xref: /openbmc/linux/fs/erofs/zdata.c (revision 10756dc5)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2018 HUAWEI, Inc.
4  *             https://www.huawei.com/
5  */
6 #include "zdata.h"
7 #include "compress.h"
8 #include <linux/prefetch.h>
9 
10 #include <trace/events/erofs.h>
11 
12 /*
13  * since pclustersize is variable for big pcluster feature, introduce slab
14  * pools implementation for different pcluster sizes.
15  */
16 struct z_erofs_pcluster_slab {
17 	struct kmem_cache *slab;
18 	unsigned int maxpages;
19 	char name[48];
20 };
21 
22 #define _PCLP(n) { .maxpages = n }
23 
24 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
25 	_PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
26 	_PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
27 };
28 
29 static void z_erofs_destroy_pcluster_pool(void)
30 {
31 	int i;
32 
33 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
34 		if (!pcluster_pool[i].slab)
35 			continue;
36 		kmem_cache_destroy(pcluster_pool[i].slab);
37 		pcluster_pool[i].slab = NULL;
38 	}
39 }
40 
41 static int z_erofs_create_pcluster_pool(void)
42 {
43 	struct z_erofs_pcluster_slab *pcs;
44 	struct z_erofs_pcluster *a;
45 	unsigned int size;
46 
47 	for (pcs = pcluster_pool;
48 	     pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
49 		size = struct_size(a, compressed_pages, pcs->maxpages);
50 
51 		sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
52 		pcs->slab = kmem_cache_create(pcs->name, size, 0,
53 					      SLAB_RECLAIM_ACCOUNT, NULL);
54 		if (pcs->slab)
55 			continue;
56 
57 		z_erofs_destroy_pcluster_pool();
58 		return -ENOMEM;
59 	}
60 	return 0;
61 }
62 
63 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
64 {
65 	int i;
66 
67 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
68 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
69 		struct z_erofs_pcluster *pcl;
70 
71 		if (nrpages > pcs->maxpages)
72 			continue;
73 
74 		pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
75 		if (!pcl)
76 			return ERR_PTR(-ENOMEM);
77 		pcl->pclusterpages = nrpages;
78 		return pcl;
79 	}
80 	return ERR_PTR(-EINVAL);
81 }
82 
83 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
84 {
85 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
86 	int i;
87 
88 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
89 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
90 
91 		if (pclusterpages > pcs->maxpages)
92 			continue;
93 
94 		kmem_cache_free(pcs->slab, pcl);
95 		return;
96 	}
97 	DBG_BUGON(1);
98 }
99 
100 /* how to allocate cached pages for a pcluster */
101 enum z_erofs_cache_alloctype {
102 	DONTALLOC,	/* don't allocate any cached pages */
103 	/*
104 	 * try to use cached I/O if page allocation succeeds or fallback
105 	 * to in-place I/O instead to avoid any direct reclaim.
106 	 */
107 	TRYALLOC,
108 };
109 
110 /*
111  * tagged pointer with 1-bit tag for all compressed pages
112  * tag 0 - the page is just found with an extra page reference
113  */
114 typedef tagptr1_t compressed_page_t;
115 
116 #define tag_compressed_page_justfound(page) \
117 	tagptr_fold(compressed_page_t, page, 1)
118 
119 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
120 
121 void z_erofs_exit_zip_subsystem(void)
122 {
123 	destroy_workqueue(z_erofs_workqueue);
124 	z_erofs_destroy_pcluster_pool();
125 }
126 
127 static inline int z_erofs_init_workqueue(void)
128 {
129 	const unsigned int onlinecpus = num_possible_cpus();
130 
131 	/*
132 	 * no need to spawn too many threads, limiting threads could minimum
133 	 * scheduling overhead, perhaps per-CPU threads should be better?
134 	 */
135 	z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
136 					    WQ_UNBOUND | WQ_HIGHPRI,
137 					    onlinecpus + onlinecpus / 4);
138 	return z_erofs_workqueue ? 0 : -ENOMEM;
139 }
140 
141 int __init z_erofs_init_zip_subsystem(void)
142 {
143 	int err = z_erofs_create_pcluster_pool();
144 
145 	if (err)
146 		return err;
147 	err = z_erofs_init_workqueue();
148 	if (err)
149 		z_erofs_destroy_pcluster_pool();
150 	return err;
151 }
152 
153 enum z_erofs_collectmode {
154 	COLLECT_SECONDARY,
155 	COLLECT_PRIMARY,
156 	/*
157 	 * The current collection was the tail of an exist chain, in addition
158 	 * that the previous processed chained collections are all decided to
159 	 * be hooked up to it.
160 	 * A new chain will be created for the remaining collections which are
161 	 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
162 	 * the next collection cannot reuse the whole page safely in
163 	 * the following scenario:
164 	 *  ________________________________________________________________
165 	 * |      tail (partial) page     |       head (partial) page       |
166 	 * |   (belongs to the next cl)   |   (belongs to the current cl)   |
167 	 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
168 	 */
169 	COLLECT_PRIMARY_HOOKED,
170 	/*
171 	 * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it
172 	 * could be dispatched into bypass queue later due to uptodated managed
173 	 * pages. All related online pages cannot be reused for inplace I/O (or
174 	 * pagevec) since it can be directly decoded without I/O submission.
175 	 */
176 	COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
177 	/*
178 	 * The current collection has been linked with the owned chain, and
179 	 * could also be linked with the remaining collections, which means
180 	 * if the processing page is the tail page of the collection, thus
181 	 * the current collection can safely use the whole page (since
182 	 * the previous collection is under control) for in-place I/O, as
183 	 * illustrated below:
184 	 *  ________________________________________________________________
185 	 * |  tail (partial) page |          head (partial) page           |
186 	 * |  (of the current cl) |      (of the previous collection)      |
187 	 * |  PRIMARY_FOLLOWED or |                                        |
188 	 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
189 	 *
190 	 * [  (*) the above page can be used as inplace I/O.               ]
191 	 */
192 	COLLECT_PRIMARY_FOLLOWED,
193 };
194 
195 struct z_erofs_collector {
196 	struct z_erofs_pagevec_ctor vector;
197 
198 	struct z_erofs_pcluster *pcl, *tailpcl;
199 	struct z_erofs_collection *cl;
200 	/* a pointer used to pick up inplace I/O pages */
201 	struct page **icpage_ptr;
202 	z_erofs_next_pcluster_t owned_head;
203 
204 	enum z_erofs_collectmode mode;
205 };
206 
207 struct z_erofs_decompress_frontend {
208 	struct inode *const inode;
209 
210 	struct z_erofs_collector clt;
211 	struct erofs_map_blocks map;
212 
213 	bool readahead;
214 	/* used for applying cache strategy on the fly */
215 	bool backmost;
216 	erofs_off_t headoffset;
217 };
218 
219 #define COLLECTOR_INIT() { \
220 	.owned_head = Z_EROFS_PCLUSTER_TAIL, \
221 	.mode = COLLECT_PRIMARY_FOLLOWED }
222 
223 #define DECOMPRESS_FRONTEND_INIT(__i) { \
224 	.inode = __i, .clt = COLLECTOR_INIT(), \
225 	.backmost = true, }
226 
227 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
228 static DEFINE_MUTEX(z_pagemap_global_lock);
229 
230 static void preload_compressed_pages(struct z_erofs_collector *clt,
231 				     struct address_space *mc,
232 				     enum z_erofs_cache_alloctype type,
233 				     struct page **pagepool)
234 {
235 	struct z_erofs_pcluster *pcl = clt->pcl;
236 	bool standalone = true;
237 	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
238 			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
239 	struct page **pages;
240 	pgoff_t index;
241 
242 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
243 		return;
244 
245 	pages = pcl->compressed_pages;
246 	index = pcl->obj.index;
247 	for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) {
248 		struct page *page;
249 		compressed_page_t t;
250 		struct page *newpage = NULL;
251 
252 		/* the compressed page was loaded before */
253 		if (READ_ONCE(*pages))
254 			continue;
255 
256 		page = find_get_page(mc, index);
257 
258 		if (page) {
259 			t = tag_compressed_page_justfound(page);
260 		} else {
261 			/* I/O is needed, no possible to decompress directly */
262 			standalone = false;
263 			switch (type) {
264 			case TRYALLOC:
265 				newpage = erofs_allocpage(pagepool, gfp);
266 				if (!newpage)
267 					continue;
268 				set_page_private(newpage,
269 						 Z_EROFS_PREALLOCATED_PAGE);
270 				t = tag_compressed_page_justfound(newpage);
271 				break;
272 			default:        /* DONTALLOC */
273 				continue;
274 			}
275 		}
276 
277 		if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
278 			continue;
279 
280 		if (page)
281 			put_page(page);
282 		else if (newpage)
283 			erofs_pagepool_add(pagepool, newpage);
284 	}
285 
286 	/*
287 	 * don't do inplace I/O if all compressed pages are available in
288 	 * managed cache since it can be moved to the bypass queue instead.
289 	 */
290 	if (standalone)
291 		clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
292 }
293 
294 /* called by erofs_shrinker to get rid of all compressed_pages */
295 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
296 				       struct erofs_workgroup *grp)
297 {
298 	struct z_erofs_pcluster *const pcl =
299 		container_of(grp, struct z_erofs_pcluster, obj);
300 	int i;
301 
302 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
303 	/*
304 	 * refcount of workgroup is now freezed as 1,
305 	 * therefore no need to worry about available decompression users.
306 	 */
307 	for (i = 0; i < pcl->pclusterpages; ++i) {
308 		struct page *page = pcl->compressed_pages[i];
309 
310 		if (!page)
311 			continue;
312 
313 		/* block other users from reclaiming or migrating the page */
314 		if (!trylock_page(page))
315 			return -EBUSY;
316 
317 		if (!erofs_page_is_managed(sbi, page))
318 			continue;
319 
320 		/* barrier is implied in the following 'unlock_page' */
321 		WRITE_ONCE(pcl->compressed_pages[i], NULL);
322 		detach_page_private(page);
323 		unlock_page(page);
324 	}
325 	return 0;
326 }
327 
328 int erofs_try_to_free_cached_page(struct page *page)
329 {
330 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
331 	int ret = 0;	/* 0 - busy */
332 
333 	if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
334 		unsigned int i;
335 
336 		DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
337 		for (i = 0; i < pcl->pclusterpages; ++i) {
338 			if (pcl->compressed_pages[i] == page) {
339 				WRITE_ONCE(pcl->compressed_pages[i], NULL);
340 				ret = 1;
341 				break;
342 			}
343 		}
344 		erofs_workgroup_unfreeze(&pcl->obj, 1);
345 
346 		if (ret)
347 			detach_page_private(page);
348 	}
349 	return ret;
350 }
351 
352 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
353 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
354 				   struct page *page)
355 {
356 	struct z_erofs_pcluster *const pcl = clt->pcl;
357 
358 	while (clt->icpage_ptr > pcl->compressed_pages)
359 		if (!cmpxchg(--clt->icpage_ptr, NULL, page))
360 			return true;
361 	return false;
362 }
363 
364 /* callers must be with collection lock held */
365 static int z_erofs_attach_page(struct z_erofs_collector *clt,
366 			       struct page *page, enum z_erofs_page_type type,
367 			       bool pvec_safereuse)
368 {
369 	int ret;
370 
371 	/* give priority for inplaceio */
372 	if (clt->mode >= COLLECT_PRIMARY &&
373 	    type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
374 	    z_erofs_try_inplace_io(clt, page))
375 		return 0;
376 
377 	ret = z_erofs_pagevec_enqueue(&clt->vector, page, type,
378 				      pvec_safereuse);
379 	clt->cl->vcnt += (unsigned int)ret;
380 	return ret ? 0 : -EAGAIN;
381 }
382 
383 static void z_erofs_try_to_claim_pcluster(struct z_erofs_collector *clt)
384 {
385 	struct z_erofs_pcluster *pcl = clt->pcl;
386 	z_erofs_next_pcluster_t *owned_head = &clt->owned_head;
387 
388 	/* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
389 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
390 		    *owned_head) == Z_EROFS_PCLUSTER_NIL) {
391 		*owned_head = &pcl->next;
392 		/* so we can attach this pcluster to our submission chain. */
393 		clt->mode = COLLECT_PRIMARY_FOLLOWED;
394 		return;
395 	}
396 
397 	/*
398 	 * type 2, link to the end of an existing open chain, be careful
399 	 * that its submission is controlled by the original attached chain.
400 	 */
401 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
402 		    *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
403 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
404 		clt->mode = COLLECT_PRIMARY_HOOKED;
405 		clt->tailpcl = NULL;
406 		return;
407 	}
408 	/* type 3, it belongs to a chain, but it isn't the end of the chain */
409 	clt->mode = COLLECT_PRIMARY;
410 }
411 
412 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
413 				     struct inode *inode,
414 				     struct erofs_map_blocks *map)
415 {
416 	struct z_erofs_pcluster *pcl = clt->pcl;
417 	struct z_erofs_collection *cl;
418 	unsigned int length;
419 
420 	/* to avoid unexpected loop formed by corrupted images */
421 	if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
422 		DBG_BUGON(1);
423 		return -EFSCORRUPTED;
424 	}
425 
426 	cl = z_erofs_primarycollection(pcl);
427 	if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
428 		DBG_BUGON(1);
429 		return -EFSCORRUPTED;
430 	}
431 
432 	length = READ_ONCE(pcl->length);
433 	if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
434 		if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
435 			DBG_BUGON(1);
436 			return -EFSCORRUPTED;
437 		}
438 	} else {
439 		unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
440 
441 		if (map->m_flags & EROFS_MAP_FULL_MAPPED)
442 			llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
443 
444 		while (llen > length &&
445 		       length != cmpxchg_relaxed(&pcl->length, length, llen)) {
446 			cpu_relax();
447 			length = READ_ONCE(pcl->length);
448 		}
449 	}
450 	mutex_lock(&cl->lock);
451 	/* used to check tail merging loop due to corrupted images */
452 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
453 		clt->tailpcl = pcl;
454 
455 	z_erofs_try_to_claim_pcluster(clt);
456 	clt->cl = cl;
457 	return 0;
458 }
459 
460 static int z_erofs_register_collection(struct z_erofs_collector *clt,
461 				       struct inode *inode,
462 				       struct erofs_map_blocks *map)
463 {
464 	bool ztailpacking = map->m_flags & EROFS_MAP_META;
465 	struct z_erofs_pcluster *pcl;
466 	struct z_erofs_collection *cl;
467 	struct erofs_workgroup *grp;
468 	int err;
469 
470 	if (!(map->m_flags & EROFS_MAP_ENCODED)) {
471 		DBG_BUGON(1);
472 		return -EFSCORRUPTED;
473 	}
474 
475 	/* no available pcluster, let's allocate one */
476 	pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
477 				     map->m_plen >> PAGE_SHIFT);
478 	if (IS_ERR(pcl))
479 		return PTR_ERR(pcl);
480 
481 	atomic_set(&pcl->obj.refcount, 1);
482 	pcl->algorithmformat = map->m_algorithmformat;
483 	pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
484 		(map->m_flags & EROFS_MAP_FULL_MAPPED ?
485 			Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
486 
487 	/* new pclusters should be claimed as type 1, primary and followed */
488 	pcl->next = clt->owned_head;
489 	clt->mode = COLLECT_PRIMARY_FOLLOWED;
490 
491 	cl = z_erofs_primarycollection(pcl);
492 	cl->pageofs = map->m_la & ~PAGE_MASK;
493 
494 	/*
495 	 * lock all primary followed works before visible to others
496 	 * and mutex_trylock *never* fails for a new pcluster.
497 	 */
498 	mutex_init(&cl->lock);
499 	DBG_BUGON(!mutex_trylock(&cl->lock));
500 
501 	if (ztailpacking) {
502 		pcl->obj.index = 0;	/* which indicates ztailpacking */
503 		pcl->pageofs_in = erofs_blkoff(map->m_pa);
504 		pcl->tailpacking_size = map->m_plen;
505 	} else {
506 		pcl->obj.index = map->m_pa >> PAGE_SHIFT;
507 
508 		grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
509 		if (IS_ERR(grp)) {
510 			err = PTR_ERR(grp);
511 			goto err_out;
512 		}
513 
514 		if (grp != &pcl->obj) {
515 			clt->pcl = container_of(grp,
516 					struct z_erofs_pcluster, obj);
517 			err = -EEXIST;
518 			goto err_out;
519 		}
520 	}
521 	/* used to check tail merging loop due to corrupted images */
522 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
523 		clt->tailpcl = pcl;
524 	clt->owned_head = &pcl->next;
525 	clt->pcl = pcl;
526 	clt->cl = cl;
527 	return 0;
528 
529 err_out:
530 	mutex_unlock(&cl->lock);
531 	z_erofs_free_pcluster(pcl);
532 	return err;
533 }
534 
535 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
536 				   struct inode *inode,
537 				   struct erofs_map_blocks *map)
538 {
539 	struct erofs_workgroup *grp;
540 	int ret;
541 
542 	DBG_BUGON(clt->cl);
543 
544 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
545 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
546 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
547 
548 	if (map->m_flags & EROFS_MAP_META) {
549 		if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
550 			DBG_BUGON(1);
551 			return -EFSCORRUPTED;
552 		}
553 		goto tailpacking;
554 	}
555 
556 	grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
557 	if (grp) {
558 		clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
559 	} else {
560 tailpacking:
561 		ret = z_erofs_register_collection(clt, inode, map);
562 		if (!ret)
563 			goto out;
564 		if (ret != -EEXIST)
565 			return ret;
566 	}
567 
568 	ret = z_erofs_lookup_collection(clt, inode, map);
569 	if (ret) {
570 		erofs_workgroup_put(&clt->pcl->obj);
571 		return ret;
572 	}
573 
574 out:
575 	z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
576 				  clt->cl->pagevec, clt->cl->vcnt);
577 	/* since file-backed online pages are traversed in reverse order */
578 	clt->icpage_ptr = clt->pcl->compressed_pages +
579 			z_erofs_pclusterpages(clt->pcl);
580 	return 0;
581 }
582 
583 /*
584  * keep in mind that no referenced pclusters will be freed
585  * only after a RCU grace period.
586  */
587 static void z_erofs_rcu_callback(struct rcu_head *head)
588 {
589 	struct z_erofs_collection *const cl =
590 		container_of(head, struct z_erofs_collection, rcu);
591 
592 	z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster,
593 					   primary_collection));
594 }
595 
596 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
597 {
598 	struct z_erofs_pcluster *const pcl =
599 		container_of(grp, struct z_erofs_pcluster, obj);
600 	struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
601 
602 	call_rcu(&cl->rcu, z_erofs_rcu_callback);
603 }
604 
605 static void z_erofs_collection_put(struct z_erofs_collection *cl)
606 {
607 	struct z_erofs_pcluster *const pcl =
608 		container_of(cl, struct z_erofs_pcluster, primary_collection);
609 
610 	erofs_workgroup_put(&pcl->obj);
611 }
612 
613 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
614 {
615 	struct z_erofs_collection *cl = clt->cl;
616 
617 	if (!cl)
618 		return false;
619 
620 	z_erofs_pagevec_ctor_exit(&clt->vector, false);
621 	mutex_unlock(&cl->lock);
622 
623 	/*
624 	 * if all pending pages are added, don't hold its reference
625 	 * any longer if the pcluster isn't hosted by ourselves.
626 	 */
627 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
628 		z_erofs_collection_put(cl);
629 
630 	clt->cl = NULL;
631 	return true;
632 }
633 
634 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
635 				       unsigned int cachestrategy,
636 				       erofs_off_t la)
637 {
638 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
639 		return false;
640 
641 	if (fe->backmost)
642 		return true;
643 
644 	return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
645 		la < fe->headoffset;
646 }
647 
648 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
649 				struct page *page, struct page **pagepool)
650 {
651 	struct inode *const inode = fe->inode;
652 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
653 	struct erofs_map_blocks *const map = &fe->map;
654 	struct z_erofs_collector *const clt = &fe->clt;
655 	const loff_t offset = page_offset(page);
656 	bool tight = true;
657 
658 	enum z_erofs_cache_alloctype cache_strategy;
659 	enum z_erofs_page_type page_type;
660 	unsigned int cur, end, spiltted, index;
661 	int err = 0;
662 
663 	/* register locked file pages as online pages in pack */
664 	z_erofs_onlinepage_init(page);
665 
666 	spiltted = 0;
667 	end = PAGE_SIZE;
668 repeat:
669 	cur = end - 1;
670 
671 	/* lucky, within the range of the current map_blocks */
672 	if (offset + cur >= map->m_la &&
673 	    offset + cur < map->m_la + map->m_llen) {
674 		/* didn't get a valid collection previously (very rare) */
675 		if (!clt->cl)
676 			goto restart_now;
677 		goto hitted;
678 	}
679 
680 	/* go ahead the next map_blocks */
681 	erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
682 
683 	if (z_erofs_collector_end(clt))
684 		fe->backmost = false;
685 
686 	map->m_la = offset + cur;
687 	map->m_llen = 0;
688 	err = z_erofs_map_blocks_iter(inode, map, 0);
689 	if (err)
690 		goto err_out;
691 
692 restart_now:
693 	if (!(map->m_flags & EROFS_MAP_MAPPED))
694 		goto hitted;
695 
696 	err = z_erofs_collector_begin(clt, inode, map);
697 	if (err)
698 		goto err_out;
699 
700 	if (z_erofs_is_inline_pcluster(clt->pcl)) {
701 		void *mp;
702 
703 		mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
704 					erofs_blknr(map->m_pa), EROFS_NO_KMAP);
705 		if (IS_ERR(mp)) {
706 			err = PTR_ERR(mp);
707 			erofs_err(inode->i_sb,
708 				  "failed to get inline page, err %d", err);
709 			goto err_out;
710 		}
711 		get_page(fe->map.buf.page);
712 		WRITE_ONCE(clt->pcl->compressed_pages[0], fe->map.buf.page);
713 		clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
714 	} else {
715 		/* preload all compressed pages (can change mode if needed) */
716 		if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy,
717 					       map->m_la))
718 			cache_strategy = TRYALLOC;
719 		else
720 			cache_strategy = DONTALLOC;
721 
722 		preload_compressed_pages(clt, MNGD_MAPPING(sbi),
723 					 cache_strategy, pagepool);
724 	}
725 
726 hitted:
727 	/*
728 	 * Ensure the current partial page belongs to this submit chain rather
729 	 * than other concurrent submit chains or the noio(bypass) chain since
730 	 * those chains are handled asynchronously thus the page cannot be used
731 	 * for inplace I/O or pagevec (should be processed in strict order.)
732 	 */
733 	tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
734 		  clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
735 
736 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
737 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
738 		zero_user_segment(page, cur, end);
739 		goto next_part;
740 	}
741 
742 	/* let's derive page type */
743 	page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
744 		(!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
745 			(tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
746 				Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
747 
748 	if (cur)
749 		tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
750 
751 retry:
752 	err = z_erofs_attach_page(clt, page, page_type,
753 				  clt->mode >= COLLECT_PRIMARY_FOLLOWED);
754 	/* should allocate an additional short-lived page for pagevec */
755 	if (err == -EAGAIN) {
756 		struct page *const newpage =
757 				alloc_page(GFP_NOFS | __GFP_NOFAIL);
758 
759 		set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE);
760 		err = z_erofs_attach_page(clt, newpage,
761 					  Z_EROFS_PAGE_TYPE_EXCLUSIVE, true);
762 		if (!err)
763 			goto retry;
764 	}
765 
766 	if (err)
767 		goto err_out;
768 
769 	index = page->index - (map->m_la >> PAGE_SHIFT);
770 
771 	z_erofs_onlinepage_fixup(page, index, true);
772 
773 	/* bump up the number of spiltted parts of a page */
774 	++spiltted;
775 	/* also update nr_pages */
776 	clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
777 next_part:
778 	/* can be used for verification */
779 	map->m_llen = offset + cur - map->m_la;
780 
781 	end = cur;
782 	if (end > 0)
783 		goto repeat;
784 
785 out:
786 	z_erofs_onlinepage_endio(page);
787 
788 	erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
789 		  __func__, page, spiltted, map->m_llen);
790 	return err;
791 
792 	/* if some error occurred while processing this page */
793 err_out:
794 	SetPageError(page);
795 	goto out;
796 }
797 
798 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi,
799 				       unsigned int readahead_pages)
800 {
801 	/* auto: enable for readpage, disable for readahead */
802 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
803 	    !readahead_pages)
804 		return true;
805 
806 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
807 	    (readahead_pages <= sbi->opt.max_sync_decompress_pages))
808 		return true;
809 
810 	return false;
811 }
812 
813 static void z_erofs_decompressqueue_work(struct work_struct *work);
814 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
815 				       bool sync, int bios)
816 {
817 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
818 
819 	/* wake up the caller thread for sync decompression */
820 	if (sync) {
821 		unsigned long flags;
822 
823 		spin_lock_irqsave(&io->u.wait.lock, flags);
824 		if (!atomic_add_return(bios, &io->pending_bios))
825 			wake_up_locked(&io->u.wait);
826 		spin_unlock_irqrestore(&io->u.wait.lock, flags);
827 		return;
828 	}
829 
830 	if (atomic_add_return(bios, &io->pending_bios))
831 		return;
832 	/* Use workqueue and sync decompression for atomic contexts only */
833 	if (in_atomic() || irqs_disabled()) {
834 		queue_work(z_erofs_workqueue, &io->u.work);
835 		/* enable sync decompression for readahead */
836 		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
837 			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
838 		return;
839 	}
840 	z_erofs_decompressqueue_work(&io->u.work);
841 }
842 
843 static bool z_erofs_page_is_invalidated(struct page *page)
844 {
845 	return !page->mapping && !z_erofs_is_shortlived_page(page);
846 }
847 
848 static void z_erofs_decompressqueue_endio(struct bio *bio)
849 {
850 	tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
851 	struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
852 	blk_status_t err = bio->bi_status;
853 	struct bio_vec *bvec;
854 	struct bvec_iter_all iter_all;
855 
856 	bio_for_each_segment_all(bvec, bio, iter_all) {
857 		struct page *page = bvec->bv_page;
858 
859 		DBG_BUGON(PageUptodate(page));
860 		DBG_BUGON(z_erofs_page_is_invalidated(page));
861 
862 		if (err)
863 			SetPageError(page);
864 
865 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
866 			if (!err)
867 				SetPageUptodate(page);
868 			unlock_page(page);
869 		}
870 	}
871 	z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
872 	bio_put(bio);
873 }
874 
875 static int z_erofs_decompress_pcluster(struct super_block *sb,
876 				       struct z_erofs_pcluster *pcl,
877 				       struct page **pagepool)
878 {
879 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
880 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
881 	struct z_erofs_pagevec_ctor ctor;
882 	unsigned int i, inputsize, outputsize, llen, nr_pages;
883 	struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
884 	struct page **pages, **compressed_pages, *page;
885 
886 	enum z_erofs_page_type page_type;
887 	bool overlapped, partial;
888 	struct z_erofs_collection *cl;
889 	int err;
890 
891 	might_sleep();
892 	cl = z_erofs_primarycollection(pcl);
893 	DBG_BUGON(!READ_ONCE(cl->nr_pages));
894 
895 	mutex_lock(&cl->lock);
896 	nr_pages = cl->nr_pages;
897 
898 	if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
899 		pages = pages_onstack;
900 	} else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
901 		   mutex_trylock(&z_pagemap_global_lock)) {
902 		pages = z_pagemap_global;
903 	} else {
904 		gfp_t gfp_flags = GFP_KERNEL;
905 
906 		if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
907 			gfp_flags |= __GFP_NOFAIL;
908 
909 		pages = kvmalloc_array(nr_pages, sizeof(struct page *),
910 				       gfp_flags);
911 
912 		/* fallback to global pagemap for the lowmem scenario */
913 		if (!pages) {
914 			mutex_lock(&z_pagemap_global_lock);
915 			pages = z_pagemap_global;
916 		}
917 	}
918 
919 	for (i = 0; i < nr_pages; ++i)
920 		pages[i] = NULL;
921 
922 	err = 0;
923 	z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
924 				  cl->pagevec, 0);
925 
926 	for (i = 0; i < cl->vcnt; ++i) {
927 		unsigned int pagenr;
928 
929 		page = z_erofs_pagevec_dequeue(&ctor, &page_type);
930 
931 		/* all pages in pagevec ought to be valid */
932 		DBG_BUGON(!page);
933 		DBG_BUGON(z_erofs_page_is_invalidated(page));
934 
935 		if (z_erofs_put_shortlivedpage(pagepool, page))
936 			continue;
937 
938 		if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
939 			pagenr = 0;
940 		else
941 			pagenr = z_erofs_onlinepage_index(page);
942 
943 		DBG_BUGON(pagenr >= nr_pages);
944 
945 		/*
946 		 * currently EROFS doesn't support multiref(dedup),
947 		 * so here erroring out one multiref page.
948 		 */
949 		if (pages[pagenr]) {
950 			DBG_BUGON(1);
951 			SetPageError(pages[pagenr]);
952 			z_erofs_onlinepage_endio(pages[pagenr]);
953 			err = -EFSCORRUPTED;
954 		}
955 		pages[pagenr] = page;
956 	}
957 	z_erofs_pagevec_ctor_exit(&ctor, true);
958 
959 	overlapped = false;
960 	compressed_pages = pcl->compressed_pages;
961 
962 	for (i = 0; i < pclusterpages; ++i) {
963 		unsigned int pagenr;
964 
965 		page = compressed_pages[i];
966 		/* all compressed pages ought to be valid */
967 		DBG_BUGON(!page);
968 
969 		if (z_erofs_is_inline_pcluster(pcl)) {
970 			if (!PageUptodate(page))
971 				err = -EIO;
972 			continue;
973 		}
974 
975 		DBG_BUGON(z_erofs_page_is_invalidated(page));
976 		if (!z_erofs_is_shortlived_page(page)) {
977 			if (erofs_page_is_managed(sbi, page)) {
978 				if (!PageUptodate(page))
979 					err = -EIO;
980 				continue;
981 			}
982 
983 			/*
984 			 * only if non-head page can be selected
985 			 * for inplace decompression
986 			 */
987 			pagenr = z_erofs_onlinepage_index(page);
988 
989 			DBG_BUGON(pagenr >= nr_pages);
990 			if (pages[pagenr]) {
991 				DBG_BUGON(1);
992 				SetPageError(pages[pagenr]);
993 				z_erofs_onlinepage_endio(pages[pagenr]);
994 				err = -EFSCORRUPTED;
995 			}
996 			pages[pagenr] = page;
997 
998 			overlapped = true;
999 		}
1000 
1001 		/* PG_error needs checking for all non-managed pages */
1002 		if (PageError(page)) {
1003 			DBG_BUGON(PageUptodate(page));
1004 			err = -EIO;
1005 		}
1006 	}
1007 
1008 	if (err)
1009 		goto out;
1010 
1011 	llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
1012 	if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
1013 		outputsize = llen;
1014 		partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
1015 	} else {
1016 		outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
1017 		partial = true;
1018 	}
1019 
1020 	if (z_erofs_is_inline_pcluster(pcl))
1021 		inputsize = pcl->tailpacking_size;
1022 	else
1023 		inputsize = pclusterpages * PAGE_SIZE;
1024 
1025 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
1026 					.sb = sb,
1027 					.in = compressed_pages,
1028 					.out = pages,
1029 					.pageofs_in = pcl->pageofs_in,
1030 					.pageofs_out = cl->pageofs,
1031 					.inputsize = inputsize,
1032 					.outputsize = outputsize,
1033 					.alg = pcl->algorithmformat,
1034 					.inplace_io = overlapped,
1035 					.partial_decoding = partial
1036 				 }, pagepool);
1037 
1038 out:
1039 	/* must handle all compressed pages before actual file pages */
1040 	if (z_erofs_is_inline_pcluster(pcl)) {
1041 		page = compressed_pages[0];
1042 		WRITE_ONCE(compressed_pages[0], NULL);
1043 		put_page(page);
1044 	} else {
1045 		for (i = 0; i < pclusterpages; ++i) {
1046 			page = compressed_pages[i];
1047 
1048 			if (erofs_page_is_managed(sbi, page))
1049 				continue;
1050 
1051 			/* recycle all individual short-lived pages */
1052 			(void)z_erofs_put_shortlivedpage(pagepool, page);
1053 			WRITE_ONCE(compressed_pages[i], NULL);
1054 		}
1055 	}
1056 
1057 	for (i = 0; i < nr_pages; ++i) {
1058 		page = pages[i];
1059 		if (!page)
1060 			continue;
1061 
1062 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1063 
1064 		/* recycle all individual short-lived pages */
1065 		if (z_erofs_put_shortlivedpage(pagepool, page))
1066 			continue;
1067 
1068 		if (err < 0)
1069 			SetPageError(page);
1070 
1071 		z_erofs_onlinepage_endio(page);
1072 	}
1073 
1074 	if (pages == z_pagemap_global)
1075 		mutex_unlock(&z_pagemap_global_lock);
1076 	else if (pages != pages_onstack)
1077 		kvfree(pages);
1078 
1079 	cl->nr_pages = 0;
1080 	cl->vcnt = 0;
1081 
1082 	/* all cl locks MUST be taken before the following line */
1083 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1084 
1085 	/* all cl locks SHOULD be released right now */
1086 	mutex_unlock(&cl->lock);
1087 
1088 	z_erofs_collection_put(cl);
1089 	return err;
1090 }
1091 
1092 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1093 				     struct page **pagepool)
1094 {
1095 	z_erofs_next_pcluster_t owned = io->head;
1096 
1097 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1098 		struct z_erofs_pcluster *pcl;
1099 
1100 		/* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1101 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1102 
1103 		/* no possible that 'owned' equals NULL */
1104 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1105 
1106 		pcl = container_of(owned, struct z_erofs_pcluster, next);
1107 		owned = READ_ONCE(pcl->next);
1108 
1109 		z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
1110 	}
1111 }
1112 
1113 static void z_erofs_decompressqueue_work(struct work_struct *work)
1114 {
1115 	struct z_erofs_decompressqueue *bgq =
1116 		container_of(work, struct z_erofs_decompressqueue, u.work);
1117 	struct page *pagepool = NULL;
1118 
1119 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1120 	z_erofs_decompress_queue(bgq, &pagepool);
1121 
1122 	erofs_release_pages(&pagepool);
1123 	kvfree(bgq);
1124 }
1125 
1126 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1127 					       unsigned int nr,
1128 					       struct page **pagepool,
1129 					       struct address_space *mc,
1130 					       gfp_t gfp)
1131 {
1132 	const pgoff_t index = pcl->obj.index;
1133 	bool tocache = false;
1134 
1135 	struct address_space *mapping;
1136 	struct page *oldpage, *page;
1137 
1138 	compressed_page_t t;
1139 	int justfound;
1140 
1141 repeat:
1142 	page = READ_ONCE(pcl->compressed_pages[nr]);
1143 	oldpage = page;
1144 
1145 	if (!page)
1146 		goto out_allocpage;
1147 
1148 	/* process the target tagged pointer */
1149 	t = tagptr_init(compressed_page_t, page);
1150 	justfound = tagptr_unfold_tags(t);
1151 	page = tagptr_unfold_ptr(t);
1152 
1153 	/*
1154 	 * preallocated cached pages, which is used to avoid direct reclaim
1155 	 * otherwise, it will go inplace I/O path instead.
1156 	 */
1157 	if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1158 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1159 		set_page_private(page, 0);
1160 		tocache = true;
1161 		goto out_tocache;
1162 	}
1163 	mapping = READ_ONCE(page->mapping);
1164 
1165 	/*
1166 	 * file-backed online pages in plcuster are all locked steady,
1167 	 * therefore it is impossible for `mapping' to be NULL.
1168 	 */
1169 	if (mapping && mapping != mc)
1170 		/* ought to be unmanaged pages */
1171 		goto out;
1172 
1173 	/* directly return for shortlived page as well */
1174 	if (z_erofs_is_shortlived_page(page))
1175 		goto out;
1176 
1177 	lock_page(page);
1178 
1179 	/* only true if page reclaim goes wrong, should never happen */
1180 	DBG_BUGON(justfound && PagePrivate(page));
1181 
1182 	/* the page is still in manage cache */
1183 	if (page->mapping == mc) {
1184 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1185 
1186 		ClearPageError(page);
1187 		if (!PagePrivate(page)) {
1188 			/*
1189 			 * impossible to be !PagePrivate(page) for
1190 			 * the current restriction as well if
1191 			 * the page is already in compressed_pages[].
1192 			 */
1193 			DBG_BUGON(!justfound);
1194 
1195 			justfound = 0;
1196 			set_page_private(page, (unsigned long)pcl);
1197 			SetPagePrivate(page);
1198 		}
1199 
1200 		/* no need to submit io if it is already up-to-date */
1201 		if (PageUptodate(page)) {
1202 			unlock_page(page);
1203 			page = NULL;
1204 		}
1205 		goto out;
1206 	}
1207 
1208 	/*
1209 	 * the managed page has been truncated, it's unsafe to
1210 	 * reuse this one, let's allocate a new cache-managed page.
1211 	 */
1212 	DBG_BUGON(page->mapping);
1213 	DBG_BUGON(!justfound);
1214 
1215 	tocache = true;
1216 	unlock_page(page);
1217 	put_page(page);
1218 out_allocpage:
1219 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1220 	if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1221 		erofs_pagepool_add(pagepool, page);
1222 		cond_resched();
1223 		goto repeat;
1224 	}
1225 out_tocache:
1226 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1227 		/* turn into temporary page if fails (1 ref) */
1228 		set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1229 		goto out;
1230 	}
1231 	attach_page_private(page, pcl);
1232 	/* drop a refcount added by allocpage (then we have 2 refs here) */
1233 	put_page(page);
1234 
1235 out:	/* the only exit (for tracing and debugging) */
1236 	return page;
1237 }
1238 
1239 static struct z_erofs_decompressqueue *
1240 jobqueue_init(struct super_block *sb,
1241 	      struct z_erofs_decompressqueue *fgq, bool *fg)
1242 {
1243 	struct z_erofs_decompressqueue *q;
1244 
1245 	if (fg && !*fg) {
1246 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1247 		if (!q) {
1248 			*fg = true;
1249 			goto fg_out;
1250 		}
1251 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1252 	} else {
1253 fg_out:
1254 		q = fgq;
1255 		init_waitqueue_head(&fgq->u.wait);
1256 		atomic_set(&fgq->pending_bios, 0);
1257 	}
1258 	q->sb = sb;
1259 	q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1260 	return q;
1261 }
1262 
1263 /* define decompression jobqueue types */
1264 enum {
1265 	JQ_BYPASS,
1266 	JQ_SUBMIT,
1267 	NR_JOBQUEUES,
1268 };
1269 
1270 static void *jobqueueset_init(struct super_block *sb,
1271 			      struct z_erofs_decompressqueue *q[],
1272 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1273 {
1274 	/*
1275 	 * if managed cache is enabled, bypass jobqueue is needed,
1276 	 * no need to read from device for all pclusters in this queue.
1277 	 */
1278 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1279 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1280 
1281 	return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1282 }
1283 
1284 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1285 				    z_erofs_next_pcluster_t qtail[],
1286 				    z_erofs_next_pcluster_t owned_head)
1287 {
1288 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1289 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1290 
1291 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1292 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1293 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1294 
1295 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1296 
1297 	WRITE_ONCE(*submit_qtail, owned_head);
1298 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1299 
1300 	qtail[JQ_BYPASS] = &pcl->next;
1301 }
1302 
1303 static void z_erofs_submit_queue(struct super_block *sb,
1304 				 struct z_erofs_decompress_frontend *f,
1305 				 struct page **pagepool,
1306 				 struct z_erofs_decompressqueue *fgq,
1307 				 bool *force_fg)
1308 {
1309 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
1310 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1311 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1312 	void *bi_private;
1313 	z_erofs_next_pcluster_t owned_head = f->clt.owned_head;
1314 	/* bio is NULL initially, so no need to initialize last_{index,bdev} */
1315 	pgoff_t last_index;
1316 	struct block_device *last_bdev;
1317 	unsigned int nr_bios = 0;
1318 	struct bio *bio = NULL;
1319 
1320 	bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1321 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1322 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1323 
1324 	/* by default, all need io submission */
1325 	q[JQ_SUBMIT]->head = owned_head;
1326 
1327 	do {
1328 		struct erofs_map_dev mdev;
1329 		struct z_erofs_pcluster *pcl;
1330 		pgoff_t cur, end;
1331 		unsigned int i = 0;
1332 		bool bypass = true;
1333 
1334 		/* no possible 'owned_head' equals the following */
1335 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1336 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1337 
1338 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1339 
1340 		/* close the main owned chain at first */
1341 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1342 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1343 		if (z_erofs_is_inline_pcluster(pcl)) {
1344 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1345 			continue;
1346 		}
1347 
1348 		/* no device id here, thus it will always succeed */
1349 		mdev = (struct erofs_map_dev) {
1350 			.m_pa = blknr_to_addr(pcl->obj.index),
1351 		};
1352 		(void)erofs_map_dev(sb, &mdev);
1353 
1354 		cur = erofs_blknr(mdev.m_pa);
1355 		end = cur + pcl->pclusterpages;
1356 
1357 		do {
1358 			struct page *page;
1359 
1360 			page = pickup_page_for_submission(pcl, i++, pagepool,
1361 							  MNGD_MAPPING(sbi),
1362 							  GFP_NOFS);
1363 			if (!page)
1364 				continue;
1365 
1366 			if (bio && (cur != last_index + 1 ||
1367 				    last_bdev != mdev.m_bdev)) {
1368 submit_bio_retry:
1369 				submit_bio(bio);
1370 				bio = NULL;
1371 			}
1372 
1373 			if (!bio) {
1374 				bio = bio_alloc(GFP_NOIO, BIO_MAX_VECS);
1375 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1376 
1377 				bio_set_dev(bio, mdev.m_bdev);
1378 				last_bdev = mdev.m_bdev;
1379 				bio->bi_iter.bi_sector = (sector_t)cur <<
1380 					LOG_SECTORS_PER_BLOCK;
1381 				bio->bi_private = bi_private;
1382 				bio->bi_opf = REQ_OP_READ;
1383 				if (f->readahead)
1384 					bio->bi_opf |= REQ_RAHEAD;
1385 				++nr_bios;
1386 			}
1387 
1388 			if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1389 				goto submit_bio_retry;
1390 
1391 			last_index = cur;
1392 			bypass = false;
1393 		} while (++cur < end);
1394 
1395 		if (!bypass)
1396 			qtail[JQ_SUBMIT] = &pcl->next;
1397 		else
1398 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1399 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1400 
1401 	if (bio)
1402 		submit_bio(bio);
1403 
1404 	/*
1405 	 * although background is preferred, no one is pending for submission.
1406 	 * don't issue workqueue for decompression but drop it directly instead.
1407 	 */
1408 	if (!*force_fg && !nr_bios) {
1409 		kvfree(q[JQ_SUBMIT]);
1410 		return;
1411 	}
1412 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1413 }
1414 
1415 static void z_erofs_runqueue(struct super_block *sb,
1416 			     struct z_erofs_decompress_frontend *f,
1417 			     struct page **pagepool, bool force_fg)
1418 {
1419 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1420 
1421 	if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL)
1422 		return;
1423 	z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1424 
1425 	/* handle bypass queue (no i/o pclusters) immediately */
1426 	z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1427 
1428 	if (!force_fg)
1429 		return;
1430 
1431 	/* wait until all bios are completed */
1432 	io_wait_event(io[JQ_SUBMIT].u.wait,
1433 		      !atomic_read(&io[JQ_SUBMIT].pending_bios));
1434 
1435 	/* handle synchronous decompress queue in the caller context */
1436 	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1437 }
1438 
1439 /*
1440  * Since partial uptodate is still unimplemented for now, we have to use
1441  * approximate readmore strategies as a start.
1442  */
1443 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1444 				      struct readahead_control *rac,
1445 				      erofs_off_t end,
1446 				      struct page **pagepool,
1447 				      bool backmost)
1448 {
1449 	struct inode *inode = f->inode;
1450 	struct erofs_map_blocks *map = &f->map;
1451 	erofs_off_t cur;
1452 	int err;
1453 
1454 	if (backmost) {
1455 		map->m_la = end;
1456 		err = z_erofs_map_blocks_iter(inode, map,
1457 					      EROFS_GET_BLOCKS_READMORE);
1458 		if (err)
1459 			return;
1460 
1461 		/* expend ra for the trailing edge if readahead */
1462 		if (rac) {
1463 			loff_t newstart = readahead_pos(rac);
1464 
1465 			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1466 			readahead_expand(rac, newstart, cur - newstart);
1467 			return;
1468 		}
1469 		end = round_up(end, PAGE_SIZE);
1470 	} else {
1471 		end = round_up(map->m_la, PAGE_SIZE);
1472 
1473 		if (!map->m_llen)
1474 			return;
1475 	}
1476 
1477 	cur = map->m_la + map->m_llen - 1;
1478 	while (cur >= end) {
1479 		pgoff_t index = cur >> PAGE_SHIFT;
1480 		struct page *page;
1481 
1482 		page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1483 		if (!page)
1484 			goto skip;
1485 
1486 		if (PageUptodate(page)) {
1487 			unlock_page(page);
1488 			put_page(page);
1489 			goto skip;
1490 		}
1491 
1492 		err = z_erofs_do_read_page(f, page, pagepool);
1493 		if (err)
1494 			erofs_err(inode->i_sb,
1495 				  "readmore error at page %lu @ nid %llu",
1496 				  index, EROFS_I(inode)->nid);
1497 		put_page(page);
1498 skip:
1499 		if (cur < PAGE_SIZE)
1500 			break;
1501 		cur = (index << PAGE_SHIFT) - 1;
1502 	}
1503 }
1504 
1505 static int z_erofs_readpage(struct file *file, struct page *page)
1506 {
1507 	struct inode *const inode = page->mapping->host;
1508 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1509 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1510 	struct page *pagepool = NULL;
1511 	int err;
1512 
1513 	trace_erofs_readpage(page, false);
1514 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1515 
1516 	z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1517 				  &pagepool, true);
1518 	err = z_erofs_do_read_page(&f, page, &pagepool);
1519 	z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1520 
1521 	(void)z_erofs_collector_end(&f.clt);
1522 
1523 	/* if some compressed cluster ready, need submit them anyway */
1524 	z_erofs_runqueue(inode->i_sb, &f, &pagepool,
1525 			 z_erofs_get_sync_decompress_policy(sbi, 0));
1526 
1527 	if (err)
1528 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1529 
1530 	erofs_put_metabuf(&f.map.buf);
1531 	erofs_release_pages(&pagepool);
1532 	return err;
1533 }
1534 
1535 static void z_erofs_readahead(struct readahead_control *rac)
1536 {
1537 	struct inode *const inode = rac->mapping->host;
1538 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1539 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1540 	struct page *pagepool = NULL, *head = NULL, *page;
1541 	unsigned int nr_pages;
1542 
1543 	f.readahead = true;
1544 	f.headoffset = readahead_pos(rac);
1545 
1546 	z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1547 				  readahead_length(rac) - 1, &pagepool, true);
1548 	nr_pages = readahead_count(rac);
1549 	trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1550 
1551 	while ((page = readahead_page(rac))) {
1552 		set_page_private(page, (unsigned long)head);
1553 		head = page;
1554 	}
1555 
1556 	while (head) {
1557 		struct page *page = head;
1558 		int err;
1559 
1560 		/* traversal in reverse order */
1561 		head = (void *)page_private(page);
1562 
1563 		err = z_erofs_do_read_page(&f, page, &pagepool);
1564 		if (err)
1565 			erofs_err(inode->i_sb,
1566 				  "readahead error at page %lu @ nid %llu",
1567 				  page->index, EROFS_I(inode)->nid);
1568 		put_page(page);
1569 	}
1570 	z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1571 	(void)z_erofs_collector_end(&f.clt);
1572 
1573 	z_erofs_runqueue(inode->i_sb, &f, &pagepool,
1574 			 z_erofs_get_sync_decompress_policy(sbi, nr_pages));
1575 	erofs_put_metabuf(&f.map.buf);
1576 	erofs_release_pages(&pagepool);
1577 }
1578 
1579 const struct address_space_operations z_erofs_aops = {
1580 	.readpage = z_erofs_readpage,
1581 	.readahead = z_erofs_readahead,
1582 };
1583