NUMA_NO_NODE;4559	int count;4560	unsigned long tmp;4561	char *p = s;4562 4563	if (!parsed_valid_hugepagesz) {4564		pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);4565		parsed_valid_hugepagesz = true;4566		return 1;4567	}4568 4569	/*4570	 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter4571	 * yet, so this hugepages= parameter goes to the "default hstate".4572	 * Otherwise, it goes with the previously parsed hugepagesz or4573	 * default_hugepagesz.4574	 */4575	else if (!hugetlb_max_hstate)4576		mhp = &default_hstate_max_huge_pages;4577	else4578		mhp = &parsed_hstate->max_huge_pages;4579 4580	if (mhp == last_mhp) {4581		pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);4582		return 1;4583	}4584 4585	while (*p) {4586		count = 0;4587		if (sscanf(p, "%lu%n", &tmp, &count) != 1)4588			goto invalid;4589		/* Parameter is node format */4590		if (p[count] == ':') {4591			if (!hugetlb_node_alloc_supported()) {4592				pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");4593				return 1;4594			}4595			if (tmp >= MAX_NUMNODES || !node_online(tmp))4596				goto invalid;4597			node = array_index_nospec(tmp, MAX_NUMNODES);4598			p += count + 1;4599			/* Parse hugepages */4600			if (sscanf(p, "%lu%n", &tmp, &count) != 1)4601				goto invalid;4602			if (!hugetlb_max_hstate)4603				default_hugepages_in_node[node] = tmp;4604			else4605				parsed_hstate->max_huge_pages_node[node] = tmp;4606			*mhp += tmp;4607			/* Go to parse next node*/4608			if (p[count] == ',')4609				p += count + 1;4610			else4611				break;4612		} else {4613			if (p != s)4614				goto invalid;4615			*mhp = tmp;4616			break;4617		}4618	}4619 4620	/*4621	 * Global state is always initialized later in hugetlb_init.4622	 * But we need to allocate gigantic hstates here early to still4623	 * use the bootmem allocator.4624	 */4625	if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))4626		hugetlb_hstate_alloc_pages(parsed_hstate);4627 4628	last_mhp = mhp;4629 4630	return 1;4631 4632invalid:4633	pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);4634	hugepages_clear_pages_in_node();4635	return 1;4636}4637__setup("hugepages=", hugepages_setup);4638 4639/*4640 * hugepagesz command line processing4641 * A specific huge page size can only be specified once with hugepagesz.4642 * hugepagesz is followed by hugepages on the command line.  The global4643 * variable 'parsed_valid_hugepagesz' is used to determine if prior4644 * hugepagesz argument was valid.4645 */4646static int __init hugepagesz_setup(char *s)4647{4648	unsigned long size;4649	struct hstate *h;4650 4651	parsed_valid_hugepagesz = false;4652	size = (unsigned long)memparse(s, NULL);4653 4654	if (!arch_hugetlb_valid_size(size)) {4655		pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);4656		return 1;4657	}4658 4659	h = size_to_hstate(size);4660	if (h) {4661		/*4662		 * hstate for this size already exists.  This is normally4663		 * an error, but is allowed if the existing hstate is the4664		 * default hstate.  More specifically, it is only allowed if4665		 * the number of huge pages for the default hstate was not4666		 * previously specified.4667		 */4668		if (!parsed_default_hugepagesz ||  h != &default_hstate ||4669		    default_hstate.max_huge_pages) {4670			pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);4671			return 1;4672		}4673 4674		/*4675		 * No need to call hugetlb_add_hstate() as hstate already4676		 * exists.  But, do set parsed_hstate so that a following4677		 * hugepages= parameter will be applied to this hstate.4678		 */4679		parsed_hstate = h;4680		parsed_valid_hugepagesz = true;4681		return 1;4682	}4683 4684	hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);4685	parsed_valid_hugepagesz = true;4686	return 1;4687}4688__setup("hugepagesz=", hugepagesz_setup);4689 4690/*4691 * default_hugepagesz command line input4692 * Only one instance of default_hugepagesz allowed on command line.4693 */4694static int __init default_hugepagesz_setup(char *s)4695{4696	unsigned long size;4697	int i;4698 4699	parsed_valid_hugepagesz = false;4700	if (parsed_default_hugepagesz) {4701		pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);4702		return 1;4703	}4704 4705	size = (unsigned long)memparse(s, NULL);4706 4707	if (!arch_hugetlb_valid_size(size)) {4708		pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);4709		return 1;4710	}4711 4712	hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);4713	parsed_valid_hugepagesz = true;4714	parsed_default_hugepagesz = true;4715	default_hstate_idx = hstate_index(size_to_hstate(size));4716 4717	/*4718	 * The number of default huge pages (for this size) could have been4719	 * specified as the first hugetlb parameter: hugepages=X.  If so,4720	 * then default_hstate_max_huge_pages is set.  If the default huge4721	 * page size is gigantic (> MAX_PAGE_ORDER), then the pages must be4722	 * allocated here from bootmem allocator.4723	 */4724	if (default_hstate_max_huge_pages) {4725		default_hstate.max_huge_pages = default_hstate_max_huge_pages;4726		for_each_online_node(i)4727			default_hstate.max_huge_pages_node[i] =4728				default_hugepages_in_node[i];4729		if (hstate_is_gigantic(&default_hstate))4730			hugetlb_hstate_alloc_pages(&default_hstate);4731		default_hstate_max_huge_pages = 0;4732	}4733 4734	return 1;4735}4736__setup("default_hugepagesz=", default_hugepagesz_setup);4737 4738static unsigned int allowed_mems_nr(struct hstate *h)4739{4740	int node;4741	unsigned int nr = 0;4742	nodemask_t *mbind_nodemask;4743	unsigned int *array = h->free_huge_pages_node;4744	gfp_t gfp_mask = htlb_alloc_mask(h);4745 4746	mbind_nodemask = policy_mbind_nodemask(gfp_mask);4747	for_each_node_mask(node, cpuset_current_mems_allowed) {4748		if (!mbind_nodemask || node_isset(node, *mbind_nodemask))4749			nr += array[node];4750	}4751 4752	return nr;4753}4754 4755#ifdef CONFIG_SYSCTL4756static int proc_hugetlb_doulongvec_minmax(const struct ctl_table *table, int write,4757					  void *buffer, size_t *length,4758					  loff_t *ppos, unsigned long *out)4759{4760	struct ctl_table dup_table;4761 4762	/*4763	 * In order to avoid races with __do_proc_doulongvec_minmax(), we4764	 * can duplicate the @table and alter the duplicate of it.4765	 */4766	dup_table = *table;4767	dup_table.data = out;4768 4769	return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);4770}4771 4772static int hugetlb_sysctl_handler_common(bool obey_mempolicy,4773			 const struct ctl_table *table, int write,4774			 void *buffer, size_t *length, loff_t *ppos)4775{4776	struct hstate *h = &default_hstate;4777	unsigned long tmp = h->max_huge_pages;4778	int ret;4779 4780	if (!hugepages_supported())4781		return -EOPNOTSUPP;4782 4783	ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,4784					     &tmp);4785	if (ret)4786		goto out;4787 4788	if (write)4789		ret = __nr_hugepages_store_common(obey_mempolicy, h,4790						  NUMA_NO_NODE, tmp, *length);4791out:4792	return ret;4793}4794 4795static int hugetlb_sysctl_handler(const struct ctl_table *table, int write,4796			  void *buffer, size_t *length, loff_t *ppos)4797{4798 4799	return hugetlb_sysctl_handler_common(false, table, write,4800							buffer, length, ppos);4801}4802 4803#ifdef CONFIG_NUMA4804static int hugetlb_mempolicy_sysctl_handler(const struct ctl_table *table, int write,4805			  void *buffer, size_t *length, loff_t *ppos)4806{4807	return hugetlb_sysctl_handler_common(true, table, write,4808							buffer, length, ppos);4809}4810#endif /* CONFIG_NUMA */4811 4812static int hugetlb_overcommit_handler(const struct ctl_table *table, int write,4813		void *buffer, size_t *length, loff_t *ppos)4814{4815	struct hstate *h = &default_hstate;4816	unsigned long tmp;4817	int ret;4818 4819	if (!hugepages_supported())4820		return -EOPNOTSUPP;4821 4822	tmp = h->nr_overcommit_huge_pages;4823 4824	if (write && hstate_is_gigantic(h))4825		return -EINVAL;4826 4827	ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,4828					     &tmp);4829	if (ret)4830		goto out;4831 4832	if (write) {4833		spin_lock_irq(&hugetlb_lock);4834		h->nr_overcommit_huge_pages = tmp;4835		spin_unlock_irq(&hugetlb_lock);4836	}4837out:4838	return ret;4839}4840 4841static struct ctl_table hugetlb_table[] = {4842	{4843		.procname	= "nr_hugepages",4844		.data		= NULL,4845		.maxlen		= sizeof(unsigned long),4846		.mode		= 0644,4847		.proc_handler	= hugetlb_sysctl_handler,4848	},4849#ifdef CONFIG_NUMA4850	{4851		.procname       = "nr_hugepages_mempolicy",4852		.data           = NULL,4853		.maxlen         = sizeof(unsigned long),4854		.mode           = 0644,4855		.proc_handler   = &hugetlb_mempolicy_sysctl_handler,4856	},4857#endif4858	{4859		.procname	= "hugetlb_shm_group",4860		.data		= &sysctl_hugetlb_shm_group,4861		.maxlen		= sizeof(gid_t),4862		.mode		= 0644,4863		.proc_handler	= proc_dointvec,4864	},4865	{4866		.procname	= "nr_overcommit_hugepages",4867		.data		= NULL,4868		.maxlen		= sizeof(unsigned long),4869		.mode		= 0644,4870		.proc_handler	= hugetlb_overcommit_handler,4871	},4872};4873 4874static void hugetlb_sysctl_init(void)4875{4876	register_sysctl_init("vm", hugetlb_table);4877}4878#endif /* CONFIG_SYSCTL */4879 4880void hugetlb_report_meminfo(struct seq_file *m)4881{4882	struct hstate *h;4883	unsigned long total = 0;4884 4885	if (!hugepages_supported())4886		return;4887 4888	for_each_hstate(h) {4889		unsigned long count = h->nr_huge_pages;4890 4891		total += huge_page_size(h) * count;4892 4893		if (h == &default_hstate)4894			seq_printf(m,4895				   "HugePages_Total:   %5lu\n"4896				   "HugePages_Free:    %5lu\n"4897				   "HugePages_Rsvd:    %5lu\n"4898				   "HugePages_Surp:    %5lu\n"4899				   "Hugepagesize:   %8lu kB\n",4900				   count,4901				   h->free_huge_pages,4902				   h->resv_huge_pages,4903				   h->surplus_huge_pages,4904				   huge_page_size(h) / SZ_1K);4905	}4906 4907	seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);4908}4909 4910int hugetlb_report_node_meminfo(char *buf, int len, int nid)4911{4912	struct hstate *h = &default_hstate;4913 4914	if (!hugepages_supported())4915		return 0;4916 4917	return sysfs_emit_at(buf, len,4918			     "Node %d HugePages_Total: %5u\n"4919			     "Node %d HugePages_Free:  %5u\n"4920			     "Node %d HugePages_Surp:  %5u\n",4921			     nid, h->nr_huge_pages_node[nid],4922			     nid, h->free_huge_pages_node[nid],4923			     nid, h->surplus_huge_pages_node[nid]);4924}4925 4926void hugetlb_show_meminfo_node(int nid)4927{4928	struct hstate *h;4929 4930	if (!hugepages_supported())4931		return;4932 4933	for_each_hstate(h)4934		printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",4935			nid,4936			h->nr_huge_pages_node[nid],4937			h->free_huge_pages_node[nid],4938			h->surplus_huge_pages_node[nid],4939			huge_page_size(h) / SZ_1K);4940}4941 4942void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)4943{4944	seq_printf(m, "HugetlbPages:\t%8lu kB\n",4945		   K(atomic_long_read(&mm->hugetlb_usage)));4946}4947 4948/* Return the number pages of memory we physically have, in PAGE_SIZE units. */4949unsigned long hugetlb_total_pages(void)4950{4951	struct hstate *h;4952	unsigned long nr_total_pages = 0;4953 4954	for_each_hstate(h)4955		nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);4956	return nr_total_pages;4957}4958 4959static int hugetlb_acct_memory(struct hstate *h, long delta)4960{4961	int ret = -ENOMEM;4962 4963	if (!delta)4964		return 0;4965 4966	spin_lock_irq(&hugetlb_lock);4967	/*4968	 * When cpuset is configured, it breaks the strict hugetlb page4969	 * reservation as the accounting is done on a global variable. Such4970	 * reservation is completely rubbish in the presence of cpuset because4971	 * the reservation is not checked against page availability for the4972	 * current cpuset. Application can still potentially OOM'ed by kernel4973	 * with lack of free htlb page in cpuset that the task is in.4974	 * Attempt to enforce strict accounting with cpuset is almost4975	 * impossible (or too ugly) because cpuset is too fluid that4976	 * task or memory node can be dynamically moved between cpusets.4977	 *4978	 * The change of semantics for shared hugetlb mapping with cpuset is4979	 * undesirable. However, in order to preserve some of the semantics,4980	 * we fall back to check against current free page availability as4981	 * a best attempt and hopefully to minimize the impact of changing4982	 * semantics that cpuset has.4983	 *4984	 * Apart from cpuset, we also have memory policy mechanism that4985	 * also determines from which node the kernel will allocate memory4986	 * in a NUMA system. So similar to cpuset, we also should consider4987	 * the memory policy of the current task. Similar to the description4988	 * above.4989	 */4990	if (delta > 0) {4991		if (gather_surplus_pages(h, delta) < 0)4992			goto out;4993 4994		if (delta > allowed_mems_nr(h)) {4995			return_unused_surplus_pages(h, delta);4996			goto out;4997		}4998	}4999 5000	ret = 0;5001	if (delta < 0)5002		return_unused_surplus_pages(h, (unsigned long) -delta);5003 5004out:5005	spin_unlock_irq(&hugetlb_lock);5006	return ret;5007}5008 5009static void hugetlb_vm_op_open(struct vm_area_struct *vma)5010{5011	struct resv_map *resv = vma_resv_map(vma);5012 5013	/*5014	 * HPAGE_RESV_OWNER indicates a private mapping.5015	 * This new VMA should share its siblings reservation map if present.5016	 * The VMA will only ever have a valid reservation map pointer where5017	 * it is being copied for another still existing VMA.  As that VMA5018	 * has a reference to the reservation map it cannot disappear until5019	 * after this open call completes.  It is therefore safe to take a5020	 * new reference here without additional locking.5021	 */5022	if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {5023		resv_map_dup_hugetlb_cgroup_uncharge_info(resv);5024		kref_get(&resv->refs);5025	}5026 5027	/*5028	 * vma_lock structure for sharable mappings is vma specific.5029	 * Clear old pointer (if copied via vm_area_dup) and allocate5030	 * new structure.  Before clearing, make sure vma_lock is not5031	 * for this vma.5032	 */5033	if (vma->vm_flags & VM_MAYSHARE) {5034		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;5035 5036		if (vma_lock) {5037			if (vma_lock->vma != vma) {5038				vma->vm_private_data = NULL;5039				hugetlb_vma_lock_alloc(vma);5040			} else5041				pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);5042		} else5043			hugetlb_vma_lock_alloc(vma);5044	}5045}5046 5047static void hugetlb_vm_op_close(struct vm_area_struct *vma)5048{5049	struct hstate *h = hstate_vma(vma);5050	struct resv_map *resv;5051	struct hugepage_subpool *spool = subpool_vma(vma);5052	unsigned long reserve, start, end;5053	long gbl_reserve;5054 5055	hugetlb_vma_lock_free(vma);5056 5057	resv = vma_resv_map(vma);5058	if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))5059		return;5060 5061	start = vma_hugecache_offset(h, vma, vma->vm_start);5062	end = vma_hugecache_offset(h, vma, vma->vm_end);5063 5064	reserve = (end - start) - region_count(resv, start, end);5065	hugetlb_cgroup_uncharge_counter(resv, start, end);5066	if (reserve) {5067		/*5068		 * Decrement reserve counts.  The global reserve count may be5069		 * adjusted if the subpool has a minimum size.5070		 */5071		gbl_reserve = hugepage_subpool_put_pages(spool, reserve);5072		hugetlb_acct_memory(h, -gbl_reserve);5073	}5074 5075	kref_put(&resv->refs, resv_map_release);5076}5077 5078static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)5079{5080	if (addr & ~(huge_page_mask(hstate_vma(vma))))5081		return -EINVAL;5082 5083	/*5084	 * PMD sharing is only possible for PUD_SIZE-aligned address ranges5085	 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this5086	 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.5087	 */5088	if (addr & ~PUD_MASK) {5089		/*5090		 * hugetlb_vm_op_split is called right before we attempt to5091		 * split the VMA. We will need to unshare PMDs in the old and5092		 * new VMAs, so let's unshare before we split.5093		 */5094		unsigned long floor = addr & PUD_MASK;5095		unsigned long ceil = floor + PUD_SIZE;5096 5097		if (floor >= vma->vm_start && ceil <= vma->vm_end)5098			hugetlb_unshare_pmds(vma, floor, ceil);5099	}5100 5101	return 0;5102}5103 5104static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)5105{5106	return huge_page_size(hstate_vma(vma));5107}5108 5109/*5110 * We cannot handle pagefaults against hugetlb pages at all.  They cause5111 * handle_mm_fault() to try to instantiate regular-sized pages in the5112 * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get5113 * this far.5114 */5115static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)5116{5117	BUG();5118	return 0;5119}5120 5121/*5122 * When a new function is introduced to vm_operations_struct and added5123 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.5124 * This is because under System V memory model, mappings created via5125 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,5126 * their original vm_ops are overwritten with shm_vm_ops.5127 */5128const struct vm_operations_struct hugetlb_vm_ops = {5129	.fault = hugetlb_vm_op_fault,5130	.open = hugetlb_vm_op_open,5131	.close = hugetlb_vm_op_close,5132	.may_split = hugetlb_vm_op_split,5133	.pagesize = hugetlb_vm_op_pagesize,5134};5135 5136static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,5137				int writable)5138{5139	pte_t entry;5140	unsigned int shift = huge_page_shift(hstate_vma(vma));5141 5142	if (writable) {5143		entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,5144					 vma->vm_page_prot)));5145	} else {5146		entry = huge_pte_wrprotect(mk_huge_pte(page,5147					   vma->vm_page_prot));5148	}5149	entry = pte_mkyoung(entry);5150	entry = arch_make_huge_pte(entry, shift, vma->vm_flags);5151 5152	return entry;5153}5154 5155static void set_huge_ptep_writable(struct vm_area_struct *vma,5156				   unsigned long address, pte_t *ptep)5157{5158	pte_t entry;5159 5160	entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(vma->vm_mm, address, ptep)));5161	if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))5162		update_mmu_cache(vma, address, ptep);5163}5164 5165bool is_hugetlb_entry_migration(pte_t pte)5166{5167	swp_entry_t swp;5168 5169	if (huge_pte_none(pte) || pte_present(pte))5170		return false;5171	swp = pte_to_swp_entry(pte);5172	if (is_migration_entry(swp))5173		return true;5174	else5175		return false;5176}5177 5178bool is_hugetlb_entry_hwpoisoned(pte_t pte)5179{5180	swp_entry_t swp;5181 5182	if (huge_pte_none(pte) || pte_present(pte))5183		return false;5184	swp = pte_to_swp_entry(pte);5185	if (is_hwpoison_entry(swp))5186		return true;5187	else5188		return false;5189}5190 5191static void5192hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,5193		      struct folio *new_folio, pte_t old, unsigned long sz)5194{5195	pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);5196 5197	__folio_mark_uptodate(new_folio);5198	hugetlb_add_new_anon_rmap(new_folio, vma, addr);5199	if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))5200		newpte = huge_pte_mkuffd_wp(newpte);5201	set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);5202	hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);5203	folio_set_hugetlb_migratable(new_folio);5204}5205 5206int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,5207			    struct vm_area_struct *dst_vma,5208			    struct vm_area_struct *src_vma)5209{5210	pte_t *src_pte, *dst_pte, entry;5211	struct folio *pte_folio;5212	unsigned long addr;5213	bool cow = is_cow_mapping(src_vma->vm_flags);5214	struct hstate *h = hstate_vma(src_vma);5215	unsigned long sz = huge_page_size(h);5216	unsigned long npages = pages_per_huge_page(h);5217	struct mmu_notifier_range range;5218	unsigned long last_addr_mask;5219	int ret = 0;5220 5221	if (cow) {5222		mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,5223					src_vma->vm_start,5224					src_vma->vm_end);5225		mmu_notifier_invalidate_range_start(&range);5226		vma_assert_write_locked(src_vma);5227		raw_write_seqcount_begin(&src->write_protect_seq);5228	} else {5229		/*5230		 * For shared mappings the vma lock must be held before5231		 * calling hugetlb_walk() in the src vma. Otherwise, the5232		 * returned ptep could go away if part of a shared pmd and5233		 * another thread calls huge_pmd_unshare.5234		 */5235		hugetlb_vma_lock_read(src_vma);5236	}5237 5238	last_addr_mask = hugetlb_mask_last_page(h);5239	for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {5240		spinlock_t *src_ptl, *dst_ptl;5241		src_pte = hugetlb_walk(src_vma, addr, sz);5242		if (!src_pte) {5243			addr |= last_addr_mask;5244			continue;5245		}5246		dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);5247		if (!dst_pte) {5248			ret = -ENOMEM;5249			break;5250		}5251 5252		/*5253		 * If the pagetables are shared don't copy or take references.5254		 *5255		 * dst_pte == src_pte is the common case of src/dest sharing.5256		 * However, src could have 'unshared' and dst shares with5257		 * another vma. So page_count of ptep page is checked instead5258		 * to reliably determine whether pte is shared.5259		 */5260		if (page_count(virt_to_page(dst_pte)) > 1) {5261			addr |= last_addr_mask;5262			continue;5263		}5264 5265		dst_ptl = huge_pte_lock(h, dst, dst_pte);5266		src_ptl = huge_pte_lockptr(h, src, src_pte);5267		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);5268		entry = huge_ptep_get(src_vma->vm_mm, addr, src_pte);5269again:5270		if (huge_pte_none(entry)) {5271			/*5272			 * Skip if src entry none.5273			 */5274			;5275		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {5276			if (!userfaultfd_wp(dst_vma))5277				entry = huge_pte_clear_uffd_wp(entry);5278			set_huge_pte_at(dst, addr, dst_pte, entry, sz);5279		} else if (unlikely(is_hugetlb_entry_migration(entry))) {5280			swp_entry_t swp_entry = pte_to_swp_entry(entry);5281			bool uffd_wp = pte_swp_uffd_wp(entry);5282 5283			if (!is_readable_migration_entry(swp_entry) && cow) {5284				/*5285				 * COW mappings require pages in both5286				 * parent and child to be set to read.5287				 */5288				swp_entry = make_readable_migration_entry(5289							swp_offset(swp_entry));5290				entry = swp_entry_to_pte(swp_entry);5291				if (userfaultfd_wp(src_vma) && uffd_wp)5292					entry = pte_swp_mkuffd_wp(entry);5293				set_huge_pte_at(src, addr, src_pte, entry, sz);5294			}5295			if (!userfaultfd_wp(dst_vma))5296				entry = huge_pte_clear_uffd_wp(entry);5297			set_huge_pte_at(dst, addr, dst_pte, entry, sz);5298		} else if (unlikely(is_pte_marker(entry))) {5299			pte_marker marker = copy_pte_marker(5300				pte_to_swp_entry(entry), dst_vma);5301 5302			if (marker)5303				set_huge_pte_at(dst, addr, dst_pte,5304						make_pte_marker(marker), sz);5305		} else {5306			entry = huge_ptep_get(src_vma->vm_mm, addr, src_pte);5307			pte_folio = page_folio(pte_page(entry));5308			folio_get(pte_folio);5309 5310			/*5311			 * Failing to duplicate the anon rmap is a rare case5312			 * where we see pinned hugetlb pages while they're5313			 * prone to COW. We need to do the COW earlier during5314			 * fork.5315			 *5316			 * When pre-allocating the page or copying data, we5317			 * need to be without the pgtable locks since we could5318			 * sleep during the process.5319			 */5320			if (!folio_test_anon(pte_folio)) {5321				hugetlb_add_file_rmap(pte_folio);5322			} else if (hugetlb_try_dup_anon_rmap(pte_folio, src_vma)) {5323				pte_t src_pte_old = entry;5324				struct folio *new_folio;5325 5326				spin_unlock(src_ptl);5327				spin_unlock(dst_ptl);5328				/* Do not use reserve as it's private owned */5329				new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);5330				if (IS_ERR(new_folio)) {5331					folio_put(pte_folio);5332					ret = PTR_ERR(new_folio);5333					break;5334				}5335				ret = copy_user_large_folio(new_folio, pte_folio,5336						ALIGN_DOWN(addr, sz), dst_vma);5337				folio_put(pte_folio);5338				if (ret) {5339					folio_put(new_folio);5340					break;5341				}5342 5343				/* Install the new hugetlb folio if src pte stable */5344				dst_ptl = huge_pte_lock(h, dst, dst_pte);5345				src_ptl = huge_pte_lockptr(h, src, src_pte);5346				spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);5347				entry = huge_ptep_get(src_vma->vm_mm, addr, src_pte);5348				if (!pte_same(src_pte_old, entry)) {5349					restore_reserve_on_error(h, dst_vma, addr,5350								new_folio);5351					folio_put(new_folio);5352					/* huge_ptep of dst_pte won't change as in child */5353					goto again;5354				}5355				hugetlb_install_folio(dst_vma, dst_pte, addr,5356						      new_folio, src_pte_old, sz);5357				spin_unlock(src_ptl);5358				spin_unlock(dst_ptl);5359				continue;5360			}5361 5362			if (cow) {5363				/*5364				 * No need to notify as we are downgrading page5365				 * table protection not changing it to point5366				 * to a new page.5367				 *5368				 * See Documentation/mm/mmu_notifier.rst5369				 */5370				huge_ptep_set_wrprotect(src, addr, src_pte);5371				entry = huge_pte_wrprotect(entry);5372			}5373 5374			if (!userfaultfd_wp(dst_vma))5375				entry = huge_pte_clear_uffd_wp(entry);5376 5377			set_huge_pte_at(dst, addr, dst_pte, entry, sz);5378			hugetlb_count_add(npages, dst);5379		}5380		spin_unlock(src_ptl);5381		spin_unlock(dst_ptl);5382	}5383 5384	if (cow) {5385		raw_write_seqcount_end(&src->write_protect_seq);5386		mmu_notifier_invalidate_range_end(&range);5387	} else {5388		hugetlb_vma_unlock_read(src_vma);5389	}5390 5391	return ret;5392}5393 5394static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,5395			  unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,5396			  unsigned long sz)5397{5398	struct hstate *h = hstate_vma(vma);5399	struct mm_struct *mm = vma->vm_mm;5400	spinlock_t *src_ptl, *dst_ptl;5401	pte_t pte;5402 5403	dst_ptl = huge_pte_lock(h, mm, dst_pte);5404	src_ptl = huge_pte_lockptr(h, mm, src_pte);5405 5406	/*5407	 * We don't have to worry about the ordering of src and dst ptlocks5408	 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.5409	 */5410	if (src_ptl != dst_ptl)5411		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);5412 5413	pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);5414	set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);5415 5416	if (src_ptl != dst_ptl)5417		spin_unlock(src_ptl);5418	spin_unlock(dst_ptl);5419}5420 5421int move_hugetlb_page_tables(struct vm_area_struct *vma,5422			     struct vm_area_struct *new_vma,5423			     unsigned long old_addr, unsigned long new_addr,5424			     unsigned long len)5425{5426	struct hstate *h = hstate_vma(vma);5427	struct address_space *mapping = vma->vm_file->f_mapping;5428	unsigned long sz = huge_page_size(h);5429	struct mm_struct *mm = vma->vm_mm;5430	unsigned long old_end = old_addr + len;5431	unsigned long last_addr_mask;5432	pte_t *src_pte, *dst_pte;5433	struct mmu_notifier_range range;5434	bool shared_pmd = false;5435 5436	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,5437				old_end);5438	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);5439	/*5440	 * In case of shared PMDs, we should cover the maximum possible5441	 * range.5442	 */5443	flush_cache_range(vma, range.start, range.end);5444 5445	mmu_notifier_invalidate_range_start(&range);5446	last_addr_mask = hugetlb_mask_last_page(h);5447	/* Prevent race with file truncation */5448	hugetlb_vma_lock_write(vma);5449	i_mmap_lock_write(mapping);5450	for (; old_addr < old_end; old_addr += sz, new_addr += sz) {5451		src_pte = hugetlb_walk(vma, old_addr, sz);5452		if (!src_pte) {5453			old_addr |= last_addr_mask;5454			new_addr |= last_addr_mask;5455			continue;5456		}5457		if (huge_pte_none(huge_ptep_get(mm, old_addr, src_pte)))5458			continue;5459 5460		if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {5461			shared_pmd = true;5462			old_addr |= last_addr_mask;5463			new_addr |= last_addr_mask;5464			continue;5465		}5466 5467		dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);5468		if (!dst_pte)5469			break;5470 5471		move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);5472	}5473 5474	if (shared_pmd)5475		flush_hugetlb_tlb_range(vma, range.start, range.end);5476	else5477		flush_hugetlb_tlb_range(vma, old_end - len, old_end);5478	mmu_notifier_invalidate_range_end(&range);5479	i_mmap_unlock_write(mapping);5480	hugetlb_vma_unlock_write(vma);5481 5482	return len + old_addr - old_end;5483}5484 5485void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,5486			    unsigned long start, unsigned long end,5487			    struct page *ref_page, zap_flags_t zap_flags)5488{5489	struct mm_struct *mm = vma->vm_mm;5490	unsigned long address;5491	pte_t *ptep;5492	pte_t pte;5493	spinlock_t *ptl;5494	struct page *page;5495	struct hstate *h = hstate_vma(vma);5496	unsigned long sz = huge_page_size(h);5497	bool adjust_reservation = false;5498	unsigned long last_addr_mask;5499	bool force_flush = false;5500 5501	WARN_ON(!is_vm_hugetlb_page(vma));5502	BUG_ON(start & ~huge_page_mask(h));5503	BUG_ON(end & ~huge_page_mask(h));5504 5505	/*5506	 * This is a hugetlb vma, all the pte entries should point5507	 * to huge page.5508	 */5509	tlb_change_page_size(tlb, sz);5510	tlb_start_vma(tlb, vma);5511 5512	last_addr_mask = hugetlb_mask_last_page(h);5513	address = start;5514	for (; address < end; address += sz) {5515		ptep = hugetlb_walk(vma, address, sz);5516		if (!ptep) {5517			address |= last_addr_mask;5518			continue;5519		}5520 5521		ptl = huge_pte_lock(h, mm, ptep);5522		if (huge_pmd_unshare(mm, vma, address, ptep)) {5523			spin_unlock(ptl);5524			tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);5525			force_flush = true;5526			address |= last_addr_mask;5527			continue;5528		}5529 5530		pte = huge_ptep_get(mm, address, ptep);5531		if (huge_pte_none(pte)) {5532			spin_unlock(ptl);5533			continue;5534		}5535 5536		/*5537		 * Migrating hugepage or HWPoisoned hugepage is already5538		 * unmapped and its refcount is dropped, so just clear pte here.5539		 */5540		if (unlikely(!pte_present(pte))) {5541			/*5542			 * If the pte was wr-protected by uffd-wp in any of the5543			 * swap forms, meanwhile the caller does not want to5544			 * drop the uffd-wp bit in this zap, then replace the5545			 * pte with a marker.5546			 */5547			if (pte_swp_uffd_wp_any(pte) &&5548			    !(zap_flags & ZAP_FLAG_DROP_MARKER))5549				set_huge_pte_at(mm, address, ptep,5550						make_pte_marker(PTE_MARKER_UFFD_WP),5551						sz);5552			else5553				huge_pte_clear(mm, address, ptep, sz);5554			spin_unlock(ptl);5555			continue;5556		}5557 5558		page = pte_page(pte);5559		/*5560		 * If a reference page is supplied, it is because a specific5561		 * page is being unmapped, not a range. Ensure the page we5562		 * are about to unmap is the actual page of interest.5563		 */5564		if (ref_page) {5565			if (page != ref_page) {5566				spin_unlock(ptl);5567				continue;5568			}5569			/*5570			 * Mark the VMA as having unmapped its page so that5571			 * future faults in this VMA will fail rather than5572			 * looking like data was lost5573			 */5574			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);5575		}5576 5577		pte = huge_ptep_get_and_clear(mm, address, ptep);5578		tlb_remove_huge_tlb_entry(h, tlb, ptep, address);5579		if (huge_pte_dirty(pte))5580			set_page_dirty(page);5581		/* Leave a uffd-wp pte marker if needed */5582		if (huge_pte_uffd_wp(pte) &&5583		    !(zap_flags & ZAP_FLAG_DROP_MARKER))5584			set_huge_pte_at(mm, address, ptep,5585					make_pte_marker(PTE_MARKER_UFFD_WP),5586					sz);5587		hugetlb_count_sub(pages_per_huge_page(h), mm);5588		hugetlb_remove_rmap(page_folio(page));5589 5590		/*5591		 * Restore the reservation for anonymous page, otherwise the5592		 * backing page could be stolen by someone.5593		 * If there we are freeing a surplus, do not set the restore5594		 * reservation bit.5595		 */5596		if (!h->surplus_huge_pages && __vma_private_lock(vma) &&5597		    folio_test_anon(page_folio(page))) {5598			folio_set_hugetlb_restore_reserve(page_folio(page));5599			/* Reservation to be adjusted after the spin lock */5600			adjust_reservation = true;5601		}5602 5603		spin_unlock(ptl);5604 5605		/*5606		 * Adjust the reservation for the region that will have the5607		 * reserve restored. Keep in mind that vma_needs_reservation() changes5608		 * resv->adds_in_progress if it succeeds. If this is not done,5609		 * do_exit() will not see it, and will keep the reservation5610		 * forever.5611		 */5612		if (adjust_reservation) {5613			int rc = vma_needs_reservation(h, vma, address);5614 5615			if (rc < 0)5616				/* Pressumably allocate_file_region_entries failed5617				 * to allocate a file_region struct. Clear5618				 * hugetlb_restore_reserve so that global reserve5619				 * count will not be incremented by free_huge_folio.5620				 * Act as if we consumed the reservation.5621				 */5622				folio_clear_hugetlb_restore_reserve(page_folio(page));5623			else if (rc)5624				vma_add_reservation(h, vma, address);5625		}5626 5627		tlb_remove_page_size(tlb, page, huge_page_size(h));5628		/*5629		 * Bail out after unmapping reference page if supplied5630		 */5631		if (ref_page)5632			break;5633	}5634	tlb_end_vma(tlb, vma);5635 5636	/*5637	 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We5638	 * could defer the flush until now, since by holding i_mmap_rwsem we5639	 * guaranteed that the last refernece would not be dropped. But we must5640	 * do the flushing before we return, as otherwise i_mmap_rwsem will be5641	 * dropped and the last reference to the shared PMDs page might be5642	 * dropped as well.5643	 *5644	 * In theory we could defer the freeing of the PMD pages as well, but5645	 * huge_pmd_unshare() relies on the exact page_count for the PMD page to5646	 * detect sharing, so we cannot defer the release of the page either.5647	 * Instead, do flush now.5648	 */5649	if (force_flush)5650		tlb_flush_mmu_tlbonly(tlb);5651}5652 5653void __hugetlb_zap_begin(struct vm_area_struct *vma,5654			 unsigned long *start, unsigned long *end)5655{5656	if (!vma->vm_file)	/* hugetlbfs_file_mmap error */5657		return;5658 5659	adjust_range_if_pmd_sharing_possible(vma, start, end);5660	hugetlb_vma_lock_write(vma);5661	if (vma->vm_file)5662		i_mmap_lock_write(vma->vm_file->f_mapping);5663}5664 5665void __hugetlb_zap_end(struct vm_area_struct *vma,5666		       struct zap_details *details)5667{5668	zap_flags_t zap_flags = details ? details->zap_flags : 0;5669 5670	if (!vma->vm_file)	/* hugetlbfs_file_mmap error */5671		return;5672 5673	if (zap_flags & ZAP_FLAG_UNMAP) {	/* final unmap */5674		/*5675		 * Unlock and free the vma lock before releasing i_mmap_rwsem.5676		 * When the vma_lock is freed, this makes the vma ineligible5677		 * for pmd sharing.  And, i_mmap_rwsem is required to set up5678		 * pmd sharing.  This is important as page tables for this5679		 * unmapped range will be asynchrously deleted.  If the page5680		 * tables are shared, there will be issues when accessed by5681		 * someone else.5682		 */5683		__hugetlb_vma_unlock_write_free(vma);5684	} else {5685		hugetlb_vma_unlock_write(vma);5686	}5687 5688	if (vma->vm_file)5689		i_mmap_unlock_write(vma->vm_file->f_mapping);5690}5691 5692void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,5693			  unsigned long end, struct page *ref_page,5694			  zap_flags_t zap_flags)5695{5696	struct mmu_notifier_range range;5697	struct mmu_gather tlb;5698 5699	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,5700				start, end);5701	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);5702	mmu_notifier_invalidate_range_start(&range);5703	tlb_gather_mmu(&tlb, vma->vm_mm);5704 5705	__unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);5706 5707	mmu_notifier_invalidate_range_end(&range);5708	tlb_finish_mmu(&tlb);5709}5710 5711/*5712 * This is called when the original mapper is failing to COW a MAP_PRIVATE5713 * mapping it owns the reserve page for. The intention is to unmap the page5714 * from other VMAs and let the children be SIGKILLed if they are faulting the5715 * same region.5716 */5717static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,5718			      struct page *page, unsigned long address)5719{5720	struct hstate *h = hstate_vma(vma);5721	struct vm_area_struct *iter_vma;5722	struct address_space *mapping;5723	pgoff_t pgoff;5724 5725	/*5726	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation5727	 * from page cache lookup which is in HPAGE_SIZE units.5728	 */5729	address = address & huge_page_mask(h);5730	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +5731			vma->vm_pgoff;5732	mapping = vma->vm_file->f_mapping;5733 5734	/*5735	 * Take the mapping lock for the duration of the table walk. As5736	 * this mapping should be shared between all the VMAs,5737	 * __unmap_hugepage_range() is called as the lock is already held5738	 */5739	i_mmap_lock_write(mapping);5740	vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {5741		/* Do not unmap the current VMA */5742		if (iter_vma == vma)5743			continue;5744 5745		/*5746		 * Shared VMAs have their own reserves and do not affect5747		 * MAP_PRIVATE accounting but it is possible that a shared5748		 * VMA is using the same page so check and skip such VMAs.5749		 */5750		if (iter_vma->vm_flags & VM_MAYSHARE)5751			continue;5752 5753		/*5754		 * Unmap the page from other VMAs without their own reserves.5755		 * They get marked to be SIGKILLed if they fault in these5756		 * areas. This is because a future no-page fault on this VMA5757		 * could insert a zeroed page instead of the data existing5758		 * from the time of fork. This would look like data corruption5759		 */5760		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))5761			unmap_hugepage_range(iter_vma, address,5762					     address + huge_page_size(h), page, 0);5763	}5764	i_mmap_unlock_write(mapping);5765}5766 5767/*5768 * hugetlb_wp() should be called with page lock of the original hugepage held.5769 * Called with hugetlb_fault_mutex_table held and pte_page locked so we5770 * cannot race with other handlers or page migration.5771 * Keep the pte_same checks anyway to make transition from the mutex easier.5772 */5773static vm_fault_t hugetlb_wp(struct folio *pagecache_folio,5774		       struct vm_fault *vmf)5775{5776	struct vm_area_struct *vma = vmf->vma;5777	struct mm_struct *mm = vma->vm_mm;5778	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;5779	pte_t pte = huge_ptep_get(mm, vmf->address, vmf->pte);5780	struct hstate *h = hstate_vma(vma);5781	struct folio *old_folio;5782	struct folio *new_folio;5783	int outside_reserve = 0;5784	vm_fault_t ret = 0;5785	struct mmu_notifier_range range;5786 5787	/*5788	 * Never handle CoW for uffd-wp protected pages.  It should be only5789	 * handled when the uffd-wp protection is removed.5790	 *5791	 * Note that only the CoW optimization path (in hugetlb_no_page())5792	 * can trigger this, because hugetlb_fault() will always resolve5793	 * uffd-wp bit first.5794	 */5795	if (!unshare && huge_pte_uffd_wp(pte))5796		return 0;5797 5798	/*5799	 * hugetlb does not support FOLL_FORCE-style write faults that keep the5800	 * PTE mapped R/O such as maybe_mkwrite() would do.5801	 */5802	if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))5803		return VM_FAULT_SIGSEGV;5804 5805	/* Let's take out MAP_SHARED mappings first. */5806	if (vma->vm_flags & VM_MAYSHARE) {5807		set_huge_ptep_writable(vma, vmf->address, vmf->pte);5808		return 0;5809	}5810 5811	old_folio = page_folio(pte_page(pte));5812 5813	delayacct_wpcopy_start();5814 5815retry_avoidcopy:5816	/*5817	 * If no-one else is actually using this page, we're the exclusive5818	 * owner and can reuse this page.5819	 *5820	 * Note that we don't rely on the (safer) folio refcount here, because5821	 * copying the hugetlb folio when there are unexpected (temporary)5822	 * folio references could harm simple fork()+exit() users when5823	 * we run out of free hugetlb folios: we would have to kill processes5824	 * in scenarios that used to work. As a side effect, there can still5825	 * be leaks between processes, for example, with FOLL_GET users.5826	 */5827	if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {5828		if (!PageAnonExclusive(&old_folio->page)) {5829			folio_move_anon_rmap(old_folio, vma);5830			SetPageAnonExclusive(&old_folio->page);5831		}5832		if (likely(!unshare))5833			set_huge_ptep_writable(vma, vmf->address, vmf->pte);5834 5835		delayacct_wpcopy_end();5836		return 0;5837	}5838	VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&5839		       PageAnonExclusive(&old_folio->page), &old_folio->page);5840 5841	/*5842	 * If the process that created a MAP_PRIVATE mapping is about to5843	 * perform a COW due to a shared page count, attempt to satisfy5844	 * the allocation without using the existing reserves. The pagecache5845	 * page is used to determine if the reserve at this address was5846	 * consumed or not. If reserves were used, a partial faulted mapping5847	 * at the time of fork() could consume its reserves on COW instead5848	 * of the full address range.5849	 */5850	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&5851			old_folio != pagecache_folio)5852		outside_reserve = 1;5853 5854	folio_get(old_folio);5855 5856	/*5857	 * Drop page table lock as buddy allocator may be called. It will5858	 * be acquired again before returning to the caller, as expected.5859	 */5860	spin_unlock(vmf->ptl);5861	new_folio = alloc_hugetlb_folio(vma, vmf->address, outside_reserve);5862 5863	if (IS_ERR(new_folio)) {5864		/*5865		 * If a process owning a MAP_PRIVATE mapping fails to COW,5866		 * it is due to references held by a child and an insufficient5867		 * huge page pool. To guarantee the original mappers5868		 * reliability, unmap the page from child processes. The child5869		 * may get SIGKILLed if it later faults.5870		 */5871		if (outside_reserve) {5872			struct address_space *mapping = vma->vm_file->f_mapping;5873			pgoff_t idx;5874			u32 hash;5875 5876			folio_put(old_folio);5877			/*5878			 * Drop hugetlb_fault_mutex and vma_lock before5879			 * unmapping.  unmapping needs to hold vma_lock5880			 * in write mode.  Dropping vma_lock in read mode5881			 * here is OK as COW mappings do not interact with5882			 * PMD sharing.5883			 *5884			 * Reacquire both after unmap operation.5885			 */5886			idx = vma_hugecache_offset(h, vma, vmf->address);5887			hash = hugetlb_fault_mutex_hash(mapping, idx);5888			hugetlb_vma_unlock_read(vma);5889			mutex_unlock(&hugetlb_fault_mutex_table[hash]);5890 5891			unmap_ref_private(mm, vma, &old_folio->page,5892					vmf->address);5893 5894			mutex_lock(&hugetlb_fault_mutex_table[hash]);5895			hugetlb_vma_lock_read(vma);5896			spin_lock(vmf->ptl);5897			vmf->pte = hugetlb_walk(vma, vmf->address,5898					huge_page_size(h));5899			if (likely(vmf->pte &&5900				   pte_same(huge_ptep_get(mm, vmf->address, vmf->pte), pte)))5901				goto retry_avoidcopy;5902			/*5903			 * race occurs while re-acquiring page table5904			 * lock, and our job is done.5905			 */5906			delayacct_wpcopy_end();5907			return 0;5908		}5909 5910		ret = vmf_error(PTR_ERR(new_folio));5911		goto out_release_old;5912	}5913 5914	/*5915	 * When the original hugepage is shared one, it does not have5916	 * anon_vma prepared.5917	 */5918	ret = __vmf_anon_prepare(vmf);5919	if (unlikely(ret))5920		goto out_release_all;5921 5922	if (copy_user_large_folio(new_folio, old_folio, vmf->real_address, vma)) {5923		ret = VM_FAULT_HWPOISON_LARGE | VM_FAULT_SET_HINDEX(hstate_index(h));5924		goto out_release_all;5925	}5926	__folio_mark_uptodate(new_folio);5927 5928	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, vmf->address,5929				vmf->address + huge_page_size(h));5930	mmu_notifier_invalidate_range_start(&range);5931 5932	/*5933	 * Retake the page table lock to check for racing updates5934	 * before the page tables are altered5935	 */5936	spin_lock(vmf->ptl);5937	vmf->pte = hugetlb_walk(vma, vmf->address, huge_page_size(h));5938	if (likely(vmf->pte && pte_same(huge_ptep_get(mm, vmf->address, vmf->pte), pte))) {5939		pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);5940 5941		/* Break COW or unshare */5942		huge_ptep_clear_flush(vma, vmf->address, vmf->pte);5943		hugetlb_remove_rmap(old_folio);5944		hugetlb_add_new_anon_rmap(new_folio, vma, vmf->address);5945		if (huge_pte_uffd_wp(pte))5946			newpte = huge_pte_mkuffd_wp(newpte);5947		set_huge_pte_at(mm, vmf->address, vmf->pte, newpte,5948				huge_page_size(h));5949		folio_set_hugetlb_migratable(new_folio);5950		/* Make the old page be freed below */5951		new_folio = old_folio;5952	}5953	spin_unlock(vmf->ptl);5954	mmu_notifier_invalidate_range_end(&range);5955out_release_all:5956	/*5957	 * No restore in case of successful pagetable update (Break COW or5958	 * unshare)5959	 */5960	if (new_folio != old_folio)5961		restore_reserve_on_error(h, vma, vmf->address, new_folio);5962	folio_put(new_folio);5963out_release_old:5964	folio_put(old_folio);5965 5966	spin_lock(vmf->ptl); /* Caller expects lock to be held */5967 5968	delayacct_wpcopy_end();5969	return ret;5970}5971 5972/*5973 * Return whether there is a pagecache page to back given address within VMA.5974 */5975bool hugetlbfs_pagecache_present(struct hstate *h,5976				 struct vm_area_struct *vma, unsigned long address)5977{5978	struct address_space *mapping = vma->vm_file->f_mapping;5979	pgoff_t idx = linear_page_index(vma, address);5980	struct folio *folio;5981 5982	folio = filemap_get_folio(mapping, idx);5983	if (IS_ERR(folio))5984		return false;5985	folio_put(folio);5986	return true;5987}5988 5989int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,5990			   pgoff_t idx)5991{5992	struct inode *inode = mapping->host;5993	struct hstate *h = hstate_inode(inode);5994	int err;5995 5996	idx <<= huge_page_order(h);5997	__folio_set_locked(folio);5998	err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);5999 6000	if (unlikely(err)) {6001		__folio_clear_locked(folio);6002		return err;6003	}6004	folio_clear_hugetlb_restore_reserve(folio);6005 6006	/*6007	 * mark folio dirty so that it will not be removed from cache/file6008	 * by non-hugetlbfs specific code paths.6009	 */6010	folio_mark_dirty(folio);6011 6012	spin_lock(&inode->i_lock);6013	inode->i_blocks += blocks_per_huge_page(h);6014	spin_unlock(&inode->i_lock);6015	return 0;6016}6017 6018static inline vm_fault_t hugetlb_handle_userfault(struct vm_fault *vmf,6019						  struct address_space *mapping,6020						  unsigned long reason)6021{6022	u32 hash;6023 6024	/*6025	 * vma_lock and hugetlb_fault_mutex must be dropped before handling6026	 * userfault. Also mmap_lock could be dropped due to handling6027	 * userfault, any vma operation should be careful from here.6028	 */6029	hugetlb_vma_unlock_read(vmf->vma);6030	hash = hugetlb_fault_mutex_hash(mapping, vmf->pgoff);6031	mutex_unlock(&hugetlb_fault_mutex_table[hash]);6032	return handle_userfault(vmf, reason);6033}6034 6035/*6036 * Recheck pte with pgtable lock.  Returns true if pte didn't change, or6037 * false if pte changed or is changing.6038 */6039static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, unsigned long addr,6040			       pte_t *ptep, pte_t old_pte)6041{6042	spinlock_t *ptl;6043	bool same;6044 6045	ptl = huge_pte_lock(h, mm, ptep);6046	same = pte_same(huge_ptep_get(mm, addr, ptep), old_pte);6047	spin_unlock(ptl);6048 6049	return same;6050}6051 6052static vm_fault_t hugetlb_no_page(struct address_space *mapping,6053			struct vm_fault *vmf)6054{6055	struct vm_area_struct *vma = vmf->vma;6056	struct mm_struct *mm = vma->vm_mm;6057	struct hstate *h = hstate_vma(vma);6058	vm_fault_t ret = VM_FAULT_SIGBUS;6059	int anon_rmap = 0;6060	unsigned long size;6061	struct folio *folio;6062	pte_t new_pte;6063	bool new_folio, new_pagecache_folio = false;6064	u32 hash = hugetlb_fault_mutex_hash(mapping, vmf->pgoff);6065 6066	/*6067	 * Currently, we are forced to kill the process in the event the6068	 * original mapper has unmapped pages from the child due to a failed6069	 * COW/unsharing. Warn that such a situation has occurred as it may not6070	 * be obvious.6071	 */6072	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {6073		pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",6074			   current->pid);6075		goto out;6076	}6077 6078	/*6079	 * Use page lock to guard against racing truncation6080	 * before we get page_table_lock.6081	 */6082	new_folio = false;6083	folio = filemap_lock_hugetlb_folio(h, mapping, vmf->pgoff);6084	if (IS_ERR(folio)) {6085		size = i_size_read(mapping->host) >> huge_page_shift(h);6086		if (vmf->pgoff >= size)6087			goto out;6088		/* Check for page in userfault range */6089		if (userfaultfd_missing(vma)) {6090			/*6091			 * Since hugetlb_no_page() was examining pte6092			 * without pgtable lock, we need to re-test under6093			 * lock because the pte may not be stable and could6094			 * have changed from under us.  Try to detect6095			 * either changed or during-changing ptes and retry6096			 * properly when needed.6097			 *6098			 * Note that userfaultfd is actually fine with6099			 * false positives (e.g. caused by pte changed),6100			 * but not wrong logical events (e.g. caused by6101			 * reading a pte during changing).  The latter can6102			 * confuse the userspace, so the strictness is very6103			 * much preferred.  E.g., MISSING event should6104			 * never happen on the page after UFFDIO_COPY has6105			 * correctly installed the page and returned.6106			 */6107			if (!hugetlb_pte_stable(h, mm, vmf->address, vmf->pte, vmf->orig_pte)) {6108				ret = 0;6109				goto out;6110			}6111 6112			return hugetlb_handle_userfault(vmf, mapping,6113							VM_UFFD_MISSING);6114		}6115 6116		if (!(vma->vm_flags & VM_MAYSHARE)) {6117			ret = __vmf_anon_prepare(vmf);6118			if (unlikely(ret))6119				goto out;6120		}6121 6122		folio = alloc_hugetlb_folio(vma, vmf->address, 0);6123		if (IS_ERR(folio)) {6124			/*6125			 * Returning error will result in faulting task being6126			 * sent SIGBUS.  The hugetlb fault mutex prevents two6127			 * tasks from racing to fault in the same page which6128			 * could result in false unable to allocate errors.6129			 * Page migration does not take the fault mutex, but6130			 * does a clear then write of pte's under page table6131			 * lock.  Page fault code could race with migration,6132			 * notice the clear pte and try to allocate a page6133			 * here.  Before returning error, get ptl and make6134			 * sure there really is no pte entry.6135			 */6136			if (hugetlb_pte_stable(h, mm, vmf->address, vmf->pte, vmf->orig_pte))6137				ret = vmf_error(PTR_ERR(folio));6138			else6139				ret = 0;6140			goto out;6141		}6142		folio_zero_user(folio, vmf->real_address);6143		__folio_mark_uptodate(folio);6144		new_folio = true;6145 6146		if (vma->vm_flags & VM_MAYSHARE) {6147			int err = hugetlb_add_to_page_cache(folio, mapping,6148							vmf->pgoff);6149			if (err) {6150				/*6151				 * err can't be -EEXIST which implies someone6152				 * else consumed the reservation since hugetlb6153				 * fault mutex is held when add a hugetlb page6154				 * to the page cache. So it's safe to call6155				 * restore_reserve_on_error() here.6156				 */6157				restore_reserve_on_error(h, vma, vmf->address,6158							folio);6159				folio_put(folio);6160				ret = VM_FAULT_SIGBUS;6161				goto out;6162			}6163			new_pagecache_folio = true;6164		} else {6165			folio_lock(folio);6166			anon_rmap = 1;6167		}6168	} else {6169		/*6170		 * If memory error occurs between mmap() and fault, some process6171		 * don't have hwpoisoned swap entry for errored virtual address.6172		 * So we need to block hugepage fault by PG_hwpoison bit check.6173		 */6174		if (unlikely(folio_test_hwpoison(folio))) {6175			ret = VM_FAULT_HWPOISON_LARGE |6176				VM_FAULT_SET_HINDEX(hstate_index(h));6177			goto backout_unlocked;6178		}6179 6180		/* Check for page in userfault range. */6181		if (userfaultfd_minor(vma)) {6182			folio_unlock(folio);6183			folio_put(folio);6184			/* See comment in userfaultfd_missing() block above */6185			if (!hugetlb_pte_stable(h, mm, vmf->address, vmf->pte, vmf->orig_pte)) {6186				ret = 0;6187				goto out;6188			}6189			return hugetlb_handle_userfault(vmf, mapping,6190							VM_UFFD_MINOR);6191		}6192	}6193 6194	/*6195	 * If we are going to COW a private mapping later, we examine the6196	 * pending reservations for this page now. This will ensure that6197	 * any allocations necessary to record that reservation occur outside6198	 * the spinlock.6199	 */6200	if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {6201		if (vma_needs_reservation(h, vma, vmf->address) < 0) {6202			ret = VM_FAULT_OOM;6203			goto backout_unlocked;6204		}6205		/* Just decrements count, does not deallocate */6206		vma_end_reservation(h, vma, vmf->address);6207	}6208 6209	vmf->ptl = huge_pte_lock(h, mm, vmf->pte);6210	ret = 0;6211	/* If pte changed from under us, retry */6212	if (!pte_same(huge_ptep_get(mm, vmf->address, vmf->pte), vmf->orig_pte))6213		goto backout;6214 6215	if (anon_rmap)6216		hugetlb_add_new_anon_rmap(folio, vma, vmf->address);6217	else6218		hugetlb_add_file_rmap(folio);6219	new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)6220				&& (vma->vm_flags & VM_SHARED)));6221	/*6222	 * If this pte was previously wr-protected, keep it wr-protected even6223	 * if populated.6224	 */6225	if (unlikely(pte_marker_uffd_wp(vmf->orig_pte)))6226		new_pte = huge_pte_mkuffd_wp(new_pte);6227	set_huge_pte_at(mm, vmf->address, vmf->pte, new_pte, huge_page_size(h));6228 6229	hugetlb_count_add(pages_per_huge_page(h), mm);6230	if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {6231		/* Optimization, do the COW without a second fault */6232		ret = hugetlb_wp(folio, vmf);6233	}6234 6235	spin_unlock(vmf->ptl);6236 6237	/*6238	 * Only set hugetlb_migratable in newly allocated pages.  Existing pages6239	 * found in the pagecache may not have hugetlb_migratable if they have6240	 * been isolated for migration.6241	 */6242	if (new_folio)6243		folio_set_hugetlb_migratable(folio);6244 6245	folio_unlock(folio);6246out:6247	hugetlb_vma_unlock_read(vma);6248 6249	/*6250	 * We must check to release the per-VMA lock. __vmf_anon_prepare() is6251	 * the only way ret can be set to VM_FAULT_RETRY.6252	 */6253	if (unlikely(ret & VM_FAULT_RETRY))6254		vma_end_read(vma);6255 6256	mutex_unlock(&hugetlb_fault_mutex_table[hash]);6257	return ret;6258 6259backout:6260	spin_unlock(vmf->ptl);6261backout_unlocked:6262	if (new_folio && !new_pagecache_folio)6263		restore_reserve_on_error(h, vma, vmf->address, folio);6264 6265	folio_unlock(folio);6266	folio_put(folio);6267	goto out;6268}6269 6270#ifdef CONFIG_SMP6271u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)6272{6273	unsigned long key[2];6274	u32 hash;6275 6276	key[0] = (unsigned long) mapping;6277	key[1] = idx;6278 6279	hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);6280 6281	return hash & (num_fault_mutexes - 1);6282}6283#else6284/*6285 * For uniprocessor systems we always use a single mutex, so just6286 * return 0 and avoid the hashing overhead.6287 */6288u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)6289{6290	return 0;6291}6292#endif6293 6294vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,6295			unsigned long address, unsigned int flags)6296{6297	vm_fault_t ret;6298	u32 hash;6299	struct folio *folio = NULL;6300	struct folio *pagecache_folio = NULL;6301	struct hstate *h = hstate_vma(vma);6302	struct address_space *mapping;6303	int need_wait_lock = 0;6304	struct vm_fault vmf = {6305		.vma = vma,6306		.address = address & huge_page_mask(h),6307		.real_address = address,6308		.flags = flags,6309		.pgoff = vma_hugecache_offset(h, vma,6310				address & huge_page_mask(h)),6311		/* TODO: Track hugetlb faults using vm_fault */6312 6313		/*6314		 * Some fields may not be initialized, be careful as it may6315		 * be hard to debug if called functions make assumptions6316		 */6317	};6318 6319	/*6320	 * Serialize hugepage allocation and instantiation, so that we don't6321	 * get spurious allocation failures if two CPUs race to instantiate6322	 * the same page in the page cache.6323	 */6324	mapping = vma->vm_file->f_mapping;6325	hash = hugetlb_fault_mutex_hash(mapping, vmf.pgoff);6326	mutex_lock(&hugetlb_fault_mutex_table[hash]);6327 6328	/*6329	 * Acquire vma lock before calling huge_pte_alloc and hold6330	 * until finished with vmf.pte.  This prevents huge_pmd_unshare from6331	 * being called elsewhere and making the vmf.pte no longer valid.6332	 */6333	hugetlb_vma_lock_read(vma);6334	vmf.pte = huge_pte_alloc(mm, vma, vmf.address, huge_page_size(h));6335	if (!vmf.pte) {6336		hugetlb_vma_unlock_read(vma);6337		mutex_unlock(&hugetlb_fault_mutex_table[hash]);6338		return VM_FAULT_OOM;6339	}6340 6341	vmf.orig_pte = huge_ptep_get(mm, vmf.address, vmf.pte);6342	if (huge_pte_none_mostly(vmf.orig_pte)) {6343		if (is_pte_marker(vmf.orig_pte)) {6344			pte_marker marker =6345				pte_marker_get(pte_to_swp_entry(vmf.orig_pte));6346 6347			if (marker & PTE_MARKER_POISONED) {6348				ret = VM_FAULT_HWPOISON_LARGE |6349				      VM_FAULT_SET_HINDEX(hstate_index(h));6350				goto out_mutex;6351			}6352		}6353 6354		/*6355		 * Other PTE markers should be handled the same way as none PTE.6356		 *6357		 * hugetlb_no_page will drop vma lock and hugetlb fault6358		 * mutex internally, which make us return immediately.6359		 */6360		return hugetlb_no_page(mapping, &vmf);6361	}6362 6363	ret = 0;6364 6365	/*6366	 * vmf.orig_pte could be a migration/hwpoison vmf.orig_pte at this6367	 * point, so this check prevents the kernel from going below assuming6368	 * that we have an active hugepage in pagecache. This goto expects6369	 * the 2nd page fault, and is_hugetlb_entry_(migration|hwpoisoned)6370	 * check will properly handle it.6371	 */6372	if (!pte_present(vmf.orig_pte)) {6373		if (unlikely(is_hugetlb_entry_migration(vmf.orig_pte))) {6374			/*6375			 * Release the hugetlb fault lock now, but retain6376			 * the vma lock, because it is needed to guard the6377			 * huge_pte_lockptr() later in6378			 * migration_entry_wait_huge(). The vma lock will6379			 * be released there.6380			 */6381			mutex_unlock(&hugetlb_fault_mutex_table[hash]);6382			migration_entry_wait_huge(vma, vmf.address, vmf.pte);6383			return 0;6384		} else if (unlikely(is_hugetlb_entry_hwpoisoned(vmf.orig_pte)))6385			ret = VM_FAULT_HWPOISON_LARGE |6386			    VM_FAULT_SET_HINDEX(hstate_index(h));6387		goto out_mutex;6388	}6389 6390	/*6391	 * If we are going to COW/unshare the mapping later, we examine the6392	 * pending reservations for this page now. This will ensure that any6393	 * allocations necessary to record that reservation occur outside the6394	 * spinlock. Also lookup the pagecache page now as it is used to6395	 * determine if a reservation has been consumed.6396	 */6397	if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&6398	    !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(vmf.orig_pte)) {6399		if (vma_needs_reservation(h, vma, vmf.address) < 0) {6400			ret = VM_FAULT_OOM;6401			goto out_mutex;6402		}6403		/* Just decrements count, does not deallocate */6404		vma_end_reservation(h, vma, vmf.address);6405 6406		pagecache_folio = filemap_lock_hugetlb_folio(h, mapping,6407							     vmf.pgoff);6408		if (IS_ERR(pagecache_folio))6409			pagecache_folio = NULL;6410	}6411 6412	vmf.ptl = huge_pte_lock(h, mm, vmf.pte);6413 6414	/* Check for a racing update before calling hugetlb_wp() */6415	if (unlikely(!pte_same(vmf.orig_pte, huge_ptep_get(mm, vmf.address, vmf.pte))))6416		goto out_ptl;6417 6418	/* Handle userfault-wp first, before trying to lock more pages */6419	if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(mm, vmf.address, vmf.pte)) &&6420	    (flags & FAULT_FLAG_WRITE) && !huge_pte_write(vmf.orig_pte)) {6421		if (!userfaultfd_wp_async(vma)) {6422			spin_unlock(vmf.ptl);6423			if (pagecache_folio) {6424				folio_unlock(pagecache_folio);6425				folio_put(pagecache_folio);6426			}6427			hugetlb_vma_unlock_read(vma);6428			mutex_unlock(&hugetlb_fault_mutex_table[hash]);6429			return handle_userfault(&vmf, VM_UFFD_WP);6430		}6431 6432		vmf.orig_pte = huge_pte_clear_uffd_wp(vmf.orig_pte);6433		set_huge_pte_at(mm, vmf.address, vmf.pte, vmf.orig_pte,6434				huge_page_size(hstate_vma(vma)));6435		/* Fallthrough to CoW */6436	}6437 6438	/*6439	 * hugetlb_wp() requires page locks of pte_page(vmf.orig_pte) and6440	 * pagecache_folio, so here we need take the former one6441	 * when folio != pagecache_folio or !pagecache_folio.6442	 */6443	folio = page_folio(pte_page(vmf.orig_pte));6444	if (folio != pagecache_folio)6445		if (!folio_trylock(folio)) {6446			need_wait_lock = 1;6447			goto out_ptl;6448		}6449 6450	folio_get(folio);6451 6452	if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {6453		if (!huge_pte_write(vmf.orig_pte)) {6454			ret = hugetlb_wp(pagecache_folio, &vmf);6455			goto out_put_page;6456		} else if (likely(flags & FAULT_FLAG_WRITE)) {6457			vmf.orig_pte = huge_pte_mkdirty(vmf.orig_pte);6458		}6459	}6460	vmf.orig_pte = pte_mkyoung(vmf.orig_pte);6461	if (huge_ptep_set_access_flags(vma, vmf.address, vmf.pte, vmf.orig_pte,6462						flags & FAULT_FLAG_WRITE))6463		update_mmu_cache(vma, vmf.address, vmf.pte);6464out_put_page:6465	if (folio != pagecache_folio)6466		folio_unlock(folio);6467	folio_put(folio);6468out_ptl:6469	spin_unlock(vmf.ptl);6470 6471	if (pagecache_folio) {6472		folio_unlock(pagecache_folio);6473		folio_put(pagecache_folio);6474	}6475out_mutex:6476	hugetlb_vma_unlock_read(vma);6477 6478	/*6479	 * We must check to release the per-VMA lock. __vmf_anon_prepare() in6480	 * hugetlb_wp() is the only way ret can be set to VM_FAULT_RETRY.6481	 */6482	if (unlikely(ret & VM_FAULT_RETRY))6483		vma_end_read(vma);6484 6485	mutex_unlock(&hugetlb_fault_mutex_table[hash]);6486	/*6487	 * Generally it's safe to hold refcount during waiting page lock. But6488	 * here we just wait to defer the next page fault to avoid busy loop and6489	 * the page is not used after unlocked before returning from the current6490	 * page fault. So we are safe from accessing freed page, even if we wait6491	 * here without taking refcount.6492	 */6493	if (need_wait_lock)6494		folio_wait_locked(folio);6495	return ret;6496}6497 6498#ifdef CONFIG_USERFAULTFD6499/*6500 * Can probably be eliminated, but still used by hugetlb_mfill_atomic_pte().6501 */6502static struct folio *alloc_hugetlb_folio_vma(struct hstate *h,6503		struct vm_area_struct *vma, unsigned long address)6504{6505	struct mempolicy *mpol;6506	nodemask_t *nodemask;6507	struct folio *folio;6508	gfp_t gfp_mask;6509	int node;6510 6511	gfp_mask = htlb_alloc_mask(h);6512	node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);6513	/*6514	 * This is used to allocate a temporary hugetlb to hold the copied6515	 * content, which will then be copied again to the final hugetlb6516	 * consuming a reservation. Set the alloc_fallback to false to indicate6517	 * that breaking the per-node hugetlb pool is not allowed in this case.6518	 */6519	folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask, false);6520	mpol_cond_put(mpol);6521 6522	return folio;6523}6524 6525/*6526 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte6527 * with modifications for hugetlb pages.6528 */6529int hugetlb_mfill_atomic_pte(pte_t *dst_pte,6530			     struct vm_area_struct *dst_vma,6531			     unsigned long dst_addr,6532			     unsigned long src_addr,6533			     uffd_flags_t flags,6534			     struct folio **foliop)6535{6536	struct mm_struct *dst_mm = dst_vma->vm_mm;6537	bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);6538	bool wp_enabled = (flags & MFILL_ATOMIC_WP);6539	struct hstate *h = hstate_vma(dst_vma);6540	struct address_space *mapping = dst_vma->vm_file->f_mapping;6541	pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);6542	unsigned long size = huge_page_size(h);6543	int vm_shared = dst_vma->vm_flags & VM_SHARED;6544	pte_t _dst_pte;6545	spinlock_t *ptl;6546	int ret = -ENOMEM;6547	struct folio *folio;6548	int writable;6549	bool folio_in_pagecache = false;6550 6551	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {6552		ptl = huge_pte_lock(h, dst_mm, dst_pte);6553 6554		/* Don't overwrite any existing PTEs (even markers) */6555		if (!huge_pte_none(huge_ptep_get(dst_mm, dst_addr, dst_pte))) {6556			spin_unlock(ptl);6557			return -EEXIST;6558		}6559 6560		_dst_pte = make_pte_marker(PTE_MARKER_POISONED);6561		set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, size);6562 6563		/* No need to invalidate - it was non-present before */6564		update_mmu_cache(dst_vma, dst_addr, dst_pte);6565 6566		spin_unlock(ptl);6567		return 0;6568	}6569 6570	if (is_continue) {6571		ret = -EFAULT;6572		folio = filemap_lock_hugetlb_folio(h, mapping, idx);6573		if (IS_ERR(folio))6574			goto out;6575		folio_in_pagecache = true;6576	} else if (!*foliop) {6577		/* If a folio already exists, then it's UFFDIO_COPY for6578		 * a non-missing case. Return -EEXIST.6579		 */6580		if (vm_shared &&6581		    hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {6582			ret = -EEXIST;6583			goto out;6584		}6585 6586		folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);6587		if (IS_ERR(folio)) {6588			ret = -ENOMEM;6589			goto out;6590		}6591 6592		ret = copy_folio_from_user(folio, (const void __user *) src_addr,6593					   false);6594 6595		/* fallback to copy_from_user outside mmap_lock */6596		if (unlikely(ret)) {6597			ret = -ENOENT;6598			/* Free the allocated folio which may have6599			 * consumed a reservation.6600			 */6601			restore_reserve_on_error(h, dst_vma, dst_addr, folio);6602			folio_put(folio);6603 6604			/* Allocate a temporary folio to hold the copied6605			 * contents.6606			 */6607			folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);6608			if (!folio) {6609				ret = -ENOMEM;6610				goto out;6611			}6612			*foliop = folio;6613			/* Set the outparam foliop and return to the caller to6614			 * copy the contents outside the lock. Don't free the6615			 * folio.6616			 */6617			goto out;6618		}6619	} else {6620		if (vm_shared &&6621		    hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {6622			folio_put(*foliop);6623			ret = -EEXIST;6624			*foliop = NULL;6625			goto out;6626		}6627 6628		folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);6629		if (IS_ERR(folio)) {6630			folio_put(*foliop);6631			ret = -ENOMEM;6632			*foliop = NULL;6633			goto out;6634		}6635		ret = copy_user_large_folio(folio, *foliop,6636					    ALIGN_DOWN(dst_addr, size), dst_vma);6637		folio_put(*foliop);6638		*foliop = NULL;6639		if (ret) {6640			folio_put(folio);6641			goto out;6642		}6643	}6644 6645	/*6646	 * If we just allocated a new page, we need a memory barrier to ensure6647	 * that preceding stores to the page become visible before the6648	 * set_pte_at() write. The memory barrier inside __folio_mark_uptodate6649	 * is what we need.6650	 *6651	 * In the case where we have not allocated a new page (is_continue),6652	 * the page must already be uptodate. UFFDIO_CONTINUE already includes6653	 * an earlier smp_wmb() to ensure that prior stores will be visible6654	 * before the set_pte_at() write.6655	 */6656	if (!is_continue)6657		__folio_mark_uptodate(folio);6658	else6659		WARN_ON_ONCE(!folio_test_uptodate(folio));6660 6661	/* Add shared, newly allocated pages to the page cache. */6662	if (vm_shared && !is_continue) {6663		ret = -EFAULT;6664		if (idx >= (i_size_read(mapping->host) >> huge_page_shift(h)))6665			goto out_release_nounlock;6666 6667		/*6668		 * Serialization between remove_inode_hugepages() and6669		 * hugetlb_add_to_page_cache() below happens through the6670		 * hugetlb_fault_mutex_table that here must be hold by6671		 * the caller.6672		 */6673		ret = hugetlb_add_to_page_cache(folio, mapping, idx);6674		if (ret)6675			goto out_release_nounlock;6676		folio_in_pagecache = true;6677	}6678 6679	ptl = huge_pte_lock(h, dst_mm, dst_pte);6680 6681	ret = -EIO;6682	if (folio_test_hwpoison(folio))6683		goto out_release_unlock;6684 6685	/*6686	 * We allow to overwrite a pte marker: consider when both MISSING|WP6687	 * registered, we firstly wr-protect a none pte which has no page cache6688	 * page backing it, then access the page.6689	 */6690	ret = -EEXIST;6691	if (!huge_pte_none_mostly(huge_ptep_get(dst_mm, dst_addr, dst_pte)))6692		goto out_release_unlock;6693 6694	if (folio_in_pagecache)6695		hugetlb_add_file_rmap(folio);6696	else6697		hugetlb_add_new_anon_rmap(folio, dst_vma, dst_addr);6698 6699	/*6700	 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY6701	 * with wp flag set, don't set pte write bit.6702	 */6703	if (wp_enabled || (is_continue && !vm_shared))6704		writable = 0;6705	else6706		writable = dst_vma->vm_flags & VM_WRITE;6707 6708	_dst_pte = make_huge_pte(dst_vma, &folio->page, writable);6709	/*6710	 * Always mark UFFDIO_COPY page dirty; note that this may not be6711	 * extremely important for hugetlbfs for now since swapping is not6712	 * supported, but we should still be clear in that this page cannot be6713	 * thrown away at will, even if write bit not set.6714	 */6715	_dst_pte = huge_pte_mkdirty(_dst_pte);6716	_dst_pte = pte_mkyoung(_dst_pte);6717 6718	if (wp_enabled)6719		_dst_pte = huge_pte_mkuffd_wp(_dst_pte);6720 6721	set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, size);6722 6723	hugetlb_count_add(pages_per_huge_page(h), dst_mm);6724 6725	/* No need to invalidate - it was non-present before */6726	update_mmu_cache(dst_vma, dst_addr, dst_pte);6727 6728	spin_unlock(ptl);6729	if (!is_continue)6730		folio_set_hugetlb_migratable(folio);6731	if (vm_shared || is_continue)6732		folio_unlock(folio);6733	ret = 0;6734out:6735	return ret;6736out_release_unlock:6737	spin_unlock(ptl);6738	if (vm_shared || is_continue)6739		folio_unlock(folio);6740out_release_nounlock:6741	if (!folio_in_pagecache)6742		restore_reserve_on_error(h, dst_vma, dst_addr, folio);6743	folio_put(folio);6744	goto out;6745}6746#endif /* CONFIG_USERFAULTFD */6747 6748long hugetlb_change_protection(struct vm_area_struct *vma,6749		unsigned long address, unsigned long end,6750		pgprot_t newprot, unsigned long cp_flags)6751{6752	struct mm_struct *mm = vma->vm_mm;6753	unsigned long start = address;6754	pte_t *ptep;6755	pte_t pte;6756	struct hstate *h = hstate_vma(vma);6757	long pages = 0, psize = huge_page_size(h);6758	bool shared_pmd = false;6759	struct mmu_notifier_range range;6760	unsigned long last_addr_mask;6761	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;6762	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;6763 6764	/*6765	 * In the case of shared PMDs, the area to flush could be beyond6766	 * start/end.  Set range.start/range.end to cover the maximum possible6767	 * range if PMD sharing is possible.6768	 */6769	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,6770				0, mm, start, end);6771	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);6772 6773	BUG_ON(address >= end);6774	flush_cache_range(vma, range.start, range.end);6775 6776	mmu_notifier_invalidate_range_start(&range);6777	hugetlb_vma_lock_write(vma);6778	i_mmap_lock_write(vma->vm_file->f_mapping);6779	last_addr_mask = hugetlb_mask_last_page(h);6780	for (; address < end; address += psize) {6781		spinlock_t *ptl;6782		ptep = hugetlb_walk(vma, address, psize);6783		if (!ptep) {6784			if (!uffd_wp) {6785				address |= last_addr_mask;6786				continue;6787			}6788			/*6789			 * Userfaultfd wr-protect requires pgtable6790			 * pre-allocations to install pte markers.6791			 */6792			ptep = huge_pte_alloc(mm, vma, address, psize);6793			if (!ptep) {6794				pages = -ENOMEM;6795				break;6796			}6797		}6798		ptl = huge_pte_lock(h, mm, ptep);6799		if (huge_pmd_unshare(mm, vma, address, ptep)) {6800			/*6801			 * When uffd-wp is enabled on the vma, unshare6802			 * shouldn't happen at all.  Warn about it if it6803			 * happened due to some reason.6804			 */6805			WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);6806			pages++;6807			spin_unlock(ptl);6808			shared_pmd = true;6809			address |= last_addr_mask;6810			continue;6811		}6812		pte = huge_ptep_get(mm, address, ptep);6813		if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {6814			/* Nothing to do. */6815		} else if (unlikely(is_hugetlb_entry_migration(pte))) {6816			swp_entry_t entry = pte_to_swp_entry(pte);6817			struct page *page = pfn_swap_entry_to_page(entry);6818			pte_t newpte = pte;6819 6820			if (is_writable_migration_entry(entry)) {6821				if (PageAnon(page))6822					entry = make_readable_exclusive_migration_entry(6823								swp_offset(entry));6824				else6825					entry = make_readable_migration_entry(6826								swp_offset(entry));6827				newpte = swp_entry_to_pte(entry);6828				pages++;6829			}6830 6831			if (uffd_wp)6832				newpte = pte_swp_mkuffd_wp(newpte);6833			else if (uffd_wp_resolve)6834				newpte = pte_swp_clear_uffd_wp(newpte);6835			if (!pte_same(pte, newpte))6836				set_huge_pte_at(mm, address, ptep, newpte, psize);6837		} else if (unlikely(is_pte_marker(pte))) {6838			/*6839			 * Do nothing on a poison marker; page is6840			 * corrupted, permissons do not apply.  Here6841			 * pte_marker_uffd_wp()==true implies !poison6842			 * because they're mutual exclusive.6843			 */6844			if (pte_marker_uffd_wp(pte) && uffd_wp_resolve)6845				/* Safe to modify directly (non-present->none). */6846				huge_pte_clear(mm, address, ptep, psize);6847		} else if (!huge_pte_none(pte)) {6848			pte_t old_pte;6849			unsigned int shift = huge_page_shift(hstate_vma(vma));6850 6851			old_pte = huge_ptep_modify_prot_start(vma, address, ptep);6852			pte = huge_pte_modify(old_pte, newprot);6853			pte = arch_make_huge_pte(pte, shift, vma->vm_flags);6854			if (uffd_wp)6855				pte = huge_pte_mkuffd_wp(pte);6856			else if (uffd_wp_resolve)6857				pte = huge_pte_clear_uffd_wp(pte);6858			huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);6859			pages++;6860		} else {6861			/* None pte */6862			if (unlikely(uffd_wp))6863				/* Safe to modify directly (none->non-present). */6864				set_huge_pte_at(mm, address, ptep,6865						make_pte_marker(PTE_MARKER_UFFD_WP),6866						psize);6867		}6868		spin_unlock(ptl);6869	}6870	/*6871	 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare6872	 * may have cleared our pud entry and done put_page on the page table:6873	 * once we release i_mmap_rwsem, another task can do the final put_page6874	 * and that page table be reused and filled with junk.  If we actually6875	 * did unshare a page of pmds, flush the range corresponding to the pud.6876	 */6877	if (shared_pmd)6878		flush_hugetlb_tlb_range(vma, range.start, range.end);6879	else6880		flush_hugetlb_tlb_range(vma, start, end);6881	/*6882	 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are6883	 * downgrading page table protection not changing it to point to a new6884	 * page.6885	 *6886	 * See Documentation/mm/mmu_notifier.rst6887	 */6888	i_mmap_unlock_write(vma->vm_file->f_mapping);6889	hugetlb_vma_unlock_write(vma);6890	mmu_notifier_invalidate_range_end(&range);6891 6892	return pages > 0 ? (pages << h->order) : pages;6893}6894 6895/* Return true if reservation was successful, false otherwise.  */6896bool hugetlb_reserve_pages(struct inode *inode,6897					long from, long to,6898					struct vm_area_struct *vma,6899					vm_flags_t vm_flags)6900{6901	long chg = -1, add = -1;6902	struct hstate *h = hstate_inode(inode);6903	struct hugepage_subpool *spool = subpool_inode(inode);6904	struct resv_map *resv_map;6905	struct hugetlb_cgroup *h_cg = NULL;6906	long gbl_reserve, regions_needed = 0;6907 6908	/* This should never happen */6909	if (from > to) {6910		VM_WARN(1, "%s called with a negative range\n", __func__);6911		return false;6912	}6913 6914	/*6915	 * vma specific semaphore used for pmd sharing and fault/truncation6916	 * synchronization6917	 */6918	hugetlb_vma_lock_alloc(vma);6919 6920	/*6921	 * Only apply hugepage reservation if asked. At fault time, an6922	 * attempt will be made for VM_NORESERVE to allocate a page6923	 * without using reserves6924	 */6925	if (vm_flags & VM_NORESERVE)6926		return true;6927 6928	/*6929	 * Shared mappings base their reservation on the number of pages that6930	 * are already allocated on behalf of the file. Private mappings need6931	 * to reserve the full area even if read-only as mprotect() may be6932	 * called to make the mapping read-write. Assume !vma is a shm mapping6933	 */6934	if (!vma || vma->vm_flags & VM_MAYSHARE) {6935		/*6936		 * resv_map can not be NULL as hugetlb_reserve_pages is only6937		 * called for inodes for which resv_maps were created (see6938		 * hugetlbfs_get_inode).6939		 */6940		resv_map = inode_resv_map(inode);6941 6942		chg = region_chg(resv_map, from, to, &regions_needed);6943	} else {6944		/* Private mapping. */6945		resv_map = resv_map_alloc();6946		if (!resv_map)6947			goto out_err;6948 6949		chg = to - from;6950 6951		set_vma_resv_map(vma, resv_map);6952		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);6953	}6954 6955	if (chg < 0)6956		goto out_err;6957 6958	if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),6959				chg * pages_per_huge_page(h), &h_cg) < 0)6960		goto out_err;6961 6962	if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {6963		/* For private mappings, the hugetlb_cgroup uncharge info hangs6964		 * of the resv_map.6965		 */6966		resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);6967	}6968 6969	/*6970	 * There must be enough pages in the subpool for the mapping. If6971	 * the subpool has a minimum size, there may be some global6972	 * reservations already in place (gbl_reserve).6973	 */6974	gbl_reserve = hugepage_subpool_get_pages(spool, chg);6975	if (gbl_reserve < 0)6976		goto out_uncharge_cgroup;6977 6978	/*6979	 * Check enough hugepages are available for the reservation.6980	 * Hand the pages back to the subpool if there are not6981	 */6982	if (hugetlb_acct_memory(h, gbl_reserve) < 0)6983		goto out_put_pages;6984 6985	/*6986	 * Account for the reservations made. Shared mappings record regions6987	 * that have reservations as they are shared by multiple VMAs.6988	 * When the last VMA disappears, the region map says how much6989	 * the reservation was and the page cache tells how much of6990	 * the reservation was consumed. Private mappings are per-VMA and6991	 * only the consumed reservations are tracked. When the VMA6992	 * disappears, the original reservation is the VMA size and the6993	 * consumed reservations are stored in the map. Hence, nothing6994	 * else has to be done for private mappings here6995	 */6996	if (!vma || vma->vm_flags & VM_MAYSHARE) {6997		add = region_add(resv_map, from, to, regions_needed, h, h_cg);6998 6999		if (unlikely(add < 0)) {7000			hugetlb_acct_memory(h, -gbl_reserve);7001			goto out_put_pages;7002		} else if (unlikely(chg > add)) {7003			/*7004			 * pages in this range were added to the reserve7005			 * map between region_chg and region_add.  This7006			 * indicates a race with alloc_hugetlb_folio.  Adjust7007			 * the subpool and reserve counts modified above7008			 * based on the difference.7009			 */7010			long rsv_adjust;7011 7012			/*7013			 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the7014			 * reference to h_cg->css. See comment below for detail.7015			 */7016			hugetlb_cgroup_uncharge_cgroup_rsvd(7017				hstate_index(h),7018				(chg - add) * pages_per_huge_page(h), h_cg);7019 7020			rsv_adjust = hugepage_subpool_put_pages(spool,7021								chg - add);7022			hugetlb_acct_memory(h, -rsv_adjust);7023		} else if (h_cg) {7024			/*7025			 * The file_regions will hold their own reference to7026			 * h_cg->css. So we should release the reference held7027			 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are7028			 * done.7029			 */7030			hugetlb_cgroup_put_rsvd_cgroup(h_cg);7031		}7032	}7033	return true;7034 7035out_put_pages:7036	/* put back original number of pages, chg */7037	(void)hugepage_subpool_put_pages(spool, chg);7038out_uncharge_cgroup:7039	hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),7040					    chg * pages_per_huge_page(h), h_cg);7041out_err:7042	hugetlb_vma_lock_free(vma);7043	if (!vma || vma->vm_flags & VM_MAYSHARE)7044		/* Only call region_abort if the region_chg succeeded but the7045		 * region_add failed or didn't run.7046		 */7047		if (chg >= 0 && add < 0)7048			region_abort(resv_map, from, to, regions_needed);7049	if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {7050		kref_put(&resv_map->refs, resv_map_release);7051		set_vma_resv_map(vma, NULL);7052	}7053	return false;7054}7055 7056long hugetlb_unreserve_pages(struct inode *inode, long start, long end,7057								long freed)7058{7059	struct hstate *h = hstate_inode(inode);7060	struct resv_map *resv_map = inode_resv_map(inode);7061	long chg = 0;7062	struct hugepage_subpool *spool = subpool_inode(inode);7063	long gbl_reserve;7064 7065	/*7066	 * Since this routine can be called in the evict inode path for all7067	 * hugetlbfs inodes, resv_map could be NULL.7068	 */7069	if (resv_map) {7070		chg = region_del(resv_map, start, end);7071		/*7072		 * region_del() can fail in the rare case where a region7073		 * must be split and another region descriptor can not be7074		 * allocated.  If end == LONG_MAX, it will not fail.7075		 */7076		if (chg < 0)7077			return chg;7078	}7079 7080	spin_lock(&inode->i_lock);7081	inode->i_blocks -= (blocks_per_huge_page(h) * freed);7082	spin_unlock(&inode->i_lock);7083 7084	/*7085	 * If the subpool has a minimum size, the number of global7086	 * reservations to be released may be adjusted.7087	 *7088	 * Note that !resv_map implies freed == 0. So (chg - freed)7089	 * won't go negative.7090	 */7091	gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));7092	hugetlb_acct_memory(h, -gbl_reserve);7093 7094	return 0;7095}7096 7097#ifdef CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING7098static unsigned long page_table_shareable(struct vm_area_struct *svma,7099				struct vm_area_struct *vma,7100				unsigned long addr, pgoff_t idx)7101{7102	unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +7103				svma->vm_start;7104	unsigned long sbase = saddr & PUD_MASK;7105	unsigned long s_end = sbase + PUD_SIZE;7106 7107	/* Allow segments to share if only one is marked locked */7108	unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;7109	unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;7110 7111	/*7112	 * match the virtual addresses, permission and the alignment of the7113	 * page table page.7114	 *7115	 * Also, vma_lock (vm_private_data) is required for sharing.7116	 */7117	if (pmd_index(addr) != pmd_index(saddr) ||7118	    vm_flags != svm_flags ||7119	    !range_in_vma(svma, sbase, s_end) ||7120	    !svma->vm_private_data)7121		return 0;7122 7123	return saddr;7124}7125 7126bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)7127{7128	unsigned long start = addr & PUD_MASK;7129	unsigned long end = start + PUD_SIZE;7130 7131#ifdef CONFIG_USERFAULTFD7132	if (uffd_disable_huge_pmd_share(vma))7133		return false;7134#endif7135	/*7136	 * check on proper vm_flags and page table alignment7137	 */7138	if (!(vma->vm_flags & VM_MAYSHARE))7139		return false;7140	if (!vma->vm_private_data)	/* vma lock required for sharing */7141		return false;7142	if (!range_in_vma(vma, start, end))7143		return false;7144	return true;7145}7146 7147/*7148 * Determine if start,end range within vma could be mapped by shared pmd.7149 * If yes, adjust start and end to cover range associated with possible7150 * shared pmd mappings.7151 */7152void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,7153				unsigned long *start, unsigned long *end)7154{7155	unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),7156		v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);7157 7158	/*7159	 * vma needs to span at least one aligned PUD size, and the range7160	 * must be at least partially within in.7161	 */7162	if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||7163		(*end <= v_start) || (*start >= v_end))7164		return;7165 7166	/* Extend the range to be PUD aligned for a worst case scenario */7167	if (*start > v_start)7168		*start = ALIGN_DOWN(*start, PUD_SIZE);7169 7170	if (*end < v_end)7171		*end = ALIGN(*end, PUD_SIZE);7172}7173 7174/*7175 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()7176 * and returns the corresponding pte. While this is not necessary for the7177 * !shared pmd case because we can allocate the pmd later as well, it makes the7178 * code much cleaner. pmd allocation is essential for the shared case because7179 * pud has to be populated inside the same i_mmap_rwsem section - otherwise7180 * racing tasks could either miss the sharing (see huge_pte_offset) or select a7181 * bad pmd for sharing.7182 */7183pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,7184		      unsigned long addr, pud_t *pud)7185{7186	struct address_space *mapping = vma->vm_file->f_mapping;7187	pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +7188			vma->vm_pgoff;7189	struct vm_area_struct *svma;7190	unsigned long saddr;7191	pte_t *spte = NULL;7192	pte_t *pte;7193 7194	i_mmap_lock_read(mapping);7195	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {7196		if (svma == vma)7197			continue;7198 7199		saddr = page_table_shareable(svma, vma, addr, idx);7200		if (saddr) {7201			spte = hugetlb_walk(svma, saddr,7202					    vma_mmu_pagesize(svma));7203			if (spte) {7204				get_page(virt_to_page(spte));7205				break;7206			}7207		}7208	}7209 7210	if (!spte)7211		goto out;7212 7213	spin_lock(&mm->page_table_lock);7214	if (pud_none(*pud)) {7215		pud_populate(mm, pud,7216				(pmd_t *)((unsigned long)spte & PAGE_MASK));7217		mm_inc_nr_pmds(mm);7218	} else {7219		put_page(virt_to_page(spte));7220	}7221	spin_unlock(&mm->page_table_lock);7222out:7223	pte = (pte_t *)pmd_alloc(mm, pud, addr);7224	i_mmap_unlock_read(mapping);7225	return pte;7226}7227 7228/*7229 * unmap huge page backed by shared pte.7230 *7231 * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared7232 * indicated by page_count > 1, unmap is achieved by clearing pud and7233 * decrementing the ref count. If count == 1, the pte page is not shared.7234 *7235 * Called with page table lock held.7236 *7237 * returns: 1 successfully unmapped a shared pte page7238 *	    0 the underlying pte page is not shared, or it is the last user7239 */7240int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,7241					unsigned long addr, pte_t *ptep)7242{7243	pgd_t *pgd = pgd_offset(mm, addr);7244	p4d_t *p4d = p4d_offset(pgd, addr);7245	pud_t *pud = pud_offset(p4d, addr);7246 7247	i_mmap_assert_write_locked(vma->vm_file->f_mapping);7248	hugetlb_vma_assert_locked(vma);7249	BUG_ON(page_count(virt_to_page(ptep)) == 0);7250	if (page_count(virt_to_page(ptep)) == 1)7251		return 0;7252 7253	pud_clear(pud);7254	put_page(virt_to_page(ptep));7255	mm_dec_nr_pmds(mm);7256	return 1;7257}7258 7259#else /* !CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING */7260 7261pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,7262		      unsigned long addr, pud_t *pud)7263{7264	return NULL;7265}7266 7267int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,7268				unsigned long addr, pte_t *ptep)7269{7270	return 0;7271}7272 7273void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,7274				unsigned long *start, unsigned long *end)7275{7276}7277 7278bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)7279{7280	return false;7281}7282#endif /* CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING */7283 7284#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB7285pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,7286			unsigned long addr, unsigned long sz)7287{7288	pgd_t *pgd;7289	p4d_t *p4d;7290	pud_t *pud;7291	pte_t *pte = NULL;7292 7293	pgd = pgd_offset(mm, addr);7294	p4d = p4d_alloc(mm, pgd, addr);7295	if (!p4d)7296		return NULL;7297	pud = pud_alloc(mm, p4d, addr);7298	if (pud) {7299		if (sz == PUD_SIZE) {7300			pte = (pte_t *)pud;7301		} else {7302			BUG_ON(sz != PMD_SIZE);7303			if (want_pmd_share(vma, addr) && pud_none(*pud))7304				pte = huge_pmd_share(mm, vma, addr, pud);7305			else7306				pte = (pte_t *)pmd_alloc(mm, pud, addr);7307		}7308	}7309 7310	if (pte) {7311		pte_t pteval = ptep_get_lockless(pte);7312 7313		BUG_ON(pte_present(pteval) && !pte_huge(pteval));7314	}7315 7316	return pte;7317}7318 7319/*7320 * huge_pte_offset() - Walk the page table to resolve the hugepage7321 * entry at address @addr7322 *7323 * Return: Pointer to page table entry (PUD or PMD) for7324 * address @addr, or NULL if a !p*d_present() entry is encountered and the7325 * size @sz doesn't match the hugepage size at this level of the page7326 * table.7327 */7328pte_t *huge_pte_offset(struct mm_struct *mm,7329		       unsigned long addr, unsigned long sz)7330{7331	pgd_t *pgd;7332	p4d_t *p4d;7333	pud_t *pud;7334	pmd_t *pmd;7335 7336	pgd = pgd_offset(mm, addr);7337	if (!pgd_present(*pgd))7338		return NULL;7339	p4d = p4d_offset(pgd, addr);7340	if (!p4d_present(*p4d))7341		return NULL;7342 7343	pud = pud_offset(p4d, addr);7344	if (sz == PUD_SIZE)7345		/* must be pud huge, non-present or none */7346		return (pte_t *)pud;7347	if (!pud_present(*pud))7348		return NULL;7349	/* must have a valid entry and size to go further */7350 7351	pmd = pmd_offset(pud, addr);7352	/* must be pmd huge, non-present or none */7353	return (pte_t *)pmd;7354}7355 7356/*7357 * Return a mask that can be used to update an address to the last huge7358 * page in a page table page mapping size.  Used to skip non-present7359 * page table entries when linearly scanning address ranges.  Architectures7360 * with unique huge page to page table relationships can define their own7361 * version of this routine.7362 */7363unsigned long hugetlb_mask_last_page(struct hstate *h)7364{7365	unsigned long hp_size = huge_page_size(h);7366 7367	if (hp_size == PUD_SIZE)7368		return P4D_SIZE - PUD_SIZE;7369	else if (hp_size == PMD_SIZE)7370		return PUD_SIZE - PMD_SIZE;7371	else7372		return 0UL;7373}7374 7375#else7376 7377/* See description above.  Architectures can provide their own version. */7378__weak unsigned long hugetlb_mask_last_page(struct hstate *h)7379{7380#ifdef CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING7381	if (huge_page_size(h) == PMD_SIZE)7382		return PUD_SIZE - PMD_SIZE;7383#endif7384	return 0UL;7385}7386 7387#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */7388 7389bool isolate_hugetlb(struct folio *folio, struct list_head *list)7390{7391	bool ret = true;7392 7393	spin_lock_irq(&hugetlb_lock);7394	if (!folio_test_hugetlb(folio) ||7395	    !folio_test_hugetlb_migratable(folio) ||7396	    !folio_try_get(folio)) {7397		ret = false;7398		goto unlock;7399	}7400	folio_clear_hugetlb_migratable(folio);7401	list_move_tail(&folio->lru, list);7402unlock:7403	spin_unlock_irq(&hugetlb_lock);7404	return ret;7405}7406 7407int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)7408{7409	int ret = 0;7410 7411	*hugetlb = false;7412	spin_lock_irq(&hugetlb_lock);7413	if (folio_test_hugetlb(folio)) {7414		*hugetlb = true;7415		if (folio_test_hugetlb_freed(folio))7416			ret = 0;7417		else if (folio_test_hugetlb_migratable(folio) || unpoison)7418			ret = folio_try_get(folio);7419		else7420			ret = -EBUSY;7421	}7422	spin_unlock_irq(&hugetlb_lock);7423	return ret;7424}7425 7426int get_huge_page_for_hwpoison(unsigned long pfn, int flags,7427				bool *migratable_cleared)7428{7429	int ret;7430 7431	spin_lock_irq(&hugetlb_lock);7432	ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);7433	spin_unlock_irq(&hugetlb_lock);7434	return ret;7435}7436 7437void folio_putback_active_hugetlb(struct folio *folio)7438{7439	spin_lock_irq(&hugetlb_lock);7440	folio_set_hugetlb_migratable(folio);7441	list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);7442	spin_unlock_irq(&hugetlb_lock);7443	folio_put(folio);7444}7445 7446void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)7447{7448	struct hstate *h = folio_hstate(old_folio);7449 7450	hugetlb_cgroup_migrate(old_folio, new_folio);7451	set_page_owner_migrate_reason(&new_folio->page, reason);7452 7453	/*7454	 * transfer temporary state of the new hugetlb folio. This is7455	 * reverse to other transitions because the newpage is going to7456	 * be final while the old one will be freed so it takes over7457	 * the temporary status.7458	 *7459	 * Also note that we have to transfer the per-node surplus state7460	 * here as well otherwise the global surplus count will not match7461	 * the per-node's.7462	 */7463	if (folio_test_hugetlb_temporary(new_folio)) {7464		int old_nid = folio_nid(old_folio);7465		int new_nid = folio_nid(new_folio);7466 7467		folio_set_hugetlb_temporary(old_folio);7468		folio_clear_hugetlb_temporary(new_folio);7469 7470 7471		/*7472		 * There is no need to transfer the per-node surplus state7473		 * when we do not cross the node.7474		 */7475		if (new_nid == old_nid)7476			return;7477		spin_lock_irq(&hugetlb_lock);7478		if (h->surplus_huge_pages_node[old_nid]) {7479			h->surplus_huge_pages_node[old_nid]--;7480			h->surplus_huge_pages_node[new_nid]++;7481		}7482		spin_unlock_irq(&hugetlb_lock);7483	}7484}7485 7486static void hugetlb_unshare_pmds(struct vm_area_struct *vma,7487				   unsigned long start,7488				   unsigned long end)7489{7490	struct hstate *h = hstate_vma(vma);7491	unsigned long sz = huge_page_size(h);7492	struct mm_struct *mm = vma->vm_mm;7493	struct mmu_notifier_range range;7494	unsigned long address;7495	spinlock_t *ptl;7496	pte_t *ptep;7497 7498	if (!(vma->vm_flags & VM_MAYSHARE))7499		return;7500 7501	if (start >= end)7502		return;7503 7504	flush_cache_range(vma, start, end);7505	/*7506	 * No need to call adjust_range_if_pmd_sharing_possible(), because7507	 * we have already done the PUD_SIZE alignment.7508	 */7509	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,7510				start, end);7511	mmu_notifier_invalidate_range_start(&range);7512	hugetlb_vma_lock_write(vma);7513	i_mmap_lock_write(vma->vm_file->f_mapping);7514	for (address = start; address < end; address += PUD_SIZE) {7515		ptep = hugetlb_walk(vma, address, sz);7516		if (!ptep)7517			continue;7518		ptl = huge_pte_lock(h, mm, ptep);7519		huge_pmd_unshare(mm, vma, address, ptep);7520		spin_unlock(ptl);7521	}7522	flush_hugetlb_tlb_range(vma, start, end);7523	i_mmap_unlock_write(vma->vm_file->f_mapping);7524	hugetlb_vma_unlock_write(vma);7525	/*7526	 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see7527	 * Documentation/mm/mmu_notifier.rst.7528	 */7529	mmu_notifier_invalidate_range_end(&range);7530}7531 7532/*7533 * This function will unconditionally remove all the shared pmd pgtable entries7534 * within the specific vma for a hugetlbfs memory range.7535 */7536void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)7537{7538	hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),7539			ALIGN_DOWN(vma->vm_end, PUD_SIZE));7540}7541 7542#ifdef CONFIG_CMA7543static bool cma_reserve_called __initdata;7544 7545static int __init cmdline_parse_hugetlb_cma(char *p)7546{7547	int nid, count = 0;7548	unsigned long tmp;7549	char *s = p;7550 7551	while (*s) {7552		if (sscanf(s, "%lu%n", &tmp, &count) != 1)7553			break;7554 7555		if (s[count] == ':') {7556			if (tmp >= MAX_NUMNODES)7557				break;7558			nid = array_index_nospec(tmp, MAX_NUMNODES);7559 7560			s += count + 1;7561			tmp = memparse(s, &s);7562			hugetlb_cma_size_in_node[nid] = tmp;7563			hugetlb_cma_size += tmp;7564 7565			/*7566			 * Skip the separator if have one, otherwise7567			 * break the parsing.7568			 */7569			if (*s == ',')7570				s++;7571			else7572				break;7573		} else {7574			hugetlb_cma_size = memparse(p, &p);7575			break;7576		}7577	}7578 7579	return 0;7580}7581 7582early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);7583 7584void __init hugetlb_cma_reserve(int order)7585{7586	unsigned long size, reserved, per_node;7587	bool node_specific_cma_alloc = false;7588	int nid;7589 7590	/*7591	 * HugeTLB CMA reservation is required for gigantic7592	 * huge pages which could not be allocated via the7593	 * page allocator. Just warn if there is any change7594	 * breaking this assumption.7595	 */7596	VM_WARN_ON(order <= MAX_PAGE_ORDER);7597	cma_reserve_called = true;7598 7599	if (!hugetlb_cma_size)7600		return;7601 7602	for (nid = 0; nid < MAX_NUMNODES; nid++) {7603		if (hugetlb_cma_size_in_node[nid] == 0)7604			continue;7605 7606		if (!node_online(nid)) {7607			pr_warn("hugetlb_cma: invalid node %d specified\n", nid);7608			hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];7609			hugetlb_cma_size_in_node[nid] = 0;7610			continue;7611		}7612 7613		if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {7614			pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",7615				nid, (PAGE_SIZE << order) / SZ_1M);7616			hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];7617			hugetlb_cma_size_in_node[nid] = 0;7618		} else {7619			node_specific_cma_alloc = true;7620		}7621	}7622 7623	/* Validate the CMA size again in case some invalid nodes specified. */7624	if (!hugetlb_cma_size)7625		return;7626 7627	if (hugetlb_cma_size < (PAGE_SIZE << order)) {7628		pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",7629			(PAGE_SIZE << order) / SZ_1M);7630		hugetlb_cma_size = 0;7631		return;7632	}7633 7634	if (!node_specific_cma_alloc) {7635		/*7636		 * If 3 GB area is requested on a machine with 4 numa nodes,7637		 * let's allocate 1 GB on first three nodes and ignore the last one.7638		 */7639		per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);7640		pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",7641			hugetlb_cma_size / SZ_1M, per_node / SZ_1M);7642	}7643 7644	reserved = 0;7645	for_each_online_node(nid) {7646		int res;7647		char name[CMA_MAX_NAME];7648 7649		if (node_specific_cma_alloc) {7650			if (hugetlb_cma_size_in_node[nid] == 0)7651				continue;7652 7653			size = hugetlb_cma_size_in_node[nid];7654		} else {7655			size = min(per_node, hugetlb_cma_size - reserved);7656		}7657 7658		size = round_up(size, PAGE_SIZE << order);7659 7660		snprintf(name, sizeof(name), "hugetlb%d", nid);7661		/*7662		 * Note that 'order per bit' is based on smallest size that7663		 * may be returned to CMA allocator in the case of7664		 * huge page demotion.7665		 */7666		res = cma_declare_contiguous_nid(0, size, 0,7667					PAGE_SIZE << order,7668					HUGETLB_PAGE_ORDER, false, name,7669					&hugetlb_cma[nid], nid);7670		if (res) {7671			pr_warn("hugetlb_cma: reservation failed: err %d, node %d",7672				res, nid);7673			continue;7674		}7675 7676		reserved += size;7677		pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",7678			size / SZ_1M, nid);7679 7680		if (reserved >= hugetlb_cma_size)7681			break;7682	}7683 7684	if (!reserved)7685		/*7686		 * hugetlb_cma_size is used to determine if allocations from7687		 * cma are possible.  Set to zero if no cma regions are set up.7688		 */7689		hugetlb_cma_size = 0;7690}7691 7692static void __init hugetlb_cma_check(void)7693{7694	if (!hugetlb_cma_size || cma_reserve_called)7695		return;7696 7697	pr_warn("hugetlb_cma: the option isn't supported by current arch\n");7698}7699 7700#endif /* CONFIG_CMA */7701