xref: /openbmc/linux/fs/erofs/zdata.c (revision f4356947)
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/prefetch.h>
9 #include <linux/psi.h>
10 #include <linux/cpuhotplug.h>
11 #include <trace/events/erofs.h>
12 
13 #define Z_EROFS_PCLUSTER_MAX_PAGES	(Z_EROFS_PCLUSTER_MAX_SIZE / PAGE_SIZE)
14 #define Z_EROFS_INLINE_BVECS		2
15 
16 /*
17  * let's leave a type here in case of introducing
18  * another tagged pointer later.
19  */
20 typedef void *z_erofs_next_pcluster_t;
21 
22 struct z_erofs_bvec {
23 	struct page *page;
24 	int offset;
25 	unsigned int end;
26 };
27 
28 #define __Z_EROFS_BVSET(name, total) \
29 struct name { \
30 	/* point to the next page which contains the following bvecs */ \
31 	struct page *nextpage; \
32 	struct z_erofs_bvec bvec[total]; \
33 }
34 __Z_EROFS_BVSET(z_erofs_bvset,);
35 __Z_EROFS_BVSET(z_erofs_bvset_inline, Z_EROFS_INLINE_BVECS);
36 
37 /*
38  * Structure fields follow one of the following exclusion rules.
39  *
40  * I: Modifiable by initialization/destruction paths and read-only
41  *    for everyone else;
42  *
43  * L: Field should be protected by the pcluster lock;
44  *
45  * A: Field should be accessed / updated in atomic for parallelized code.
46  */
47 struct z_erofs_pcluster {
48 	struct erofs_workgroup obj;
49 	struct mutex lock;
50 
51 	/* A: point to next chained pcluster or TAILs */
52 	z_erofs_next_pcluster_t next;
53 
54 	/* L: the maximum decompression size of this round */
55 	unsigned int length;
56 
57 	/* L: total number of bvecs */
58 	unsigned int vcnt;
59 
60 	/* I: page offset of start position of decompression */
61 	unsigned short pageofs_out;
62 
63 	/* I: page offset of inline compressed data */
64 	unsigned short pageofs_in;
65 
66 	union {
67 		/* L: inline a certain number of bvec for bootstrap */
68 		struct z_erofs_bvset_inline bvset;
69 
70 		/* I: can be used to free the pcluster by RCU. */
71 		struct rcu_head rcu;
72 	};
73 
74 	union {
75 		/* I: physical cluster size in pages */
76 		unsigned short pclusterpages;
77 
78 		/* I: tailpacking inline compressed size */
79 		unsigned short tailpacking_size;
80 	};
81 
82 	/* I: compression algorithm format */
83 	unsigned char algorithmformat;
84 
85 	/* L: whether partial decompression or not */
86 	bool partial;
87 
88 	/* L: indicate several pageofs_outs or not */
89 	bool multibases;
90 
91 	/* A: compressed bvecs (can be cached or inplaced pages) */
92 	struct z_erofs_bvec compressed_bvecs[];
93 };
94 
95 /* let's avoid the valid 32-bit kernel addresses */
96 
97 /* the chained workgroup has't submitted io (still open) */
98 #define Z_EROFS_PCLUSTER_TAIL           ((void *)0x5F0ECAFE)
99 /* the chained workgroup has already submitted io */
100 #define Z_EROFS_PCLUSTER_TAIL_CLOSED    ((void *)0x5F0EDEAD)
101 
102 #define Z_EROFS_PCLUSTER_NIL            (NULL)
103 
104 struct z_erofs_decompressqueue {
105 	struct super_block *sb;
106 	atomic_t pending_bios;
107 	z_erofs_next_pcluster_t head;
108 
109 	union {
110 		struct completion done;
111 		struct work_struct work;
112 		struct kthread_work kthread_work;
113 	} u;
114 	bool eio, sync;
115 };
116 
117 static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl)
118 {
119 	return !pcl->obj.index;
120 }
121 
122 static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl)
123 {
124 	if (z_erofs_is_inline_pcluster(pcl))
125 		return 1;
126 	return pcl->pclusterpages;
127 }
128 
129 /*
130  * bit 30: I/O error occurred on this page
131  * bit 0 - 29: remaining parts to complete this page
132  */
133 #define Z_EROFS_PAGE_EIO			(1 << 30)
134 
135 static inline void z_erofs_onlinepage_init(struct page *page)
136 {
137 	union {
138 		atomic_t o;
139 		unsigned long v;
140 	} u = { .o = ATOMIC_INIT(1) };
141 
142 	set_page_private(page, u.v);
143 	smp_wmb();
144 	SetPagePrivate(page);
145 }
146 
147 static inline void z_erofs_onlinepage_split(struct page *page)
148 {
149 	atomic_inc((atomic_t *)&page->private);
150 }
151 
152 static inline void z_erofs_page_mark_eio(struct page *page)
153 {
154 	int orig;
155 
156 	do {
157 		orig = atomic_read((atomic_t *)&page->private);
158 	} while (atomic_cmpxchg((atomic_t *)&page->private, orig,
159 				orig | Z_EROFS_PAGE_EIO) != orig);
160 }
161 
162 static inline void z_erofs_onlinepage_endio(struct page *page)
163 {
164 	unsigned int v;
165 
166 	DBG_BUGON(!PagePrivate(page));
167 	v = atomic_dec_return((atomic_t *)&page->private);
168 	if (!(v & ~Z_EROFS_PAGE_EIO)) {
169 		set_page_private(page, 0);
170 		ClearPagePrivate(page);
171 		if (!(v & Z_EROFS_PAGE_EIO))
172 			SetPageUptodate(page);
173 		unlock_page(page);
174 	}
175 }
176 
177 #define Z_EROFS_ONSTACK_PAGES		32
178 
179 /*
180  * since pclustersize is variable for big pcluster feature, introduce slab
181  * pools implementation for different pcluster sizes.
182  */
183 struct z_erofs_pcluster_slab {
184 	struct kmem_cache *slab;
185 	unsigned int maxpages;
186 	char name[48];
187 };
188 
189 #define _PCLP(n) { .maxpages = n }
190 
191 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
192 	_PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
193 	_PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
194 };
195 
196 struct z_erofs_bvec_iter {
197 	struct page *bvpage;
198 	struct z_erofs_bvset *bvset;
199 	unsigned int nr, cur;
200 };
201 
202 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter)
203 {
204 	if (iter->bvpage)
205 		kunmap_local(iter->bvset);
206 	return iter->bvpage;
207 }
208 
209 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter)
210 {
211 	unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec;
212 	/* have to access nextpage in advance, otherwise it will be unmapped */
213 	struct page *nextpage = iter->bvset->nextpage;
214 	struct page *oldpage;
215 
216 	DBG_BUGON(!nextpage);
217 	oldpage = z_erofs_bvec_iter_end(iter);
218 	iter->bvpage = nextpage;
219 	iter->bvset = kmap_local_page(nextpage);
220 	iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec);
221 	iter->cur = 0;
222 	return oldpage;
223 }
224 
225 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter,
226 				    struct z_erofs_bvset_inline *bvset,
227 				    unsigned int bootstrap_nr,
228 				    unsigned int cur)
229 {
230 	*iter = (struct z_erofs_bvec_iter) {
231 		.nr = bootstrap_nr,
232 		.bvset = (struct z_erofs_bvset *)bvset,
233 	};
234 
235 	while (cur > iter->nr) {
236 		cur -= iter->nr;
237 		z_erofs_bvset_flip(iter);
238 	}
239 	iter->cur = cur;
240 }
241 
242 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter,
243 				struct z_erofs_bvec *bvec,
244 				struct page **candidate_bvpage)
245 {
246 	if (iter->cur == iter->nr) {
247 		if (!*candidate_bvpage)
248 			return -EAGAIN;
249 
250 		DBG_BUGON(iter->bvset->nextpage);
251 		iter->bvset->nextpage = *candidate_bvpage;
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 	 * The current pclusters was the tail of an exist chain, in addition
504 	 * that the previous processed chained pclusters are all decided to
505 	 * be hooked up to it.
506 	 * A new chain will be created for the remaining pclusters which are
507 	 * not processed yet, so different from Z_EROFS_PCLUSTER_FOLLOWED,
508 	 * the next pcluster cannot reuse the whole page safely for inplace I/O
509 	 * in the following scenario:
510 	 *  ________________________________________________________________
511 	 * |      tail (partial) page     |       head (partial) page       |
512 	 * |   (belongs to the next pcl)  |   (belongs to the current pcl)  |
513 	 * |_______PCLUSTER_FOLLOWED______|________PCLUSTER_HOOKED__________|
514 	 */
515 	Z_EROFS_PCLUSTER_HOOKED,
516 	/*
517 	 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it
518 	 * could be dispatched into bypass queue later due to uptodated managed
519 	 * pages. All related online pages cannot be reused for inplace I/O (or
520 	 * bvpage) since it can be directly decoded without I/O submission.
521 	 */
522 	Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE,
523 	/*
524 	 * The current collection has been linked with the owned chain, and
525 	 * could also be linked with the remaining collections, which means
526 	 * if the processing page is the tail page of the collection, thus
527 	 * the current collection can safely use the whole page (since
528 	 * the previous collection is under control) for in-place I/O, as
529 	 * illustrated below:
530 	 *  ________________________________________________________________
531 	 * |  tail (partial) page |          head (partial) page           |
532 	 * |  (of the current cl) |      (of the previous collection)      |
533 	 * | PCLUSTER_FOLLOWED or |                                        |
534 	 * |_____PCLUSTER_HOOKED__|___________PCLUSTER_FOLLOWED____________|
535 	 *
536 	 * [  (*) the above page can be used as inplace I/O.               ]
537 	 */
538 	Z_EROFS_PCLUSTER_FOLLOWED,
539 };
540 
541 struct z_erofs_decompress_frontend {
542 	struct inode *const inode;
543 	struct erofs_map_blocks map;
544 	struct z_erofs_bvec_iter biter;
545 
546 	struct page *candidate_bvpage;
547 	struct z_erofs_pcluster *pcl, *tailpcl;
548 	z_erofs_next_pcluster_t owned_head;
549 	enum z_erofs_pclustermode mode;
550 
551 	bool readahead;
552 	/* used for applying cache strategy on the fly */
553 	bool backmost;
554 	erofs_off_t headoffset;
555 
556 	/* a pointer used to pick up inplace I/O pages */
557 	unsigned int icur;
558 };
559 
560 #define DECOMPRESS_FRONTEND_INIT(__i) { \
561 	.inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \
562 	.mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true }
563 
564 static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe)
565 {
566 	unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy;
567 
568 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
569 		return false;
570 
571 	if (fe->backmost)
572 		return true;
573 
574 	if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
575 	    fe->map.m_la < fe->headoffset)
576 		return true;
577 
578 	return false;
579 }
580 
581 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe,
582 			       struct page **pagepool)
583 {
584 	struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));
585 	struct z_erofs_pcluster *pcl = fe->pcl;
586 	bool shouldalloc = z_erofs_should_alloc_cache(fe);
587 	bool standalone = true;
588 	/*
589 	 * optimistic allocation without direct reclaim since inplace I/O
590 	 * can be used if low memory otherwise.
591 	 */
592 	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
593 			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
594 	unsigned int i;
595 
596 	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED)
597 		return;
598 
599 	for (i = 0; i < pcl->pclusterpages; ++i) {
600 		struct page *page;
601 		void *t;	/* mark pages just found for debugging */
602 		struct page *newpage = NULL;
603 
604 		/* the compressed page was loaded before */
605 		if (READ_ONCE(pcl->compressed_bvecs[i].page))
606 			continue;
607 
608 		page = find_get_page(mc, pcl->obj.index + i);
609 
610 		if (page) {
611 			t = (void *)((unsigned long)page | 1);
612 		} else {
613 			/* I/O is needed, no possible to decompress directly */
614 			standalone = false;
615 			if (!shouldalloc)
616 				continue;
617 
618 			/*
619 			 * try to use cached I/O if page allocation
620 			 * succeeds or fallback to in-place I/O instead
621 			 * to avoid any direct reclaim.
622 			 */
623 			newpage = erofs_allocpage(pagepool, gfp);
624 			if (!newpage)
625 				continue;
626 			set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE);
627 			t = (void *)((unsigned long)newpage | 1);
628 		}
629 
630 		if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL, t))
631 			continue;
632 
633 		if (page)
634 			put_page(page);
635 		else if (newpage)
636 			erofs_pagepool_add(pagepool, newpage);
637 	}
638 
639 	/*
640 	 * don't do inplace I/O if all compressed pages are available in
641 	 * managed cache since it can be moved to the bypass queue instead.
642 	 */
643 	if (standalone)
644 		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
645 }
646 
647 /* called by erofs_shrinker to get rid of all compressed_pages */
648 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
649 				       struct erofs_workgroup *grp)
650 {
651 	struct z_erofs_pcluster *const pcl =
652 		container_of(grp, struct z_erofs_pcluster, obj);
653 	int i;
654 
655 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
656 	/*
657 	 * refcount of workgroup is now freezed as 1,
658 	 * therefore no need to worry about available decompression users.
659 	 */
660 	for (i = 0; i < pcl->pclusterpages; ++i) {
661 		struct page *page = pcl->compressed_bvecs[i].page;
662 
663 		if (!page)
664 			continue;
665 
666 		/* block other users from reclaiming or migrating the page */
667 		if (!trylock_page(page))
668 			return -EBUSY;
669 
670 		if (!erofs_page_is_managed(sbi, page))
671 			continue;
672 
673 		/* barrier is implied in the following 'unlock_page' */
674 		WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
675 		detach_page_private(page);
676 		unlock_page(page);
677 	}
678 	return 0;
679 }
680 
681 int erofs_try_to_free_cached_page(struct page *page)
682 {
683 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
684 	int ret, i;
685 
686 	if (!erofs_workgroup_try_to_freeze(&pcl->obj, 1))
687 		return 0;
688 
689 	ret = 0;
690 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
691 	for (i = 0; i < pcl->pclusterpages; ++i) {
692 		if (pcl->compressed_bvecs[i].page == page) {
693 			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
694 			ret = 1;
695 			break;
696 		}
697 	}
698 	erofs_workgroup_unfreeze(&pcl->obj, 1);
699 	if (ret)
700 		detach_page_private(page);
701 	return ret;
702 }
703 
704 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
705 				   struct z_erofs_bvec *bvec)
706 {
707 	struct z_erofs_pcluster *const pcl = fe->pcl;
708 
709 	while (fe->icur > 0) {
710 		if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
711 			     NULL, bvec->page)) {
712 			pcl->compressed_bvecs[fe->icur] = *bvec;
713 			return true;
714 		}
715 	}
716 	return false;
717 }
718 
719 /* callers must be with pcluster lock held */
720 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,
721 			       struct z_erofs_bvec *bvec, bool exclusive)
722 {
723 	int ret;
724 
725 	if (exclusive) {
726 		/* give priority for inplaceio to use file pages first */
727 		if (z_erofs_try_inplace_io(fe, bvec))
728 			return 0;
729 		/* otherwise, check if it can be used as a bvpage */
730 		if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED &&
731 		    !fe->candidate_bvpage)
732 			fe->candidate_bvpage = bvec->page;
733 	}
734 	ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage);
735 	fe->pcl->vcnt += (ret >= 0);
736 	return ret;
737 }
738 
739 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
740 {
741 	struct z_erofs_pcluster *pcl = f->pcl;
742 	z_erofs_next_pcluster_t *owned_head = &f->owned_head;
743 
744 	/* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
745 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
746 		    *owned_head) == Z_EROFS_PCLUSTER_NIL) {
747 		*owned_head = &pcl->next;
748 		/* so we can attach this pcluster to our submission chain. */
749 		f->mode = Z_EROFS_PCLUSTER_FOLLOWED;
750 		return;
751 	}
752 
753 	/*
754 	 * type 2, link to the end of an existing open chain, be careful
755 	 * that its submission is controlled by the original attached chain.
756 	 */
757 	if (*owned_head != &pcl->next && pcl != f->tailpcl &&
758 	    cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
759 		    *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
760 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
761 		f->mode = Z_EROFS_PCLUSTER_HOOKED;
762 		f->tailpcl = NULL;
763 		return;
764 	}
765 	/* type 3, it belongs to a chain, but it isn't the end of the chain */
766 	f->mode = Z_EROFS_PCLUSTER_INFLIGHT;
767 }
768 
769 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)
770 {
771 	struct erofs_map_blocks *map = &fe->map;
772 	bool ztailpacking = map->m_flags & EROFS_MAP_META;
773 	struct z_erofs_pcluster *pcl;
774 	struct erofs_workgroup *grp;
775 	int err;
776 
777 	if (!(map->m_flags & EROFS_MAP_ENCODED) ||
778 	    (!ztailpacking && !(map->m_pa >> PAGE_SHIFT))) {
779 		DBG_BUGON(1);
780 		return -EFSCORRUPTED;
781 	}
782 
783 	/* no available pcluster, let's allocate one */
784 	pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
785 				     map->m_plen >> PAGE_SHIFT);
786 	if (IS_ERR(pcl))
787 		return PTR_ERR(pcl);
788 
789 	atomic_set(&pcl->obj.refcount, 1);
790 	pcl->algorithmformat = map->m_algorithmformat;
791 	pcl->length = 0;
792 	pcl->partial = true;
793 
794 	/* new pclusters should be claimed as type 1, primary and followed */
795 	pcl->next = fe->owned_head;
796 	pcl->pageofs_out = map->m_la & ~PAGE_MASK;
797 	fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;
798 
799 	/*
800 	 * lock all primary followed works before visible to others
801 	 * and mutex_trylock *never* fails for a new pcluster.
802 	 */
803 	mutex_init(&pcl->lock);
804 	DBG_BUGON(!mutex_trylock(&pcl->lock));
805 
806 	if (ztailpacking) {
807 		pcl->obj.index = 0;	/* which indicates ztailpacking */
808 		pcl->pageofs_in = erofs_blkoff(fe->inode->i_sb, map->m_pa);
809 		pcl->tailpacking_size = map->m_plen;
810 	} else {
811 		pcl->obj.index = map->m_pa >> PAGE_SHIFT;
812 
813 		grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);
814 		if (IS_ERR(grp)) {
815 			err = PTR_ERR(grp);
816 			goto err_out;
817 		}
818 
819 		if (grp != &pcl->obj) {
820 			fe->pcl = container_of(grp,
821 					struct z_erofs_pcluster, obj);
822 			err = -EEXIST;
823 			goto err_out;
824 		}
825 	}
826 	/* used to check tail merging loop due to corrupted images */
827 	if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
828 		fe->tailpcl = pcl;
829 	fe->owned_head = &pcl->next;
830 	fe->pcl = pcl;
831 	return 0;
832 
833 err_out:
834 	mutex_unlock(&pcl->lock);
835 	z_erofs_free_pcluster(pcl);
836 	return err;
837 }
838 
839 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
840 {
841 	struct erofs_map_blocks *map = &fe->map;
842 	struct erofs_workgroup *grp = NULL;
843 	int ret;
844 
845 	DBG_BUGON(fe->pcl);
846 
847 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */
848 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
849 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
850 
851 	if (!(map->m_flags & EROFS_MAP_META)) {
852 		grp = erofs_find_workgroup(fe->inode->i_sb,
853 					   map->m_pa >> PAGE_SHIFT);
854 	} else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
855 		DBG_BUGON(1);
856 		return -EFSCORRUPTED;
857 	}
858 
859 	if (grp) {
860 		fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
861 		ret = -EEXIST;
862 	} else {
863 		ret = z_erofs_register_pcluster(fe);
864 	}
865 
866 	if (ret == -EEXIST) {
867 		mutex_lock(&fe->pcl->lock);
868 		/* used to check tail merging loop due to corrupted images */
869 		if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
870 			fe->tailpcl = fe->pcl;
871 
872 		z_erofs_try_to_claim_pcluster(fe);
873 	} else if (ret) {
874 		return ret;
875 	}
876 	z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
877 				Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
878 	/* since file-backed online pages are traversed in reverse order */
879 	fe->icur = z_erofs_pclusterpages(fe->pcl);
880 	return 0;
881 }
882 
883 /*
884  * keep in mind that no referenced pclusters will be freed
885  * only after a RCU grace period.
886  */
887 static void z_erofs_rcu_callback(struct rcu_head *head)
888 {
889 	z_erofs_free_pcluster(container_of(head,
890 			struct z_erofs_pcluster, rcu));
891 }
892 
893 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
894 {
895 	struct z_erofs_pcluster *const pcl =
896 		container_of(grp, struct z_erofs_pcluster, obj);
897 
898 	call_rcu(&pcl->rcu, z_erofs_rcu_callback);
899 }
900 
901 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe)
902 {
903 	struct z_erofs_pcluster *pcl = fe->pcl;
904 
905 	if (!pcl)
906 		return false;
907 
908 	z_erofs_bvec_iter_end(&fe->biter);
909 	mutex_unlock(&pcl->lock);
910 
911 	if (fe->candidate_bvpage) {
912 		DBG_BUGON(z_erofs_is_shortlived_page(fe->candidate_bvpage));
913 		fe->candidate_bvpage = NULL;
914 	}
915 
916 	/*
917 	 * if all pending pages are added, don't hold its reference
918 	 * any longer if the pcluster isn't hosted by ourselves.
919 	 */
920 	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE)
921 		erofs_workgroup_put(&pcl->obj);
922 
923 	fe->pcl = NULL;
924 	return true;
925 }
926 
927 static int z_erofs_read_fragment(struct inode *inode, erofs_off_t pos,
928 				 struct page *page, unsigned int pageofs,
929 				 unsigned int len)
930 {
931 	struct super_block *sb = inode->i_sb;
932 	struct inode *packed_inode = EROFS_I_SB(inode)->packed_inode;
933 	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
934 	u8 *src, *dst;
935 	unsigned int i, cnt;
936 
937 	if (!packed_inode)
938 		return -EFSCORRUPTED;
939 
940 	buf.inode = packed_inode;
941 	pos += EROFS_I(inode)->z_fragmentoff;
942 	for (i = 0; i < len; i += cnt) {
943 		cnt = min_t(unsigned int, len - i,
944 			    sb->s_blocksize - erofs_blkoff(sb, pos));
945 		src = erofs_bread(&buf, erofs_blknr(sb, pos), EROFS_KMAP);
946 		if (IS_ERR(src)) {
947 			erofs_put_metabuf(&buf);
948 			return PTR_ERR(src);
949 		}
950 
951 		dst = kmap_local_page(page);
952 		memcpy(dst + pageofs + i, src + erofs_blkoff(sb, pos), cnt);
953 		kunmap_local(dst);
954 		pos += cnt;
955 	}
956 	erofs_put_metabuf(&buf);
957 	return 0;
958 }
959 
960 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
961 				struct page *page, struct page **pagepool)
962 {
963 	struct inode *const inode = fe->inode;
964 	struct erofs_map_blocks *const map = &fe->map;
965 	const loff_t offset = page_offset(page);
966 	bool tight = true, exclusive;
967 	unsigned int cur, end, spiltted;
968 	int err = 0;
969 
970 	/* register locked file pages as online pages in pack */
971 	z_erofs_onlinepage_init(page);
972 
973 	spiltted = 0;
974 	end = PAGE_SIZE;
975 repeat:
976 	cur = end - 1;
977 
978 	if (offset + cur < map->m_la ||
979 	    offset + cur >= map->m_la + map->m_llen) {
980 		if (z_erofs_collector_end(fe))
981 			fe->backmost = false;
982 		map->m_la = offset + cur;
983 		map->m_llen = 0;
984 		err = z_erofs_map_blocks_iter(inode, map, 0);
985 		if (err)
986 			goto out;
987 	} else {
988 		if (fe->pcl)
989 			goto hitted;
990 		/* didn't get a valid pcluster previously (very rare) */
991 	}
992 
993 	if (!(map->m_flags & EROFS_MAP_MAPPED) ||
994 	    map->m_flags & EROFS_MAP_FRAGMENT)
995 		goto hitted;
996 
997 	err = z_erofs_collector_begin(fe);
998 	if (err)
999 		goto out;
1000 
1001 	if (z_erofs_is_inline_pcluster(fe->pcl)) {
1002 		void *mp;
1003 
1004 		mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
1005 					erofs_blknr(inode->i_sb, map->m_pa),
1006 					EROFS_NO_KMAP);
1007 		if (IS_ERR(mp)) {
1008 			err = PTR_ERR(mp);
1009 			erofs_err(inode->i_sb,
1010 				  "failed to get inline page, err %d", err);
1011 			goto out;
1012 		}
1013 		get_page(fe->map.buf.page);
1014 		WRITE_ONCE(fe->pcl->compressed_bvecs[0].page,
1015 			   fe->map.buf.page);
1016 		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
1017 	} else {
1018 		/* bind cache first when cached decompression is preferred */
1019 		z_erofs_bind_cache(fe, pagepool);
1020 	}
1021 hitted:
1022 	/*
1023 	 * Ensure the current partial page belongs to this submit chain rather
1024 	 * than other concurrent submit chains or the noio(bypass) chain since
1025 	 * those chains are handled asynchronously thus the page cannot be used
1026 	 * for inplace I/O or bvpage (should be processed in a strict order.)
1027 	 */
1028 	tight &= (fe->mode >= Z_EROFS_PCLUSTER_HOOKED &&
1029 		  fe->mode != Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE);
1030 
1031 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
1032 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
1033 		zero_user_segment(page, cur, end);
1034 		goto next_part;
1035 	}
1036 	if (map->m_flags & EROFS_MAP_FRAGMENT) {
1037 		unsigned int pageofs, skip, len;
1038 
1039 		if (offset > map->m_la) {
1040 			pageofs = 0;
1041 			skip = offset - map->m_la;
1042 		} else {
1043 			pageofs = map->m_la & ~PAGE_MASK;
1044 			skip = 0;
1045 		}
1046 		len = min_t(unsigned int, map->m_llen - skip, end - cur);
1047 		err = z_erofs_read_fragment(inode, skip, page, pageofs, len);
1048 		if (err)
1049 			goto out;
1050 		++spiltted;
1051 		tight = false;
1052 		goto next_part;
1053 	}
1054 
1055 	exclusive = (!cur && (!spiltted || tight));
1056 	if (cur)
1057 		tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED);
1058 
1059 retry:
1060 	err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) {
1061 					.page = page,
1062 					.offset = offset - map->m_la,
1063 					.end = end,
1064 				  }), exclusive);
1065 	/* should allocate an additional short-lived page for bvset */
1066 	if (err == -EAGAIN && !fe->candidate_bvpage) {
1067 		fe->candidate_bvpage = alloc_page(GFP_NOFS | __GFP_NOFAIL);
1068 		set_page_private(fe->candidate_bvpage,
1069 				 Z_EROFS_SHORTLIVED_PAGE);
1070 		goto retry;
1071 	}
1072 
1073 	if (err) {
1074 		DBG_BUGON(err == -EAGAIN && fe->candidate_bvpage);
1075 		goto out;
1076 	}
1077 
1078 	z_erofs_onlinepage_split(page);
1079 	/* bump up the number of spiltted parts of a page */
1080 	++spiltted;
1081 	if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))
1082 		fe->pcl->multibases = true;
1083 	if (fe->pcl->length < offset + end - map->m_la) {
1084 		fe->pcl->length = offset + end - map->m_la;
1085 		fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK;
1086 	}
1087 	if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&
1088 	    !(map->m_flags & EROFS_MAP_PARTIAL_REF) &&
1089 	    fe->pcl->length == map->m_llen)
1090 		fe->pcl->partial = false;
1091 next_part:
1092 	/* shorten the remaining extent to update progress */
1093 	map->m_llen = offset + cur - map->m_la;
1094 	map->m_flags &= ~EROFS_MAP_FULL_MAPPED;
1095 
1096 	end = cur;
1097 	if (end > 0)
1098 		goto repeat;
1099 
1100 out:
1101 	if (err)
1102 		z_erofs_page_mark_eio(page);
1103 	z_erofs_onlinepage_endio(page);
1104 	return err;
1105 }
1106 
1107 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi,
1108 				       unsigned int readahead_pages)
1109 {
1110 	/* auto: enable for read_folio, disable for readahead */
1111 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
1112 	    !readahead_pages)
1113 		return true;
1114 
1115 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
1116 	    (readahead_pages <= sbi->opt.max_sync_decompress_pages))
1117 		return true;
1118 
1119 	return false;
1120 }
1121 
1122 static bool z_erofs_page_is_invalidated(struct page *page)
1123 {
1124 	return !page->mapping && !z_erofs_is_shortlived_page(page);
1125 }
1126 
1127 struct z_erofs_decompress_backend {
1128 	struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];
1129 	struct super_block *sb;
1130 	struct z_erofs_pcluster *pcl;
1131 
1132 	/* pages with the longest decompressed length for deduplication */
1133 	struct page **decompressed_pages;
1134 	/* pages to keep the compressed data */
1135 	struct page **compressed_pages;
1136 
1137 	struct list_head decompressed_secondary_bvecs;
1138 	struct page **pagepool;
1139 	unsigned int onstack_used, nr_pages;
1140 };
1141 
1142 struct z_erofs_bvec_item {
1143 	struct z_erofs_bvec bvec;
1144 	struct list_head list;
1145 };
1146 
1147 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,
1148 					 struct z_erofs_bvec *bvec)
1149 {
1150 	struct z_erofs_bvec_item *item;
1151 
1152 	if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK)) {
1153 		unsigned int pgnr;
1154 
1155 		pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;
1156 		DBG_BUGON(pgnr >= be->nr_pages);
1157 		if (!be->decompressed_pages[pgnr]) {
1158 			be->decompressed_pages[pgnr] = bvec->page;
1159 			return;
1160 		}
1161 	}
1162 
1163 	/* (cold path) one pcluster is requested multiple times */
1164 	item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);
1165 	item->bvec = *bvec;
1166 	list_add(&item->list, &be->decompressed_secondary_bvecs);
1167 }
1168 
1169 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,
1170 				      int err)
1171 {
1172 	unsigned int off0 = be->pcl->pageofs_out;
1173 	struct list_head *p, *n;
1174 
1175 	list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {
1176 		struct z_erofs_bvec_item *bvi;
1177 		unsigned int end, cur;
1178 		void *dst, *src;
1179 
1180 		bvi = container_of(p, struct z_erofs_bvec_item, list);
1181 		cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;
1182 		end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,
1183 			    bvi->bvec.end);
1184 		dst = kmap_local_page(bvi->bvec.page);
1185 		while (cur < end) {
1186 			unsigned int pgnr, scur, len;
1187 
1188 			pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;
1189 			DBG_BUGON(pgnr >= be->nr_pages);
1190 
1191 			scur = bvi->bvec.offset + cur -
1192 					((pgnr << PAGE_SHIFT) - off0);
1193 			len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);
1194 			if (!be->decompressed_pages[pgnr]) {
1195 				err = -EFSCORRUPTED;
1196 				cur += len;
1197 				continue;
1198 			}
1199 			src = kmap_local_page(be->decompressed_pages[pgnr]);
1200 			memcpy(dst + cur, src + scur, len);
1201 			kunmap_local(src);
1202 			cur += len;
1203 		}
1204 		kunmap_local(dst);
1205 		if (err)
1206 			z_erofs_page_mark_eio(bvi->bvec.page);
1207 		z_erofs_onlinepage_endio(bvi->bvec.page);
1208 		list_del(p);
1209 		kfree(bvi);
1210 	}
1211 }
1212 
1213 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)
1214 {
1215 	struct z_erofs_pcluster *pcl = be->pcl;
1216 	struct z_erofs_bvec_iter biter;
1217 	struct page *old_bvpage;
1218 	int i;
1219 
1220 	z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);
1221 	for (i = 0; i < pcl->vcnt; ++i) {
1222 		struct z_erofs_bvec bvec;
1223 
1224 		z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);
1225 
1226 		if (old_bvpage)
1227 			z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1228 
1229 		DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));
1230 		z_erofs_do_decompressed_bvec(be, &bvec);
1231 	}
1232 
1233 	old_bvpage = z_erofs_bvec_iter_end(&biter);
1234 	if (old_bvpage)
1235 		z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1236 }
1237 
1238 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,
1239 				  bool *overlapped)
1240 {
1241 	struct z_erofs_pcluster *pcl = be->pcl;
1242 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1243 	int i, err = 0;
1244 
1245 	*overlapped = false;
1246 	for (i = 0; i < pclusterpages; ++i) {
1247 		struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];
1248 		struct page *page = bvec->page;
1249 
1250 		/* compressed pages ought to be present before decompressing */
1251 		if (!page) {
1252 			DBG_BUGON(1);
1253 			continue;
1254 		}
1255 		be->compressed_pages[i] = page;
1256 
1257 		if (z_erofs_is_inline_pcluster(pcl)) {
1258 			if (!PageUptodate(page))
1259 				err = -EIO;
1260 			continue;
1261 		}
1262 
1263 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1264 		if (!z_erofs_is_shortlived_page(page)) {
1265 			if (erofs_page_is_managed(EROFS_SB(be->sb), page)) {
1266 				if (!PageUptodate(page))
1267 					err = -EIO;
1268 				continue;
1269 			}
1270 			z_erofs_do_decompressed_bvec(be, bvec);
1271 			*overlapped = true;
1272 		}
1273 	}
1274 
1275 	if (err)
1276 		return err;
1277 	return 0;
1278 }
1279 
1280 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,
1281 				       int err)
1282 {
1283 	struct erofs_sb_info *const sbi = EROFS_SB(be->sb);
1284 	struct z_erofs_pcluster *pcl = be->pcl;
1285 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1286 	unsigned int i, inputsize;
1287 	int err2;
1288 	struct page *page;
1289 	bool overlapped;
1290 
1291 	mutex_lock(&pcl->lock);
1292 	be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;
1293 
1294 	/* allocate (de)compressed page arrays if cannot be kept on stack */
1295 	be->decompressed_pages = NULL;
1296 	be->compressed_pages = NULL;
1297 	be->onstack_used = 0;
1298 	if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {
1299 		be->decompressed_pages = be->onstack_pages;
1300 		be->onstack_used = be->nr_pages;
1301 		memset(be->decompressed_pages, 0,
1302 		       sizeof(struct page *) * be->nr_pages);
1303 	}
1304 
1305 	if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)
1306 		be->compressed_pages = be->onstack_pages + be->onstack_used;
1307 
1308 	if (!be->decompressed_pages)
1309 		be->decompressed_pages =
1310 			kvcalloc(be->nr_pages, sizeof(struct page *),
1311 				 GFP_KERNEL | __GFP_NOFAIL);
1312 	if (!be->compressed_pages)
1313 		be->compressed_pages =
1314 			kvcalloc(pclusterpages, sizeof(struct page *),
1315 				 GFP_KERNEL | __GFP_NOFAIL);
1316 
1317 	z_erofs_parse_out_bvecs(be);
1318 	err2 = z_erofs_parse_in_bvecs(be, &overlapped);
1319 	if (err2)
1320 		err = err2;
1321 	if (err)
1322 		goto out;
1323 
1324 	if (z_erofs_is_inline_pcluster(pcl))
1325 		inputsize = pcl->tailpacking_size;
1326 	else
1327 		inputsize = pclusterpages * PAGE_SIZE;
1328 
1329 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
1330 					.sb = be->sb,
1331 					.in = be->compressed_pages,
1332 					.out = be->decompressed_pages,
1333 					.pageofs_in = pcl->pageofs_in,
1334 					.pageofs_out = pcl->pageofs_out,
1335 					.inputsize = inputsize,
1336 					.outputsize = pcl->length,
1337 					.alg = pcl->algorithmformat,
1338 					.inplace_io = overlapped,
1339 					.partial_decoding = pcl->partial,
1340 					.fillgaps = pcl->multibases,
1341 				 }, be->pagepool);
1342 
1343 out:
1344 	/* must handle all compressed pages before actual file pages */
1345 	if (z_erofs_is_inline_pcluster(pcl)) {
1346 		page = pcl->compressed_bvecs[0].page;
1347 		WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);
1348 		put_page(page);
1349 	} else {
1350 		for (i = 0; i < pclusterpages; ++i) {
1351 			page = pcl->compressed_bvecs[i].page;
1352 
1353 			if (erofs_page_is_managed(sbi, page))
1354 				continue;
1355 
1356 			/* recycle all individual short-lived pages */
1357 			(void)z_erofs_put_shortlivedpage(be->pagepool, page);
1358 			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
1359 		}
1360 	}
1361 	if (be->compressed_pages < be->onstack_pages ||
1362 	    be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)
1363 		kvfree(be->compressed_pages);
1364 	z_erofs_fill_other_copies(be, err);
1365 
1366 	for (i = 0; i < be->nr_pages; ++i) {
1367 		page = be->decompressed_pages[i];
1368 		if (!page)
1369 			continue;
1370 
1371 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1372 
1373 		/* recycle all individual short-lived pages */
1374 		if (z_erofs_put_shortlivedpage(be->pagepool, page))
1375 			continue;
1376 		if (err)
1377 			z_erofs_page_mark_eio(page);
1378 		z_erofs_onlinepage_endio(page);
1379 	}
1380 
1381 	if (be->decompressed_pages != be->onstack_pages)
1382 		kvfree(be->decompressed_pages);
1383 
1384 	pcl->length = 0;
1385 	pcl->partial = true;
1386 	pcl->multibases = false;
1387 	pcl->bvset.nextpage = NULL;
1388 	pcl->vcnt = 0;
1389 
1390 	/* pcluster lock MUST be taken before the following line */
1391 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1392 	mutex_unlock(&pcl->lock);
1393 	return err;
1394 }
1395 
1396 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1397 				     struct page **pagepool)
1398 {
1399 	struct z_erofs_decompress_backend be = {
1400 		.sb = io->sb,
1401 		.pagepool = pagepool,
1402 		.decompressed_secondary_bvecs =
1403 			LIST_HEAD_INIT(be.decompressed_secondary_bvecs),
1404 	};
1405 	z_erofs_next_pcluster_t owned = io->head;
1406 
1407 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1408 		/* impossible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1409 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1410 		/* impossible that 'owned' equals Z_EROFS_PCLUSTER_NIL */
1411 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1412 
1413 		be.pcl = container_of(owned, struct z_erofs_pcluster, next);
1414 		owned = READ_ONCE(be.pcl->next);
1415 
1416 		z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0);
1417 		erofs_workgroup_put(&be.pcl->obj);
1418 	}
1419 }
1420 
1421 static void z_erofs_decompressqueue_work(struct work_struct *work)
1422 {
1423 	struct z_erofs_decompressqueue *bgq =
1424 		container_of(work, struct z_erofs_decompressqueue, u.work);
1425 	struct page *pagepool = NULL;
1426 
1427 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1428 	z_erofs_decompress_queue(bgq, &pagepool);
1429 	erofs_release_pages(&pagepool);
1430 	kvfree(bgq);
1431 }
1432 
1433 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1434 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
1435 {
1436 	z_erofs_decompressqueue_work((struct work_struct *)work);
1437 }
1438 #endif
1439 
1440 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1441 				       int bios)
1442 {
1443 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1444 
1445 	/* wake up the caller thread for sync decompression */
1446 	if (io->sync) {
1447 		if (!atomic_add_return(bios, &io->pending_bios))
1448 			complete(&io->u.done);
1449 		return;
1450 	}
1451 
1452 	if (atomic_add_return(bios, &io->pending_bios))
1453 		return;
1454 	/* Use (kthread_)work and sync decompression for atomic contexts only */
1455 	if (in_atomic() || irqs_disabled()) {
1456 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1457 		struct kthread_worker *worker;
1458 
1459 		rcu_read_lock();
1460 		worker = rcu_dereference(
1461 				z_erofs_pcpu_workers[raw_smp_processor_id()]);
1462 		if (!worker) {
1463 			INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);
1464 			queue_work(z_erofs_workqueue, &io->u.work);
1465 		} else {
1466 			kthread_queue_work(worker, &io->u.kthread_work);
1467 		}
1468 		rcu_read_unlock();
1469 #else
1470 		queue_work(z_erofs_workqueue, &io->u.work);
1471 #endif
1472 		/* enable sync decompression for readahead */
1473 		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1474 			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1475 		return;
1476 	}
1477 	z_erofs_decompressqueue_work(&io->u.work);
1478 }
1479 
1480 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1481 					       unsigned int nr,
1482 					       struct page **pagepool,
1483 					       struct address_space *mc)
1484 {
1485 	const pgoff_t index = pcl->obj.index;
1486 	gfp_t gfp = mapping_gfp_mask(mc);
1487 	bool tocache = false;
1488 
1489 	struct address_space *mapping;
1490 	struct page *oldpage, *page;
1491 	int justfound;
1492 
1493 repeat:
1494 	page = READ_ONCE(pcl->compressed_bvecs[nr].page);
1495 	oldpage = page;
1496 
1497 	if (!page)
1498 		goto out_allocpage;
1499 
1500 	justfound = (unsigned long)page & 1UL;
1501 	page = (struct page *)((unsigned long)page & ~1UL);
1502 
1503 	/*
1504 	 * preallocated cached pages, which is used to avoid direct reclaim
1505 	 * otherwise, it will go inplace I/O path instead.
1506 	 */
1507 	if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1508 		WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1509 		set_page_private(page, 0);
1510 		tocache = true;
1511 		goto out_tocache;
1512 	}
1513 	mapping = READ_ONCE(page->mapping);
1514 
1515 	/*
1516 	 * file-backed online pages in plcuster are all locked steady,
1517 	 * therefore it is impossible for `mapping' to be NULL.
1518 	 */
1519 	if (mapping && mapping != mc)
1520 		/* ought to be unmanaged pages */
1521 		goto out;
1522 
1523 	/* directly return for shortlived page as well */
1524 	if (z_erofs_is_shortlived_page(page))
1525 		goto out;
1526 
1527 	lock_page(page);
1528 
1529 	/* only true if page reclaim goes wrong, should never happen */
1530 	DBG_BUGON(justfound && PagePrivate(page));
1531 
1532 	/* the page is still in manage cache */
1533 	if (page->mapping == mc) {
1534 		WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1535 
1536 		if (!PagePrivate(page)) {
1537 			/*
1538 			 * impossible to be !PagePrivate(page) for
1539 			 * the current restriction as well if
1540 			 * the page is already in compressed_bvecs[].
1541 			 */
1542 			DBG_BUGON(!justfound);
1543 
1544 			justfound = 0;
1545 			set_page_private(page, (unsigned long)pcl);
1546 			SetPagePrivate(page);
1547 		}
1548 
1549 		/* no need to submit io if it is already up-to-date */
1550 		if (PageUptodate(page)) {
1551 			unlock_page(page);
1552 			page = NULL;
1553 		}
1554 		goto out;
1555 	}
1556 
1557 	/*
1558 	 * the managed page has been truncated, it's unsafe to
1559 	 * reuse this one, let's allocate a new cache-managed page.
1560 	 */
1561 	DBG_BUGON(page->mapping);
1562 	DBG_BUGON(!justfound);
1563 
1564 	tocache = true;
1565 	unlock_page(page);
1566 	put_page(page);
1567 out_allocpage:
1568 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1569 	if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page,
1570 			       oldpage, page)) {
1571 		erofs_pagepool_add(pagepool, page);
1572 		cond_resched();
1573 		goto repeat;
1574 	}
1575 out_tocache:
1576 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1577 		/* turn into temporary page if fails (1 ref) */
1578 		set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1579 		goto out;
1580 	}
1581 	attach_page_private(page, pcl);
1582 	/* drop a refcount added by allocpage (then we have 2 refs here) */
1583 	put_page(page);
1584 
1585 out:	/* the only exit (for tracing and debugging) */
1586 	return page;
1587 }
1588 
1589 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb,
1590 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1591 {
1592 	struct z_erofs_decompressqueue *q;
1593 
1594 	if (fg && !*fg) {
1595 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1596 		if (!q) {
1597 			*fg = true;
1598 			goto fg_out;
1599 		}
1600 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1601 		kthread_init_work(&q->u.kthread_work,
1602 				  z_erofs_decompressqueue_kthread_work);
1603 #else
1604 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1605 #endif
1606 	} else {
1607 fg_out:
1608 		q = fgq;
1609 		init_completion(&fgq->u.done);
1610 		atomic_set(&fgq->pending_bios, 0);
1611 		q->eio = false;
1612 		q->sync = true;
1613 	}
1614 	q->sb = sb;
1615 	q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1616 	return q;
1617 }
1618 
1619 /* define decompression jobqueue types */
1620 enum {
1621 	JQ_BYPASS,
1622 	JQ_SUBMIT,
1623 	NR_JOBQUEUES,
1624 };
1625 
1626 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1627 				    z_erofs_next_pcluster_t qtail[],
1628 				    z_erofs_next_pcluster_t owned_head)
1629 {
1630 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1631 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1632 
1633 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1634 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1635 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1636 
1637 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1638 
1639 	WRITE_ONCE(*submit_qtail, owned_head);
1640 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1641 
1642 	qtail[JQ_BYPASS] = &pcl->next;
1643 }
1644 
1645 static void z_erofs_decompressqueue_endio(struct bio *bio)
1646 {
1647 	struct z_erofs_decompressqueue *q = bio->bi_private;
1648 	blk_status_t err = bio->bi_status;
1649 	struct bio_vec *bvec;
1650 	struct bvec_iter_all iter_all;
1651 
1652 	bio_for_each_segment_all(bvec, bio, iter_all) {
1653 		struct page *page = bvec->bv_page;
1654 
1655 		DBG_BUGON(PageUptodate(page));
1656 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1657 
1658 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1659 			if (!err)
1660 				SetPageUptodate(page);
1661 			unlock_page(page);
1662 		}
1663 	}
1664 	if (err)
1665 		q->eio = true;
1666 	z_erofs_decompress_kickoff(q, -1);
1667 	bio_put(bio);
1668 }
1669 
1670 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
1671 				 struct page **pagepool,
1672 				 struct z_erofs_decompressqueue *fgq,
1673 				 bool *force_fg)
1674 {
1675 	struct super_block *sb = f->inode->i_sb;
1676 	struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));
1677 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1678 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1679 	z_erofs_next_pcluster_t owned_head = f->owned_head;
1680 	/* bio is NULL initially, so no need to initialize last_{index,bdev} */
1681 	pgoff_t last_index;
1682 	struct block_device *last_bdev;
1683 	unsigned int nr_bios = 0;
1684 	struct bio *bio = NULL;
1685 	unsigned long pflags;
1686 	int memstall = 0;
1687 
1688 	/*
1689 	 * if managed cache is enabled, bypass jobqueue is needed,
1690 	 * no need to read from device for all pclusters in this queue.
1691 	 */
1692 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1693 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg);
1694 
1695 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1696 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1697 
1698 	/* by default, all need io submission */
1699 	q[JQ_SUBMIT]->head = owned_head;
1700 
1701 	do {
1702 		struct erofs_map_dev mdev;
1703 		struct z_erofs_pcluster *pcl;
1704 		pgoff_t cur, end;
1705 		unsigned int i = 0;
1706 		bool bypass = true;
1707 
1708 		/* no possible 'owned_head' equals the following */
1709 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1710 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1711 
1712 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1713 
1714 		/* close the main owned chain at first */
1715 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1716 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1717 		if (z_erofs_is_inline_pcluster(pcl)) {
1718 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1719 			continue;
1720 		}
1721 
1722 		/* no device id here, thus it will always succeed */
1723 		mdev = (struct erofs_map_dev) {
1724 			.m_pa = erofs_pos(sb, pcl->obj.index),
1725 		};
1726 		(void)erofs_map_dev(sb, &mdev);
1727 
1728 		cur = erofs_blknr(sb, mdev.m_pa);
1729 		end = cur + pcl->pclusterpages;
1730 
1731 		do {
1732 			struct page *page;
1733 
1734 			page = pickup_page_for_submission(pcl, i++, pagepool,
1735 							  mc);
1736 			if (!page)
1737 				continue;
1738 
1739 			if (bio && (cur != last_index + 1 ||
1740 				    last_bdev != mdev.m_bdev)) {
1741 submit_bio_retry:
1742 				submit_bio(bio);
1743 				if (memstall) {
1744 					psi_memstall_leave(&pflags);
1745 					memstall = 0;
1746 				}
1747 				bio = NULL;
1748 			}
1749 
1750 			if (unlikely(PageWorkingset(page)) && !memstall) {
1751 				psi_memstall_enter(&pflags);
1752 				memstall = 1;
1753 			}
1754 
1755 			if (!bio) {
1756 				bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1757 						REQ_OP_READ, GFP_NOIO);
1758 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1759 
1760 				last_bdev = mdev.m_bdev;
1761 				bio->bi_iter.bi_sector = (sector_t)cur <<
1762 					(sb->s_blocksize_bits - 9);
1763 				bio->bi_private = q[JQ_SUBMIT];
1764 				if (f->readahead)
1765 					bio->bi_opf |= REQ_RAHEAD;
1766 				++nr_bios;
1767 			}
1768 
1769 			if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1770 				goto submit_bio_retry;
1771 
1772 			last_index = cur;
1773 			bypass = false;
1774 		} while (++cur < end);
1775 
1776 		if (!bypass)
1777 			qtail[JQ_SUBMIT] = &pcl->next;
1778 		else
1779 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1780 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1781 
1782 	if (bio) {
1783 		submit_bio(bio);
1784 		if (memstall)
1785 			psi_memstall_leave(&pflags);
1786 	}
1787 
1788 	/*
1789 	 * although background is preferred, no one is pending for submission.
1790 	 * don't issue decompression but drop it directly instead.
1791 	 */
1792 	if (!*force_fg && !nr_bios) {
1793 		kvfree(q[JQ_SUBMIT]);
1794 		return;
1795 	}
1796 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios);
1797 }
1798 
1799 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f,
1800 			     struct page **pagepool, bool force_fg)
1801 {
1802 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1803 
1804 	if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1805 		return;
1806 	z_erofs_submit_queue(f, pagepool, io, &force_fg);
1807 
1808 	/* handle bypass queue (no i/o pclusters) immediately */
1809 	z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1810 
1811 	if (!force_fg)
1812 		return;
1813 
1814 	/* wait until all bios are completed */
1815 	wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1816 
1817 	/* handle synchronous decompress queue in the caller context */
1818 	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1819 }
1820 
1821 /*
1822  * Since partial uptodate is still unimplemented for now, we have to use
1823  * approximate readmore strategies as a start.
1824  */
1825 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1826 				      struct readahead_control *rac,
1827 				      erofs_off_t end,
1828 				      struct page **pagepool,
1829 				      bool backmost)
1830 {
1831 	struct inode *inode = f->inode;
1832 	struct erofs_map_blocks *map = &f->map;
1833 	erofs_off_t cur;
1834 	int err;
1835 
1836 	if (backmost) {
1837 		map->m_la = end;
1838 		err = z_erofs_map_blocks_iter(inode, map,
1839 					      EROFS_GET_BLOCKS_READMORE);
1840 		if (err)
1841 			return;
1842 
1843 		/* expend ra for the trailing edge if readahead */
1844 		if (rac) {
1845 			loff_t newstart = readahead_pos(rac);
1846 
1847 			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1848 			readahead_expand(rac, newstart, cur - newstart);
1849 			return;
1850 		}
1851 		end = round_up(end, PAGE_SIZE);
1852 	} else {
1853 		end = round_up(map->m_la, PAGE_SIZE);
1854 
1855 		if (!map->m_llen)
1856 			return;
1857 	}
1858 
1859 	cur = map->m_la + map->m_llen - 1;
1860 	while (cur >= end) {
1861 		pgoff_t index = cur >> PAGE_SHIFT;
1862 		struct page *page;
1863 
1864 		page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1865 		if (page) {
1866 			if (PageUptodate(page)) {
1867 				unlock_page(page);
1868 			} else {
1869 				err = z_erofs_do_read_page(f, page, pagepool);
1870 				if (err)
1871 					erofs_err(inode->i_sb,
1872 						  "readmore error at page %lu @ nid %llu",
1873 						  index, EROFS_I(inode)->nid);
1874 			}
1875 			put_page(page);
1876 		}
1877 
1878 		if (cur < PAGE_SIZE)
1879 			break;
1880 		cur = (index << PAGE_SHIFT) - 1;
1881 	}
1882 }
1883 
1884 static int z_erofs_read_folio(struct file *file, struct folio *folio)
1885 {
1886 	struct page *page = &folio->page;
1887 	struct inode *const inode = page->mapping->host;
1888 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1889 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1890 	struct page *pagepool = NULL;
1891 	int err;
1892 
1893 	trace_erofs_readpage(page, false);
1894 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1895 
1896 	z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1897 				  &pagepool, true);
1898 	err = z_erofs_do_read_page(&f, page, &pagepool);
1899 	z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1900 
1901 	(void)z_erofs_collector_end(&f);
1902 
1903 	/* if some compressed cluster ready, need submit them anyway */
1904 	z_erofs_runqueue(&f, &pagepool,
1905 			 z_erofs_get_sync_decompress_policy(sbi, 0));
1906 
1907 	if (err)
1908 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1909 
1910 	erofs_put_metabuf(&f.map.buf);
1911 	erofs_release_pages(&pagepool);
1912 	return err;
1913 }
1914 
1915 static void z_erofs_readahead(struct readahead_control *rac)
1916 {
1917 	struct inode *const inode = rac->mapping->host;
1918 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1919 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1920 	struct page *pagepool = NULL, *head = NULL, *page;
1921 	unsigned int nr_pages;
1922 
1923 	f.readahead = true;
1924 	f.headoffset = readahead_pos(rac);
1925 
1926 	z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1927 				  readahead_length(rac) - 1, &pagepool, true);
1928 	nr_pages = readahead_count(rac);
1929 	trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1930 
1931 	while ((page = readahead_page(rac))) {
1932 		set_page_private(page, (unsigned long)head);
1933 		head = page;
1934 	}
1935 
1936 	while (head) {
1937 		struct page *page = head;
1938 		int err;
1939 
1940 		/* traversal in reverse order */
1941 		head = (void *)page_private(page);
1942 
1943 		err = z_erofs_do_read_page(&f, page, &pagepool);
1944 		if (err)
1945 			erofs_err(inode->i_sb,
1946 				  "readahead error at page %lu @ nid %llu",
1947 				  page->index, EROFS_I(inode)->nid);
1948 		put_page(page);
1949 	}
1950 	z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1951 	(void)z_erofs_collector_end(&f);
1952 
1953 	z_erofs_runqueue(&f, &pagepool,
1954 			 z_erofs_get_sync_decompress_policy(sbi, nr_pages));
1955 	erofs_put_metabuf(&f.map.buf);
1956 	erofs_release_pages(&pagepool);
1957 }
1958 
1959 const struct address_space_operations z_erofs_aops = {
1960 	.read_folio = z_erofs_read_folio,
1961 	.readahead = z_erofs_readahead,
1962 };
1963