xref: /linux/fs/erofs/zdata.c (revision 6e17c6de3ddf3073741d9c91a796ee696914d8a0)
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: page offset of start position of decompression */
60 	unsigned short pageofs_out;
61 
62 	/* I: page offset of inline compressed data */
63 	unsigned short pageofs_in;
64 
65 	union {
66 		/* L: inline a certain number of bvec for bootstrap */
67 		struct z_erofs_bvset_inline bvset;
68 
69 		/* I: can be used to free the pcluster by RCU. */
70 		struct rcu_head rcu;
71 	};
72 
73 	union {
74 		/* I: physical cluster size in pages */
75 		unsigned short pclusterpages;
76 
77 		/* I: tailpacking inline compressed size */
78 		unsigned short tailpacking_size;
79 	};
80 
81 	/* I: compression algorithm format */
82 	unsigned char algorithmformat;
83 
84 	/* L: whether partial decompression or not */
85 	bool partial;
86 
87 	/* L: indicate several pageofs_outs or not */
88 	bool multibases;
89 
90 	/* A: compressed bvecs (can be cached or inplaced pages) */
91 	struct z_erofs_bvec compressed_bvecs[];
92 };
93 
94 /* the end of a chain of pclusters */
95 #define Z_EROFS_PCLUSTER_TAIL           ((void *) 0x700 + POISON_POINTER_DELTA)
96 #define Z_EROFS_PCLUSTER_NIL            (NULL)
97 
98 struct z_erofs_decompressqueue {
99 	struct super_block *sb;
100 	atomic_t pending_bios;
101 	z_erofs_next_pcluster_t head;
102 
103 	union {
104 		struct completion done;
105 		struct work_struct work;
106 		struct kthread_work kthread_work;
107 	} u;
108 	bool eio, sync;
109 };
110 
111 static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl)
112 {
113 	return !pcl->obj.index;
114 }
115 
116 static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl)
117 {
118 	if (z_erofs_is_inline_pcluster(pcl))
119 		return 1;
120 	return pcl->pclusterpages;
121 }
122 
123 /*
124  * bit 30: I/O error occurred on this page
125  * bit 0 - 29: remaining parts to complete this page
126  */
127 #define Z_EROFS_PAGE_EIO			(1 << 30)
128 
129 static inline void z_erofs_onlinepage_init(struct page *page)
130 {
131 	union {
132 		atomic_t o;
133 		unsigned long v;
134 	} u = { .o = ATOMIC_INIT(1) };
135 
136 	set_page_private(page, u.v);
137 	smp_wmb();
138 	SetPagePrivate(page);
139 }
140 
141 static inline void z_erofs_onlinepage_split(struct page *page)
142 {
143 	atomic_inc((atomic_t *)&page->private);
144 }
145 
146 static inline void z_erofs_page_mark_eio(struct page *page)
147 {
148 	int orig;
149 
150 	do {
151 		orig = atomic_read((atomic_t *)&page->private);
152 	} while (atomic_cmpxchg((atomic_t *)&page->private, orig,
153 				orig | Z_EROFS_PAGE_EIO) != orig);
154 }
155 
156 static inline void z_erofs_onlinepage_endio(struct page *page)
157 {
158 	unsigned int v;
159 
160 	DBG_BUGON(!PagePrivate(page));
161 	v = atomic_dec_return((atomic_t *)&page->private);
162 	if (!(v & ~Z_EROFS_PAGE_EIO)) {
163 		set_page_private(page, 0);
164 		ClearPagePrivate(page);
165 		if (!(v & Z_EROFS_PAGE_EIO))
166 			SetPageUptodate(page);
167 		unlock_page(page);
168 	}
169 }
170 
171 #define Z_EROFS_ONSTACK_PAGES		32
172 
173 /*
174  * since pclustersize is variable for big pcluster feature, introduce slab
175  * pools implementation for different pcluster sizes.
176  */
177 struct z_erofs_pcluster_slab {
178 	struct kmem_cache *slab;
179 	unsigned int maxpages;
180 	char name[48];
181 };
182 
183 #define _PCLP(n) { .maxpages = n }
184 
185 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
186 	_PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
187 	_PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
188 };
189 
190 struct z_erofs_bvec_iter {
191 	struct page *bvpage;
192 	struct z_erofs_bvset *bvset;
193 	unsigned int nr, cur;
194 };
195 
196 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter)
197 {
198 	if (iter->bvpage)
199 		kunmap_local(iter->bvset);
200 	return iter->bvpage;
201 }
202 
203 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter)
204 {
205 	unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec;
206 	/* have to access nextpage in advance, otherwise it will be unmapped */
207 	struct page *nextpage = iter->bvset->nextpage;
208 	struct page *oldpage;
209 
210 	DBG_BUGON(!nextpage);
211 	oldpage = z_erofs_bvec_iter_end(iter);
212 	iter->bvpage = nextpage;
213 	iter->bvset = kmap_local_page(nextpage);
214 	iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec);
215 	iter->cur = 0;
216 	return oldpage;
217 }
218 
219 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter,
220 				    struct z_erofs_bvset_inline *bvset,
221 				    unsigned int bootstrap_nr,
222 				    unsigned int cur)
223 {
224 	*iter = (struct z_erofs_bvec_iter) {
225 		.nr = bootstrap_nr,
226 		.bvset = (struct z_erofs_bvset *)bvset,
227 	};
228 
229 	while (cur > iter->nr) {
230 		cur -= iter->nr;
231 		z_erofs_bvset_flip(iter);
232 	}
233 	iter->cur = cur;
234 }
235 
236 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter,
237 				struct z_erofs_bvec *bvec,
238 				struct page **candidate_bvpage,
239 				struct page **pagepool)
240 {
241 	if (iter->cur >= iter->nr) {
242 		struct page *nextpage = *candidate_bvpage;
243 
244 		if (!nextpage) {
245 			nextpage = erofs_allocpage(pagepool, GFP_NOFS);
246 			if (!nextpage)
247 				return -ENOMEM;
248 			set_page_private(nextpage, Z_EROFS_SHORTLIVED_PAGE);
249 		}
250 		DBG_BUGON(iter->bvset->nextpage);
251 		iter->bvset->nextpage = nextpage;
252 		z_erofs_bvset_flip(iter);
253 
254 		iter->bvset->nextpage = NULL;
255 		*candidate_bvpage = NULL;
256 	}
257 	iter->bvset->bvec[iter->cur++] = *bvec;
258 	return 0;
259 }
260 
261 static void z_erofs_bvec_dequeue(struct z_erofs_bvec_iter *iter,
262 				 struct z_erofs_bvec *bvec,
263 				 struct page **old_bvpage)
264 {
265 	if (iter->cur == iter->nr)
266 		*old_bvpage = z_erofs_bvset_flip(iter);
267 	else
268 		*old_bvpage = NULL;
269 	*bvec = iter->bvset->bvec[iter->cur++];
270 }
271 
272 static void z_erofs_destroy_pcluster_pool(void)
273 {
274 	int i;
275 
276 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
277 		if (!pcluster_pool[i].slab)
278 			continue;
279 		kmem_cache_destroy(pcluster_pool[i].slab);
280 		pcluster_pool[i].slab = NULL;
281 	}
282 }
283 
284 static int z_erofs_create_pcluster_pool(void)
285 {
286 	struct z_erofs_pcluster_slab *pcs;
287 	struct z_erofs_pcluster *a;
288 	unsigned int size;
289 
290 	for (pcs = pcluster_pool;
291 	     pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
292 		size = struct_size(a, compressed_bvecs, pcs->maxpages);
293 
294 		sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
295 		pcs->slab = kmem_cache_create(pcs->name, size, 0,
296 					      SLAB_RECLAIM_ACCOUNT, NULL);
297 		if (pcs->slab)
298 			continue;
299 
300 		z_erofs_destroy_pcluster_pool();
301 		return -ENOMEM;
302 	}
303 	return 0;
304 }
305 
306 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
307 {
308 	int i;
309 
310 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
311 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
312 		struct z_erofs_pcluster *pcl;
313 
314 		if (nrpages > pcs->maxpages)
315 			continue;
316 
317 		pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
318 		if (!pcl)
319 			return ERR_PTR(-ENOMEM);
320 		pcl->pclusterpages = nrpages;
321 		return pcl;
322 	}
323 	return ERR_PTR(-EINVAL);
324 }
325 
326 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
327 {
328 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
329 	int i;
330 
331 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
332 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
333 
334 		if (pclusterpages > pcs->maxpages)
335 			continue;
336 
337 		kmem_cache_free(pcs->slab, pcl);
338 		return;
339 	}
340 	DBG_BUGON(1);
341 }
342 
343 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
344 
345 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
346 static struct kthread_worker __rcu **z_erofs_pcpu_workers;
347 
348 static void erofs_destroy_percpu_workers(void)
349 {
350 	struct kthread_worker *worker;
351 	unsigned int cpu;
352 
353 	for_each_possible_cpu(cpu) {
354 		worker = rcu_dereference_protected(
355 					z_erofs_pcpu_workers[cpu], 1);
356 		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
357 		if (worker)
358 			kthread_destroy_worker(worker);
359 	}
360 	kfree(z_erofs_pcpu_workers);
361 }
362 
363 static struct kthread_worker *erofs_init_percpu_worker(int cpu)
364 {
365 	struct kthread_worker *worker =
366 		kthread_create_worker_on_cpu(cpu, 0, "erofs_worker/%u", cpu);
367 
368 	if (IS_ERR(worker))
369 		return worker;
370 	if (IS_ENABLED(CONFIG_EROFS_FS_PCPU_KTHREAD_HIPRI))
371 		sched_set_fifo_low(worker->task);
372 	return worker;
373 }
374 
375 static int erofs_init_percpu_workers(void)
376 {
377 	struct kthread_worker *worker;
378 	unsigned int cpu;
379 
380 	z_erofs_pcpu_workers = kcalloc(num_possible_cpus(),
381 			sizeof(struct kthread_worker *), GFP_ATOMIC);
382 	if (!z_erofs_pcpu_workers)
383 		return -ENOMEM;
384 
385 	for_each_online_cpu(cpu) {	/* could miss cpu{off,on}line? */
386 		worker = erofs_init_percpu_worker(cpu);
387 		if (!IS_ERR(worker))
388 			rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
389 	}
390 	return 0;
391 }
392 #else
393 static inline void erofs_destroy_percpu_workers(void) {}
394 static inline int erofs_init_percpu_workers(void) { return 0; }
395 #endif
396 
397 #if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_EROFS_FS_PCPU_KTHREAD)
398 static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock);
399 static enum cpuhp_state erofs_cpuhp_state;
400 
401 static int erofs_cpu_online(unsigned int cpu)
402 {
403 	struct kthread_worker *worker, *old;
404 
405 	worker = erofs_init_percpu_worker(cpu);
406 	if (IS_ERR(worker))
407 		return PTR_ERR(worker);
408 
409 	spin_lock(&z_erofs_pcpu_worker_lock);
410 	old = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
411 			lockdep_is_held(&z_erofs_pcpu_worker_lock));
412 	if (!old)
413 		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
414 	spin_unlock(&z_erofs_pcpu_worker_lock);
415 	if (old)
416 		kthread_destroy_worker(worker);
417 	return 0;
418 }
419 
420 static int erofs_cpu_offline(unsigned int cpu)
421 {
422 	struct kthread_worker *worker;
423 
424 	spin_lock(&z_erofs_pcpu_worker_lock);
425 	worker = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
426 			lockdep_is_held(&z_erofs_pcpu_worker_lock));
427 	rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
428 	spin_unlock(&z_erofs_pcpu_worker_lock);
429 
430 	synchronize_rcu();
431 	if (worker)
432 		kthread_destroy_worker(worker);
433 	return 0;
434 }
435 
436 static int erofs_cpu_hotplug_init(void)
437 {
438 	int state;
439 
440 	state = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
441 			"fs/erofs:online", erofs_cpu_online, erofs_cpu_offline);
442 	if (state < 0)
443 		return state;
444 
445 	erofs_cpuhp_state = state;
446 	return 0;
447 }
448 
449 static void erofs_cpu_hotplug_destroy(void)
450 {
451 	if (erofs_cpuhp_state)
452 		cpuhp_remove_state_nocalls(erofs_cpuhp_state);
453 }
454 #else /* !CONFIG_HOTPLUG_CPU || !CONFIG_EROFS_FS_PCPU_KTHREAD */
455 static inline int erofs_cpu_hotplug_init(void) { return 0; }
456 static inline void erofs_cpu_hotplug_destroy(void) {}
457 #endif
458 
459 void z_erofs_exit_zip_subsystem(void)
460 {
461 	erofs_cpu_hotplug_destroy();
462 	erofs_destroy_percpu_workers();
463 	destroy_workqueue(z_erofs_workqueue);
464 	z_erofs_destroy_pcluster_pool();
465 }
466 
467 int __init z_erofs_init_zip_subsystem(void)
468 {
469 	int err = z_erofs_create_pcluster_pool();
470 
471 	if (err)
472 		goto out_error_pcluster_pool;
473 
474 	z_erofs_workqueue = alloc_workqueue("erofs_worker",
475 			WQ_UNBOUND | WQ_HIGHPRI, num_possible_cpus());
476 	if (!z_erofs_workqueue) {
477 		err = -ENOMEM;
478 		goto out_error_workqueue_init;
479 	}
480 
481 	err = erofs_init_percpu_workers();
482 	if (err)
483 		goto out_error_pcpu_worker;
484 
485 	err = erofs_cpu_hotplug_init();
486 	if (err < 0)
487 		goto out_error_cpuhp_init;
488 	return err;
489 
490 out_error_cpuhp_init:
491 	erofs_destroy_percpu_workers();
492 out_error_pcpu_worker:
493 	destroy_workqueue(z_erofs_workqueue);
494 out_error_workqueue_init:
495 	z_erofs_destroy_pcluster_pool();
496 out_error_pcluster_pool:
497 	return err;
498 }
499 
500 enum z_erofs_pclustermode {
501 	Z_EROFS_PCLUSTER_INFLIGHT,
502 	/*
503 	 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it
504 	 * could be dispatched into bypass queue later due to uptodated managed
505 	 * pages. All related online pages cannot be reused for inplace I/O (or
506 	 * bvpage) since it can be directly decoded without I/O submission.
507 	 */
508 	Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE,
509 	/*
510 	 * The current collection has been linked with the owned chain, and
511 	 * could also be linked with the remaining collections, which means
512 	 * if the processing page is the tail page of the collection, thus
513 	 * the current collection can safely use the whole page (since
514 	 * the previous collection is under control) for in-place I/O, as
515 	 * illustrated below:
516 	 *  ________________________________________________________________
517 	 * |  tail (partial) page |          head (partial) page           |
518 	 * |  (of the current cl) |      (of the previous collection)      |
519 	 * |                      |                                        |
520 	 * |__PCLUSTER_FOLLOWED___|___________PCLUSTER_FOLLOWED____________|
521 	 *
522 	 * [  (*) the above page can be used as inplace I/O.               ]
523 	 */
524 	Z_EROFS_PCLUSTER_FOLLOWED,
525 };
526 
527 struct z_erofs_decompress_frontend {
528 	struct inode *const inode;
529 	struct erofs_map_blocks map;
530 	struct z_erofs_bvec_iter biter;
531 
532 	struct page *pagepool;
533 	struct page *candidate_bvpage;
534 	struct z_erofs_pcluster *pcl;
535 	z_erofs_next_pcluster_t owned_head;
536 	enum z_erofs_pclustermode mode;
537 
538 	/* used for applying cache strategy on the fly */
539 	bool backmost;
540 	erofs_off_t headoffset;
541 
542 	/* a pointer used to pick up inplace I/O pages */
543 	unsigned int icur;
544 };
545 
546 #define DECOMPRESS_FRONTEND_INIT(__i) { \
547 	.inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \
548 	.mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true }
549 
550 static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe)
551 {
552 	unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy;
553 
554 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
555 		return false;
556 
557 	if (fe->backmost)
558 		return true;
559 
560 	if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
561 	    fe->map.m_la < fe->headoffset)
562 		return true;
563 
564 	return false;
565 }
566 
567 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe)
568 {
569 	struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));
570 	struct z_erofs_pcluster *pcl = fe->pcl;
571 	bool shouldalloc = z_erofs_should_alloc_cache(fe);
572 	bool standalone = true;
573 	/*
574 	 * optimistic allocation without direct reclaim since inplace I/O
575 	 * can be used if low memory otherwise.
576 	 */
577 	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
578 			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
579 	unsigned int i;
580 
581 	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED)
582 		return;
583 
584 	for (i = 0; i < pcl->pclusterpages; ++i) {
585 		struct page *page;
586 		void *t;	/* mark pages just found for debugging */
587 		struct page *newpage = NULL;
588 
589 		/* the compressed page was loaded before */
590 		if (READ_ONCE(pcl->compressed_bvecs[i].page))
591 			continue;
592 
593 		page = find_get_page(mc, pcl->obj.index + i);
594 
595 		if (page) {
596 			t = (void *)((unsigned long)page | 1);
597 		} else {
598 			/* I/O is needed, no possible to decompress directly */
599 			standalone = false;
600 			if (!shouldalloc)
601 				continue;
602 
603 			/*
604 			 * try to use cached I/O if page allocation
605 			 * succeeds or fallback to in-place I/O instead
606 			 * to avoid any direct reclaim.
607 			 */
608 			newpage = erofs_allocpage(&fe->pagepool, gfp);
609 			if (!newpage)
610 				continue;
611 			set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE);
612 			t = (void *)((unsigned long)newpage | 1);
613 		}
614 
615 		if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL, t))
616 			continue;
617 
618 		if (page)
619 			put_page(page);
620 		else if (newpage)
621 			erofs_pagepool_add(&fe->pagepool, newpage);
622 	}
623 
624 	/*
625 	 * don't do inplace I/O if all compressed pages are available in
626 	 * managed cache since it can be moved to the bypass queue instead.
627 	 */
628 	if (standalone)
629 		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
630 }
631 
632 /* called by erofs_shrinker to get rid of all compressed_pages */
633 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
634 				       struct erofs_workgroup *grp)
635 {
636 	struct z_erofs_pcluster *const pcl =
637 		container_of(grp, struct z_erofs_pcluster, obj);
638 	int i;
639 
640 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
641 	/*
642 	 * refcount of workgroup is now freezed as 0,
643 	 * therefore no need to worry about available decompression users.
644 	 */
645 	for (i = 0; i < pcl->pclusterpages; ++i) {
646 		struct page *page = pcl->compressed_bvecs[i].page;
647 
648 		if (!page)
649 			continue;
650 
651 		/* block other users from reclaiming or migrating the page */
652 		if (!trylock_page(page))
653 			return -EBUSY;
654 
655 		if (!erofs_page_is_managed(sbi, page))
656 			continue;
657 
658 		/* barrier is implied in the following 'unlock_page' */
659 		WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
660 		detach_page_private(page);
661 		unlock_page(page);
662 	}
663 	return 0;
664 }
665 
666 static bool z_erofs_cache_release_folio(struct folio *folio, gfp_t gfp)
667 {
668 	struct z_erofs_pcluster *pcl = folio_get_private(folio);
669 	bool ret;
670 	int i;
671 
672 	if (!folio_test_private(folio))
673 		return true;
674 
675 	ret = false;
676 	spin_lock(&pcl->obj.lockref.lock);
677 	if (pcl->obj.lockref.count > 0)
678 		goto out;
679 
680 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
681 	for (i = 0; i < pcl->pclusterpages; ++i) {
682 		if (pcl->compressed_bvecs[i].page == &folio->page) {
683 			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
684 			ret = true;
685 			break;
686 		}
687 	}
688 	if (ret)
689 		folio_detach_private(folio);
690 out:
691 	spin_unlock(&pcl->obj.lockref.lock);
692 	return ret;
693 }
694 
695 /*
696  * It will be called only on inode eviction. In case that there are still some
697  * decompression requests in progress, wait with rescheduling for a bit here.
698  * An extra lock could be introduced instead but it seems unnecessary.
699  */
700 static void z_erofs_cache_invalidate_folio(struct folio *folio,
701 					   size_t offset, size_t length)
702 {
703 	const size_t stop = length + offset;
704 
705 	/* Check for potential overflow in debug mode */
706 	DBG_BUGON(stop > folio_size(folio) || stop < length);
707 
708 	if (offset == 0 && stop == folio_size(folio))
709 		while (!z_erofs_cache_release_folio(folio, GFP_NOFS))
710 			cond_resched();
711 }
712 
713 static const struct address_space_operations z_erofs_cache_aops = {
714 	.release_folio = z_erofs_cache_release_folio,
715 	.invalidate_folio = z_erofs_cache_invalidate_folio,
716 };
717 
718 int erofs_init_managed_cache(struct super_block *sb)
719 {
720 	struct inode *const inode = new_inode(sb);
721 
722 	if (!inode)
723 		return -ENOMEM;
724 
725 	set_nlink(inode, 1);
726 	inode->i_size = OFFSET_MAX;
727 	inode->i_mapping->a_ops = &z_erofs_cache_aops;
728 	mapping_set_gfp_mask(inode->i_mapping, GFP_NOFS);
729 	EROFS_SB(sb)->managed_cache = inode;
730 	return 0;
731 }
732 
733 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
734 				   struct z_erofs_bvec *bvec)
735 {
736 	struct z_erofs_pcluster *const pcl = fe->pcl;
737 
738 	while (fe->icur > 0) {
739 		if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
740 			     NULL, bvec->page)) {
741 			pcl->compressed_bvecs[fe->icur] = *bvec;
742 			return true;
743 		}
744 	}
745 	return false;
746 }
747 
748 /* callers must be with pcluster lock held */
749 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,
750 			       struct z_erofs_bvec *bvec, bool exclusive)
751 {
752 	int ret;
753 
754 	if (exclusive) {
755 		/* give priority for inplaceio to use file pages first */
756 		if (z_erofs_try_inplace_io(fe, bvec))
757 			return 0;
758 		/* otherwise, check if it can be used as a bvpage */
759 		if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED &&
760 		    !fe->candidate_bvpage)
761 			fe->candidate_bvpage = bvec->page;
762 	}
763 	ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage,
764 				   &fe->pagepool);
765 	fe->pcl->vcnt += (ret >= 0);
766 	return ret;
767 }
768 
769 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
770 {
771 	struct z_erofs_pcluster *pcl = f->pcl;
772 	z_erofs_next_pcluster_t *owned_head = &f->owned_head;
773 
774 	/* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
775 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
776 		    *owned_head) == Z_EROFS_PCLUSTER_NIL) {
777 		*owned_head = &pcl->next;
778 		/* so we can attach this pcluster to our submission chain. */
779 		f->mode = Z_EROFS_PCLUSTER_FOLLOWED;
780 		return;
781 	}
782 
783 	/* type 2, it belongs to an ongoing chain */
784 	f->mode = Z_EROFS_PCLUSTER_INFLIGHT;
785 }
786 
787 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)
788 {
789 	struct erofs_map_blocks *map = &fe->map;
790 	bool ztailpacking = map->m_flags & EROFS_MAP_META;
791 	struct z_erofs_pcluster *pcl;
792 	struct erofs_workgroup *grp;
793 	int err;
794 
795 	if (!(map->m_flags & EROFS_MAP_ENCODED) ||
796 	    (!ztailpacking && !(map->m_pa >> PAGE_SHIFT))) {
797 		DBG_BUGON(1);
798 		return -EFSCORRUPTED;
799 	}
800 
801 	/* no available pcluster, let's allocate one */
802 	pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
803 				     map->m_plen >> PAGE_SHIFT);
804 	if (IS_ERR(pcl))
805 		return PTR_ERR(pcl);
806 
807 	spin_lock_init(&pcl->obj.lockref.lock);
808 	pcl->algorithmformat = map->m_algorithmformat;
809 	pcl->length = 0;
810 	pcl->partial = true;
811 
812 	/* new pclusters should be claimed as type 1, primary and followed */
813 	pcl->next = fe->owned_head;
814 	pcl->pageofs_out = map->m_la & ~PAGE_MASK;
815 	fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;
816 
817 	/*
818 	 * lock all primary followed works before visible to others
819 	 * and mutex_trylock *never* fails for a new pcluster.
820 	 */
821 	mutex_init(&pcl->lock);
822 	DBG_BUGON(!mutex_trylock(&pcl->lock));
823 
824 	if (ztailpacking) {
825 		pcl->obj.index = 0;	/* which indicates ztailpacking */
826 		pcl->pageofs_in = erofs_blkoff(fe->inode->i_sb, map->m_pa);
827 		pcl->tailpacking_size = map->m_plen;
828 	} else {
829 		pcl->obj.index = map->m_pa >> PAGE_SHIFT;
830 
831 		grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);
832 		if (IS_ERR(grp)) {
833 			err = PTR_ERR(grp);
834 			goto err_out;
835 		}
836 
837 		if (grp != &pcl->obj) {
838 			fe->pcl = container_of(grp,
839 					struct z_erofs_pcluster, obj);
840 			err = -EEXIST;
841 			goto err_out;
842 		}
843 	}
844 	fe->owned_head = &pcl->next;
845 	fe->pcl = pcl;
846 	return 0;
847 
848 err_out:
849 	mutex_unlock(&pcl->lock);
850 	z_erofs_free_pcluster(pcl);
851 	return err;
852 }
853 
854 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
855 {
856 	struct erofs_map_blocks *map = &fe->map;
857 	struct erofs_workgroup *grp = NULL;
858 	int ret;
859 
860 	DBG_BUGON(fe->pcl);
861 
862 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */
863 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
864 
865 	if (!(map->m_flags & EROFS_MAP_META)) {
866 		grp = erofs_find_workgroup(fe->inode->i_sb,
867 					   map->m_pa >> PAGE_SHIFT);
868 	} else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
869 		DBG_BUGON(1);
870 		return -EFSCORRUPTED;
871 	}
872 
873 	if (grp) {
874 		fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
875 		ret = -EEXIST;
876 	} else {
877 		ret = z_erofs_register_pcluster(fe);
878 	}
879 
880 	if (ret == -EEXIST) {
881 		mutex_lock(&fe->pcl->lock);
882 		z_erofs_try_to_claim_pcluster(fe);
883 	} else if (ret) {
884 		return ret;
885 	}
886 	z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
887 				Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
888 	/* since file-backed online pages are traversed in reverse order */
889 	fe->icur = z_erofs_pclusterpages(fe->pcl);
890 	return 0;
891 }
892 
893 /*
894  * keep in mind that no referenced pclusters will be freed
895  * only after a RCU grace period.
896  */
897 static void z_erofs_rcu_callback(struct rcu_head *head)
898 {
899 	z_erofs_free_pcluster(container_of(head,
900 			struct z_erofs_pcluster, rcu));
901 }
902 
903 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
904 {
905 	struct z_erofs_pcluster *const pcl =
906 		container_of(grp, struct z_erofs_pcluster, obj);
907 
908 	call_rcu(&pcl->rcu, z_erofs_rcu_callback);
909 }
910 
911 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe)
912 {
913 	struct z_erofs_pcluster *pcl = fe->pcl;
914 
915 	if (!pcl)
916 		return false;
917 
918 	z_erofs_bvec_iter_end(&fe->biter);
919 	mutex_unlock(&pcl->lock);
920 
921 	if (fe->candidate_bvpage)
922 		fe->candidate_bvpage = NULL;
923 
924 	/*
925 	 * if all pending pages are added, don't hold its reference
926 	 * any longer if the pcluster isn't hosted by ourselves.
927 	 */
928 	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE)
929 		erofs_workgroup_put(&pcl->obj);
930 
931 	fe->pcl = NULL;
932 	return true;
933 }
934 
935 static int z_erofs_read_fragment(struct inode *inode, erofs_off_t pos,
936 				 struct page *page, unsigned int pageofs,
937 				 unsigned int len)
938 {
939 	struct super_block *sb = inode->i_sb;
940 	struct inode *packed_inode = EROFS_I_SB(inode)->packed_inode;
941 	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
942 	u8 *src, *dst;
943 	unsigned int i, cnt;
944 
945 	if (!packed_inode)
946 		return -EFSCORRUPTED;
947 
948 	buf.inode = packed_inode;
949 	pos += EROFS_I(inode)->z_fragmentoff;
950 	for (i = 0; i < len; i += cnt) {
951 		cnt = min_t(unsigned int, len - i,
952 			    sb->s_blocksize - erofs_blkoff(sb, pos));
953 		src = erofs_bread(&buf, erofs_blknr(sb, pos), EROFS_KMAP);
954 		if (IS_ERR(src)) {
955 			erofs_put_metabuf(&buf);
956 			return PTR_ERR(src);
957 		}
958 
959 		dst = kmap_local_page(page);
960 		memcpy(dst + pageofs + i, src + erofs_blkoff(sb, pos), cnt);
961 		kunmap_local(dst);
962 		pos += cnt;
963 	}
964 	erofs_put_metabuf(&buf);
965 	return 0;
966 }
967 
968 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
969 				struct page *page)
970 {
971 	struct inode *const inode = fe->inode;
972 	struct erofs_map_blocks *const map = &fe->map;
973 	const loff_t offset = page_offset(page);
974 	bool tight = true, exclusive;
975 	unsigned int cur, end, spiltted;
976 	int err = 0;
977 
978 	/* register locked file pages as online pages in pack */
979 	z_erofs_onlinepage_init(page);
980 
981 	spiltted = 0;
982 	end = PAGE_SIZE;
983 repeat:
984 	cur = end - 1;
985 
986 	if (offset + cur < map->m_la ||
987 	    offset + cur >= map->m_la + map->m_llen) {
988 		if (z_erofs_collector_end(fe))
989 			fe->backmost = false;
990 		map->m_la = offset + cur;
991 		map->m_llen = 0;
992 		err = z_erofs_map_blocks_iter(inode, map, 0);
993 		if (err)
994 			goto out;
995 	} else {
996 		if (fe->pcl)
997 			goto hitted;
998 		/* didn't get a valid pcluster previously (very rare) */
999 	}
1000 
1001 	if (!(map->m_flags & EROFS_MAP_MAPPED) ||
1002 	    map->m_flags & EROFS_MAP_FRAGMENT)
1003 		goto hitted;
1004 
1005 	err = z_erofs_collector_begin(fe);
1006 	if (err)
1007 		goto out;
1008 
1009 	if (z_erofs_is_inline_pcluster(fe->pcl)) {
1010 		void *mp;
1011 
1012 		mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
1013 					erofs_blknr(inode->i_sb, map->m_pa),
1014 					EROFS_NO_KMAP);
1015 		if (IS_ERR(mp)) {
1016 			err = PTR_ERR(mp);
1017 			erofs_err(inode->i_sb,
1018 				  "failed to get inline page, err %d", err);
1019 			goto out;
1020 		}
1021 		get_page(fe->map.buf.page);
1022 		WRITE_ONCE(fe->pcl->compressed_bvecs[0].page,
1023 			   fe->map.buf.page);
1024 		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
1025 	} else {
1026 		/* bind cache first when cached decompression is preferred */
1027 		z_erofs_bind_cache(fe);
1028 	}
1029 hitted:
1030 	/*
1031 	 * Ensure the current partial page belongs to this submit chain rather
1032 	 * than other concurrent submit chains or the noio(bypass) chain since
1033 	 * those chains are handled asynchronously thus the page cannot be used
1034 	 * for inplace I/O or bvpage (should be processed in a strict order.)
1035 	 */
1036 	tight &= (fe->mode > Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE);
1037 
1038 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
1039 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
1040 		zero_user_segment(page, cur, end);
1041 		goto next_part;
1042 	}
1043 	if (map->m_flags & EROFS_MAP_FRAGMENT) {
1044 		unsigned int pageofs, skip, len;
1045 
1046 		if (offset > map->m_la) {
1047 			pageofs = 0;
1048 			skip = offset - map->m_la;
1049 		} else {
1050 			pageofs = map->m_la & ~PAGE_MASK;
1051 			skip = 0;
1052 		}
1053 		len = min_t(unsigned int, map->m_llen - skip, end - cur);
1054 		err = z_erofs_read_fragment(inode, skip, page, pageofs, len);
1055 		if (err)
1056 			goto out;
1057 		++spiltted;
1058 		tight = false;
1059 		goto next_part;
1060 	}
1061 
1062 	exclusive = (!cur && (!spiltted || tight));
1063 	if (cur)
1064 		tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED);
1065 
1066 	err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) {
1067 					.page = page,
1068 					.offset = offset - map->m_la,
1069 					.end = end,
1070 				  }), exclusive);
1071 	if (err)
1072 		goto out;
1073 
1074 	z_erofs_onlinepage_split(page);
1075 	/* bump up the number of spiltted parts of a page */
1076 	++spiltted;
1077 	if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))
1078 		fe->pcl->multibases = true;
1079 	if (fe->pcl->length < offset + end - map->m_la) {
1080 		fe->pcl->length = offset + end - map->m_la;
1081 		fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK;
1082 	}
1083 	if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&
1084 	    !(map->m_flags & EROFS_MAP_PARTIAL_REF) &&
1085 	    fe->pcl->length == map->m_llen)
1086 		fe->pcl->partial = false;
1087 next_part:
1088 	/* shorten the remaining extent to update progress */
1089 	map->m_llen = offset + cur - map->m_la;
1090 	map->m_flags &= ~EROFS_MAP_FULL_MAPPED;
1091 
1092 	end = cur;
1093 	if (end > 0)
1094 		goto repeat;
1095 
1096 out:
1097 	if (err)
1098 		z_erofs_page_mark_eio(page);
1099 	z_erofs_onlinepage_endio(page);
1100 	return err;
1101 }
1102 
1103 static bool z_erofs_is_sync_decompress(struct erofs_sb_info *sbi,
1104 				       unsigned int readahead_pages)
1105 {
1106 	/* auto: enable for read_folio, disable for readahead */
1107 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
1108 	    !readahead_pages)
1109 		return true;
1110 
1111 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
1112 	    (readahead_pages <= sbi->opt.max_sync_decompress_pages))
1113 		return true;
1114 
1115 	return false;
1116 }
1117 
1118 static bool z_erofs_page_is_invalidated(struct page *page)
1119 {
1120 	return !page->mapping && !z_erofs_is_shortlived_page(page);
1121 }
1122 
1123 struct z_erofs_decompress_backend {
1124 	struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];
1125 	struct super_block *sb;
1126 	struct z_erofs_pcluster *pcl;
1127 
1128 	/* pages with the longest decompressed length for deduplication */
1129 	struct page **decompressed_pages;
1130 	/* pages to keep the compressed data */
1131 	struct page **compressed_pages;
1132 
1133 	struct list_head decompressed_secondary_bvecs;
1134 	struct page **pagepool;
1135 	unsigned int onstack_used, nr_pages;
1136 };
1137 
1138 struct z_erofs_bvec_item {
1139 	struct z_erofs_bvec bvec;
1140 	struct list_head list;
1141 };
1142 
1143 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,
1144 					 struct z_erofs_bvec *bvec)
1145 {
1146 	struct z_erofs_bvec_item *item;
1147 
1148 	if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK)) {
1149 		unsigned int pgnr;
1150 
1151 		pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;
1152 		DBG_BUGON(pgnr >= be->nr_pages);
1153 		if (!be->decompressed_pages[pgnr]) {
1154 			be->decompressed_pages[pgnr] = bvec->page;
1155 			return;
1156 		}
1157 	}
1158 
1159 	/* (cold path) one pcluster is requested multiple times */
1160 	item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);
1161 	item->bvec = *bvec;
1162 	list_add(&item->list, &be->decompressed_secondary_bvecs);
1163 }
1164 
1165 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,
1166 				      int err)
1167 {
1168 	unsigned int off0 = be->pcl->pageofs_out;
1169 	struct list_head *p, *n;
1170 
1171 	list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {
1172 		struct z_erofs_bvec_item *bvi;
1173 		unsigned int end, cur;
1174 		void *dst, *src;
1175 
1176 		bvi = container_of(p, struct z_erofs_bvec_item, list);
1177 		cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;
1178 		end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,
1179 			    bvi->bvec.end);
1180 		dst = kmap_local_page(bvi->bvec.page);
1181 		while (cur < end) {
1182 			unsigned int pgnr, scur, len;
1183 
1184 			pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;
1185 			DBG_BUGON(pgnr >= be->nr_pages);
1186 
1187 			scur = bvi->bvec.offset + cur -
1188 					((pgnr << PAGE_SHIFT) - off0);
1189 			len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);
1190 			if (!be->decompressed_pages[pgnr]) {
1191 				err = -EFSCORRUPTED;
1192 				cur += len;
1193 				continue;
1194 			}
1195 			src = kmap_local_page(be->decompressed_pages[pgnr]);
1196 			memcpy(dst + cur, src + scur, len);
1197 			kunmap_local(src);
1198 			cur += len;
1199 		}
1200 		kunmap_local(dst);
1201 		if (err)
1202 			z_erofs_page_mark_eio(bvi->bvec.page);
1203 		z_erofs_onlinepage_endio(bvi->bvec.page);
1204 		list_del(p);
1205 		kfree(bvi);
1206 	}
1207 }
1208 
1209 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)
1210 {
1211 	struct z_erofs_pcluster *pcl = be->pcl;
1212 	struct z_erofs_bvec_iter biter;
1213 	struct page *old_bvpage;
1214 	int i;
1215 
1216 	z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);
1217 	for (i = 0; i < pcl->vcnt; ++i) {
1218 		struct z_erofs_bvec bvec;
1219 
1220 		z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);
1221 
1222 		if (old_bvpage)
1223 			z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1224 
1225 		DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));
1226 		z_erofs_do_decompressed_bvec(be, &bvec);
1227 	}
1228 
1229 	old_bvpage = z_erofs_bvec_iter_end(&biter);
1230 	if (old_bvpage)
1231 		z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1232 }
1233 
1234 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,
1235 				  bool *overlapped)
1236 {
1237 	struct z_erofs_pcluster *pcl = be->pcl;
1238 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1239 	int i, err = 0;
1240 
1241 	*overlapped = false;
1242 	for (i = 0; i < pclusterpages; ++i) {
1243 		struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];
1244 		struct page *page = bvec->page;
1245 
1246 		/* compressed pages ought to be present before decompressing */
1247 		if (!page) {
1248 			DBG_BUGON(1);
1249 			continue;
1250 		}
1251 		be->compressed_pages[i] = page;
1252 
1253 		if (z_erofs_is_inline_pcluster(pcl)) {
1254 			if (!PageUptodate(page))
1255 				err = -EIO;
1256 			continue;
1257 		}
1258 
1259 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1260 		if (!z_erofs_is_shortlived_page(page)) {
1261 			if (erofs_page_is_managed(EROFS_SB(be->sb), page)) {
1262 				if (!PageUptodate(page))
1263 					err = -EIO;
1264 				continue;
1265 			}
1266 			z_erofs_do_decompressed_bvec(be, bvec);
1267 			*overlapped = true;
1268 		}
1269 	}
1270 
1271 	if (err)
1272 		return err;
1273 	return 0;
1274 }
1275 
1276 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,
1277 				       int err)
1278 {
1279 	struct erofs_sb_info *const sbi = EROFS_SB(be->sb);
1280 	struct z_erofs_pcluster *pcl = be->pcl;
1281 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1282 	const struct z_erofs_decompressor *decompressor =
1283 				&erofs_decompressors[pcl->algorithmformat];
1284 	unsigned int i, inputsize;
1285 	int err2;
1286 	struct page *page;
1287 	bool overlapped;
1288 
1289 	mutex_lock(&pcl->lock);
1290 	be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;
1291 
1292 	/* allocate (de)compressed page arrays if cannot be kept on stack */
1293 	be->decompressed_pages = NULL;
1294 	be->compressed_pages = NULL;
1295 	be->onstack_used = 0;
1296 	if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {
1297 		be->decompressed_pages = be->onstack_pages;
1298 		be->onstack_used = be->nr_pages;
1299 		memset(be->decompressed_pages, 0,
1300 		       sizeof(struct page *) * be->nr_pages);
1301 	}
1302 
1303 	if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)
1304 		be->compressed_pages = be->onstack_pages + be->onstack_used;
1305 
1306 	if (!be->decompressed_pages)
1307 		be->decompressed_pages =
1308 			kvcalloc(be->nr_pages, sizeof(struct page *),
1309 				 GFP_KERNEL | __GFP_NOFAIL);
1310 	if (!be->compressed_pages)
1311 		be->compressed_pages =
1312 			kvcalloc(pclusterpages, sizeof(struct page *),
1313 				 GFP_KERNEL | __GFP_NOFAIL);
1314 
1315 	z_erofs_parse_out_bvecs(be);
1316 	err2 = z_erofs_parse_in_bvecs(be, &overlapped);
1317 	if (err2)
1318 		err = err2;
1319 	if (err)
1320 		goto out;
1321 
1322 	if (z_erofs_is_inline_pcluster(pcl))
1323 		inputsize = pcl->tailpacking_size;
1324 	else
1325 		inputsize = pclusterpages * PAGE_SIZE;
1326 
1327 	err = decompressor->decompress(&(struct z_erofs_decompress_req) {
1328 					.sb = be->sb,
1329 					.in = be->compressed_pages,
1330 					.out = be->decompressed_pages,
1331 					.pageofs_in = pcl->pageofs_in,
1332 					.pageofs_out = pcl->pageofs_out,
1333 					.inputsize = inputsize,
1334 					.outputsize = pcl->length,
1335 					.alg = pcl->algorithmformat,
1336 					.inplace_io = overlapped,
1337 					.partial_decoding = pcl->partial,
1338 					.fillgaps = pcl->multibases,
1339 				 }, be->pagepool);
1340 
1341 out:
1342 	/* must handle all compressed pages before actual file pages */
1343 	if (z_erofs_is_inline_pcluster(pcl)) {
1344 		page = pcl->compressed_bvecs[0].page;
1345 		WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);
1346 		put_page(page);
1347 	} else {
1348 		for (i = 0; i < pclusterpages; ++i) {
1349 			page = pcl->compressed_bvecs[i].page;
1350 
1351 			if (erofs_page_is_managed(sbi, page))
1352 				continue;
1353 
1354 			/* recycle all individual short-lived pages */
1355 			(void)z_erofs_put_shortlivedpage(be->pagepool, page);
1356 			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
1357 		}
1358 	}
1359 	if (be->compressed_pages < be->onstack_pages ||
1360 	    be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)
1361 		kvfree(be->compressed_pages);
1362 	z_erofs_fill_other_copies(be, err);
1363 
1364 	for (i = 0; i < be->nr_pages; ++i) {
1365 		page = be->decompressed_pages[i];
1366 		if (!page)
1367 			continue;
1368 
1369 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1370 
1371 		/* recycle all individual short-lived pages */
1372 		if (z_erofs_put_shortlivedpage(be->pagepool, page))
1373 			continue;
1374 		if (err)
1375 			z_erofs_page_mark_eio(page);
1376 		z_erofs_onlinepage_endio(page);
1377 	}
1378 
1379 	if (be->decompressed_pages != be->onstack_pages)
1380 		kvfree(be->decompressed_pages);
1381 
1382 	pcl->length = 0;
1383 	pcl->partial = true;
1384 	pcl->multibases = false;
1385 	pcl->bvset.nextpage = NULL;
1386 	pcl->vcnt = 0;
1387 
1388 	/* pcluster lock MUST be taken before the following line */
1389 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1390 	mutex_unlock(&pcl->lock);
1391 	return err;
1392 }
1393 
1394 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1395 				     struct page **pagepool)
1396 {
1397 	struct z_erofs_decompress_backend be = {
1398 		.sb = io->sb,
1399 		.pagepool = pagepool,
1400 		.decompressed_secondary_bvecs =
1401 			LIST_HEAD_INIT(be.decompressed_secondary_bvecs),
1402 	};
1403 	z_erofs_next_pcluster_t owned = io->head;
1404 
1405 	while (owned != Z_EROFS_PCLUSTER_TAIL) {
1406 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1407 
1408 		be.pcl = container_of(owned, struct z_erofs_pcluster, next);
1409 		owned = READ_ONCE(be.pcl->next);
1410 
1411 		z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0);
1412 		erofs_workgroup_put(&be.pcl->obj);
1413 	}
1414 }
1415 
1416 static void z_erofs_decompressqueue_work(struct work_struct *work)
1417 {
1418 	struct z_erofs_decompressqueue *bgq =
1419 		container_of(work, struct z_erofs_decompressqueue, u.work);
1420 	struct page *pagepool = NULL;
1421 
1422 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL);
1423 	z_erofs_decompress_queue(bgq, &pagepool);
1424 	erofs_release_pages(&pagepool);
1425 	kvfree(bgq);
1426 }
1427 
1428 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1429 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
1430 {
1431 	z_erofs_decompressqueue_work((struct work_struct *)work);
1432 }
1433 #endif
1434 
1435 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1436 				       int bios)
1437 {
1438 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1439 
1440 	/* wake up the caller thread for sync decompression */
1441 	if (io->sync) {
1442 		if (!atomic_add_return(bios, &io->pending_bios))
1443 			complete(&io->u.done);
1444 		return;
1445 	}
1446 
1447 	if (atomic_add_return(bios, &io->pending_bios))
1448 		return;
1449 	/* Use (kthread_)work and sync decompression for atomic contexts only */
1450 	if (!in_task() || irqs_disabled() || rcu_read_lock_any_held()) {
1451 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1452 		struct kthread_worker *worker;
1453 
1454 		rcu_read_lock();
1455 		worker = rcu_dereference(
1456 				z_erofs_pcpu_workers[raw_smp_processor_id()]);
1457 		if (!worker) {
1458 			INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);
1459 			queue_work(z_erofs_workqueue, &io->u.work);
1460 		} else {
1461 			kthread_queue_work(worker, &io->u.kthread_work);
1462 		}
1463 		rcu_read_unlock();
1464 #else
1465 		queue_work(z_erofs_workqueue, &io->u.work);
1466 #endif
1467 		/* enable sync decompression for readahead */
1468 		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1469 			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1470 		return;
1471 	}
1472 	z_erofs_decompressqueue_work(&io->u.work);
1473 }
1474 
1475 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1476 					       unsigned int nr,
1477 					       struct page **pagepool,
1478 					       struct address_space *mc)
1479 {
1480 	const pgoff_t index = pcl->obj.index;
1481 	gfp_t gfp = mapping_gfp_mask(mc);
1482 	bool tocache = false;
1483 
1484 	struct address_space *mapping;
1485 	struct page *oldpage, *page;
1486 	int justfound;
1487 
1488 repeat:
1489 	page = READ_ONCE(pcl->compressed_bvecs[nr].page);
1490 	oldpage = page;
1491 
1492 	if (!page)
1493 		goto out_allocpage;
1494 
1495 	justfound = (unsigned long)page & 1UL;
1496 	page = (struct page *)((unsigned long)page & ~1UL);
1497 
1498 	/*
1499 	 * preallocated cached pages, which is used to avoid direct reclaim
1500 	 * otherwise, it will go inplace I/O path instead.
1501 	 */
1502 	if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1503 		WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1504 		set_page_private(page, 0);
1505 		tocache = true;
1506 		goto out_tocache;
1507 	}
1508 	mapping = READ_ONCE(page->mapping);
1509 
1510 	/*
1511 	 * file-backed online pages in plcuster are all locked steady,
1512 	 * therefore it is impossible for `mapping' to be NULL.
1513 	 */
1514 	if (mapping && mapping != mc)
1515 		/* ought to be unmanaged pages */
1516 		goto out;
1517 
1518 	/* directly return for shortlived page as well */
1519 	if (z_erofs_is_shortlived_page(page))
1520 		goto out;
1521 
1522 	lock_page(page);
1523 
1524 	/* only true if page reclaim goes wrong, should never happen */
1525 	DBG_BUGON(justfound && PagePrivate(page));
1526 
1527 	/* the page is still in manage cache */
1528 	if (page->mapping == mc) {
1529 		WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1530 
1531 		if (!PagePrivate(page)) {
1532 			/*
1533 			 * impossible to be !PagePrivate(page) for
1534 			 * the current restriction as well if
1535 			 * the page is already in compressed_bvecs[].
1536 			 */
1537 			DBG_BUGON(!justfound);
1538 
1539 			justfound = 0;
1540 			set_page_private(page, (unsigned long)pcl);
1541 			SetPagePrivate(page);
1542 		}
1543 
1544 		/* no need to submit io if it is already up-to-date */
1545 		if (PageUptodate(page)) {
1546 			unlock_page(page);
1547 			page = NULL;
1548 		}
1549 		goto out;
1550 	}
1551 
1552 	/*
1553 	 * the managed page has been truncated, it's unsafe to
1554 	 * reuse this one, let's allocate a new cache-managed page.
1555 	 */
1556 	DBG_BUGON(page->mapping);
1557 	DBG_BUGON(!justfound);
1558 
1559 	tocache = true;
1560 	unlock_page(page);
1561 	put_page(page);
1562 out_allocpage:
1563 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1564 	if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page,
1565 			       oldpage, page)) {
1566 		erofs_pagepool_add(pagepool, page);
1567 		cond_resched();
1568 		goto repeat;
1569 	}
1570 out_tocache:
1571 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1572 		/* turn into temporary page if fails (1 ref) */
1573 		set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1574 		goto out;
1575 	}
1576 	attach_page_private(page, pcl);
1577 	/* drop a refcount added by allocpage (then we have 2 refs here) */
1578 	put_page(page);
1579 
1580 out:	/* the only exit (for tracing and debugging) */
1581 	return page;
1582 }
1583 
1584 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb,
1585 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1586 {
1587 	struct z_erofs_decompressqueue *q;
1588 
1589 	if (fg && !*fg) {
1590 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1591 		if (!q) {
1592 			*fg = true;
1593 			goto fg_out;
1594 		}
1595 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1596 		kthread_init_work(&q->u.kthread_work,
1597 				  z_erofs_decompressqueue_kthread_work);
1598 #else
1599 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1600 #endif
1601 	} else {
1602 fg_out:
1603 		q = fgq;
1604 		init_completion(&fgq->u.done);
1605 		atomic_set(&fgq->pending_bios, 0);
1606 		q->eio = false;
1607 		q->sync = true;
1608 	}
1609 	q->sb = sb;
1610 	q->head = Z_EROFS_PCLUSTER_TAIL;
1611 	return q;
1612 }
1613 
1614 /* define decompression jobqueue types */
1615 enum {
1616 	JQ_BYPASS,
1617 	JQ_SUBMIT,
1618 	NR_JOBQUEUES,
1619 };
1620 
1621 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1622 				    z_erofs_next_pcluster_t qtail[],
1623 				    z_erofs_next_pcluster_t owned_head)
1624 {
1625 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1626 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1627 
1628 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL);
1629 
1630 	WRITE_ONCE(*submit_qtail, owned_head);
1631 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1632 
1633 	qtail[JQ_BYPASS] = &pcl->next;
1634 }
1635 
1636 static void z_erofs_decompressqueue_endio(struct bio *bio)
1637 {
1638 	struct z_erofs_decompressqueue *q = bio->bi_private;
1639 	blk_status_t err = bio->bi_status;
1640 	struct bio_vec *bvec;
1641 	struct bvec_iter_all iter_all;
1642 
1643 	bio_for_each_segment_all(bvec, bio, iter_all) {
1644 		struct page *page = bvec->bv_page;
1645 
1646 		DBG_BUGON(PageUptodate(page));
1647 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1648 
1649 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1650 			if (!err)
1651 				SetPageUptodate(page);
1652 			unlock_page(page);
1653 		}
1654 	}
1655 	if (err)
1656 		q->eio = true;
1657 	z_erofs_decompress_kickoff(q, -1);
1658 	bio_put(bio);
1659 }
1660 
1661 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
1662 				 struct z_erofs_decompressqueue *fgq,
1663 				 bool *force_fg, bool readahead)
1664 {
1665 	struct super_block *sb = f->inode->i_sb;
1666 	struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));
1667 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1668 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1669 	z_erofs_next_pcluster_t owned_head = f->owned_head;
1670 	/* bio is NULL initially, so no need to initialize last_{index,bdev} */
1671 	pgoff_t last_index;
1672 	struct block_device *last_bdev;
1673 	unsigned int nr_bios = 0;
1674 	struct bio *bio = NULL;
1675 	unsigned long pflags;
1676 	int memstall = 0;
1677 
1678 	/*
1679 	 * if managed cache is enabled, bypass jobqueue is needed,
1680 	 * no need to read from device for all pclusters in this queue.
1681 	 */
1682 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1683 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg);
1684 
1685 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1686 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1687 
1688 	/* by default, all need io submission */
1689 	q[JQ_SUBMIT]->head = owned_head;
1690 
1691 	do {
1692 		struct erofs_map_dev mdev;
1693 		struct z_erofs_pcluster *pcl;
1694 		pgoff_t cur, end;
1695 		unsigned int i = 0;
1696 		bool bypass = true;
1697 
1698 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1699 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1700 		owned_head = READ_ONCE(pcl->next);
1701 
1702 		if (z_erofs_is_inline_pcluster(pcl)) {
1703 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1704 			continue;
1705 		}
1706 
1707 		/* no device id here, thus it will always succeed */
1708 		mdev = (struct erofs_map_dev) {
1709 			.m_pa = erofs_pos(sb, pcl->obj.index),
1710 		};
1711 		(void)erofs_map_dev(sb, &mdev);
1712 
1713 		cur = erofs_blknr(sb, mdev.m_pa);
1714 		end = cur + pcl->pclusterpages;
1715 
1716 		do {
1717 			struct page *page;
1718 
1719 			page = pickup_page_for_submission(pcl, i++,
1720 					&f->pagepool, mc);
1721 			if (!page)
1722 				continue;
1723 
1724 			if (bio && (cur != last_index + 1 ||
1725 				    last_bdev != mdev.m_bdev)) {
1726 submit_bio_retry:
1727 				submit_bio(bio);
1728 				if (memstall) {
1729 					psi_memstall_leave(&pflags);
1730 					memstall = 0;
1731 				}
1732 				bio = NULL;
1733 			}
1734 
1735 			if (unlikely(PageWorkingset(page)) && !memstall) {
1736 				psi_memstall_enter(&pflags);
1737 				memstall = 1;
1738 			}
1739 
1740 			if (!bio) {
1741 				bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1742 						REQ_OP_READ, GFP_NOIO);
1743 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1744 
1745 				last_bdev = mdev.m_bdev;
1746 				bio->bi_iter.bi_sector = (sector_t)cur <<
1747 					(sb->s_blocksize_bits - 9);
1748 				bio->bi_private = q[JQ_SUBMIT];
1749 				if (readahead)
1750 					bio->bi_opf |= REQ_RAHEAD;
1751 				++nr_bios;
1752 			}
1753 
1754 			if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1755 				goto submit_bio_retry;
1756 
1757 			last_index = cur;
1758 			bypass = false;
1759 		} while (++cur < end);
1760 
1761 		if (!bypass)
1762 			qtail[JQ_SUBMIT] = &pcl->next;
1763 		else
1764 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1765 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1766 
1767 	if (bio) {
1768 		submit_bio(bio);
1769 		if (memstall)
1770 			psi_memstall_leave(&pflags);
1771 	}
1772 
1773 	/*
1774 	 * although background is preferred, no one is pending for submission.
1775 	 * don't issue decompression but drop it directly instead.
1776 	 */
1777 	if (!*force_fg && !nr_bios) {
1778 		kvfree(q[JQ_SUBMIT]);
1779 		return;
1780 	}
1781 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios);
1782 }
1783 
1784 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f,
1785 			     bool force_fg, bool ra)
1786 {
1787 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1788 
1789 	if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1790 		return;
1791 	z_erofs_submit_queue(f, io, &force_fg, ra);
1792 
1793 	/* handle bypass queue (no i/o pclusters) immediately */
1794 	z_erofs_decompress_queue(&io[JQ_BYPASS], &f->pagepool);
1795 
1796 	if (!force_fg)
1797 		return;
1798 
1799 	/* wait until all bios are completed */
1800 	wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1801 
1802 	/* handle synchronous decompress queue in the caller context */
1803 	z_erofs_decompress_queue(&io[JQ_SUBMIT], &f->pagepool);
1804 }
1805 
1806 /*
1807  * Since partial uptodate is still unimplemented for now, we have to use
1808  * approximate readmore strategies as a start.
1809  */
1810 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1811 		struct readahead_control *rac, bool backmost)
1812 {
1813 	struct inode *inode = f->inode;
1814 	struct erofs_map_blocks *map = &f->map;
1815 	erofs_off_t cur, end, headoffset = f->headoffset;
1816 	int err;
1817 
1818 	if (backmost) {
1819 		if (rac)
1820 			end = headoffset + readahead_length(rac) - 1;
1821 		else
1822 			end = headoffset + PAGE_SIZE - 1;
1823 		map->m_la = end;
1824 		err = z_erofs_map_blocks_iter(inode, map,
1825 					      EROFS_GET_BLOCKS_READMORE);
1826 		if (err)
1827 			return;
1828 
1829 		/* expand ra for the trailing edge if readahead */
1830 		if (rac) {
1831 			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1832 			readahead_expand(rac, headoffset, cur - headoffset);
1833 			return;
1834 		}
1835 		end = round_up(end, PAGE_SIZE);
1836 	} else {
1837 		end = round_up(map->m_la, PAGE_SIZE);
1838 
1839 		if (!map->m_llen)
1840 			return;
1841 	}
1842 
1843 	cur = map->m_la + map->m_llen - 1;
1844 	while (cur >= end) {
1845 		pgoff_t index = cur >> PAGE_SHIFT;
1846 		struct page *page;
1847 
1848 		page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1849 		if (page) {
1850 			if (PageUptodate(page)) {
1851 				unlock_page(page);
1852 			} else {
1853 				err = z_erofs_do_read_page(f, page);
1854 				if (err)
1855 					erofs_err(inode->i_sb,
1856 						  "readmore error at page %lu @ nid %llu",
1857 						  index, EROFS_I(inode)->nid);
1858 			}
1859 			put_page(page);
1860 		}
1861 
1862 		if (cur < PAGE_SIZE)
1863 			break;
1864 		cur = (index << PAGE_SHIFT) - 1;
1865 	}
1866 }
1867 
1868 static int z_erofs_read_folio(struct file *file, struct folio *folio)
1869 {
1870 	struct page *page = &folio->page;
1871 	struct inode *const inode = page->mapping->host;
1872 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1873 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1874 	int err;
1875 
1876 	trace_erofs_readpage(page, false);
1877 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1878 
1879 	z_erofs_pcluster_readmore(&f, NULL, true);
1880 	err = z_erofs_do_read_page(&f, page);
1881 	z_erofs_pcluster_readmore(&f, NULL, false);
1882 	(void)z_erofs_collector_end(&f);
1883 
1884 	/* if some compressed cluster ready, need submit them anyway */
1885 	z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, 0), false);
1886 
1887 	if (err)
1888 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1889 
1890 	erofs_put_metabuf(&f.map.buf);
1891 	erofs_release_pages(&f.pagepool);
1892 	return err;
1893 }
1894 
1895 static void z_erofs_readahead(struct readahead_control *rac)
1896 {
1897 	struct inode *const inode = rac->mapping->host;
1898 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1899 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1900 	struct page *head = NULL, *page;
1901 	unsigned int nr_pages;
1902 
1903 	f.headoffset = readahead_pos(rac);
1904 
1905 	z_erofs_pcluster_readmore(&f, rac, true);
1906 	nr_pages = readahead_count(rac);
1907 	trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1908 
1909 	while ((page = readahead_page(rac))) {
1910 		set_page_private(page, (unsigned long)head);
1911 		head = page;
1912 	}
1913 
1914 	while (head) {
1915 		struct page *page = head;
1916 		int err;
1917 
1918 		/* traversal in reverse order */
1919 		head = (void *)page_private(page);
1920 
1921 		err = z_erofs_do_read_page(&f, page);
1922 		if (err)
1923 			erofs_err(inode->i_sb,
1924 				  "readahead error at page %lu @ nid %llu",
1925 				  page->index, EROFS_I(inode)->nid);
1926 		put_page(page);
1927 	}
1928 	z_erofs_pcluster_readmore(&f, rac, false);
1929 	(void)z_erofs_collector_end(&f);
1930 
1931 	z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, nr_pages), true);
1932 	erofs_put_metabuf(&f.map.buf);
1933 	erofs_release_pages(&f.pagepool);
1934 }
1935 
1936 const struct address_space_operations z_erofs_aops = {
1937 	.read_folio = z_erofs_read_folio,
1938 	.readahead = z_erofs_readahead,
1939 };
1940