xref: /linux/fs/erofs/zdata.c (revision 3630400697a3d334a391c1dba1b601d852145f2c)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2018 HUAWEI, Inc.
4  *             https://www.huawei.com/
5  * Copyright (C) 2022 Alibaba Cloud
6  */
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		2
14 
15 /*
16  * let's leave a type here in case of introducing
17  * another tagged pointer later.
18  */
19 typedef void *z_erofs_next_pcluster_t;
20 
21 struct z_erofs_bvec {
22 	struct page *page;
23 	int offset;
24 	unsigned int end;
25 };
26 
27 #define __Z_EROFS_BVSET(name, total) \
28 struct 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-only
40  *    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  */
46 struct 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 
96 struct 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 
109 static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl)
110 {
111 	return !pcl->obj.index;
112 }
113 
114 static 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)
120 static 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		32
126 
127 /*
128  * since pclustersize is variable for big pcluster feature, introduce slab
129  * pools implementation for different pcluster sizes.
130  */
131 struct 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 
139 static 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 
144 struct z_erofs_bvec_iter {
145 	struct page *bvpage;
146 	struct z_erofs_bvset *bvset;
147 	unsigned int nr, cur;
148 };
149 
150 static 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 
157 static 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 
173 static 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 
190 static 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 
216 static 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 	else
223 		*old_bvpage = NULL;
224 	*bvec = iter->bvset->bvec[iter->cur++];
225 }
226 
227 static 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 
239 static 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 
261 static 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 
281 static 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 
298 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
299 
300 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
301 static struct kthread_worker __rcu **z_erofs_pcpu_workers;
302 
303 static 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 
318 static 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 
330 static 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 #else
348 static inline void erofs_destroy_percpu_workers(void) {}
349 static inline int erofs_init_percpu_workers(void) { return 0; }
350 #endif
351 
352 #if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_EROFS_FS_PCPU_KTHREAD)
353 static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock);
354 static enum cpuhp_state erofs_cpuhp_state;
355 
356 static 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 
375 static 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 
391 static 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 
404 static 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 */
410 static inline int erofs_cpu_hotplug_init(void) { return 0; }
411 static inline void erofs_cpu_hotplug_destroy(void) {}
412 #endif
413 
414 void 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 
423 int __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 
450 err_cpuhp_init:
451 	erofs_destroy_percpu_workers();
452 err_pcpu_worker:
453 	destroy_workqueue(z_erofs_workqueue);
454 err_workqueue_init:
455 	z_erofs_destroy_pcluster_pool();
456 err_pcluster_pool:
457 	z_erofs_exit_decompressor();
458 err_decompressor:
459 	return err;
460 }
461 
462 enum z_erofs_pclustermode {
463 	Z_EROFS_PCLUSTER_INFLIGHT,
464 	/*
465 	 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it
466 	 * could be dispatched into bypass queue later due to uptodated managed
467 	 * pages. All related online pages cannot be reused for inplace I/O (or
468 	 * 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 can
473 	 * also be linked with the remaining pclusters, which means if the
474 	 * processing page is the tail page of a pcluster, this pcluster can
475 	 * safely use the whole page (since the previous pcluster is within the
476 	 * 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 
487 struct 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 
508 static 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 
525 static 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/O
534 	 * 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 to
560 			 * 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 in
583 	 * 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 */
590 int 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 
618 static 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 some
647  * decompression requests in progress, wait with rescheduling for a bit here.
648  * An extra lock could be introduced instead but it seems unnecessary.
649  */
650 static 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 
663 static 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 
668 int 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 */
684 static 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 
713 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
714 {
715 	struct z_erofs_pcluster *pcl = f->pcl;
716 	z_erofs_next_pcluster_t *owned_head = &f->owned_head;
717 
718 	/* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
719 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
720 		    *owned_head) == Z_EROFS_PCLUSTER_NIL) {
721 		*owned_head = &pcl->next;
722 		/* so we can attach this pcluster to our submission chain. */
723 		f->mode = Z_EROFS_PCLUSTER_FOLLOWED;
724 		return;
725 	}
726 
727 	/* type 2, it belongs to an ongoing chain */
728 	f->mode = Z_EROFS_PCLUSTER_INFLIGHT;
729 }
730 
731 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)
732 {
733 	struct erofs_map_blocks *map = &fe->map;
734 	struct super_block *sb = fe->inode->i_sb;
735 	bool ztailpacking = map->m_flags & EROFS_MAP_META;
736 	struct z_erofs_pcluster *pcl;
737 	struct erofs_workgroup *grp;
738 	int err;
739 
740 	if (!(map->m_flags & EROFS_MAP_ENCODED) ||
741 	    (!ztailpacking && !erofs_blknr(sb, map->m_pa))) {
742 		DBG_BUGON(1);
743 		return -EFSCORRUPTED;
744 	}
745 
746 	/* no available pcluster, let's allocate one */
747 	pcl = z_erofs_alloc_pcluster(map->m_plen);
748 	if (IS_ERR(pcl))
749 		return PTR_ERR(pcl);
750 
751 	spin_lock_init(&pcl->obj.lockref.lock);
752 	pcl->obj.lockref.count = 1;	/* one ref for this request */
753 	pcl->algorithmformat = map->m_algorithmformat;
754 	pcl->length = 0;
755 	pcl->partial = true;
756 
757 	/* new pclusters should be claimed as type 1, primary and followed */
758 	pcl->next = fe->owned_head;
759 	pcl->pageofs_out = map->m_la & ~PAGE_MASK;
760 	fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;
761 
762 	/*
763 	 * lock all primary followed works before visible to others
764 	 * and mutex_trylock *never* fails for a new pcluster.
765 	 */
766 	mutex_init(&pcl->lock);
767 	DBG_BUGON(!mutex_trylock(&pcl->lock));
768 
769 	if (ztailpacking) {
770 		pcl->obj.index = 0;	/* which indicates ztailpacking */
771 	} else {
772 		pcl->obj.index = erofs_blknr(sb, map->m_pa);
773 
774 		grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);
775 		if (IS_ERR(grp)) {
776 			err = PTR_ERR(grp);
777 			goto err_out;
778 		}
779 
780 		if (grp != &pcl->obj) {
781 			fe->pcl = container_of(grp,
782 					struct z_erofs_pcluster, obj);
783 			err = -EEXIST;
784 			goto err_out;
785 		}
786 	}
787 	fe->owned_head = &pcl->next;
788 	fe->pcl = pcl;
789 	return 0;
790 
791 err_out:
792 	mutex_unlock(&pcl->lock);
793 	z_erofs_free_pcluster(pcl);
794 	return err;
795 }
796 
797 static int z_erofs_pcluster_begin(struct z_erofs_decompress_frontend *fe)
798 {
799 	struct erofs_map_blocks *map = &fe->map;
800 	struct super_block *sb = fe->inode->i_sb;
801 	erofs_blk_t blknr = erofs_blknr(sb, map->m_pa);
802 	struct erofs_workgroup *grp = NULL;
803 	int ret;
804 
805 	DBG_BUGON(fe->pcl);
806 
807 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */
808 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
809 
810 	if (!(map->m_flags & EROFS_MAP_META)) {
811 		grp = erofs_find_workgroup(sb, blknr);
812 	} else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
813 		DBG_BUGON(1);
814 		return -EFSCORRUPTED;
815 	}
816 
817 	if (grp) {
818 		fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
819 		ret = -EEXIST;
820 	} else {
821 		ret = z_erofs_register_pcluster(fe);
822 	}
823 
824 	if (ret == -EEXIST) {
825 		mutex_lock(&fe->pcl->lock);
826 		z_erofs_try_to_claim_pcluster(fe);
827 	} else if (ret) {
828 		return ret;
829 	}
830 
831 	z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
832 				Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
833 	if (!z_erofs_is_inline_pcluster(fe->pcl)) {
834 		/* bind cache first when cached decompression is preferred */
835 		z_erofs_bind_cache(fe);
836 	} else {
837 		void *mptr;
838 
839 		mptr = erofs_read_metabuf(&map->buf, sb, map->m_pa, EROFS_NO_KMAP);
840 		if (IS_ERR(mptr)) {
841 			ret = PTR_ERR(mptr);
842 			erofs_err(sb, "failed to get inline data %d", ret);
843 			return ret;
844 		}
845 		get_page(map->buf.page);
846 		WRITE_ONCE(fe->pcl->compressed_bvecs[0].page, map->buf.page);
847 		fe->pcl->pageofs_in = map->m_pa & ~PAGE_MASK;
848 		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
849 	}
850 	/* file-backed inplace I/O pages are traversed in reverse order */
851 	fe->icur = z_erofs_pclusterpages(fe->pcl);
852 	return 0;
853 }
854 
855 /*
856  * keep in mind that no referenced pclusters will be freed
857  * only after a RCU grace period.
858  */
859 static void z_erofs_rcu_callback(struct rcu_head *head)
860 {
861 	z_erofs_free_pcluster(container_of(head,
862 			struct z_erofs_pcluster, rcu));
863 }
864 
865 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
866 {
867 	struct z_erofs_pcluster *const pcl =
868 		container_of(grp, struct z_erofs_pcluster, obj);
869 
870 	call_rcu(&pcl->rcu, z_erofs_rcu_callback);
871 }
872 
873 static void z_erofs_pcluster_end(struct z_erofs_decompress_frontend *fe)
874 {
875 	struct z_erofs_pcluster *pcl = fe->pcl;
876 
877 	if (!pcl)
878 		return;
879 
880 	z_erofs_bvec_iter_end(&fe->biter);
881 	mutex_unlock(&pcl->lock);
882 
883 	if (fe->candidate_bvpage)
884 		fe->candidate_bvpage = NULL;
885 
886 	/*
887 	 * if all pending pages are added, don't hold its reference
888 	 * any longer if the pcluster isn't hosted by ourselves.
889 	 */
890 	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE)
891 		erofs_workgroup_put(&pcl->obj);
892 
893 	fe->pcl = NULL;
894 }
895 
896 static int z_erofs_read_fragment(struct super_block *sb, struct folio *folio,
897 			unsigned int cur, unsigned int end, erofs_off_t pos)
898 {
899 	struct inode *packed_inode = EROFS_SB(sb)->packed_inode;
900 	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
901 	unsigned int cnt;
902 	u8 *src;
903 
904 	if (!packed_inode)
905 		return -EFSCORRUPTED;
906 
907 	buf.mapping = packed_inode->i_mapping;
908 	for (; cur < end; cur += cnt, pos += cnt) {
909 		cnt = min(end - cur, sb->s_blocksize - erofs_blkoff(sb, pos));
910 		src = erofs_bread(&buf, pos, EROFS_KMAP);
911 		if (IS_ERR(src)) {
912 			erofs_put_metabuf(&buf);
913 			return PTR_ERR(src);
914 		}
915 		memcpy_to_folio(folio, cur, src, cnt);
916 	}
917 	erofs_put_metabuf(&buf);
918 	return 0;
919 }
920 
921 static int z_erofs_scan_folio(struct z_erofs_decompress_frontend *f,
922 			      struct folio *folio, bool ra)
923 {
924 	struct inode *const inode = f->inode;
925 	struct erofs_map_blocks *const map = &f->map;
926 	const loff_t offset = folio_pos(folio);
927 	const unsigned int bs = i_blocksize(inode);
928 	unsigned int end = folio_size(folio), split = 0, cur, pgs;
929 	bool tight, excl;
930 	int err = 0;
931 
932 	tight = (bs == PAGE_SIZE);
933 	erofs_onlinefolio_init(folio);
934 	do {
935 		if (offset + end - 1 < map->m_la ||
936 		    offset + end - 1 >= map->m_la + map->m_llen) {
937 			z_erofs_pcluster_end(f);
938 			map->m_la = offset + end - 1;
939 			map->m_llen = 0;
940 			err = z_erofs_map_blocks_iter(inode, map, 0);
941 			if (err)
942 				break;
943 		}
944 
945 		cur = offset > map->m_la ? 0 : map->m_la - offset;
946 		pgs = round_down(cur, PAGE_SIZE);
947 		/* bump split parts first to avoid several separate cases */
948 		++split;
949 
950 		if (!(map->m_flags & EROFS_MAP_MAPPED)) {
951 			folio_zero_segment(folio, cur, end);
952 			tight = false;
953 		} else if (map->m_flags & EROFS_MAP_FRAGMENT) {
954 			erofs_off_t fpos = offset + cur - map->m_la;
955 
956 			err = z_erofs_read_fragment(inode->i_sb, folio, cur,
957 					cur + min(map->m_llen - fpos, end - cur),
958 					EROFS_I(inode)->z_fragmentoff + fpos);
959 			if (err)
960 				break;
961 			tight = false;
962 		} else {
963 			if (!f->pcl) {
964 				err = z_erofs_pcluster_begin(f);
965 				if (err)
966 					break;
967 				f->pcl->besteffort |= !ra;
968 			}
969 
970 			pgs = round_down(end - 1, PAGE_SIZE);
971 			/*
972 			 * Ensure this partial page belongs to this submit chain
973 			 * rather than other concurrent submit chains or
974 			 * noio(bypass) chains since those chains are handled
975 			 * asynchronously thus it cannot be used for inplace I/O
976 			 * or bvpage (should be processed in the strict order.)
977 			 */
978 			tight &= (f->mode >= Z_EROFS_PCLUSTER_FOLLOWED);
979 			excl = false;
980 			if (cur <= pgs) {
981 				excl = (split <= 1) || tight;
982 				cur = pgs;
983 			}
984 
985 			err = z_erofs_attach_page(f, &((struct z_erofs_bvec) {
986 				.page = folio_page(folio, pgs >> PAGE_SHIFT),
987 				.offset = offset + pgs - map->m_la,
988 				.end = end - pgs, }), excl);
989 			if (err)
990 				break;
991 
992 			erofs_onlinefolio_split(folio);
993 			if (f->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))
994 				f->pcl->multibases = true;
995 			if (f->pcl->length < offset + end - map->m_la) {
996 				f->pcl->length = offset + end - map->m_la;
997 				f->pcl->pageofs_out = map->m_la & ~PAGE_MASK;
998 			}
999 			if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&
1000 			    !(map->m_flags & EROFS_MAP_PARTIAL_REF) &&
1001 			    f->pcl->length == map->m_llen)
1002 				f->pcl->partial = false;
1003 		}
1004 		/* shorten the remaining extent to update progress */
1005 		map->m_llen = offset + cur - map->m_la;
1006 		map->m_flags &= ~EROFS_MAP_FULL_MAPPED;
1007 		if (cur <= pgs) {
1008 			split = cur < pgs;
1009 			tight = (bs == PAGE_SIZE);
1010 		}
1011 	} while ((end = cur) > 0);
1012 	erofs_onlinefolio_end(folio, err);
1013 	return err;
1014 }
1015 
1016 static bool z_erofs_is_sync_decompress(struct erofs_sb_info *sbi,
1017 				       unsigned int readahead_pages)
1018 {
1019 	/* auto: enable for read_folio, disable for readahead */
1020 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
1021 	    !readahead_pages)
1022 		return true;
1023 
1024 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
1025 	    (readahead_pages <= sbi->opt.max_sync_decompress_pages))
1026 		return true;
1027 
1028 	return false;
1029 }
1030 
1031 static bool z_erofs_page_is_invalidated(struct page *page)
1032 {
1033 	return !page_folio(page)->mapping && !z_erofs_is_shortlived_page(page);
1034 }
1035 
1036 struct z_erofs_decompress_backend {
1037 	struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];
1038 	struct super_block *sb;
1039 	struct z_erofs_pcluster *pcl;
1040 
1041 	/* pages with the longest decompressed length for deduplication */
1042 	struct page **decompressed_pages;
1043 	/* pages to keep the compressed data */
1044 	struct page **compressed_pages;
1045 
1046 	struct list_head decompressed_secondary_bvecs;
1047 	struct page **pagepool;
1048 	unsigned int onstack_used, nr_pages;
1049 };
1050 
1051 struct z_erofs_bvec_item {
1052 	struct z_erofs_bvec bvec;
1053 	struct list_head list;
1054 };
1055 
1056 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,
1057 					 struct z_erofs_bvec *bvec)
1058 {
1059 	struct z_erofs_bvec_item *item;
1060 	unsigned int pgnr;
1061 
1062 	if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK) &&
1063 	    (bvec->end == PAGE_SIZE ||
1064 	     bvec->offset + bvec->end == be->pcl->length)) {
1065 		pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;
1066 		DBG_BUGON(pgnr >= be->nr_pages);
1067 		if (!be->decompressed_pages[pgnr]) {
1068 			be->decompressed_pages[pgnr] = bvec->page;
1069 			return;
1070 		}
1071 	}
1072 
1073 	/* (cold path) one pcluster is requested multiple times */
1074 	item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);
1075 	item->bvec = *bvec;
1076 	list_add(&item->list, &be->decompressed_secondary_bvecs);
1077 }
1078 
1079 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,
1080 				      int err)
1081 {
1082 	unsigned int off0 = be->pcl->pageofs_out;
1083 	struct list_head *p, *n;
1084 
1085 	list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {
1086 		struct z_erofs_bvec_item *bvi;
1087 		unsigned int end, cur;
1088 		void *dst, *src;
1089 
1090 		bvi = container_of(p, struct z_erofs_bvec_item, list);
1091 		cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;
1092 		end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,
1093 			    bvi->bvec.end);
1094 		dst = kmap_local_page(bvi->bvec.page);
1095 		while (cur < end) {
1096 			unsigned int pgnr, scur, len;
1097 
1098 			pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;
1099 			DBG_BUGON(pgnr >= be->nr_pages);
1100 
1101 			scur = bvi->bvec.offset + cur -
1102 					((pgnr << PAGE_SHIFT) - off0);
1103 			len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);
1104 			if (!be->decompressed_pages[pgnr]) {
1105 				err = -EFSCORRUPTED;
1106 				cur += len;
1107 				continue;
1108 			}
1109 			src = kmap_local_page(be->decompressed_pages[pgnr]);
1110 			memcpy(dst + cur, src + scur, len);
1111 			kunmap_local(src);
1112 			cur += len;
1113 		}
1114 		kunmap_local(dst);
1115 		erofs_onlinefolio_end(page_folio(bvi->bvec.page), err);
1116 		list_del(p);
1117 		kfree(bvi);
1118 	}
1119 }
1120 
1121 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)
1122 {
1123 	struct z_erofs_pcluster *pcl = be->pcl;
1124 	struct z_erofs_bvec_iter biter;
1125 	struct page *old_bvpage;
1126 	int i;
1127 
1128 	z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);
1129 	for (i = 0; i < pcl->vcnt; ++i) {
1130 		struct z_erofs_bvec bvec;
1131 
1132 		z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);
1133 
1134 		if (old_bvpage)
1135 			z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1136 
1137 		DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));
1138 		z_erofs_do_decompressed_bvec(be, &bvec);
1139 	}
1140 
1141 	old_bvpage = z_erofs_bvec_iter_end(&biter);
1142 	if (old_bvpage)
1143 		z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1144 }
1145 
1146 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,
1147 				  bool *overlapped)
1148 {
1149 	struct z_erofs_pcluster *pcl = be->pcl;
1150 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1151 	int i, err = 0;
1152 
1153 	*overlapped = false;
1154 	for (i = 0; i < pclusterpages; ++i) {
1155 		struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];
1156 		struct page *page = bvec->page;
1157 
1158 		/* compressed data ought to be valid when decompressing */
1159 		if (IS_ERR(page) || !page) {
1160 			bvec->page = NULL;	/* clear the failure reason */
1161 			err = page ? PTR_ERR(page) : -EIO;
1162 			continue;
1163 		}
1164 		be->compressed_pages[i] = page;
1165 
1166 		if (z_erofs_is_inline_pcluster(pcl) ||
1167 		    erofs_folio_is_managed(EROFS_SB(be->sb), page_folio(page))) {
1168 			if (!PageUptodate(page))
1169 				err = -EIO;
1170 			continue;
1171 		}
1172 
1173 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1174 		if (z_erofs_is_shortlived_page(page))
1175 			continue;
1176 		z_erofs_do_decompressed_bvec(be, bvec);
1177 		*overlapped = true;
1178 	}
1179 	return err;
1180 }
1181 
1182 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,
1183 				       int err)
1184 {
1185 	struct erofs_sb_info *const sbi = EROFS_SB(be->sb);
1186 	struct z_erofs_pcluster *pcl = be->pcl;
1187 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1188 	const struct z_erofs_decompressor *decomp =
1189 				z_erofs_decomp[pcl->algorithmformat];
1190 	int i, j, jtop, err2;
1191 	struct page *page;
1192 	bool overlapped;
1193 
1194 	mutex_lock(&pcl->lock);
1195 	be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;
1196 
1197 	/* allocate (de)compressed page arrays if cannot be kept on stack */
1198 	be->decompressed_pages = NULL;
1199 	be->compressed_pages = NULL;
1200 	be->onstack_used = 0;
1201 	if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {
1202 		be->decompressed_pages = be->onstack_pages;
1203 		be->onstack_used = be->nr_pages;
1204 		memset(be->decompressed_pages, 0,
1205 		       sizeof(struct page *) * be->nr_pages);
1206 	}
1207 
1208 	if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)
1209 		be->compressed_pages = be->onstack_pages + be->onstack_used;
1210 
1211 	if (!be->decompressed_pages)
1212 		be->decompressed_pages =
1213 			kvcalloc(be->nr_pages, sizeof(struct page *),
1214 				 GFP_KERNEL | __GFP_NOFAIL);
1215 	if (!be->compressed_pages)
1216 		be->compressed_pages =
1217 			kvcalloc(pclusterpages, sizeof(struct page *),
1218 				 GFP_KERNEL | __GFP_NOFAIL);
1219 
1220 	z_erofs_parse_out_bvecs(be);
1221 	err2 = z_erofs_parse_in_bvecs(be, &overlapped);
1222 	if (err2)
1223 		err = err2;
1224 	if (!err)
1225 		err = decomp->decompress(&(struct z_erofs_decompress_req) {
1226 					.sb = be->sb,
1227 					.in = be->compressed_pages,
1228 					.out = be->decompressed_pages,
1229 					.pageofs_in = pcl->pageofs_in,
1230 					.pageofs_out = pcl->pageofs_out,
1231 					.inputsize = pcl->pclustersize,
1232 					.outputsize = pcl->length,
1233 					.alg = pcl->algorithmformat,
1234 					.inplace_io = overlapped,
1235 					.partial_decoding = pcl->partial,
1236 					.fillgaps = pcl->multibases,
1237 					.gfp = pcl->besteffort ? GFP_KERNEL :
1238 						GFP_NOWAIT | __GFP_NORETRY
1239 				 }, be->pagepool);
1240 
1241 	/* must handle all compressed pages before actual file pages */
1242 	if (z_erofs_is_inline_pcluster(pcl)) {
1243 		page = pcl->compressed_bvecs[0].page;
1244 		WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);
1245 		put_page(page);
1246 	} else {
1247 		/* managed folios are still left in compressed_bvecs[] */
1248 		for (i = 0; i < pclusterpages; ++i) {
1249 			page = be->compressed_pages[i];
1250 			if (!page ||
1251 			    erofs_folio_is_managed(sbi, page_folio(page)))
1252 				continue;
1253 			(void)z_erofs_put_shortlivedpage(be->pagepool, page);
1254 			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
1255 		}
1256 	}
1257 	if (be->compressed_pages < be->onstack_pages ||
1258 	    be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)
1259 		kvfree(be->compressed_pages);
1260 
1261 	jtop = 0;
1262 	z_erofs_fill_other_copies(be, err);
1263 	for (i = 0; i < be->nr_pages; ++i) {
1264 		page = be->decompressed_pages[i];
1265 		if (!page)
1266 			continue;
1267 
1268 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1269 		if (!z_erofs_is_shortlived_page(page)) {
1270 			erofs_onlinefolio_end(page_folio(page), err);
1271 			continue;
1272 		}
1273 		if (pcl->algorithmformat != Z_EROFS_COMPRESSION_LZ4) {
1274 			erofs_pagepool_add(be->pagepool, page);
1275 			continue;
1276 		}
1277 		for (j = 0; j < jtop && be->decompressed_pages[j] != page; ++j)
1278 			;
1279 		if (j >= jtop)	/* this bounce page is newly detected */
1280 			be->decompressed_pages[jtop++] = page;
1281 	}
1282 	while (jtop)
1283 		erofs_pagepool_add(be->pagepool,
1284 				   be->decompressed_pages[--jtop]);
1285 	if (be->decompressed_pages != be->onstack_pages)
1286 		kvfree(be->decompressed_pages);
1287 
1288 	pcl->length = 0;
1289 	pcl->partial = true;
1290 	pcl->multibases = false;
1291 	pcl->besteffort = false;
1292 	pcl->bvset.nextpage = NULL;
1293 	pcl->vcnt = 0;
1294 
1295 	/* pcluster lock MUST be taken before the following line */
1296 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1297 	mutex_unlock(&pcl->lock);
1298 	return err;
1299 }
1300 
1301 static int z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1302 				    struct page **pagepool)
1303 {
1304 	struct z_erofs_decompress_backend be = {
1305 		.sb = io->sb,
1306 		.pagepool = pagepool,
1307 		.decompressed_secondary_bvecs =
1308 			LIST_HEAD_INIT(be.decompressed_secondary_bvecs),
1309 	};
1310 	z_erofs_next_pcluster_t owned = io->head;
1311 	int err = io->eio ? -EIO : 0;
1312 
1313 	while (owned != Z_EROFS_PCLUSTER_TAIL) {
1314 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1315 
1316 		be.pcl = container_of(owned, struct z_erofs_pcluster, next);
1317 		owned = READ_ONCE(be.pcl->next);
1318 
1319 		err = z_erofs_decompress_pcluster(&be, err) ?: err;
1320 		if (z_erofs_is_inline_pcluster(be.pcl))
1321 			z_erofs_free_pcluster(be.pcl);
1322 		else
1323 			erofs_workgroup_put(&be.pcl->obj);
1324 	}
1325 	return err;
1326 }
1327 
1328 static void z_erofs_decompressqueue_work(struct work_struct *work)
1329 {
1330 	struct z_erofs_decompressqueue *bgq =
1331 		container_of(work, struct z_erofs_decompressqueue, u.work);
1332 	struct page *pagepool = NULL;
1333 
1334 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL);
1335 	z_erofs_decompress_queue(bgq, &pagepool);
1336 	erofs_release_pages(&pagepool);
1337 	kvfree(bgq);
1338 }
1339 
1340 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1341 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
1342 {
1343 	z_erofs_decompressqueue_work((struct work_struct *)work);
1344 }
1345 #endif
1346 
1347 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1348 				       int bios)
1349 {
1350 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1351 
1352 	/* wake up the caller thread for sync decompression */
1353 	if (io->sync) {
1354 		if (!atomic_add_return(bios, &io->pending_bios))
1355 			complete(&io->u.done);
1356 		return;
1357 	}
1358 
1359 	if (atomic_add_return(bios, &io->pending_bios))
1360 		return;
1361 	/* Use (kthread_)work and sync decompression for atomic contexts only */
1362 	if (!in_task() || irqs_disabled() || rcu_read_lock_any_held()) {
1363 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1364 		struct kthread_worker *worker;
1365 
1366 		rcu_read_lock();
1367 		worker = rcu_dereference(
1368 				z_erofs_pcpu_workers[raw_smp_processor_id()]);
1369 		if (!worker) {
1370 			INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);
1371 			queue_work(z_erofs_workqueue, &io->u.work);
1372 		} else {
1373 			kthread_queue_work(worker, &io->u.kthread_work);
1374 		}
1375 		rcu_read_unlock();
1376 #else
1377 		queue_work(z_erofs_workqueue, &io->u.work);
1378 #endif
1379 		/* enable sync decompression for readahead */
1380 		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1381 			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1382 		return;
1383 	}
1384 	z_erofs_decompressqueue_work(&io->u.work);
1385 }
1386 
1387 static void z_erofs_fill_bio_vec(struct bio_vec *bvec,
1388 				 struct z_erofs_decompress_frontend *f,
1389 				 struct z_erofs_pcluster *pcl,
1390 				 unsigned int nr,
1391 				 struct address_space *mc)
1392 {
1393 	gfp_t gfp = mapping_gfp_mask(mc);
1394 	bool tocache = false;
1395 	struct z_erofs_bvec zbv;
1396 	struct address_space *mapping;
1397 	struct folio *folio;
1398 	struct page *page;
1399 	int bs = i_blocksize(f->inode);
1400 
1401 	/* Except for inplace folios, the entire folio can be used for I/Os */
1402 	bvec->bv_offset = 0;
1403 	bvec->bv_len = PAGE_SIZE;
1404 repeat:
1405 	spin_lock(&pcl->obj.lockref.lock);
1406 	zbv = pcl->compressed_bvecs[nr];
1407 	spin_unlock(&pcl->obj.lockref.lock);
1408 	if (!zbv.page)
1409 		goto out_allocfolio;
1410 
1411 	bvec->bv_page = zbv.page;
1412 	DBG_BUGON(z_erofs_is_shortlived_page(bvec->bv_page));
1413 
1414 	folio = page_folio(zbv.page);
1415 	/*
1416 	 * Handle preallocated cached folios.  We tried to allocate such folios
1417 	 * without triggering direct reclaim.  If allocation failed, inplace
1418 	 * file-backed folios will be used instead.
1419 	 */
1420 	if (folio->private == (void *)Z_EROFS_PREALLOCATED_PAGE) {
1421 		tocache = true;
1422 		goto out_tocache;
1423 	}
1424 
1425 	mapping = READ_ONCE(folio->mapping);
1426 	/*
1427 	 * File-backed folios for inplace I/Os are all locked steady,
1428 	 * therefore it is impossible for `mapping` to be NULL.
1429 	 */
1430 	if (mapping && mapping != mc) {
1431 		if (zbv.offset < 0)
1432 			bvec->bv_offset = round_up(-zbv.offset, bs);
1433 		bvec->bv_len = round_up(zbv.end, bs) - bvec->bv_offset;
1434 		return;
1435 	}
1436 
1437 	folio_lock(folio);
1438 	if (likely(folio->mapping == mc)) {
1439 		/*
1440 		 * The cached folio is still in managed cache but without
1441 		 * a valid `->private` pcluster hint.  Let's reconnect them.
1442 		 */
1443 		if (!folio_test_private(folio)) {
1444 			folio_attach_private(folio, pcl);
1445 			/* compressed_bvecs[] already takes a ref before */
1446 			folio_put(folio);
1447 		}
1448 		if (likely(folio->private == pcl))  {
1449 			/* don't submit cache I/Os again if already uptodate */
1450 			if (folio_test_uptodate(folio)) {
1451 				folio_unlock(folio);
1452 				bvec->bv_page = NULL;
1453 			}
1454 			return;
1455 		}
1456 		/*
1457 		 * Already linked with another pcluster, which only appears in
1458 		 * crafted images by fuzzers for now.  But handle this anyway.
1459 		 */
1460 		tocache = false;	/* use temporary short-lived pages */
1461 	} else {
1462 		DBG_BUGON(1); /* referenced managed folios can't be truncated */
1463 		tocache = true;
1464 	}
1465 	folio_unlock(folio);
1466 	folio_put(folio);
1467 out_allocfolio:
1468 	page = __erofs_allocpage(&f->pagepool, gfp, true);
1469 	spin_lock(&pcl->obj.lockref.lock);
1470 	if (unlikely(pcl->compressed_bvecs[nr].page != zbv.page)) {
1471 		if (page)
1472 			erofs_pagepool_add(&f->pagepool, page);
1473 		spin_unlock(&pcl->obj.lockref.lock);
1474 		cond_resched();
1475 		goto repeat;
1476 	}
1477 	pcl->compressed_bvecs[nr].page = page ? page : ERR_PTR(-ENOMEM);
1478 	spin_unlock(&pcl->obj.lockref.lock);
1479 	bvec->bv_page = page;
1480 	if (!page)
1481 		return;
1482 	folio = page_folio(page);
1483 out_tocache:
1484 	if (!tocache || bs != PAGE_SIZE ||
1485 	    filemap_add_folio(mc, folio, pcl->obj.index + nr, gfp)) {
1486 		/* turn into a temporary shortlived folio (1 ref) */
1487 		folio->private = (void *)Z_EROFS_SHORTLIVED_PAGE;
1488 		return;
1489 	}
1490 	folio_attach_private(folio, pcl);
1491 	/* drop a refcount added by allocpage (then 2 refs in total here) */
1492 	folio_put(folio);
1493 }
1494 
1495 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb,
1496 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1497 {
1498 	struct z_erofs_decompressqueue *q;
1499 
1500 	if (fg && !*fg) {
1501 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1502 		if (!q) {
1503 			*fg = true;
1504 			goto fg_out;
1505 		}
1506 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1507 		kthread_init_work(&q->u.kthread_work,
1508 				  z_erofs_decompressqueue_kthread_work);
1509 #else
1510 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1511 #endif
1512 	} else {
1513 fg_out:
1514 		q = fgq;
1515 		init_completion(&fgq->u.done);
1516 		atomic_set(&fgq->pending_bios, 0);
1517 		q->eio = false;
1518 		q->sync = true;
1519 	}
1520 	q->sb = sb;
1521 	q->head = Z_EROFS_PCLUSTER_TAIL;
1522 	return q;
1523 }
1524 
1525 /* define decompression jobqueue types */
1526 enum {
1527 	JQ_BYPASS,
1528 	JQ_SUBMIT,
1529 	NR_JOBQUEUES,
1530 };
1531 
1532 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1533 				    z_erofs_next_pcluster_t qtail[],
1534 				    z_erofs_next_pcluster_t owned_head)
1535 {
1536 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1537 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1538 
1539 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL);
1540 
1541 	WRITE_ONCE(*submit_qtail, owned_head);
1542 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1543 
1544 	qtail[JQ_BYPASS] = &pcl->next;
1545 }
1546 
1547 static void z_erofs_endio(struct bio *bio)
1548 {
1549 	struct z_erofs_decompressqueue *q = bio->bi_private;
1550 	blk_status_t err = bio->bi_status;
1551 	struct folio_iter fi;
1552 
1553 	bio_for_each_folio_all(fi, bio) {
1554 		struct folio *folio = fi.folio;
1555 
1556 		DBG_BUGON(folio_test_uptodate(folio));
1557 		DBG_BUGON(z_erofs_page_is_invalidated(&folio->page));
1558 		if (!erofs_folio_is_managed(EROFS_SB(q->sb), folio))
1559 			continue;
1560 
1561 		if (!err)
1562 			folio_mark_uptodate(folio);
1563 		folio_unlock(folio);
1564 	}
1565 	if (err)
1566 		q->eio = true;
1567 	z_erofs_decompress_kickoff(q, -1);
1568 	if (bio->bi_bdev)
1569 		bio_put(bio);
1570 }
1571 
1572 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
1573 				 struct z_erofs_decompressqueue *fgq,
1574 				 bool *force_fg, bool readahead)
1575 {
1576 	struct super_block *sb = f->inode->i_sb;
1577 	struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));
1578 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1579 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1580 	z_erofs_next_pcluster_t owned_head = f->owned_head;
1581 	/* bio is NULL initially, so no need to initialize last_{index,bdev} */
1582 	erofs_off_t last_pa;
1583 	unsigned int nr_bios = 0;
1584 	struct bio *bio = NULL;
1585 	unsigned long pflags;
1586 	int memstall = 0;
1587 
1588 	/* No need to read from device for pclusters in the bypass queue. */
1589 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1590 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg);
1591 
1592 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1593 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1594 
1595 	/* by default, all need io submission */
1596 	q[JQ_SUBMIT]->head = owned_head;
1597 
1598 	do {
1599 		struct erofs_map_dev mdev;
1600 		struct z_erofs_pcluster *pcl;
1601 		erofs_off_t cur, end;
1602 		struct bio_vec bvec;
1603 		unsigned int i = 0;
1604 		bool bypass = true;
1605 
1606 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1607 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1608 		owned_head = READ_ONCE(pcl->next);
1609 
1610 		if (z_erofs_is_inline_pcluster(pcl)) {
1611 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1612 			continue;
1613 		}
1614 
1615 		/* no device id here, thus it will always succeed */
1616 		mdev = (struct erofs_map_dev) {
1617 			.m_pa = erofs_pos(sb, pcl->obj.index),
1618 		};
1619 		(void)erofs_map_dev(sb, &mdev);
1620 
1621 		cur = mdev.m_pa;
1622 		end = cur + pcl->pclustersize;
1623 		do {
1624 			bvec.bv_page = NULL;
1625 			if (bio && (cur != last_pa ||
1626 				    bio->bi_bdev != mdev.m_bdev)) {
1627 drain_io:
1628 				if (erofs_is_fileio_mode(EROFS_SB(sb)))
1629 					erofs_fileio_submit_bio(bio);
1630 				else if (erofs_is_fscache_mode(sb))
1631 					erofs_fscache_submit_bio(bio);
1632 				else
1633 					submit_bio(bio);
1634 
1635 				if (memstall) {
1636 					psi_memstall_leave(&pflags);
1637 					memstall = 0;
1638 				}
1639 				bio = NULL;
1640 			}
1641 
1642 			if (!bvec.bv_page) {
1643 				z_erofs_fill_bio_vec(&bvec, f, pcl, i++, mc);
1644 				if (!bvec.bv_page)
1645 					continue;
1646 				if (cur + bvec.bv_len > end)
1647 					bvec.bv_len = end - cur;
1648 				DBG_BUGON(bvec.bv_len < sb->s_blocksize);
1649 			}
1650 
1651 			if (unlikely(PageWorkingset(bvec.bv_page)) &&
1652 			    !memstall) {
1653 				psi_memstall_enter(&pflags);
1654 				memstall = 1;
1655 			}
1656 
1657 			if (!bio) {
1658 				if (erofs_is_fileio_mode(EROFS_SB(sb)))
1659 					bio = erofs_fileio_bio_alloc(&mdev);
1660 				else if (erofs_is_fscache_mode(sb))
1661 					bio = erofs_fscache_bio_alloc(&mdev);
1662 				else
1663 					bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1664 							REQ_OP_READ, GFP_NOIO);
1665 				bio->bi_end_io = z_erofs_endio;
1666 				bio->bi_iter.bi_sector = cur >> 9;
1667 				bio->bi_private = q[JQ_SUBMIT];
1668 				if (readahead)
1669 					bio->bi_opf |= REQ_RAHEAD;
1670 				++nr_bios;
1671 			}
1672 
1673 			if (!bio_add_page(bio, bvec.bv_page, bvec.bv_len,
1674 					  bvec.bv_offset))
1675 				goto drain_io;
1676 			last_pa = cur + bvec.bv_len;
1677 			bypass = false;
1678 		} while ((cur += bvec.bv_len) < end);
1679 
1680 		if (!bypass)
1681 			qtail[JQ_SUBMIT] = &pcl->next;
1682 		else
1683 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1684 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1685 
1686 	if (bio) {
1687 		if (erofs_is_fileio_mode(EROFS_SB(sb)))
1688 			erofs_fileio_submit_bio(bio);
1689 		else if (erofs_is_fscache_mode(sb))
1690 			erofs_fscache_submit_bio(bio);
1691 		else
1692 			submit_bio(bio);
1693 		if (memstall)
1694 			psi_memstall_leave(&pflags);
1695 	}
1696 
1697 	/*
1698 	 * although background is preferred, no one is pending for submission.
1699 	 * don't issue decompression but drop it directly instead.
1700 	 */
1701 	if (!*force_fg && !nr_bios) {
1702 		kvfree(q[JQ_SUBMIT]);
1703 		return;
1704 	}
1705 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios);
1706 }
1707 
1708 static int z_erofs_runqueue(struct z_erofs_decompress_frontend *f,
1709 			    unsigned int ra_folios)
1710 {
1711 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1712 	struct erofs_sb_info *sbi = EROFS_I_SB(f->inode);
1713 	bool force_fg = z_erofs_is_sync_decompress(sbi, ra_folios);
1714 	int err;
1715 
1716 	if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1717 		return 0;
1718 	z_erofs_submit_queue(f, io, &force_fg, !!ra_folios);
1719 
1720 	/* handle bypass queue (no i/o pclusters) immediately */
1721 	err = z_erofs_decompress_queue(&io[JQ_BYPASS], &f->pagepool);
1722 	if (!force_fg)
1723 		return err;
1724 
1725 	/* wait until all bios are completed */
1726 	wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1727 
1728 	/* handle synchronous decompress queue in the caller context */
1729 	return z_erofs_decompress_queue(&io[JQ_SUBMIT], &f->pagepool) ?: err;
1730 }
1731 
1732 /*
1733  * Since partial uptodate is still unimplemented for now, we have to use
1734  * approximate readmore strategies as a start.
1735  */
1736 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1737 		struct readahead_control *rac, bool backmost)
1738 {
1739 	struct inode *inode = f->inode;
1740 	struct erofs_map_blocks *map = &f->map;
1741 	erofs_off_t cur, end, headoffset = f->headoffset;
1742 	int err;
1743 
1744 	if (backmost) {
1745 		if (rac)
1746 			end = headoffset + readahead_length(rac) - 1;
1747 		else
1748 			end = headoffset + PAGE_SIZE - 1;
1749 		map->m_la = end;
1750 		err = z_erofs_map_blocks_iter(inode, map,
1751 					      EROFS_GET_BLOCKS_READMORE);
1752 		if (err)
1753 			return;
1754 
1755 		/* expand ra for the trailing edge if readahead */
1756 		if (rac) {
1757 			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1758 			readahead_expand(rac, headoffset, cur - headoffset);
1759 			return;
1760 		}
1761 		end = round_up(end, PAGE_SIZE);
1762 	} else {
1763 		end = round_up(map->m_la, PAGE_SIZE);
1764 		if (!map->m_llen)
1765 			return;
1766 	}
1767 
1768 	cur = map->m_la + map->m_llen - 1;
1769 	while ((cur >= end) && (cur < i_size_read(inode))) {
1770 		pgoff_t index = cur >> PAGE_SHIFT;
1771 		struct folio *folio;
1772 
1773 		folio = erofs_grab_folio_nowait(inode->i_mapping, index);
1774 		if (!IS_ERR_OR_NULL(folio)) {
1775 			if (folio_test_uptodate(folio))
1776 				folio_unlock(folio);
1777 			else
1778 				z_erofs_scan_folio(f, folio, !!rac);
1779 			folio_put(folio);
1780 		}
1781 
1782 		if (cur < PAGE_SIZE)
1783 			break;
1784 		cur = (index << PAGE_SHIFT) - 1;
1785 	}
1786 }
1787 
1788 static int z_erofs_read_folio(struct file *file, struct folio *folio)
1789 {
1790 	struct inode *const inode = folio->mapping->host;
1791 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1792 	int err;
1793 
1794 	trace_erofs_read_folio(folio, false);
1795 	f.headoffset = (erofs_off_t)folio->index << PAGE_SHIFT;
1796 
1797 	z_erofs_pcluster_readmore(&f, NULL, true);
1798 	err = z_erofs_scan_folio(&f, folio, false);
1799 	z_erofs_pcluster_readmore(&f, NULL, false);
1800 	z_erofs_pcluster_end(&f);
1801 
1802 	/* if some pclusters are ready, need submit them anyway */
1803 	err = z_erofs_runqueue(&f, 0) ?: err;
1804 	if (err && err != -EINTR)
1805 		erofs_err(inode->i_sb, "read error %d @ %lu of nid %llu",
1806 			  err, folio->index, EROFS_I(inode)->nid);
1807 
1808 	erofs_put_metabuf(&f.map.buf);
1809 	erofs_release_pages(&f.pagepool);
1810 	return err;
1811 }
1812 
1813 static void z_erofs_readahead(struct readahead_control *rac)
1814 {
1815 	struct inode *const inode = rac->mapping->host;
1816 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1817 	struct folio *head = NULL, *folio;
1818 	unsigned int nr_folios;
1819 	int err;
1820 
1821 	f.headoffset = readahead_pos(rac);
1822 
1823 	z_erofs_pcluster_readmore(&f, rac, true);
1824 	nr_folios = readahead_count(rac);
1825 	trace_erofs_readpages(inode, readahead_index(rac), nr_folios, false);
1826 
1827 	while ((folio = readahead_folio(rac))) {
1828 		folio->private = head;
1829 		head = folio;
1830 	}
1831 
1832 	/* traverse in reverse order for best metadata I/O performance */
1833 	while (head) {
1834 		folio = head;
1835 		head = folio_get_private(folio);
1836 
1837 		err = z_erofs_scan_folio(&f, folio, true);
1838 		if (err && err != -EINTR)
1839 			erofs_err(inode->i_sb, "readahead error at folio %lu @ nid %llu",
1840 				  folio->index, EROFS_I(inode)->nid);
1841 	}
1842 	z_erofs_pcluster_readmore(&f, rac, false);
1843 	z_erofs_pcluster_end(&f);
1844 
1845 	(void)z_erofs_runqueue(&f, nr_folios);
1846 	erofs_put_metabuf(&f.map.buf);
1847 	erofs_release_pages(&f.pagepool);
1848 }
1849 
1850 const struct address_space_operations z_erofs_aops = {
1851 	.read_folio = z_erofs_read_folio,
1852 	.readahead = z_erofs_readahead,
1853 };
1854