xref: /openbmc/linux/fs/erofs/zdata.c (revision 305c8388)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2018 HUAWEI, Inc.
4  *             http://www.huawei.com/
5  * Created by Gao Xiang <gaoxiang25@huawei.com>
6  */
7 #include "zdata.h"
8 #include "compress.h"
9 #include <linux/prefetch.h>
10 
11 #include <trace/events/erofs.h>
12 
13 /*
14  * a compressed_pages[] placeholder in order to avoid
15  * being filled with file pages for in-place decompression.
16  */
17 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
18 
19 /* how to allocate cached pages for a pcluster */
20 enum z_erofs_cache_alloctype {
21 	DONTALLOC,	/* don't allocate any cached pages */
22 	DELAYEDALLOC,	/* delayed allocation (at the time of submitting io) */
23 };
24 
25 /*
26  * tagged pointer with 1-bit tag for all compressed pages
27  * tag 0 - the page is just found with an extra page reference
28  */
29 typedef tagptr1_t compressed_page_t;
30 
31 #define tag_compressed_page_justfound(page) \
32 	tagptr_fold(compressed_page_t, page, 1)
33 
34 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
35 static struct kmem_cache *pcluster_cachep __read_mostly;
36 
37 void z_erofs_exit_zip_subsystem(void)
38 {
39 	destroy_workqueue(z_erofs_workqueue);
40 	kmem_cache_destroy(pcluster_cachep);
41 }
42 
43 static inline int z_erofs_init_workqueue(void)
44 {
45 	const unsigned int onlinecpus = num_possible_cpus();
46 	const unsigned int flags = WQ_UNBOUND | WQ_HIGHPRI | WQ_CPU_INTENSIVE;
47 
48 	/*
49 	 * no need to spawn too many threads, limiting threads could minimum
50 	 * scheduling overhead, perhaps per-CPU threads should be better?
51 	 */
52 	z_erofs_workqueue = alloc_workqueue("erofs_unzipd", flags,
53 					    onlinecpus + onlinecpus / 4);
54 	return z_erofs_workqueue ? 0 : -ENOMEM;
55 }
56 
57 static void z_erofs_pcluster_init_once(void *ptr)
58 {
59 	struct z_erofs_pcluster *pcl = ptr;
60 	struct z_erofs_collection *cl = z_erofs_primarycollection(pcl);
61 	unsigned int i;
62 
63 	mutex_init(&cl->lock);
64 	cl->nr_pages = 0;
65 	cl->vcnt = 0;
66 	for (i = 0; i < Z_EROFS_CLUSTER_MAX_PAGES; ++i)
67 		pcl->compressed_pages[i] = NULL;
68 }
69 
70 int __init z_erofs_init_zip_subsystem(void)
71 {
72 	pcluster_cachep = kmem_cache_create("erofs_compress",
73 					    Z_EROFS_WORKGROUP_SIZE, 0,
74 					    SLAB_RECLAIM_ACCOUNT,
75 					    z_erofs_pcluster_init_once);
76 	if (pcluster_cachep) {
77 		if (!z_erofs_init_workqueue())
78 			return 0;
79 
80 		kmem_cache_destroy(pcluster_cachep);
81 	}
82 	return -ENOMEM;
83 }
84 
85 enum z_erofs_collectmode {
86 	COLLECT_SECONDARY,
87 	COLLECT_PRIMARY,
88 	/*
89 	 * The current collection was the tail of an exist chain, in addition
90 	 * that the previous processed chained collections are all decided to
91 	 * be hooked up to it.
92 	 * A new chain will be created for the remaining collections which are
93 	 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
94 	 * the next collection cannot reuse the whole page safely in
95 	 * the following scenario:
96 	 *  ________________________________________________________________
97 	 * |      tail (partial) page     |       head (partial) page       |
98 	 * |   (belongs to the next cl)   |   (belongs to the current cl)   |
99 	 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
100 	 */
101 	COLLECT_PRIMARY_HOOKED,
102 	COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
103 	/*
104 	 * The current collection has been linked with the owned chain, and
105 	 * could also be linked with the remaining collections, which means
106 	 * if the processing page is the tail page of the collection, thus
107 	 * the current collection can safely use the whole page (since
108 	 * the previous collection is under control) for in-place I/O, as
109 	 * illustrated below:
110 	 *  ________________________________________________________________
111 	 * |  tail (partial) page |          head (partial) page           |
112 	 * |  (of the current cl) |      (of the previous collection)      |
113 	 * |  PRIMARY_FOLLOWED or |                                        |
114 	 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
115 	 *
116 	 * [  (*) the above page can be used as inplace I/O.               ]
117 	 */
118 	COLLECT_PRIMARY_FOLLOWED,
119 };
120 
121 struct z_erofs_collector {
122 	struct z_erofs_pagevec_ctor vector;
123 
124 	struct z_erofs_pcluster *pcl, *tailpcl;
125 	struct z_erofs_collection *cl;
126 	struct page **compressedpages;
127 	z_erofs_next_pcluster_t owned_head;
128 
129 	enum z_erofs_collectmode mode;
130 };
131 
132 struct z_erofs_decompress_frontend {
133 	struct inode *const inode;
134 
135 	struct z_erofs_collector clt;
136 	struct erofs_map_blocks map;
137 
138 	/* used for applying cache strategy on the fly */
139 	bool backmost;
140 	erofs_off_t headoffset;
141 };
142 
143 #define COLLECTOR_INIT() { \
144 	.owned_head = Z_EROFS_PCLUSTER_TAIL, \
145 	.mode = COLLECT_PRIMARY_FOLLOWED }
146 
147 #define DECOMPRESS_FRONTEND_INIT(__i) { \
148 	.inode = __i, .clt = COLLECTOR_INIT(), \
149 	.backmost = true, }
150 
151 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
152 static DEFINE_MUTEX(z_pagemap_global_lock);
153 
154 static void preload_compressed_pages(struct z_erofs_collector *clt,
155 				     struct address_space *mc,
156 				     enum z_erofs_cache_alloctype type,
157 				     struct list_head *pagepool)
158 {
159 	const struct z_erofs_pcluster *pcl = clt->pcl;
160 	const unsigned int clusterpages = BIT(pcl->clusterbits);
161 	struct page **pages = clt->compressedpages;
162 	pgoff_t index = pcl->obj.index + (pages - pcl->compressed_pages);
163 	bool standalone = true;
164 
165 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
166 		return;
167 
168 	for (; pages < pcl->compressed_pages + clusterpages; ++pages) {
169 		struct page *page;
170 		compressed_page_t t;
171 
172 		/* the compressed page was loaded before */
173 		if (READ_ONCE(*pages))
174 			continue;
175 
176 		page = find_get_page(mc, index);
177 
178 		if (page) {
179 			t = tag_compressed_page_justfound(page);
180 		} else if (type == DELAYEDALLOC) {
181 			t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED);
182 		} else {	/* DONTALLOC */
183 			if (standalone)
184 				clt->compressedpages = pages;
185 			standalone = false;
186 			continue;
187 		}
188 
189 		if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
190 			continue;
191 
192 		if (page)
193 			put_page(page);
194 	}
195 
196 	if (standalone)		/* downgrade to PRIMARY_FOLLOWED_NOINPLACE */
197 		clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
198 }
199 
200 /* called by erofs_shrinker to get rid of all compressed_pages */
201 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
202 				       struct erofs_workgroup *grp)
203 {
204 	struct z_erofs_pcluster *const pcl =
205 		container_of(grp, struct z_erofs_pcluster, obj);
206 	struct address_space *const mapping = MNGD_MAPPING(sbi);
207 	const unsigned int clusterpages = BIT(pcl->clusterbits);
208 	int i;
209 
210 	/*
211 	 * refcount of workgroup is now freezed as 1,
212 	 * therefore no need to worry about available decompression users.
213 	 */
214 	for (i = 0; i < clusterpages; ++i) {
215 		struct page *page = pcl->compressed_pages[i];
216 
217 		if (!page)
218 			continue;
219 
220 		/* block other users from reclaiming or migrating the page */
221 		if (!trylock_page(page))
222 			return -EBUSY;
223 
224 		if (page->mapping != mapping)
225 			continue;
226 
227 		/* barrier is implied in the following 'unlock_page' */
228 		WRITE_ONCE(pcl->compressed_pages[i], NULL);
229 		set_page_private(page, 0);
230 		ClearPagePrivate(page);
231 
232 		unlock_page(page);
233 		put_page(page);
234 	}
235 	return 0;
236 }
237 
238 int erofs_try_to_free_cached_page(struct address_space *mapping,
239 				  struct page *page)
240 {
241 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
242 	const unsigned int clusterpages = BIT(pcl->clusterbits);
243 	int ret = 0;	/* 0 - busy */
244 
245 	if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
246 		unsigned int i;
247 
248 		for (i = 0; i < clusterpages; ++i) {
249 			if (pcl->compressed_pages[i] == page) {
250 				WRITE_ONCE(pcl->compressed_pages[i], NULL);
251 				ret = 1;
252 				break;
253 			}
254 		}
255 		erofs_workgroup_unfreeze(&pcl->obj, 1);
256 
257 		if (ret) {
258 			ClearPagePrivate(page);
259 			put_page(page);
260 		}
261 	}
262 	return ret;
263 }
264 
265 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
266 static inline bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
267 					  struct page *page)
268 {
269 	struct z_erofs_pcluster *const pcl = clt->pcl;
270 	const unsigned int clusterpages = BIT(pcl->clusterbits);
271 
272 	while (clt->compressedpages < pcl->compressed_pages + clusterpages) {
273 		if (!cmpxchg(clt->compressedpages++, NULL, page))
274 			return true;
275 	}
276 	return false;
277 }
278 
279 /* callers must be with collection lock held */
280 static int z_erofs_attach_page(struct z_erofs_collector *clt,
281 			       struct page *page,
282 			       enum z_erofs_page_type type)
283 {
284 	int ret;
285 	bool occupied;
286 
287 	/* give priority for inplaceio */
288 	if (clt->mode >= COLLECT_PRIMARY &&
289 	    type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
290 	    z_erofs_try_inplace_io(clt, page))
291 		return 0;
292 
293 	ret = z_erofs_pagevec_enqueue(&clt->vector,
294 				      page, type, &occupied);
295 	clt->cl->vcnt += (unsigned int)ret;
296 
297 	return ret ? 0 : -EAGAIN;
298 }
299 
300 static enum z_erofs_collectmode
301 try_to_claim_pcluster(struct z_erofs_pcluster *pcl,
302 		      z_erofs_next_pcluster_t *owned_head)
303 {
304 	/* let's claim these following types of pclusters */
305 retry:
306 	if (pcl->next == Z_EROFS_PCLUSTER_NIL) {
307 		/* type 1, nil pcluster */
308 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
309 			    *owned_head) != Z_EROFS_PCLUSTER_NIL)
310 			goto retry;
311 
312 		*owned_head = &pcl->next;
313 		/* lucky, I am the followee :) */
314 		return COLLECT_PRIMARY_FOLLOWED;
315 	} else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) {
316 		/*
317 		 * type 2, link to the end of a existing open chain,
318 		 * be careful that its submission itself is governed
319 		 * by the original owned chain.
320 		 */
321 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
322 			    *owned_head) != Z_EROFS_PCLUSTER_TAIL)
323 			goto retry;
324 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
325 		return COLLECT_PRIMARY_HOOKED;
326 	}
327 	return COLLECT_PRIMARY;	/* :( better luck next time */
328 }
329 
330 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
331 				     struct inode *inode,
332 				     struct erofs_map_blocks *map)
333 {
334 	struct z_erofs_pcluster *pcl = clt->pcl;
335 	struct z_erofs_collection *cl;
336 	unsigned int length;
337 
338 	/* to avoid unexpected loop formed by corrupted images */
339 	if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
340 		DBG_BUGON(1);
341 		return -EFSCORRUPTED;
342 	}
343 
344 	cl = z_erofs_primarycollection(pcl);
345 	if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
346 		DBG_BUGON(1);
347 		return -EFSCORRUPTED;
348 	}
349 
350 	length = READ_ONCE(pcl->length);
351 	if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
352 		if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
353 			DBG_BUGON(1);
354 			return -EFSCORRUPTED;
355 		}
356 	} else {
357 		unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
358 
359 		if (map->m_flags & EROFS_MAP_FULL_MAPPED)
360 			llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
361 
362 		while (llen > length &&
363 		       length != cmpxchg_relaxed(&pcl->length, length, llen)) {
364 			cpu_relax();
365 			length = READ_ONCE(pcl->length);
366 		}
367 	}
368 	mutex_lock(&cl->lock);
369 	/* used to check tail merging loop due to corrupted images */
370 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
371 		clt->tailpcl = pcl;
372 	clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head);
373 	/* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */
374 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
375 		clt->tailpcl = NULL;
376 	clt->cl = cl;
377 	return 0;
378 }
379 
380 static int z_erofs_register_collection(struct z_erofs_collector *clt,
381 				       struct inode *inode,
382 				       struct erofs_map_blocks *map)
383 {
384 	struct z_erofs_pcluster *pcl;
385 	struct z_erofs_collection *cl;
386 	struct erofs_workgroup *grp;
387 	int err;
388 
389 	/* no available workgroup, let's allocate one */
390 	pcl = kmem_cache_alloc(pcluster_cachep, GFP_NOFS);
391 	if (!pcl)
392 		return -ENOMEM;
393 
394 	atomic_set(&pcl->obj.refcount, 1);
395 	pcl->obj.index = map->m_pa >> PAGE_SHIFT;
396 
397 	pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
398 		(map->m_flags & EROFS_MAP_FULL_MAPPED ?
399 			Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
400 
401 	if (map->m_flags & EROFS_MAP_ZIPPED)
402 		pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
403 	else
404 		pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
405 
406 	pcl->clusterbits = EROFS_I(inode)->z_physical_clusterbits[0];
407 	pcl->clusterbits -= PAGE_SHIFT;
408 
409 	/* new pclusters should be claimed as type 1, primary and followed */
410 	pcl->next = clt->owned_head;
411 	clt->mode = COLLECT_PRIMARY_FOLLOWED;
412 
413 	cl = z_erofs_primarycollection(pcl);
414 
415 	/* must be cleaned before freeing to slab */
416 	DBG_BUGON(cl->nr_pages);
417 	DBG_BUGON(cl->vcnt);
418 
419 	cl->pageofs = map->m_la & ~PAGE_MASK;
420 
421 	/*
422 	 * lock all primary followed works before visible to others
423 	 * and mutex_trylock *never* fails for a new pcluster.
424 	 */
425 	DBG_BUGON(!mutex_trylock(&cl->lock));
426 
427 	grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
428 	if (IS_ERR(grp)) {
429 		err = PTR_ERR(grp);
430 		goto err_out;
431 	}
432 
433 	if (grp != &pcl->obj) {
434 		clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
435 		err = -EEXIST;
436 		goto err_out;
437 	}
438 	/* used to check tail merging loop due to corrupted images */
439 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
440 		clt->tailpcl = pcl;
441 	clt->owned_head = &pcl->next;
442 	clt->pcl = pcl;
443 	clt->cl = cl;
444 	return 0;
445 
446 err_out:
447 	mutex_unlock(&cl->lock);
448 	kmem_cache_free(pcluster_cachep, pcl);
449 	return err;
450 }
451 
452 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
453 				   struct inode *inode,
454 				   struct erofs_map_blocks *map)
455 {
456 	struct erofs_workgroup *grp;
457 	int ret;
458 
459 	DBG_BUGON(clt->cl);
460 
461 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
462 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
463 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
464 
465 	if (!PAGE_ALIGNED(map->m_pa)) {
466 		DBG_BUGON(1);
467 		return -EINVAL;
468 	}
469 
470 	grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
471 	if (grp) {
472 		clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
473 	} else {
474 		ret = z_erofs_register_collection(clt, inode, map);
475 
476 		if (!ret)
477 			goto out;
478 		if (ret != -EEXIST)
479 			return ret;
480 	}
481 
482 	ret = z_erofs_lookup_collection(clt, inode, map);
483 	if (ret) {
484 		erofs_workgroup_put(&clt->pcl->obj);
485 		return ret;
486 	}
487 
488 out:
489 	z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
490 				  clt->cl->pagevec, clt->cl->vcnt);
491 
492 	clt->compressedpages = clt->pcl->compressed_pages;
493 	if (clt->mode <= COLLECT_PRIMARY) /* cannot do in-place I/O */
494 		clt->compressedpages += Z_EROFS_CLUSTER_MAX_PAGES;
495 	return 0;
496 }
497 
498 /*
499  * keep in mind that no referenced pclusters will be freed
500  * only after a RCU grace period.
501  */
502 static void z_erofs_rcu_callback(struct rcu_head *head)
503 {
504 	struct z_erofs_collection *const cl =
505 		container_of(head, struct z_erofs_collection, rcu);
506 
507 	kmem_cache_free(pcluster_cachep,
508 			container_of(cl, struct z_erofs_pcluster,
509 				     primary_collection));
510 }
511 
512 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
513 {
514 	struct z_erofs_pcluster *const pcl =
515 		container_of(grp, struct z_erofs_pcluster, obj);
516 	struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
517 
518 	call_rcu(&cl->rcu, z_erofs_rcu_callback);
519 }
520 
521 static void z_erofs_collection_put(struct z_erofs_collection *cl)
522 {
523 	struct z_erofs_pcluster *const pcl =
524 		container_of(cl, struct z_erofs_pcluster, primary_collection);
525 
526 	erofs_workgroup_put(&pcl->obj);
527 }
528 
529 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
530 {
531 	struct z_erofs_collection *cl = clt->cl;
532 
533 	if (!cl)
534 		return false;
535 
536 	z_erofs_pagevec_ctor_exit(&clt->vector, false);
537 	mutex_unlock(&cl->lock);
538 
539 	/*
540 	 * if all pending pages are added, don't hold its reference
541 	 * any longer if the pcluster isn't hosted by ourselves.
542 	 */
543 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
544 		z_erofs_collection_put(cl);
545 
546 	clt->cl = NULL;
547 	return true;
548 }
549 
550 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
551 				       unsigned int cachestrategy,
552 				       erofs_off_t la)
553 {
554 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
555 		return false;
556 
557 	if (fe->backmost)
558 		return true;
559 
560 	return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
561 		la < fe->headoffset;
562 }
563 
564 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
565 				struct page *page,
566 				struct list_head *pagepool)
567 {
568 	struct inode *const inode = fe->inode;
569 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
570 	struct erofs_map_blocks *const map = &fe->map;
571 	struct z_erofs_collector *const clt = &fe->clt;
572 	const loff_t offset = page_offset(page);
573 	bool tight = true;
574 
575 	enum z_erofs_cache_alloctype cache_strategy;
576 	enum z_erofs_page_type page_type;
577 	unsigned int cur, end, spiltted, index;
578 	int err = 0;
579 
580 	/* register locked file pages as online pages in pack */
581 	z_erofs_onlinepage_init(page);
582 
583 	spiltted = 0;
584 	end = PAGE_SIZE;
585 repeat:
586 	cur = end - 1;
587 
588 	/* lucky, within the range of the current map_blocks */
589 	if (offset + cur >= map->m_la &&
590 	    offset + cur < map->m_la + map->m_llen) {
591 		/* didn't get a valid collection previously (very rare) */
592 		if (!clt->cl)
593 			goto restart_now;
594 		goto hitted;
595 	}
596 
597 	/* go ahead the next map_blocks */
598 	erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
599 
600 	if (z_erofs_collector_end(clt))
601 		fe->backmost = false;
602 
603 	map->m_la = offset + cur;
604 	map->m_llen = 0;
605 	err = z_erofs_map_blocks_iter(inode, map, 0);
606 	if (err)
607 		goto err_out;
608 
609 restart_now:
610 	if (!(map->m_flags & EROFS_MAP_MAPPED))
611 		goto hitted;
612 
613 	err = z_erofs_collector_begin(clt, inode, map);
614 	if (err)
615 		goto err_out;
616 
617 	/* preload all compressed pages (maybe downgrade role if necessary) */
618 	if (should_alloc_managed_pages(fe, sbi->ctx.cache_strategy, map->m_la))
619 		cache_strategy = DELAYEDALLOC;
620 	else
621 		cache_strategy = DONTALLOC;
622 
623 	preload_compressed_pages(clt, MNGD_MAPPING(sbi),
624 				 cache_strategy, pagepool);
625 
626 hitted:
627 	/*
628 	 * Ensure the current partial page belongs to this submit chain rather
629 	 * than other concurrent submit chains or the noio(bypass) chain since
630 	 * those chains are handled asynchronously thus the page cannot be used
631 	 * for inplace I/O or pagevec (should be processed in strict order.)
632 	 */
633 	tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
634 		  clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
635 
636 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
637 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
638 		zero_user_segment(page, cur, end);
639 		goto next_part;
640 	}
641 
642 	/* let's derive page type */
643 	page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
644 		(!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
645 			(tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
646 				Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
647 
648 	if (cur)
649 		tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
650 
651 retry:
652 	err = z_erofs_attach_page(clt, page, page_type);
653 	/* should allocate an additional staging page for pagevec */
654 	if (err == -EAGAIN) {
655 		struct page *const newpage =
656 			erofs_allocpage(pagepool, GFP_NOFS | __GFP_NOFAIL);
657 
658 		newpage->mapping = Z_EROFS_MAPPING_STAGING;
659 		err = z_erofs_attach_page(clt, newpage,
660 					  Z_EROFS_PAGE_TYPE_EXCLUSIVE);
661 		if (!err)
662 			goto retry;
663 	}
664 
665 	if (err)
666 		goto err_out;
667 
668 	index = page->index - (map->m_la >> PAGE_SHIFT);
669 
670 	z_erofs_onlinepage_fixup(page, index, true);
671 
672 	/* bump up the number of spiltted parts of a page */
673 	++spiltted;
674 	/* also update nr_pages */
675 	clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
676 next_part:
677 	/* can be used for verification */
678 	map->m_llen = offset + cur - map->m_la;
679 
680 	end = cur;
681 	if (end > 0)
682 		goto repeat;
683 
684 out:
685 	z_erofs_onlinepage_endio(page);
686 
687 	erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
688 		  __func__, page, spiltted, map->m_llen);
689 	return err;
690 
691 	/* if some error occurred while processing this page */
692 err_out:
693 	SetPageError(page);
694 	goto out;
695 }
696 
697 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
698 				       bool sync, int bios)
699 {
700 	/* wake up the caller thread for sync decompression */
701 	if (sync) {
702 		unsigned long flags;
703 
704 		spin_lock_irqsave(&io->u.wait.lock, flags);
705 		if (!atomic_add_return(bios, &io->pending_bios))
706 			wake_up_locked(&io->u.wait);
707 		spin_unlock_irqrestore(&io->u.wait.lock, flags);
708 		return;
709 	}
710 
711 	if (!atomic_add_return(bios, &io->pending_bios))
712 		queue_work(z_erofs_workqueue, &io->u.work);
713 }
714 
715 static void z_erofs_decompressqueue_endio(struct bio *bio)
716 {
717 	tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
718 	struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
719 	blk_status_t err = bio->bi_status;
720 	struct bio_vec *bvec;
721 	struct bvec_iter_all iter_all;
722 
723 	bio_for_each_segment_all(bvec, bio, iter_all) {
724 		struct page *page = bvec->bv_page;
725 
726 		DBG_BUGON(PageUptodate(page));
727 		DBG_BUGON(!page->mapping);
728 
729 		if (err)
730 			SetPageError(page);
731 
732 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
733 			if (!err)
734 				SetPageUptodate(page);
735 			unlock_page(page);
736 		}
737 	}
738 	z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
739 	bio_put(bio);
740 }
741 
742 static int z_erofs_decompress_pcluster(struct super_block *sb,
743 				       struct z_erofs_pcluster *pcl,
744 				       struct list_head *pagepool)
745 {
746 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
747 	const unsigned int clusterpages = BIT(pcl->clusterbits);
748 	struct z_erofs_pagevec_ctor ctor;
749 	unsigned int i, outputsize, llen, nr_pages;
750 	struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
751 	struct page **pages, **compressed_pages, *page;
752 
753 	enum z_erofs_page_type page_type;
754 	bool overlapped, partial;
755 	struct z_erofs_collection *cl;
756 	int err;
757 
758 	might_sleep();
759 	cl = z_erofs_primarycollection(pcl);
760 	DBG_BUGON(!READ_ONCE(cl->nr_pages));
761 
762 	mutex_lock(&cl->lock);
763 	nr_pages = cl->nr_pages;
764 
765 	if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
766 		pages = pages_onstack;
767 	} else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
768 		   mutex_trylock(&z_pagemap_global_lock)) {
769 		pages = z_pagemap_global;
770 	} else {
771 		gfp_t gfp_flags = GFP_KERNEL;
772 
773 		if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
774 			gfp_flags |= __GFP_NOFAIL;
775 
776 		pages = kvmalloc_array(nr_pages, sizeof(struct page *),
777 				       gfp_flags);
778 
779 		/* fallback to global pagemap for the lowmem scenario */
780 		if (!pages) {
781 			mutex_lock(&z_pagemap_global_lock);
782 			pages = z_pagemap_global;
783 		}
784 	}
785 
786 	for (i = 0; i < nr_pages; ++i)
787 		pages[i] = NULL;
788 
789 	err = 0;
790 	z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
791 				  cl->pagevec, 0);
792 
793 	for (i = 0; i < cl->vcnt; ++i) {
794 		unsigned int pagenr;
795 
796 		page = z_erofs_pagevec_dequeue(&ctor, &page_type);
797 
798 		/* all pages in pagevec ought to be valid */
799 		DBG_BUGON(!page);
800 		DBG_BUGON(!page->mapping);
801 
802 		if (z_erofs_put_stagingpage(pagepool, page))
803 			continue;
804 
805 		if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
806 			pagenr = 0;
807 		else
808 			pagenr = z_erofs_onlinepage_index(page);
809 
810 		DBG_BUGON(pagenr >= nr_pages);
811 
812 		/*
813 		 * currently EROFS doesn't support multiref(dedup),
814 		 * so here erroring out one multiref page.
815 		 */
816 		if (pages[pagenr]) {
817 			DBG_BUGON(1);
818 			SetPageError(pages[pagenr]);
819 			z_erofs_onlinepage_endio(pages[pagenr]);
820 			err = -EFSCORRUPTED;
821 		}
822 		pages[pagenr] = page;
823 	}
824 	z_erofs_pagevec_ctor_exit(&ctor, true);
825 
826 	overlapped = false;
827 	compressed_pages = pcl->compressed_pages;
828 
829 	for (i = 0; i < clusterpages; ++i) {
830 		unsigned int pagenr;
831 
832 		page = compressed_pages[i];
833 
834 		/* all compressed pages ought to be valid */
835 		DBG_BUGON(!page);
836 		DBG_BUGON(!page->mapping);
837 
838 		if (!z_erofs_page_is_staging(page)) {
839 			if (erofs_page_is_managed(sbi, page)) {
840 				if (!PageUptodate(page))
841 					err = -EIO;
842 				continue;
843 			}
844 
845 			/*
846 			 * only if non-head page can be selected
847 			 * for inplace decompression
848 			 */
849 			pagenr = z_erofs_onlinepage_index(page);
850 
851 			DBG_BUGON(pagenr >= nr_pages);
852 			if (pages[pagenr]) {
853 				DBG_BUGON(1);
854 				SetPageError(pages[pagenr]);
855 				z_erofs_onlinepage_endio(pages[pagenr]);
856 				err = -EFSCORRUPTED;
857 			}
858 			pages[pagenr] = page;
859 
860 			overlapped = true;
861 		}
862 
863 		/* PG_error needs checking for inplaced and staging pages */
864 		if (PageError(page)) {
865 			DBG_BUGON(PageUptodate(page));
866 			err = -EIO;
867 		}
868 	}
869 
870 	if (err)
871 		goto out;
872 
873 	llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
874 	if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
875 		outputsize = llen;
876 		partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
877 	} else {
878 		outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
879 		partial = true;
880 	}
881 
882 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
883 					.sb = sb,
884 					.in = compressed_pages,
885 					.out = pages,
886 					.pageofs_out = cl->pageofs,
887 					.inputsize = PAGE_SIZE,
888 					.outputsize = outputsize,
889 					.alg = pcl->algorithmformat,
890 					.inplace_io = overlapped,
891 					.partial_decoding = partial
892 				 }, pagepool);
893 
894 out:
895 	/* must handle all compressed pages before endding pages */
896 	for (i = 0; i < clusterpages; ++i) {
897 		page = compressed_pages[i];
898 
899 		if (erofs_page_is_managed(sbi, page))
900 			continue;
901 
902 		/* recycle all individual staging pages */
903 		(void)z_erofs_put_stagingpage(pagepool, page);
904 
905 		WRITE_ONCE(compressed_pages[i], NULL);
906 	}
907 
908 	for (i = 0; i < nr_pages; ++i) {
909 		page = pages[i];
910 		if (!page)
911 			continue;
912 
913 		DBG_BUGON(!page->mapping);
914 
915 		/* recycle all individual staging pages */
916 		if (z_erofs_put_stagingpage(pagepool, page))
917 			continue;
918 
919 		if (err < 0)
920 			SetPageError(page);
921 
922 		z_erofs_onlinepage_endio(page);
923 	}
924 
925 	if (pages == z_pagemap_global)
926 		mutex_unlock(&z_pagemap_global_lock);
927 	else if (pages != pages_onstack)
928 		kvfree(pages);
929 
930 	cl->nr_pages = 0;
931 	cl->vcnt = 0;
932 
933 	/* all cl locks MUST be taken before the following line */
934 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
935 
936 	/* all cl locks SHOULD be released right now */
937 	mutex_unlock(&cl->lock);
938 
939 	z_erofs_collection_put(cl);
940 	return err;
941 }
942 
943 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
944 				     struct list_head *pagepool)
945 {
946 	z_erofs_next_pcluster_t owned = io->head;
947 
948 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
949 		struct z_erofs_pcluster *pcl;
950 
951 		/* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
952 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
953 
954 		/* no possible that 'owned' equals NULL */
955 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
956 
957 		pcl = container_of(owned, struct z_erofs_pcluster, next);
958 		owned = READ_ONCE(pcl->next);
959 
960 		z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
961 	}
962 }
963 
964 static void z_erofs_decompressqueue_work(struct work_struct *work)
965 {
966 	struct z_erofs_decompressqueue *bgq =
967 		container_of(work, struct z_erofs_decompressqueue, u.work);
968 	LIST_HEAD(pagepool);
969 
970 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
971 	z_erofs_decompress_queue(bgq, &pagepool);
972 
973 	put_pages_list(&pagepool);
974 	kvfree(bgq);
975 }
976 
977 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
978 					       unsigned int nr,
979 					       struct list_head *pagepool,
980 					       struct address_space *mc,
981 					       gfp_t gfp)
982 {
983 	const pgoff_t index = pcl->obj.index;
984 	bool tocache = false;
985 
986 	struct address_space *mapping;
987 	struct page *oldpage, *page;
988 
989 	compressed_page_t t;
990 	int justfound;
991 
992 repeat:
993 	page = READ_ONCE(pcl->compressed_pages[nr]);
994 	oldpage = page;
995 
996 	if (!page)
997 		goto out_allocpage;
998 
999 	/*
1000 	 * the cached page has not been allocated and
1001 	 * an placeholder is out there, prepare it now.
1002 	 */
1003 	if (page == PAGE_UNALLOCATED) {
1004 		tocache = true;
1005 		goto out_allocpage;
1006 	}
1007 
1008 	/* process the target tagged pointer */
1009 	t = tagptr_init(compressed_page_t, page);
1010 	justfound = tagptr_unfold_tags(t);
1011 	page = tagptr_unfold_ptr(t);
1012 
1013 	mapping = READ_ONCE(page->mapping);
1014 
1015 	/*
1016 	 * unmanaged (file) pages are all locked solidly,
1017 	 * therefore it is impossible for `mapping' to be NULL.
1018 	 */
1019 	if (mapping && mapping != mc)
1020 		/* ought to be unmanaged pages */
1021 		goto out;
1022 
1023 	lock_page(page);
1024 
1025 	/* only true if page reclaim goes wrong, should never happen */
1026 	DBG_BUGON(justfound && PagePrivate(page));
1027 
1028 	/* the page is still in manage cache */
1029 	if (page->mapping == mc) {
1030 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1031 
1032 		ClearPageError(page);
1033 		if (!PagePrivate(page)) {
1034 			/*
1035 			 * impossible to be !PagePrivate(page) for
1036 			 * the current restriction as well if
1037 			 * the page is already in compressed_pages[].
1038 			 */
1039 			DBG_BUGON(!justfound);
1040 
1041 			justfound = 0;
1042 			set_page_private(page, (unsigned long)pcl);
1043 			SetPagePrivate(page);
1044 		}
1045 
1046 		/* no need to submit io if it is already up-to-date */
1047 		if (PageUptodate(page)) {
1048 			unlock_page(page);
1049 			page = NULL;
1050 		}
1051 		goto out;
1052 	}
1053 
1054 	/*
1055 	 * the managed page has been truncated, it's unsafe to
1056 	 * reuse this one, let's allocate a new cache-managed page.
1057 	 */
1058 	DBG_BUGON(page->mapping);
1059 	DBG_BUGON(!justfound);
1060 
1061 	tocache = true;
1062 	unlock_page(page);
1063 	put_page(page);
1064 out_allocpage:
1065 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1066 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1067 		/* non-LRU / non-movable temporary page is needed */
1068 		page->mapping = Z_EROFS_MAPPING_STAGING;
1069 		tocache = false;
1070 	}
1071 
1072 	if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1073 		if (tocache) {
1074 			/* since it added to managed cache successfully */
1075 			unlock_page(page);
1076 			put_page(page);
1077 		} else {
1078 			list_add(&page->lru, pagepool);
1079 		}
1080 		cond_resched();
1081 		goto repeat;
1082 	}
1083 	set_page_private(page, (unsigned long)pcl);
1084 	SetPagePrivate(page);
1085 out:	/* the only exit (for tracing and debugging) */
1086 	return page;
1087 }
1088 
1089 static struct z_erofs_decompressqueue *
1090 jobqueue_init(struct super_block *sb,
1091 	      struct z_erofs_decompressqueue *fgq, bool *fg)
1092 {
1093 	struct z_erofs_decompressqueue *q;
1094 
1095 	if (fg && !*fg) {
1096 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1097 		if (!q) {
1098 			*fg = true;
1099 			goto fg_out;
1100 		}
1101 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1102 	} else {
1103 fg_out:
1104 		q = fgq;
1105 		init_waitqueue_head(&fgq->u.wait);
1106 		atomic_set(&fgq->pending_bios, 0);
1107 	}
1108 	q->sb = sb;
1109 	q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1110 	return q;
1111 }
1112 
1113 /* define decompression jobqueue types */
1114 enum {
1115 	JQ_BYPASS,
1116 	JQ_SUBMIT,
1117 	NR_JOBQUEUES,
1118 };
1119 
1120 static void *jobqueueset_init(struct super_block *sb,
1121 			      struct z_erofs_decompressqueue *q[],
1122 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1123 {
1124 	/*
1125 	 * if managed cache is enabled, bypass jobqueue is needed,
1126 	 * no need to read from device for all pclusters in this queue.
1127 	 */
1128 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1129 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1130 
1131 	return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1132 }
1133 
1134 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1135 				    z_erofs_next_pcluster_t qtail[],
1136 				    z_erofs_next_pcluster_t owned_head)
1137 {
1138 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1139 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1140 
1141 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1142 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1143 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1144 
1145 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1146 
1147 	WRITE_ONCE(*submit_qtail, owned_head);
1148 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1149 
1150 	qtail[JQ_BYPASS] = &pcl->next;
1151 }
1152 
1153 static void z_erofs_submit_queue(struct super_block *sb,
1154 				 z_erofs_next_pcluster_t owned_head,
1155 				 struct list_head *pagepool,
1156 				 struct z_erofs_decompressqueue *fgq,
1157 				 bool *force_fg)
1158 {
1159 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
1160 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1161 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1162 	void *bi_private;
1163 	/* since bio will be NULL, no need to initialize last_index */
1164 	pgoff_t uninitialized_var(last_index);
1165 	unsigned int nr_bios = 0;
1166 	struct bio *bio = NULL;
1167 
1168 	bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1169 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1170 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1171 
1172 	/* by default, all need io submission */
1173 	q[JQ_SUBMIT]->head = owned_head;
1174 
1175 	do {
1176 		struct z_erofs_pcluster *pcl;
1177 		pgoff_t cur, end;
1178 		unsigned int i = 0;
1179 		bool bypass = true;
1180 
1181 		/* no possible 'owned_head' equals the following */
1182 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1183 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1184 
1185 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1186 
1187 		cur = pcl->obj.index;
1188 		end = cur + BIT(pcl->clusterbits);
1189 
1190 		/* close the main owned chain at first */
1191 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1192 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1193 
1194 		do {
1195 			struct page *page;
1196 			int err;
1197 
1198 			page = pickup_page_for_submission(pcl, i++, pagepool,
1199 							  MNGD_MAPPING(sbi),
1200 							  GFP_NOFS);
1201 			if (!page)
1202 				continue;
1203 
1204 			if (bio && cur != last_index + 1) {
1205 submit_bio_retry:
1206 				submit_bio(bio);
1207 				bio = NULL;
1208 			}
1209 
1210 			if (!bio) {
1211 				bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1212 
1213 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1214 				bio_set_dev(bio, sb->s_bdev);
1215 				bio->bi_iter.bi_sector = (sector_t)cur <<
1216 					LOG_SECTORS_PER_BLOCK;
1217 				bio->bi_private = bi_private;
1218 				bio->bi_opf = REQ_OP_READ;
1219 				++nr_bios;
1220 			}
1221 
1222 			err = bio_add_page(bio, page, PAGE_SIZE, 0);
1223 			if (err < PAGE_SIZE)
1224 				goto submit_bio_retry;
1225 
1226 			last_index = cur;
1227 			bypass = false;
1228 		} while (++cur < end);
1229 
1230 		if (!bypass)
1231 			qtail[JQ_SUBMIT] = &pcl->next;
1232 		else
1233 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1234 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1235 
1236 	if (bio)
1237 		submit_bio(bio);
1238 
1239 	/*
1240 	 * although background is preferred, no one is pending for submission.
1241 	 * don't issue workqueue for decompression but drop it directly instead.
1242 	 */
1243 	if (!*force_fg && !nr_bios) {
1244 		kvfree(q[JQ_SUBMIT]);
1245 		return;
1246 	}
1247 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1248 }
1249 
1250 static void z_erofs_runqueue(struct super_block *sb,
1251 			     struct z_erofs_collector *clt,
1252 			     struct list_head *pagepool, bool force_fg)
1253 {
1254 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1255 
1256 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
1257 		return;
1258 	z_erofs_submit_queue(sb, clt->owned_head, pagepool, io, &force_fg);
1259 
1260 	/* handle bypass queue (no i/o pclusters) immediately */
1261 	z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1262 
1263 	if (!force_fg)
1264 		return;
1265 
1266 	/* wait until all bios are completed */
1267 	io_wait_event(io[JQ_SUBMIT].u.wait,
1268 		      !atomic_read(&io[JQ_SUBMIT].pending_bios));
1269 
1270 	/* handle synchronous decompress queue in the caller context */
1271 	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1272 }
1273 
1274 static int z_erofs_readpage(struct file *file, struct page *page)
1275 {
1276 	struct inode *const inode = page->mapping->host;
1277 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1278 	int err;
1279 	LIST_HEAD(pagepool);
1280 
1281 	trace_erofs_readpage(page, false);
1282 
1283 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1284 
1285 	err = z_erofs_do_read_page(&f, page, &pagepool);
1286 	(void)z_erofs_collector_end(&f.clt);
1287 
1288 	/* if some compressed cluster ready, need submit them anyway */
1289 	z_erofs_runqueue(inode->i_sb, &f.clt, &pagepool, true);
1290 
1291 	if (err)
1292 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1293 
1294 	if (f.map.mpage)
1295 		put_page(f.map.mpage);
1296 
1297 	/* clean up the remaining free pages */
1298 	put_pages_list(&pagepool);
1299 	return err;
1300 }
1301 
1302 static bool should_decompress_synchronously(struct erofs_sb_info *sbi,
1303 					    unsigned int nr)
1304 {
1305 	return nr <= sbi->ctx.max_sync_decompress_pages;
1306 }
1307 
1308 static void z_erofs_readahead(struct readahead_control *rac)
1309 {
1310 	struct inode *const inode = rac->mapping->host;
1311 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1312 
1313 	bool sync = should_decompress_synchronously(sbi, readahead_count(rac));
1314 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1315 	struct page *page, *head = NULL;
1316 	LIST_HEAD(pagepool);
1317 
1318 	trace_erofs_readpages(inode, readahead_index(rac),
1319 			readahead_count(rac), false);
1320 
1321 	f.headoffset = readahead_pos(rac);
1322 
1323 	while ((page = readahead_page(rac))) {
1324 		prefetchw(&page->flags);
1325 
1326 		/*
1327 		 * A pure asynchronous readahead is indicated if
1328 		 * a PG_readahead marked page is hitted at first.
1329 		 * Let's also do asynchronous decompression for this case.
1330 		 */
1331 		sync &= !(PageReadahead(page) && !head);
1332 
1333 		set_page_private(page, (unsigned long)head);
1334 		head = page;
1335 	}
1336 
1337 	while (head) {
1338 		struct page *page = head;
1339 		int err;
1340 
1341 		/* traversal in reverse order */
1342 		head = (void *)page_private(page);
1343 
1344 		err = z_erofs_do_read_page(&f, page, &pagepool);
1345 		if (err)
1346 			erofs_err(inode->i_sb,
1347 				  "readahead error at page %lu @ nid %llu",
1348 				  page->index, EROFS_I(inode)->nid);
1349 		put_page(page);
1350 	}
1351 
1352 	(void)z_erofs_collector_end(&f.clt);
1353 
1354 	z_erofs_runqueue(inode->i_sb, &f.clt, &pagepool, sync);
1355 
1356 	if (f.map.mpage)
1357 		put_page(f.map.mpage);
1358 
1359 	/* clean up the remaining free pages */
1360 	put_pages_list(&pagepool);
1361 }
1362 
1363 const struct address_space_operations z_erofs_aops = {
1364 	.readpage = z_erofs_readpage,
1365 	.readahead = z_erofs_readahead,
1366 };
1367 
1368