brintos

brintos / linux-shallow public Read only

0
0
Text · 49.0 KiB · a569ff9 Raw
1843 lines · c
1// SPDX-License-Identifier: GPL-2.0-only2/*3 * Copyright (C) 2018 HUAWEI, Inc.4 *             https://www.huawei.com/5 * Copyright (C) 2022 Alibaba Cloud6 */7#include "compress.h"8#include <linux/psi.h>9#include <linux/cpuhotplug.h>10#include <trace/events/erofs.h>11 12#define Z_EROFS_PCLUSTER_MAX_PAGES	(Z_EROFS_PCLUSTER_MAX_SIZE / PAGE_SIZE)13#define Z_EROFS_INLINE_BVECS		214 15/*16 * let's leave a type here in case of introducing17 * another tagged pointer later.18 */19typedef void *z_erofs_next_pcluster_t;20 21struct z_erofs_bvec {22	struct page *page;23	int offset;24	unsigned int end;25};26 27#define __Z_EROFS_BVSET(name, total) \28struct name { \29	/* point to the next page which contains the following bvecs */ \30	struct page *nextpage; \31	struct z_erofs_bvec bvec[total]; \32}33__Z_EROFS_BVSET(z_erofs_bvset,);34__Z_EROFS_BVSET(z_erofs_bvset_inline, Z_EROFS_INLINE_BVECS);35 36/*37 * Structure fields follow one of the following exclusion rules.38 *39 * I: Modifiable by initialization/destruction paths and read-only40 *    for everyone else;41 *42 * L: Field should be protected by the pcluster lock;43 *44 * A: Field should be accessed / updated in atomic for parallelized code.45 */46struct z_erofs_pcluster {47	struct erofs_workgroup obj;48	struct mutex lock;49 50	/* A: point to next chained pcluster or TAILs */51	z_erofs_next_pcluster_t next;52 53	/* L: the maximum decompression size of this round */54	unsigned int length;55 56	/* L: total number of bvecs */57	unsigned int vcnt;58 59	/* I: pcluster size (compressed size) in bytes */60	unsigned int pclustersize;61 62	/* I: page offset of start position of decompression */63	unsigned short pageofs_out;64 65	/* I: page offset of inline compressed data */66	unsigned short pageofs_in;67 68	union {69		/* L: inline a certain number of bvec for bootstrap */70		struct z_erofs_bvset_inline bvset;71 72		/* I: can be used to free the pcluster by RCU. */73		struct rcu_head rcu;74	};75 76	/* I: compression algorithm format */77	unsigned char algorithmformat;78 79	/* L: whether partial decompression or not */80	bool partial;81 82	/* L: indicate several pageofs_outs or not */83	bool multibases;84 85	/* L: whether extra buffer allocations are best-effort */86	bool besteffort;87 88	/* A: compressed bvecs (can be cached or inplaced pages) */89	struct z_erofs_bvec compressed_bvecs[];90};91 92/* the end of a chain of pclusters */93#define Z_EROFS_PCLUSTER_TAIL           ((void *) 0x700 + POISON_POINTER_DELTA)94#define Z_EROFS_PCLUSTER_NIL            (NULL)95 96struct z_erofs_decompressqueue {97	struct super_block *sb;98	atomic_t pending_bios;99	z_erofs_next_pcluster_t head;100 101	union {102		struct completion done;103		struct work_struct work;104		struct kthread_work kthread_work;105	} u;106	bool eio, sync;107};108 109static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl)110{111	return !pcl->obj.index;112}113 114static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl)115{116	return PAGE_ALIGN(pcl->pclustersize) >> PAGE_SHIFT;117}118 119#define MNGD_MAPPING(sbi)	((sbi)->managed_cache->i_mapping)120static bool erofs_folio_is_managed(struct erofs_sb_info *sbi, struct folio *fo)121{122	return fo->mapping == MNGD_MAPPING(sbi);123}124 125#define Z_EROFS_ONSTACK_PAGES		32126 127/*128 * since pclustersize is variable for big pcluster feature, introduce slab129 * pools implementation for different pcluster sizes.130 */131struct z_erofs_pcluster_slab {132	struct kmem_cache *slab;133	unsigned int maxpages;134	char name[48];135};136 137#define _PCLP(n) { .maxpages = n }138 139static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {140	_PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),141	_PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)142};143 144struct z_erofs_bvec_iter {145	struct page *bvpage;146	struct z_erofs_bvset *bvset;147	unsigned int nr, cur;148};149 150static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter)151{152	if (iter->bvpage)153		kunmap_local(iter->bvset);154	return iter->bvpage;155}156 157static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter)158{159	unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec;160	/* have to access nextpage in advance, otherwise it will be unmapped */161	struct page *nextpage = iter->bvset->nextpage;162	struct page *oldpage;163 164	DBG_BUGON(!nextpage);165	oldpage = z_erofs_bvec_iter_end(iter);166	iter->bvpage = nextpage;167	iter->bvset = kmap_local_page(nextpage);168	iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec);169	iter->cur = 0;170	return oldpage;171}172 173static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter,174				    struct z_erofs_bvset_inline *bvset,175				    unsigned int bootstrap_nr,176				    unsigned int cur)177{178	*iter = (struct z_erofs_bvec_iter) {179		.nr = bootstrap_nr,180		.bvset = (struct z_erofs_bvset *)bvset,181	};182 183	while (cur > iter->nr) {184		cur -= iter->nr;185		z_erofs_bvset_flip(iter);186	}187	iter->cur = cur;188}189 190static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter,191				struct z_erofs_bvec *bvec,192				struct page **candidate_bvpage,193				struct page **pagepool)194{195	if (iter->cur >= iter->nr) {196		struct page *nextpage = *candidate_bvpage;197 198		if (!nextpage) {199			nextpage = __erofs_allocpage(pagepool, GFP_KERNEL,200					true);201			if (!nextpage)202				return -ENOMEM;203			set_page_private(nextpage, Z_EROFS_SHORTLIVED_PAGE);204		}205		DBG_BUGON(iter->bvset->nextpage);206		iter->bvset->nextpage = nextpage;207		z_erofs_bvset_flip(iter);208 209		iter->bvset->nextpage = NULL;210		*candidate_bvpage = NULL;211	}212	iter->bvset->bvec[iter->cur++] = *bvec;213	return 0;214}215 216static void z_erofs_bvec_dequeue(struct z_erofs_bvec_iter *iter,217				 struct z_erofs_bvec *bvec,218				 struct page **old_bvpage)219{220	if (iter->cur == iter->nr)221		*old_bvpage = z_erofs_bvset_flip(iter);222	else223		*old_bvpage = NULL;224	*bvec = iter->bvset->bvec[iter->cur++];225}226 227static void z_erofs_destroy_pcluster_pool(void)228{229	int i;230 231	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {232		if (!pcluster_pool[i].slab)233			continue;234		kmem_cache_destroy(pcluster_pool[i].slab);235		pcluster_pool[i].slab = NULL;236	}237}238 239static int z_erofs_create_pcluster_pool(void)240{241	struct z_erofs_pcluster_slab *pcs;242	struct z_erofs_pcluster *a;243	unsigned int size;244 245	for (pcs = pcluster_pool;246	     pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {247		size = struct_size(a, compressed_bvecs, pcs->maxpages);248 249		sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);250		pcs->slab = kmem_cache_create(pcs->name, size, 0,251					      SLAB_RECLAIM_ACCOUNT, NULL);252		if (pcs->slab)253			continue;254 255		z_erofs_destroy_pcluster_pool();256		return -ENOMEM;257	}258	return 0;259}260 261static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int size)262{263	unsigned int nrpages = PAGE_ALIGN(size) >> PAGE_SHIFT;264	struct z_erofs_pcluster_slab *pcs = pcluster_pool;265 266	for (; pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {267		struct z_erofs_pcluster *pcl;268 269		if (nrpages > pcs->maxpages)270			continue;271 272		pcl = kmem_cache_zalloc(pcs->slab, GFP_KERNEL);273		if (!pcl)274			return ERR_PTR(-ENOMEM);275		pcl->pclustersize = size;276		return pcl;277	}278	return ERR_PTR(-EINVAL);279}280 281static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)282{283	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);284	int i;285 286	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {287		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;288 289		if (pclusterpages > pcs->maxpages)290			continue;291 292		kmem_cache_free(pcs->slab, pcl);293		return;294	}295	DBG_BUGON(1);296}297 298static struct workqueue_struct *z_erofs_workqueue __read_mostly;299 300#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD301static struct kthread_worker __rcu **z_erofs_pcpu_workers;302 303static void erofs_destroy_percpu_workers(void)304{305	struct kthread_worker *worker;306	unsigned int cpu;307 308	for_each_possible_cpu(cpu) {309		worker = rcu_dereference_protected(310					z_erofs_pcpu_workers[cpu], 1);311		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);312		if (worker)313			kthread_destroy_worker(worker);314	}315	kfree(z_erofs_pcpu_workers);316}317 318static struct kthread_worker *erofs_init_percpu_worker(int cpu)319{320	struct kthread_worker *worker =321		kthread_create_worker_on_cpu(cpu, 0, "erofs_worker/%u", cpu);322 323	if (IS_ERR(worker))324		return worker;325	if (IS_ENABLED(CONFIG_EROFS_FS_PCPU_KTHREAD_HIPRI))326		sched_set_fifo_low(worker->task);327	return worker;328}329 330static int erofs_init_percpu_workers(void)331{332	struct kthread_worker *worker;333	unsigned int cpu;334 335	z_erofs_pcpu_workers = kcalloc(num_possible_cpus(),336			sizeof(struct kthread_worker *), GFP_ATOMIC);337	if (!z_erofs_pcpu_workers)338		return -ENOMEM;339 340	for_each_online_cpu(cpu) {	/* could miss cpu{off,on}line? */341		worker = erofs_init_percpu_worker(cpu);342		if (!IS_ERR(worker))343			rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);344	}345	return 0;346}347#else348static inline void erofs_destroy_percpu_workers(void) {}349static inline int erofs_init_percpu_workers(void) { return 0; }350#endif351 352#if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_EROFS_FS_PCPU_KTHREAD)353static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock);354static enum cpuhp_state erofs_cpuhp_state;355 356static int erofs_cpu_online(unsigned int cpu)357{358	struct kthread_worker *worker, *old;359 360	worker = erofs_init_percpu_worker(cpu);361	if (IS_ERR(worker))362		return PTR_ERR(worker);363 364	spin_lock(&z_erofs_pcpu_worker_lock);365	old = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],366			lockdep_is_held(&z_erofs_pcpu_worker_lock));367	if (!old)368		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);369	spin_unlock(&z_erofs_pcpu_worker_lock);370	if (old)371		kthread_destroy_worker(worker);372	return 0;373}374 375static int erofs_cpu_offline(unsigned int cpu)376{377	struct kthread_worker *worker;378 379	spin_lock(&z_erofs_pcpu_worker_lock);380	worker = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],381			lockdep_is_held(&z_erofs_pcpu_worker_lock));382	rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);383	spin_unlock(&z_erofs_pcpu_worker_lock);384 385	synchronize_rcu();386	if (worker)387		kthread_destroy_worker(worker);388	return 0;389}390 391static int erofs_cpu_hotplug_init(void)392{393	int state;394 395	state = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,396			"fs/erofs:online", erofs_cpu_online, erofs_cpu_offline);397	if (state < 0)398		return state;399 400	erofs_cpuhp_state = state;401	return 0;402}403 404static void erofs_cpu_hotplug_destroy(void)405{406	if (erofs_cpuhp_state)407		cpuhp_remove_state_nocalls(erofs_cpuhp_state);408}409#else /* !CONFIG_HOTPLUG_CPU || !CONFIG_EROFS_FS_PCPU_KTHREAD */410static inline int erofs_cpu_hotplug_init(void) { return 0; }411static inline void erofs_cpu_hotplug_destroy(void) {}412#endif413 414void z_erofs_exit_subsystem(void)415{416	erofs_cpu_hotplug_destroy();417	erofs_destroy_percpu_workers();418	destroy_workqueue(z_erofs_workqueue);419	z_erofs_destroy_pcluster_pool();420	z_erofs_exit_decompressor();421}422 423int __init z_erofs_init_subsystem(void)424{425	int err = z_erofs_init_decompressor();426 427	if (err)428		goto err_decompressor;429 430	err = z_erofs_create_pcluster_pool();431	if (err)432		goto err_pcluster_pool;433 434	z_erofs_workqueue = alloc_workqueue("erofs_worker",435			WQ_UNBOUND | WQ_HIGHPRI, num_possible_cpus());436	if (!z_erofs_workqueue) {437		err = -ENOMEM;438		goto err_workqueue_init;439	}440 441	err = erofs_init_percpu_workers();442	if (err)443		goto err_pcpu_worker;444 445	err = erofs_cpu_hotplug_init();446	if (err < 0)447		goto err_cpuhp_init;448	return err;449 450err_cpuhp_init:451	erofs_destroy_percpu_workers();452err_pcpu_worker:453	destroy_workqueue(z_erofs_workqueue);454err_workqueue_init:455	z_erofs_destroy_pcluster_pool();456err_pcluster_pool:457	z_erofs_exit_decompressor();458err_decompressor:459	return err;460}461 462enum z_erofs_pclustermode {463	Z_EROFS_PCLUSTER_INFLIGHT,464	/*465	 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it466	 * could be dispatched into bypass queue later due to uptodated managed467	 * pages. All related online pages cannot be reused for inplace I/O (or468	 * bvpage) since it can be directly decoded without I/O submission.469	 */470	Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE,471	/*472	 * The pcluster was just linked to a decompression chain by us.  It can473	 * also be linked with the remaining pclusters, which means if the474	 * processing page is the tail page of a pcluster, this pcluster can475	 * safely use the whole page (since the previous pcluster is within the476	 * same chain) for in-place I/O, as illustrated below:477	 *  ___________________________________________________478	 * |  tail (partial) page  |    head (partial) page    |479	 * |  (of the current pcl) |   (of the previous pcl)   |480	 * |___PCLUSTER_FOLLOWED___|_____PCLUSTER_FOLLOWED_____|481	 *482	 * [  (*) the page above can be used as inplace I/O.   ]483	 */484	Z_EROFS_PCLUSTER_FOLLOWED,485};486 487struct z_erofs_decompress_frontend {488	struct inode *const inode;489	struct erofs_map_blocks map;490	struct z_erofs_bvec_iter biter;491 492	struct page *pagepool;493	struct page *candidate_bvpage;494	struct z_erofs_pcluster *pcl;495	z_erofs_next_pcluster_t owned_head;496	enum z_erofs_pclustermode mode;497 498	erofs_off_t headoffset;499 500	/* a pointer used to pick up inplace I/O pages */501	unsigned int icur;502};503 504#define DECOMPRESS_FRONTEND_INIT(__i) { \505	.inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \506	.mode = Z_EROFS_PCLUSTER_FOLLOWED }507 508static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe)509{510	unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy;511 512	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)513		return false;514 515	if (!(fe->map.m_flags & EROFS_MAP_FULL_MAPPED))516		return true;517 518	if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&519	    fe->map.m_la < fe->headoffset)520		return true;521 522	return false;523}524 525static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe)526{527	struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));528	struct z_erofs_pcluster *pcl = fe->pcl;529	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);530	bool shouldalloc = z_erofs_should_alloc_cache(fe);531	bool standalone = true;532	/*533	 * optimistic allocation without direct reclaim since inplace I/O534	 * can be used if low memory otherwise.535	 */536	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |537			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;538	unsigned int i;539 540	if (i_blocksize(fe->inode) != PAGE_SIZE ||541	    fe->mode < Z_EROFS_PCLUSTER_FOLLOWED)542		return;543 544	for (i = 0; i < pclusterpages; ++i) {545		struct page *page, *newpage;546 547		/* Inaccurate check w/o locking to avoid unneeded lookups */548		if (READ_ONCE(pcl->compressed_bvecs[i].page))549			continue;550 551		page = find_get_page(mc, pcl->obj.index + i);552		if (!page) {553			/* I/O is needed, no possible to decompress directly */554			standalone = false;555			if (!shouldalloc)556				continue;557 558			/*559			 * Try cached I/O if allocation succeeds or fallback to560			 * in-place I/O instead to avoid any direct reclaim.561			 */562			newpage = erofs_allocpage(&fe->pagepool, gfp);563			if (!newpage)564				continue;565			set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE);566		}567		spin_lock(&pcl->obj.lockref.lock);568		if (!pcl->compressed_bvecs[i].page) {569			pcl->compressed_bvecs[i].page = page ? page : newpage;570			spin_unlock(&pcl->obj.lockref.lock);571			continue;572		}573		spin_unlock(&pcl->obj.lockref.lock);574 575		if (page)576			put_page(page);577		else if (newpage)578			erofs_pagepool_add(&fe->pagepool, newpage);579	}580 581	/*582	 * don't do inplace I/O if all compressed pages are available in583	 * managed cache since it can be moved to the bypass queue instead.584	 */585	if (standalone)586		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;587}588 589/* (erofs_shrinker) disconnect cached encoded data with pclusters */590int erofs_try_to_free_all_cached_folios(struct erofs_sb_info *sbi,591					struct erofs_workgroup *grp)592{593	struct z_erofs_pcluster *const pcl =594		container_of(grp, struct z_erofs_pcluster, obj);595	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);596	struct folio *folio;597	int i;598 599	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));600	/* Each cached folio contains one page unless bs > ps is supported */601	for (i = 0; i < pclusterpages; ++i) {602		if (pcl->compressed_bvecs[i].page) {603			folio = page_folio(pcl->compressed_bvecs[i].page);604			/* Avoid reclaiming or migrating this folio */605			if (!folio_trylock(folio))606				return -EBUSY;607 608			if (!erofs_folio_is_managed(sbi, folio))609				continue;610			pcl->compressed_bvecs[i].page = NULL;611			folio_detach_private(folio);612			folio_unlock(folio);613		}614	}615	return 0;616}617 618static bool z_erofs_cache_release_folio(struct folio *folio, gfp_t gfp)619{620	struct z_erofs_pcluster *pcl = folio_get_private(folio);621	struct z_erofs_bvec *bvec = pcl->compressed_bvecs;622	struct z_erofs_bvec *end = bvec + z_erofs_pclusterpages(pcl);623	bool ret;624 625	if (!folio_test_private(folio))626		return true;627 628	ret = false;629	spin_lock(&pcl->obj.lockref.lock);630	if (pcl->obj.lockref.count <= 0) {631		DBG_BUGON(z_erofs_is_inline_pcluster(pcl));632		for (; bvec < end; ++bvec) {633			if (bvec->page && page_folio(bvec->page) == folio) {634				bvec->page = NULL;635				folio_detach_private(folio);636				ret = true;637				break;638			}639		}640	}641	spin_unlock(&pcl->obj.lockref.lock);642	return ret;643}644 645/*646 * It will be called only on inode eviction. In case that there are still some647 * decompression requests in progress, wait with rescheduling for a bit here.648 * An extra lock could be introduced instead but it seems unnecessary.649 */650static void z_erofs_cache_invalidate_folio(struct folio *folio,651					   size_t offset, size_t length)652{653	const size_t stop = length + offset;654 655	/* Check for potential overflow in debug mode */656	DBG_BUGON(stop > folio_size(folio) || stop < length);657 658	if (offset == 0 && stop == folio_size(folio))659		while (!z_erofs_cache_release_folio(folio, 0))660			cond_resched();661}662 663static const struct address_space_operations z_erofs_cache_aops = {664	.release_folio = z_erofs_cache_release_folio,665	.invalidate_folio = z_erofs_cache_invalidate_folio,666};667 668int erofs_init_managed_cache(struct super_block *sb)669{670	struct inode *const inode = new_inode(sb);671 672	if (!inode)673		return -ENOMEM;674 675	set_nlink(inode, 1);676	inode->i_size = OFFSET_MAX;677	inode->i_mapping->a_ops = &z_erofs_cache_aops;678	mapping_set_gfp_mask(inode->i_mapping, GFP_KERNEL);679	EROFS_SB(sb)->managed_cache = inode;680	return 0;681}682 683/* callers must be with pcluster lock held */684static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,685			       struct z_erofs_bvec *bvec, bool exclusive)686{687	struct z_erofs_pcluster *pcl = fe->pcl;688	int ret;689 690	if (exclusive) {691		/* give priority for inplaceio to use file pages first */692		spin_lock(&pcl->obj.lockref.lock);693		while (fe->icur > 0) {694			if (pcl->compressed_bvecs[--fe->icur].page)695				continue;696			pcl->compressed_bvecs[fe->icur] = *bvec;697			spin_unlock(&pcl->obj.lockref.lock);698			return 0;699		}700		spin_unlock(&pcl->obj.lockref.lock);701 702		/* otherwise, check if it can be used as a bvpage */703		if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED &&704		    !fe->candidate_bvpage)705			fe->candidate_bvpage = bvec->page;706	}707	ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage,708				   &fe->pagepool);709	fe->pcl->vcnt += (ret >= 0);710	return ret;711}712 713static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)714{715	struct erofs_map_blocks *map = &fe->map;716	struct super_block *sb = fe->inode->i_sb;717	bool ztailpacking = map->m_flags & EROFS_MAP_META;718	struct z_erofs_pcluster *pcl;719	struct erofs_workgroup *grp;720	int err;721 722	if (!(map->m_flags & EROFS_MAP_ENCODED) ||723	    (!ztailpacking && !erofs_blknr(sb, map->m_pa))) {724		DBG_BUGON(1);725		return -EFSCORRUPTED;726	}727 728	/* no available pcluster, let's allocate one */729	pcl = z_erofs_alloc_pcluster(map->m_plen);730	if (IS_ERR(pcl))731		return PTR_ERR(pcl);732 733	spin_lock_init(&pcl->obj.lockref.lock);734	pcl->obj.lockref.count = 1;	/* one ref for this request */735	pcl->algorithmformat = map->m_algorithmformat;736	pcl->length = 0;737	pcl->partial = true;738 739	/* new pclusters should be claimed as type 1, primary and followed */740	pcl->next = fe->owned_head;741	pcl->pageofs_out = map->m_la & ~PAGE_MASK;742	fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;743 744	/*745	 * lock all primary followed works before visible to others746	 * and mutex_trylock *never* fails for a new pcluster.747	 */748	mutex_init(&pcl->lock);749	DBG_BUGON(!mutex_trylock(&pcl->lock));750 751	if (ztailpacking) {752		pcl->obj.index = 0;	/* which indicates ztailpacking */753	} else {754		pcl->obj.index = erofs_blknr(sb, map->m_pa);755 756		grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);757		if (IS_ERR(grp)) {758			err = PTR_ERR(grp);759			goto err_out;760		}761 762		if (grp != &pcl->obj) {763			fe->pcl = container_of(grp,764					struct z_erofs_pcluster, obj);765			err = -EEXIST;766			goto err_out;767		}768	}769	fe->owned_head = &pcl->next;770	fe->pcl = pcl;771	return 0;772 773err_out:774	mutex_unlock(&pcl->lock);775	z_erofs_free_pcluster(pcl);776	return err;777}778 779static int z_erofs_pcluster_begin(struct z_erofs_decompress_frontend *fe)780{781	struct erofs_map_blocks *map = &fe->map;782	struct super_block *sb = fe->inode->i_sb;783	erofs_blk_t blknr = erofs_blknr(sb, map->m_pa);784	struct erofs_workgroup *grp = NULL;785	int ret;786 787	DBG_BUGON(fe->pcl);788	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */789	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);790 791	if (!(map->m_flags & EROFS_MAP_META)) {792		grp = erofs_find_workgroup(sb, blknr);793	} else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {794		DBG_BUGON(1);795		return -EFSCORRUPTED;796	}797 798	if (grp) {799		fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);800		ret = -EEXIST;801	} else {802		ret = z_erofs_register_pcluster(fe);803	}804 805	if (ret == -EEXIST) {806		mutex_lock(&fe->pcl->lock);807		/* check if this pcluster hasn't been linked into any chain. */808		if (cmpxchg(&fe->pcl->next, Z_EROFS_PCLUSTER_NIL,809			    fe->owned_head) == Z_EROFS_PCLUSTER_NIL) {810			/* .. so it can be attached to our submission chain */811			fe->owned_head = &fe->pcl->next;812			fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;813		} else {	/* otherwise, it belongs to an inflight chain */814			fe->mode = Z_EROFS_PCLUSTER_INFLIGHT;815		}816	} else if (ret) {817		return ret;818	}819 820	z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,821				Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);822	if (!z_erofs_is_inline_pcluster(fe->pcl)) {823		/* bind cache first when cached decompression is preferred */824		z_erofs_bind_cache(fe);825	} else {826		void *mptr;827 828		mptr = erofs_read_metabuf(&map->buf, sb, map->m_pa, EROFS_NO_KMAP);829		if (IS_ERR(mptr)) {830			ret = PTR_ERR(mptr);831			erofs_err(sb, "failed to get inline data %d", ret);832			return ret;833		}834		get_page(map->buf.page);835		WRITE_ONCE(fe->pcl->compressed_bvecs[0].page, map->buf.page);836		fe->pcl->pageofs_in = map->m_pa & ~PAGE_MASK;837		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;838	}839	/* file-backed inplace I/O pages are traversed in reverse order */840	fe->icur = z_erofs_pclusterpages(fe->pcl);841	return 0;842}843 844/*845 * keep in mind that no referenced pclusters will be freed846 * only after a RCU grace period.847 */848static void z_erofs_rcu_callback(struct rcu_head *head)849{850	z_erofs_free_pcluster(container_of(head,851			struct z_erofs_pcluster, rcu));852}853 854void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)855{856	struct z_erofs_pcluster *const pcl =857		container_of(grp, struct z_erofs_pcluster, obj);858 859	call_rcu(&pcl->rcu, z_erofs_rcu_callback);860}861 862static void z_erofs_pcluster_end(struct z_erofs_decompress_frontend *fe)863{864	struct z_erofs_pcluster *pcl = fe->pcl;865 866	if (!pcl)867		return;868 869	z_erofs_bvec_iter_end(&fe->biter);870	mutex_unlock(&pcl->lock);871 872	if (fe->candidate_bvpage)873		fe->candidate_bvpage = NULL;874 875	/*876	 * if all pending pages are added, don't hold its reference877	 * any longer if the pcluster isn't hosted by ourselves.878	 */879	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE)880		erofs_workgroup_put(&pcl->obj);881 882	fe->pcl = NULL;883}884 885static int z_erofs_read_fragment(struct super_block *sb, struct folio *folio,886			unsigned int cur, unsigned int end, erofs_off_t pos)887{888	struct inode *packed_inode = EROFS_SB(sb)->packed_inode;889	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;890	unsigned int cnt;891	u8 *src;892 893	if (!packed_inode)894		return -EFSCORRUPTED;895 896	buf.mapping = packed_inode->i_mapping;897	for (; cur < end; cur += cnt, pos += cnt) {898		cnt = min(end - cur, sb->s_blocksize - erofs_blkoff(sb, pos));899		src = erofs_bread(&buf, pos, EROFS_KMAP);900		if (IS_ERR(src)) {901			erofs_put_metabuf(&buf);902			return PTR_ERR(src);903		}904		memcpy_to_folio(folio, cur, src, cnt);905	}906	erofs_put_metabuf(&buf);907	return 0;908}909 910static int z_erofs_scan_folio(struct z_erofs_decompress_frontend *f,911			      struct folio *folio, bool ra)912{913	struct inode *const inode = f->inode;914	struct erofs_map_blocks *const map = &f->map;915	const loff_t offset = folio_pos(folio);916	const unsigned int bs = i_blocksize(inode);917	unsigned int end = folio_size(folio), split = 0, cur, pgs;918	bool tight, excl;919	int err = 0;920 921	tight = (bs == PAGE_SIZE);922	erofs_onlinefolio_init(folio);923	do {924		if (offset + end - 1 < map->m_la ||925		    offset + end - 1 >= map->m_la + map->m_llen) {926			z_erofs_pcluster_end(f);927			map->m_la = offset + end - 1;928			map->m_llen = 0;929			err = z_erofs_map_blocks_iter(inode, map, 0);930			if (err)931				break;932		}933 934		cur = offset > map->m_la ? 0 : map->m_la - offset;935		pgs = round_down(cur, PAGE_SIZE);936		/* bump split parts first to avoid several separate cases */937		++split;938 939		if (!(map->m_flags & EROFS_MAP_MAPPED)) {940			folio_zero_segment(folio, cur, end);941			tight = false;942		} else if (map->m_flags & EROFS_MAP_FRAGMENT) {943			erofs_off_t fpos = offset + cur - map->m_la;944 945			err = z_erofs_read_fragment(inode->i_sb, folio, cur,946					cur + min(map->m_llen - fpos, end - cur),947					EROFS_I(inode)->z_fragmentoff + fpos);948			if (err)949				break;950			tight = false;951		} else {952			if (!f->pcl) {953				err = z_erofs_pcluster_begin(f);954				if (err)955					break;956				f->pcl->besteffort |= !ra;957			}958 959			pgs = round_down(end - 1, PAGE_SIZE);960			/*961			 * Ensure this partial page belongs to this submit chain962			 * rather than other concurrent submit chains or963			 * noio(bypass) chains since those chains are handled964			 * asynchronously thus it cannot be used for inplace I/O965			 * or bvpage (should be processed in the strict order.)966			 */967			tight &= (f->mode >= Z_EROFS_PCLUSTER_FOLLOWED);968			excl = false;969			if (cur <= pgs) {970				excl = (split <= 1) || tight;971				cur = pgs;972			}973 974			err = z_erofs_attach_page(f, &((struct z_erofs_bvec) {975				.page = folio_page(folio, pgs >> PAGE_SHIFT),976				.offset = offset + pgs - map->m_la,977				.end = end - pgs, }), excl);978			if (err)979				break;980 981			erofs_onlinefolio_split(folio);982			if (f->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))983				f->pcl->multibases = true;984			if (f->pcl->length < offset + end - map->m_la) {985				f->pcl->length = offset + end - map->m_la;986				f->pcl->pageofs_out = map->m_la & ~PAGE_MASK;987			}988			if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&989			    !(map->m_flags & EROFS_MAP_PARTIAL_REF) &&990			    f->pcl->length == map->m_llen)991				f->pcl->partial = false;992		}993		/* shorten the remaining extent to update progress */994		map->m_llen = offset + cur - map->m_la;995		map->m_flags &= ~EROFS_MAP_FULL_MAPPED;996		if (cur <= pgs) {997			split = cur < pgs;998			tight = (bs == PAGE_SIZE);999		}1000	} while ((end = cur) > 0);1001	erofs_onlinefolio_end(folio, err);1002	return err;1003}1004 1005static bool z_erofs_is_sync_decompress(struct erofs_sb_info *sbi,1006				       unsigned int readahead_pages)1007{1008	/* auto: enable for read_folio, disable for readahead */1009	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&1010	    !readahead_pages)1011		return true;1012 1013	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&1014	    (readahead_pages <= sbi->opt.max_sync_decompress_pages))1015		return true;1016 1017	return false;1018}1019 1020static bool z_erofs_page_is_invalidated(struct page *page)1021{1022	return !page_folio(page)->mapping && !z_erofs_is_shortlived_page(page);1023}1024 1025struct z_erofs_decompress_backend {1026	struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];1027	struct super_block *sb;1028	struct z_erofs_pcluster *pcl;1029 1030	/* pages with the longest decompressed length for deduplication */1031	struct page **decompressed_pages;1032	/* pages to keep the compressed data */1033	struct page **compressed_pages;1034 1035	struct list_head decompressed_secondary_bvecs;1036	struct page **pagepool;1037	unsigned int onstack_used, nr_pages;1038};1039 1040struct z_erofs_bvec_item {1041	struct z_erofs_bvec bvec;1042	struct list_head list;1043};1044 1045static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,1046					 struct z_erofs_bvec *bvec)1047{1048	struct z_erofs_bvec_item *item;1049	unsigned int pgnr;1050 1051	if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK) &&1052	    (bvec->end == PAGE_SIZE ||1053	     bvec->offset + bvec->end == be->pcl->length)) {1054		pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;1055		DBG_BUGON(pgnr >= be->nr_pages);1056		if (!be->decompressed_pages[pgnr]) {1057			be->decompressed_pages[pgnr] = bvec->page;1058			return;1059		}1060	}1061 1062	/* (cold path) one pcluster is requested multiple times */1063	item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);1064	item->bvec = *bvec;1065	list_add(&item->list, &be->decompressed_secondary_bvecs);1066}1067 1068static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,1069				      int err)1070{1071	unsigned int off0 = be->pcl->pageofs_out;1072	struct list_head *p, *n;1073 1074	list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {1075		struct z_erofs_bvec_item *bvi;1076		unsigned int end, cur;1077		void *dst, *src;1078 1079		bvi = container_of(p, struct z_erofs_bvec_item, list);1080		cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;1081		end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,1082			    bvi->bvec.end);1083		dst = kmap_local_page(bvi->bvec.page);1084		while (cur < end) {1085			unsigned int pgnr, scur, len;1086 1087			pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;1088			DBG_BUGON(pgnr >= be->nr_pages);1089 1090			scur = bvi->bvec.offset + cur -1091					((pgnr << PAGE_SHIFT) - off0);1092			len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);1093			if (!be->decompressed_pages[pgnr]) {1094				err = -EFSCORRUPTED;1095				cur += len;1096				continue;1097			}1098			src = kmap_local_page(be->decompressed_pages[pgnr]);1099			memcpy(dst + cur, src + scur, len);1100			kunmap_local(src);1101			cur += len;1102		}1103		kunmap_local(dst);1104		erofs_onlinefolio_end(page_folio(bvi->bvec.page), err);1105		list_del(p);1106		kfree(bvi);1107	}1108}1109 1110static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)1111{1112	struct z_erofs_pcluster *pcl = be->pcl;1113	struct z_erofs_bvec_iter biter;1114	struct page *old_bvpage;1115	int i;1116 1117	z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);1118	for (i = 0; i < pcl->vcnt; ++i) {1119		struct z_erofs_bvec bvec;1120 1121		z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);1122 1123		if (old_bvpage)1124			z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);1125 1126		DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));1127		z_erofs_do_decompressed_bvec(be, &bvec);1128	}1129 1130	old_bvpage = z_erofs_bvec_iter_end(&biter);1131	if (old_bvpage)1132		z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);1133}1134 1135static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,1136				  bool *overlapped)1137{1138	struct z_erofs_pcluster *pcl = be->pcl;1139	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);1140	int i, err = 0;1141 1142	*overlapped = false;1143	for (i = 0; i < pclusterpages; ++i) {1144		struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];1145		struct page *page = bvec->page;1146 1147		/* compressed data ought to be valid when decompressing */1148		if (IS_ERR(page) || !page) {1149			bvec->page = NULL;	/* clear the failure reason */1150			err = page ? PTR_ERR(page) : -EIO;1151			continue;1152		}1153		be->compressed_pages[i] = page;1154 1155		if (z_erofs_is_inline_pcluster(pcl) ||1156		    erofs_folio_is_managed(EROFS_SB(be->sb), page_folio(page))) {1157			if (!PageUptodate(page))1158				err = -EIO;1159			continue;1160		}1161 1162		DBG_BUGON(z_erofs_page_is_invalidated(page));1163		if (z_erofs_is_shortlived_page(page))1164			continue;1165		z_erofs_do_decompressed_bvec(be, bvec);1166		*overlapped = true;1167	}1168	return err;1169}1170 1171static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,1172				       int err)1173{1174	struct erofs_sb_info *const sbi = EROFS_SB(be->sb);1175	struct z_erofs_pcluster *pcl = be->pcl;1176	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);1177	const struct z_erofs_decompressor *decomp =1178				z_erofs_decomp[pcl->algorithmformat];1179	int i, j, jtop, err2;1180	struct page *page;1181	bool overlapped;1182 1183	mutex_lock(&pcl->lock);1184	be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;1185 1186	/* allocate (de)compressed page arrays if cannot be kept on stack */1187	be->decompressed_pages = NULL;1188	be->compressed_pages = NULL;1189	be->onstack_used = 0;1190	if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {1191		be->decompressed_pages = be->onstack_pages;1192		be->onstack_used = be->nr_pages;1193		memset(be->decompressed_pages, 0,1194		       sizeof(struct page *) * be->nr_pages);1195	}1196 1197	if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)1198		be->compressed_pages = be->onstack_pages + be->onstack_used;1199 1200	if (!be->decompressed_pages)1201		be->decompressed_pages =1202			kvcalloc(be->nr_pages, sizeof(struct page *),1203				 GFP_KERNEL | __GFP_NOFAIL);1204	if (!be->compressed_pages)1205		be->compressed_pages =1206			kvcalloc(pclusterpages, sizeof(struct page *),1207				 GFP_KERNEL | __GFP_NOFAIL);1208 1209	z_erofs_parse_out_bvecs(be);1210	err2 = z_erofs_parse_in_bvecs(be, &overlapped);1211	if (err2)1212		err = err2;1213	if (!err)1214		err = decomp->decompress(&(struct z_erofs_decompress_req) {1215					.sb = be->sb,1216					.in = be->compressed_pages,1217					.out = be->decompressed_pages,1218					.pageofs_in = pcl->pageofs_in,1219					.pageofs_out = pcl->pageofs_out,1220					.inputsize = pcl->pclustersize,1221					.outputsize = pcl->length,1222					.alg = pcl->algorithmformat,1223					.inplace_io = overlapped,1224					.partial_decoding = pcl->partial,1225					.fillgaps = pcl->multibases,1226					.gfp = pcl->besteffort ? GFP_KERNEL :1227						GFP_NOWAIT | __GFP_NORETRY1228				 }, be->pagepool);1229 1230	/* must handle all compressed pages before actual file pages */1231	if (z_erofs_is_inline_pcluster(pcl)) {1232		page = pcl->compressed_bvecs[0].page;1233		WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);1234		put_page(page);1235	} else {1236		/* managed folios are still left in compressed_bvecs[] */1237		for (i = 0; i < pclusterpages; ++i) {1238			page = be->compressed_pages[i];1239			if (!page ||1240			    erofs_folio_is_managed(sbi, page_folio(page)))1241				continue;1242			(void)z_erofs_put_shortlivedpage(be->pagepool, page);1243			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);1244		}1245	}1246	if (be->compressed_pages < be->onstack_pages ||1247	    be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)1248		kvfree(be->compressed_pages);1249 1250	jtop = 0;1251	z_erofs_fill_other_copies(be, err);1252	for (i = 0; i < be->nr_pages; ++i) {1253		page = be->decompressed_pages[i];1254		if (!page)1255			continue;1256 1257		DBG_BUGON(z_erofs_page_is_invalidated(page));1258		if (!z_erofs_is_shortlived_page(page)) {1259			erofs_onlinefolio_end(page_folio(page), err);1260			continue;1261		}1262		if (pcl->algorithmformat != Z_EROFS_COMPRESSION_LZ4) {1263			erofs_pagepool_add(be->pagepool, page);1264			continue;1265		}1266		for (j = 0; j < jtop && be->decompressed_pages[j] != page; ++j)1267			;1268		if (j >= jtop)	/* this bounce page is newly detected */1269			be->decompressed_pages[jtop++] = page;1270	}1271	while (jtop)1272		erofs_pagepool_add(be->pagepool,1273				   be->decompressed_pages[--jtop]);1274	if (be->decompressed_pages != be->onstack_pages)1275		kvfree(be->decompressed_pages);1276 1277	pcl->length = 0;1278	pcl->partial = true;1279	pcl->multibases = false;1280	pcl->besteffort = false;1281	pcl->bvset.nextpage = NULL;1282	pcl->vcnt = 0;1283 1284	/* pcluster lock MUST be taken before the following line */1285	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);1286	mutex_unlock(&pcl->lock);1287	return err;1288}1289 1290static int z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,1291				    struct page **pagepool)1292{1293	struct z_erofs_decompress_backend be = {1294		.sb = io->sb,1295		.pagepool = pagepool,1296		.decompressed_secondary_bvecs =1297			LIST_HEAD_INIT(be.decompressed_secondary_bvecs),1298	};1299	z_erofs_next_pcluster_t owned = io->head;1300	int err = io->eio ? -EIO : 0;1301 1302	while (owned != Z_EROFS_PCLUSTER_TAIL) {1303		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);1304 1305		be.pcl = container_of(owned, struct z_erofs_pcluster, next);1306		owned = READ_ONCE(be.pcl->next);1307 1308		err = z_erofs_decompress_pcluster(&be, err) ?: err;1309		if (z_erofs_is_inline_pcluster(be.pcl))1310			z_erofs_free_pcluster(be.pcl);1311		else1312			erofs_workgroup_put(&be.pcl->obj);1313	}1314	return err;1315}1316 1317static void z_erofs_decompressqueue_work(struct work_struct *work)1318{1319	struct z_erofs_decompressqueue *bgq =1320		container_of(work, struct z_erofs_decompressqueue, u.work);1321	struct page *pagepool = NULL;1322 1323	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL);1324	z_erofs_decompress_queue(bgq, &pagepool);1325	erofs_release_pages(&pagepool);1326	kvfree(bgq);1327}1328 1329#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD1330static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)1331{1332	z_erofs_decompressqueue_work((struct work_struct *)work);1333}1334#endif1335 1336static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,1337				       int bios)1338{1339	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);1340 1341	/* wake up the caller thread for sync decompression */1342	if (io->sync) {1343		if (!atomic_add_return(bios, &io->pending_bios))1344			complete(&io->u.done);1345		return;1346	}1347 1348	if (atomic_add_return(bios, &io->pending_bios))1349		return;1350	/* Use (kthread_)work and sync decompression for atomic contexts only */1351	if (!in_task() || irqs_disabled() || rcu_read_lock_any_held()) {1352#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD1353		struct kthread_worker *worker;1354 1355		rcu_read_lock();1356		worker = rcu_dereference(1357				z_erofs_pcpu_workers[raw_smp_processor_id()]);1358		if (!worker) {1359			INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);1360			queue_work(z_erofs_workqueue, &io->u.work);1361		} else {1362			kthread_queue_work(worker, &io->u.kthread_work);1363		}1364		rcu_read_unlock();1365#else1366		queue_work(z_erofs_workqueue, &io->u.work);1367#endif1368		/* enable sync decompression for readahead */1369		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)1370			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;1371		return;1372	}1373	z_erofs_decompressqueue_work(&io->u.work);1374}1375 1376static void z_erofs_fill_bio_vec(struct bio_vec *bvec,1377				 struct z_erofs_decompress_frontend *f,1378				 struct z_erofs_pcluster *pcl,1379				 unsigned int nr,1380				 struct address_space *mc)1381{1382	gfp_t gfp = mapping_gfp_mask(mc);1383	bool tocache = false;1384	struct z_erofs_bvec zbv;1385	struct address_space *mapping;1386	struct folio *folio;1387	struct page *page;1388	int bs = i_blocksize(f->inode);1389 1390	/* Except for inplace folios, the entire folio can be used for I/Os */1391	bvec->bv_offset = 0;1392	bvec->bv_len = PAGE_SIZE;1393repeat:1394	spin_lock(&pcl->obj.lockref.lock);1395	zbv = pcl->compressed_bvecs[nr];1396	spin_unlock(&pcl->obj.lockref.lock);1397	if (!zbv.page)1398		goto out_allocfolio;1399 1400	bvec->bv_page = zbv.page;1401	DBG_BUGON(z_erofs_is_shortlived_page(bvec->bv_page));1402 1403	folio = page_folio(zbv.page);1404	/*1405	 * Handle preallocated cached folios.  We tried to allocate such folios1406	 * without triggering direct reclaim.  If allocation failed, inplace1407	 * file-backed folios will be used instead.1408	 */1409	if (folio->private == (void *)Z_EROFS_PREALLOCATED_PAGE) {1410		tocache = true;1411		goto out_tocache;1412	}1413 1414	mapping = READ_ONCE(folio->mapping);1415	/*1416	 * File-backed folios for inplace I/Os are all locked steady,1417	 * therefore it is impossible for `mapping` to be NULL.1418	 */1419	if (mapping && mapping != mc) {1420		if (zbv.offset < 0)1421			bvec->bv_offset = round_up(-zbv.offset, bs);1422		bvec->bv_len = round_up(zbv.end, bs) - bvec->bv_offset;1423		return;1424	}1425 1426	folio_lock(folio);1427	if (likely(folio->mapping == mc)) {1428		/*1429		 * The cached folio is still in managed cache but without1430		 * a valid `->private` pcluster hint.  Let's reconnect them.1431		 */1432		if (!folio_test_private(folio)) {1433			folio_attach_private(folio, pcl);1434			/* compressed_bvecs[] already takes a ref before */1435			folio_put(folio);1436		}1437		if (likely(folio->private == pcl))  {1438			/* don't submit cache I/Os again if already uptodate */1439			if (folio_test_uptodate(folio)) {1440				folio_unlock(folio);1441				bvec->bv_page = NULL;1442			}1443			return;1444		}1445		/*1446		 * Already linked with another pcluster, which only appears in1447		 * crafted images by fuzzers for now.  But handle this anyway.1448		 */1449		tocache = false;	/* use temporary short-lived pages */1450	} else {1451		DBG_BUGON(1); /* referenced managed folios can't be truncated */1452		tocache = true;1453	}1454	folio_unlock(folio);1455	folio_put(folio);1456out_allocfolio:1457	page = __erofs_allocpage(&f->pagepool, gfp, true);1458	spin_lock(&pcl->obj.lockref.lock);1459	if (unlikely(pcl->compressed_bvecs[nr].page != zbv.page)) {1460		if (page)1461			erofs_pagepool_add(&f->pagepool, page);1462		spin_unlock(&pcl->obj.lockref.lock);1463		cond_resched();1464		goto repeat;1465	}1466	pcl->compressed_bvecs[nr].page = page ? page : ERR_PTR(-ENOMEM);1467	spin_unlock(&pcl->obj.lockref.lock);1468	bvec->bv_page = page;1469	if (!page)1470		return;1471	folio = page_folio(page);1472out_tocache:1473	if (!tocache || bs != PAGE_SIZE ||1474	    filemap_add_folio(mc, folio, pcl->obj.index + nr, gfp)) {1475		/* turn into a temporary shortlived folio (1 ref) */1476		folio->private = (void *)Z_EROFS_SHORTLIVED_PAGE;1477		return;1478	}1479	folio_attach_private(folio, pcl);1480	/* drop a refcount added by allocpage (then 2 refs in total here) */1481	folio_put(folio);1482}1483 1484static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb,1485			      struct z_erofs_decompressqueue *fgq, bool *fg)1486{1487	struct z_erofs_decompressqueue *q;1488 1489	if (fg && !*fg) {1490		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);1491		if (!q) {1492			*fg = true;1493			goto fg_out;1494		}1495#ifdef CONFIG_EROFS_FS_PCPU_KTHREAD1496		kthread_init_work(&q->u.kthread_work,1497				  z_erofs_decompressqueue_kthread_work);1498#else1499		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);1500#endif1501	} else {1502fg_out:1503		q = fgq;1504		init_completion(&fgq->u.done);1505		atomic_set(&fgq->pending_bios, 0);1506		q->eio = false;1507		q->sync = true;1508	}1509	q->sb = sb;1510	q->head = Z_EROFS_PCLUSTER_TAIL;1511	return q;1512}1513 1514/* define decompression jobqueue types */1515enum {1516	JQ_BYPASS,1517	JQ_SUBMIT,1518	NR_JOBQUEUES,1519};1520 1521static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,1522				    z_erofs_next_pcluster_t qtail[],1523				    z_erofs_next_pcluster_t owned_head)1524{1525	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];1526	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];1527 1528	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL);1529 1530	WRITE_ONCE(*submit_qtail, owned_head);1531	WRITE_ONCE(*bypass_qtail, &pcl->next);1532 1533	qtail[JQ_BYPASS] = &pcl->next;1534}1535 1536static void z_erofs_endio(struct bio *bio)1537{1538	struct z_erofs_decompressqueue *q = bio->bi_private;1539	blk_status_t err = bio->bi_status;1540	struct folio_iter fi;1541 1542	bio_for_each_folio_all(fi, bio) {1543		struct folio *folio = fi.folio;1544 1545		DBG_BUGON(folio_test_uptodate(folio));1546		DBG_BUGON(z_erofs_page_is_invalidated(&folio->page));1547		if (!erofs_folio_is_managed(EROFS_SB(q->sb), folio))1548			continue;1549 1550		if (!err)1551			folio_mark_uptodate(folio);1552		folio_unlock(folio);1553	}1554	if (err)1555		q->eio = true;1556	z_erofs_decompress_kickoff(q, -1);1557	if (bio->bi_bdev)1558		bio_put(bio);1559}1560 1561static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,1562				 struct z_erofs_decompressqueue *fgq,1563				 bool *force_fg, bool readahead)1564{1565	struct super_block *sb = f->inode->i_sb;1566	struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));1567	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];1568	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];1569	z_erofs_next_pcluster_t owned_head = f->owned_head;1570	/* bio is NULL initially, so no need to initialize last_{index,bdev} */1571	erofs_off_t last_pa;1572	unsigned int nr_bios = 0;1573	struct bio *bio = NULL;1574	unsigned long pflags;1575	int memstall = 0;1576 1577	/* No need to read from device for pclusters in the bypass queue. */1578	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);1579	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg);1580 1581	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;1582	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;1583 1584	/* by default, all need io submission */1585	q[JQ_SUBMIT]->head = owned_head;1586 1587	do {1588		struct erofs_map_dev mdev;1589		struct z_erofs_pcluster *pcl;1590		erofs_off_t cur, end;1591		struct bio_vec bvec;1592		unsigned int i = 0;1593		bool bypass = true;1594 1595		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);1596		pcl = container_of(owned_head, struct z_erofs_pcluster, next);1597		owned_head = READ_ONCE(pcl->next);1598 1599		if (z_erofs_is_inline_pcluster(pcl)) {1600			move_to_bypass_jobqueue(pcl, qtail, owned_head);1601			continue;1602		}1603 1604		/* no device id here, thus it will always succeed */1605		mdev = (struct erofs_map_dev) {1606			.m_pa = erofs_pos(sb, pcl->obj.index),1607		};1608		(void)erofs_map_dev(sb, &mdev);1609 1610		cur = mdev.m_pa;1611		end = cur + pcl->pclustersize;1612		do {1613			bvec.bv_page = NULL;1614			if (bio && (cur != last_pa ||1615				    bio->bi_bdev != mdev.m_bdev)) {1616drain_io:1617				if (erofs_is_fileio_mode(EROFS_SB(sb)))1618					erofs_fileio_submit_bio(bio);1619				else if (erofs_is_fscache_mode(sb))1620					erofs_fscache_submit_bio(bio);1621				else1622					submit_bio(bio);1623 1624				if (memstall) {1625					psi_memstall_leave(&pflags);1626					memstall = 0;1627				}1628				bio = NULL;1629			}1630 1631			if (!bvec.bv_page) {1632				z_erofs_fill_bio_vec(&bvec, f, pcl, i++, mc);1633				if (!bvec.bv_page)1634					continue;1635				if (cur + bvec.bv_len > end)1636					bvec.bv_len = end - cur;1637				DBG_BUGON(bvec.bv_len < sb->s_blocksize);1638			}1639 1640			if (unlikely(PageWorkingset(bvec.bv_page)) &&1641			    !memstall) {1642				psi_memstall_enter(&pflags);1643				memstall = 1;1644			}1645 1646			if (!bio) {1647				if (erofs_is_fileio_mode(EROFS_SB(sb)))1648					bio = erofs_fileio_bio_alloc(&mdev);1649				else if (erofs_is_fscache_mode(sb))1650					bio = erofs_fscache_bio_alloc(&mdev);1651				else1652					bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,1653							REQ_OP_READ, GFP_NOIO);1654				bio->bi_end_io = z_erofs_endio;1655				bio->bi_iter.bi_sector = cur >> 9;1656				bio->bi_private = q[JQ_SUBMIT];1657				if (readahead)1658					bio->bi_opf |= REQ_RAHEAD;1659				++nr_bios;1660			}1661 1662			if (!bio_add_page(bio, bvec.bv_page, bvec.bv_len,1663					  bvec.bv_offset))1664				goto drain_io;1665			last_pa = cur + bvec.bv_len;1666			bypass = false;1667		} while ((cur += bvec.bv_len) < end);1668 1669		if (!bypass)1670			qtail[JQ_SUBMIT] = &pcl->next;1671		else1672			move_to_bypass_jobqueue(pcl, qtail, owned_head);1673	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);1674 1675	if (bio) {1676		if (erofs_is_fileio_mode(EROFS_SB(sb)))1677			erofs_fileio_submit_bio(bio);1678		else if (erofs_is_fscache_mode(sb))1679			erofs_fscache_submit_bio(bio);1680		else1681			submit_bio(bio);1682		if (memstall)1683			psi_memstall_leave(&pflags);1684	}1685 1686	/*1687	 * although background is preferred, no one is pending for submission.1688	 * don't issue decompression but drop it directly instead.1689	 */1690	if (!*force_fg && !nr_bios) {1691		kvfree(q[JQ_SUBMIT]);1692		return;1693	}1694	z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios);1695}1696 1697static int z_erofs_runqueue(struct z_erofs_decompress_frontend *f,1698			    unsigned int ra_folios)1699{1700	struct z_erofs_decompressqueue io[NR_JOBQUEUES];1701	struct erofs_sb_info *sbi = EROFS_I_SB(f->inode);1702	bool force_fg = z_erofs_is_sync_decompress(sbi, ra_folios);1703	int err;1704 1705	if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)1706		return 0;1707	z_erofs_submit_queue(f, io, &force_fg, !!ra_folios);1708 1709	/* handle bypass queue (no i/o pclusters) immediately */1710	err = z_erofs_decompress_queue(&io[JQ_BYPASS], &f->pagepool);1711	if (!force_fg)1712		return err;1713 1714	/* wait until all bios are completed */1715	wait_for_completion_io(&io[JQ_SUBMIT].u.done);1716 1717	/* handle synchronous decompress queue in the caller context */1718	return z_erofs_decompress_queue(&io[JQ_SUBMIT], &f->pagepool) ?: err;1719}1720 1721/*1722 * Since partial uptodate is still unimplemented for now, we have to use1723 * approximate readmore strategies as a start.1724 */1725static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,1726		struct readahead_control *rac, bool backmost)1727{1728	struct inode *inode = f->inode;1729	struct erofs_map_blocks *map = &f->map;1730	erofs_off_t cur, end, headoffset = f->headoffset;1731	int err;1732 1733	if (backmost) {1734		if (rac)1735			end = headoffset + readahead_length(rac) - 1;1736		else1737			end = headoffset + PAGE_SIZE - 1;1738		map->m_la = end;1739		err = z_erofs_map_blocks_iter(inode, map,1740					      EROFS_GET_BLOCKS_READMORE);1741		if (err)1742			return;1743 1744		/* expand ra for the trailing edge if readahead */1745		if (rac) {1746			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);1747			readahead_expand(rac, headoffset, cur - headoffset);1748			return;1749		}1750		end = round_up(end, PAGE_SIZE);1751	} else {1752		end = round_up(map->m_la, PAGE_SIZE);1753		if (!map->m_llen)1754			return;1755	}1756 1757	cur = map->m_la + map->m_llen - 1;1758	while ((cur >= end) && (cur < i_size_read(inode))) {1759		pgoff_t index = cur >> PAGE_SHIFT;1760		struct folio *folio;1761 1762		folio = erofs_grab_folio_nowait(inode->i_mapping, index);1763		if (!IS_ERR_OR_NULL(folio)) {1764			if (folio_test_uptodate(folio))1765				folio_unlock(folio);1766			else1767				z_erofs_scan_folio(f, folio, !!rac);1768			folio_put(folio);1769		}1770 1771		if (cur < PAGE_SIZE)1772			break;1773		cur = (index << PAGE_SHIFT) - 1;1774	}1775}1776 1777static int z_erofs_read_folio(struct file *file, struct folio *folio)1778{1779	struct inode *const inode = folio->mapping->host;1780	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);1781	int err;1782 1783	trace_erofs_read_folio(folio, false);1784	f.headoffset = (erofs_off_t)folio->index << PAGE_SHIFT;1785 1786	z_erofs_pcluster_readmore(&f, NULL, true);1787	err = z_erofs_scan_folio(&f, folio, false);1788	z_erofs_pcluster_readmore(&f, NULL, false);1789	z_erofs_pcluster_end(&f);1790 1791	/* if some pclusters are ready, need submit them anyway */1792	err = z_erofs_runqueue(&f, 0) ?: err;1793	if (err && err != -EINTR)1794		erofs_err(inode->i_sb, "read error %d @ %lu of nid %llu",1795			  err, folio->index, EROFS_I(inode)->nid);1796 1797	erofs_put_metabuf(&f.map.buf);1798	erofs_release_pages(&f.pagepool);1799	return err;1800}1801 1802static void z_erofs_readahead(struct readahead_control *rac)1803{1804	struct inode *const inode = rac->mapping->host;1805	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);1806	struct folio *head = NULL, *folio;1807	unsigned int nr_folios;1808	int err;1809 1810	f.headoffset = readahead_pos(rac);1811 1812	z_erofs_pcluster_readmore(&f, rac, true);1813	nr_folios = readahead_count(rac);1814	trace_erofs_readpages(inode, readahead_index(rac), nr_folios, false);1815 1816	while ((folio = readahead_folio(rac))) {1817		folio->private = head;1818		head = folio;1819	}1820 1821	/* traverse in reverse order for best metadata I/O performance */1822	while (head) {1823		folio = head;1824		head = folio_get_private(folio);1825 1826		err = z_erofs_scan_folio(&f, folio, true);1827		if (err && err != -EINTR)1828			erofs_err(inode->i_sb, "readahead error at folio %lu @ nid %llu",1829				  folio->index, EROFS_I(inode)->nid);1830	}1831	z_erofs_pcluster_readmore(&f, rac, false);1832	z_erofs_pcluster_end(&f);1833 1834	(void)z_erofs_runqueue(&f, nr_folios);1835	erofs_put_metabuf(&f.map.buf);1836	erofs_release_pages(&f.pagepool);1837}1838 1839const struct address_space_operations z_erofs_aops = {1840	.read_folio = z_erofs_read_folio,1841	.readahead = z_erofs_readahead,1842};1843