xref: /openbmc/linux/fs/erofs/zdata.c (revision cef69974)
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_decompress_frontend {
196 	struct inode *const inode;
197 	struct erofs_map_blocks map;
198 
199 	struct z_erofs_pagevec_ctor vector;
200 
201 	struct z_erofs_pcluster *pcl, *tailpcl;
202 	struct z_erofs_collection *cl;
203 	/* a pointer used to pick up inplace I/O pages */
204 	struct page **icpage_ptr;
205 	z_erofs_next_pcluster_t owned_head;
206 
207 	enum z_erofs_collectmode mode;
208 
209 	bool readahead;
210 	/* used for applying cache strategy on the fly */
211 	bool backmost;
212 	erofs_off_t headoffset;
213 };
214 
215 #define DECOMPRESS_FRONTEND_INIT(__i) { \
216 	.inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \
217 	.mode = COLLECT_PRIMARY_FOLLOWED }
218 
219 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
220 static DEFINE_MUTEX(z_pagemap_global_lock);
221 
222 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe,
223 			       enum z_erofs_cache_alloctype type,
224 			       struct page **pagepool)
225 {
226 	struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));
227 	struct z_erofs_pcluster *pcl = fe->pcl;
228 	bool standalone = true;
229 	/*
230 	 * optimistic allocation without direct reclaim since inplace I/O
231 	 * can be used if low memory otherwise.
232 	 */
233 	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
234 			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
235 	struct page **pages;
236 	pgoff_t index;
237 
238 	if (fe->mode < COLLECT_PRIMARY_FOLLOWED)
239 		return;
240 
241 	pages = pcl->compressed_pages;
242 	index = pcl->obj.index;
243 	for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) {
244 		struct page *page;
245 		compressed_page_t t;
246 		struct page *newpage = NULL;
247 
248 		/* the compressed page was loaded before */
249 		if (READ_ONCE(*pages))
250 			continue;
251 
252 		page = find_get_page(mc, index);
253 
254 		if (page) {
255 			t = tag_compressed_page_justfound(page);
256 		} else {
257 			/* I/O is needed, no possible to decompress directly */
258 			standalone = false;
259 			switch (type) {
260 			case TRYALLOC:
261 				newpage = erofs_allocpage(pagepool, gfp);
262 				if (!newpage)
263 					continue;
264 				set_page_private(newpage,
265 						 Z_EROFS_PREALLOCATED_PAGE);
266 				t = tag_compressed_page_justfound(newpage);
267 				break;
268 			default:        /* DONTALLOC */
269 				continue;
270 			}
271 		}
272 
273 		if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
274 			continue;
275 
276 		if (page)
277 			put_page(page);
278 		else if (newpage)
279 			erofs_pagepool_add(pagepool, newpage);
280 	}
281 
282 	/*
283 	 * don't do inplace I/O if all compressed pages are available in
284 	 * managed cache since it can be moved to the bypass queue instead.
285 	 */
286 	if (standalone)
287 		fe->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
288 }
289 
290 /* called by erofs_shrinker to get rid of all compressed_pages */
291 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
292 				       struct erofs_workgroup *grp)
293 {
294 	struct z_erofs_pcluster *const pcl =
295 		container_of(grp, struct z_erofs_pcluster, obj);
296 	int i;
297 
298 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
299 	/*
300 	 * refcount of workgroup is now freezed as 1,
301 	 * therefore no need to worry about available decompression users.
302 	 */
303 	for (i = 0; i < pcl->pclusterpages; ++i) {
304 		struct page *page = pcl->compressed_pages[i];
305 
306 		if (!page)
307 			continue;
308 
309 		/* block other users from reclaiming or migrating the page */
310 		if (!trylock_page(page))
311 			return -EBUSY;
312 
313 		if (!erofs_page_is_managed(sbi, page))
314 			continue;
315 
316 		/* barrier is implied in the following 'unlock_page' */
317 		WRITE_ONCE(pcl->compressed_pages[i], NULL);
318 		detach_page_private(page);
319 		unlock_page(page);
320 	}
321 	return 0;
322 }
323 
324 int erofs_try_to_free_cached_page(struct page *page)
325 {
326 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
327 	int ret = 0;	/* 0 - busy */
328 
329 	if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
330 		unsigned int i;
331 
332 		DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
333 		for (i = 0; i < pcl->pclusterpages; ++i) {
334 			if (pcl->compressed_pages[i] == page) {
335 				WRITE_ONCE(pcl->compressed_pages[i], NULL);
336 				ret = 1;
337 				break;
338 			}
339 		}
340 		erofs_workgroup_unfreeze(&pcl->obj, 1);
341 
342 		if (ret)
343 			detach_page_private(page);
344 	}
345 	return ret;
346 }
347 
348 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
349 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
350 				   struct page *page)
351 {
352 	struct z_erofs_pcluster *const pcl = fe->pcl;
353 
354 	while (fe->icpage_ptr > pcl->compressed_pages)
355 		if (!cmpxchg(--fe->icpage_ptr, NULL, page))
356 			return true;
357 	return false;
358 }
359 
360 /* callers must be with collection lock held */
361 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,
362 			       struct page *page, enum z_erofs_page_type type,
363 			       bool pvec_safereuse)
364 {
365 	int ret;
366 
367 	/* give priority for inplaceio */
368 	if (fe->mode >= COLLECT_PRIMARY &&
369 	    type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
370 	    z_erofs_try_inplace_io(fe, page))
371 		return 0;
372 
373 	ret = z_erofs_pagevec_enqueue(&fe->vector, page, type,
374 				      pvec_safereuse);
375 	fe->cl->vcnt += (unsigned int)ret;
376 	return ret ? 0 : -EAGAIN;
377 }
378 
379 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
380 {
381 	struct z_erofs_pcluster *pcl = f->pcl;
382 	z_erofs_next_pcluster_t *owned_head = &f->owned_head;
383 
384 	/* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
385 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
386 		    *owned_head) == Z_EROFS_PCLUSTER_NIL) {
387 		*owned_head = &pcl->next;
388 		/* so we can attach this pcluster to our submission chain. */
389 		f->mode = COLLECT_PRIMARY_FOLLOWED;
390 		return;
391 	}
392 
393 	/*
394 	 * type 2, link to the end of an existing open chain, be careful
395 	 * that its submission is controlled by the original attached chain.
396 	 */
397 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
398 		    *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
399 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
400 		f->mode = COLLECT_PRIMARY_HOOKED;
401 		f->tailpcl = NULL;
402 		return;
403 	}
404 	/* type 3, it belongs to a chain, but it isn't the end of the chain */
405 	f->mode = COLLECT_PRIMARY;
406 }
407 
408 static int z_erofs_lookup_collection(struct z_erofs_decompress_frontend *fe,
409 				     struct inode *inode,
410 				     struct erofs_map_blocks *map)
411 {
412 	struct z_erofs_pcluster *pcl = fe->pcl;
413 	struct z_erofs_collection *cl;
414 	unsigned int length;
415 
416 	/* to avoid unexpected loop formed by corrupted images */
417 	if (fe->owned_head == &pcl->next || pcl == fe->tailpcl) {
418 		DBG_BUGON(1);
419 		return -EFSCORRUPTED;
420 	}
421 
422 	cl = z_erofs_primarycollection(pcl);
423 	if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
424 		DBG_BUGON(1);
425 		return -EFSCORRUPTED;
426 	}
427 
428 	length = READ_ONCE(pcl->length);
429 	if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
430 		if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
431 			DBG_BUGON(1);
432 			return -EFSCORRUPTED;
433 		}
434 	} else {
435 		unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
436 
437 		if (map->m_flags & EROFS_MAP_FULL_MAPPED)
438 			llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
439 
440 		while (llen > length &&
441 		       length != cmpxchg_relaxed(&pcl->length, length, llen)) {
442 			cpu_relax();
443 			length = READ_ONCE(pcl->length);
444 		}
445 	}
446 	mutex_lock(&cl->lock);
447 	/* used to check tail merging loop due to corrupted images */
448 	if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
449 		fe->tailpcl = pcl;
450 
451 	z_erofs_try_to_claim_pcluster(fe);
452 	fe->cl = cl;
453 	return 0;
454 }
455 
456 static int z_erofs_register_collection(struct z_erofs_decompress_frontend *fe,
457 				       struct inode *inode,
458 				       struct erofs_map_blocks *map)
459 {
460 	bool ztailpacking = map->m_flags & EROFS_MAP_META;
461 	struct z_erofs_pcluster *pcl;
462 	struct z_erofs_collection *cl;
463 	struct erofs_workgroup *grp;
464 	int err;
465 
466 	if (!(map->m_flags & EROFS_MAP_ENCODED)) {
467 		DBG_BUGON(1);
468 		return -EFSCORRUPTED;
469 	}
470 
471 	/* no available pcluster, let's allocate one */
472 	pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
473 				     map->m_plen >> PAGE_SHIFT);
474 	if (IS_ERR(pcl))
475 		return PTR_ERR(pcl);
476 
477 	atomic_set(&pcl->obj.refcount, 1);
478 	pcl->algorithmformat = map->m_algorithmformat;
479 	pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
480 		(map->m_flags & EROFS_MAP_FULL_MAPPED ?
481 			Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
482 
483 	/* new pclusters should be claimed as type 1, primary and followed */
484 	pcl->next = fe->owned_head;
485 	fe->mode = COLLECT_PRIMARY_FOLLOWED;
486 
487 	cl = z_erofs_primarycollection(pcl);
488 	cl->pageofs = map->m_la & ~PAGE_MASK;
489 
490 	/*
491 	 * lock all primary followed works before visible to others
492 	 * and mutex_trylock *never* fails for a new pcluster.
493 	 */
494 	mutex_init(&cl->lock);
495 	DBG_BUGON(!mutex_trylock(&cl->lock));
496 
497 	if (ztailpacking) {
498 		pcl->obj.index = 0;	/* which indicates ztailpacking */
499 		pcl->pageofs_in = erofs_blkoff(map->m_pa);
500 		pcl->tailpacking_size = map->m_plen;
501 	} else {
502 		pcl->obj.index = map->m_pa >> PAGE_SHIFT;
503 
504 		grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
505 		if (IS_ERR(grp)) {
506 			err = PTR_ERR(grp);
507 			goto err_out;
508 		}
509 
510 		if (grp != &pcl->obj) {
511 			fe->pcl = container_of(grp,
512 					struct z_erofs_pcluster, obj);
513 			err = -EEXIST;
514 			goto err_out;
515 		}
516 	}
517 	/* used to check tail merging loop due to corrupted images */
518 	if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
519 		fe->tailpcl = pcl;
520 	fe->owned_head = &pcl->next;
521 	fe->pcl = pcl;
522 	fe->cl = cl;
523 	return 0;
524 
525 err_out:
526 	mutex_unlock(&cl->lock);
527 	z_erofs_free_pcluster(pcl);
528 	return err;
529 }
530 
531 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe,
532 				   struct inode *inode,
533 				   struct erofs_map_blocks *map)
534 {
535 	struct erofs_workgroup *grp;
536 	int ret;
537 
538 	DBG_BUGON(fe->cl);
539 
540 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
541 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
542 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
543 
544 	if (map->m_flags & EROFS_MAP_META) {
545 		if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
546 			DBG_BUGON(1);
547 			return -EFSCORRUPTED;
548 		}
549 		goto tailpacking;
550 	}
551 
552 	grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
553 	if (grp) {
554 		fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
555 	} else {
556 tailpacking:
557 		ret = z_erofs_register_collection(fe, inode, map);
558 		if (!ret)
559 			goto out;
560 		if (ret != -EEXIST)
561 			return ret;
562 	}
563 
564 	ret = z_erofs_lookup_collection(fe, inode, map);
565 	if (ret) {
566 		erofs_workgroup_put(&fe->pcl->obj);
567 		return ret;
568 	}
569 
570 out:
571 	z_erofs_pagevec_ctor_init(&fe->vector, Z_EROFS_NR_INLINE_PAGEVECS,
572 				  fe->cl->pagevec, fe->cl->vcnt);
573 	/* since file-backed online pages are traversed in reverse order */
574 	fe->icpage_ptr = fe->pcl->compressed_pages +
575 			z_erofs_pclusterpages(fe->pcl);
576 	return 0;
577 }
578 
579 /*
580  * keep in mind that no referenced pclusters will be freed
581  * only after a RCU grace period.
582  */
583 static void z_erofs_rcu_callback(struct rcu_head *head)
584 {
585 	struct z_erofs_collection *const cl =
586 		container_of(head, struct z_erofs_collection, rcu);
587 
588 	z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster,
589 					   primary_collection));
590 }
591 
592 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
593 {
594 	struct z_erofs_pcluster *const pcl =
595 		container_of(grp, struct z_erofs_pcluster, obj);
596 	struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
597 
598 	call_rcu(&cl->rcu, z_erofs_rcu_callback);
599 }
600 
601 static void z_erofs_collection_put(struct z_erofs_collection *cl)
602 {
603 	struct z_erofs_pcluster *const pcl =
604 		container_of(cl, struct z_erofs_pcluster, primary_collection);
605 
606 	erofs_workgroup_put(&pcl->obj);
607 }
608 
609 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe)
610 {
611 	struct z_erofs_collection *cl = fe->cl;
612 
613 	if (!cl)
614 		return false;
615 
616 	z_erofs_pagevec_ctor_exit(&fe->vector, false);
617 	mutex_unlock(&cl->lock);
618 
619 	/*
620 	 * if all pending pages are added, don't hold its reference
621 	 * any longer if the pcluster isn't hosted by ourselves.
622 	 */
623 	if (fe->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
624 		z_erofs_collection_put(cl);
625 
626 	fe->cl = NULL;
627 	return true;
628 }
629 
630 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
631 				       unsigned int cachestrategy,
632 				       erofs_off_t la)
633 {
634 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
635 		return false;
636 
637 	if (fe->backmost)
638 		return true;
639 
640 	return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
641 		la < fe->headoffset;
642 }
643 
644 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
645 				struct page *page, struct page **pagepool)
646 {
647 	struct inode *const inode = fe->inode;
648 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
649 	struct erofs_map_blocks *const map = &fe->map;
650 	const loff_t offset = page_offset(page);
651 	bool tight = true;
652 
653 	enum z_erofs_cache_alloctype cache_strategy;
654 	enum z_erofs_page_type page_type;
655 	unsigned int cur, end, spiltted, index;
656 	int err = 0;
657 
658 	/* register locked file pages as online pages in pack */
659 	z_erofs_onlinepage_init(page);
660 
661 	spiltted = 0;
662 	end = PAGE_SIZE;
663 repeat:
664 	cur = end - 1;
665 
666 	/* lucky, within the range of the current map_blocks */
667 	if (offset + cur >= map->m_la &&
668 	    offset + cur < map->m_la + map->m_llen) {
669 		/* didn't get a valid collection previously (very rare) */
670 		if (!fe->cl)
671 			goto restart_now;
672 		goto hitted;
673 	}
674 
675 	/* go ahead the next map_blocks */
676 	erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
677 
678 	if (z_erofs_collector_end(fe))
679 		fe->backmost = false;
680 
681 	map->m_la = offset + cur;
682 	map->m_llen = 0;
683 	err = z_erofs_map_blocks_iter(inode, map, 0);
684 	if (err)
685 		goto err_out;
686 
687 restart_now:
688 	if (!(map->m_flags & EROFS_MAP_MAPPED))
689 		goto hitted;
690 
691 	err = z_erofs_collector_begin(fe, inode, map);
692 	if (err)
693 		goto err_out;
694 
695 	if (z_erofs_is_inline_pcluster(fe->pcl)) {
696 		void *mp;
697 
698 		mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
699 					erofs_blknr(map->m_pa), EROFS_NO_KMAP);
700 		if (IS_ERR(mp)) {
701 			err = PTR_ERR(mp);
702 			erofs_err(inode->i_sb,
703 				  "failed to get inline page, err %d", err);
704 			goto err_out;
705 		}
706 		get_page(fe->map.buf.page);
707 		WRITE_ONCE(fe->pcl->compressed_pages[0], fe->map.buf.page);
708 		fe->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
709 	} else {
710 		/* bind cache first when cached decompression is preferred */
711 		if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy,
712 					       map->m_la))
713 			cache_strategy = TRYALLOC;
714 		else
715 			cache_strategy = DONTALLOC;
716 
717 		z_erofs_bind_cache(fe, cache_strategy, pagepool);
718 	}
719 hitted:
720 	/*
721 	 * Ensure the current partial page belongs to this submit chain rather
722 	 * than other concurrent submit chains or the noio(bypass) chain since
723 	 * those chains are handled asynchronously thus the page cannot be used
724 	 * for inplace I/O or pagevec (should be processed in strict order.)
725 	 */
726 	tight &= (fe->mode >= COLLECT_PRIMARY_HOOKED &&
727 		  fe->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
728 
729 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
730 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
731 		zero_user_segment(page, cur, end);
732 		goto next_part;
733 	}
734 
735 	/* let's derive page type */
736 	page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
737 		(!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
738 			(tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
739 				Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
740 
741 	if (cur)
742 		tight &= (fe->mode >= COLLECT_PRIMARY_FOLLOWED);
743 
744 retry:
745 	err = z_erofs_attach_page(fe, page, page_type,
746 				  fe->mode >= COLLECT_PRIMARY_FOLLOWED);
747 	/* should allocate an additional short-lived page for pagevec */
748 	if (err == -EAGAIN) {
749 		struct page *const newpage =
750 				alloc_page(GFP_NOFS | __GFP_NOFAIL);
751 
752 		set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE);
753 		err = z_erofs_attach_page(fe, newpage,
754 					  Z_EROFS_PAGE_TYPE_EXCLUSIVE, true);
755 		if (!err)
756 			goto retry;
757 	}
758 
759 	if (err)
760 		goto err_out;
761 
762 	index = page->index - (map->m_la >> PAGE_SHIFT);
763 
764 	z_erofs_onlinepage_fixup(page, index, true);
765 
766 	/* bump up the number of spiltted parts of a page */
767 	++spiltted;
768 	/* also update nr_pages */
769 	fe->cl->nr_pages = max_t(pgoff_t, fe->cl->nr_pages, index + 1);
770 next_part:
771 	/* can be used for verification */
772 	map->m_llen = offset + cur - map->m_la;
773 
774 	end = cur;
775 	if (end > 0)
776 		goto repeat;
777 
778 out:
779 	z_erofs_onlinepage_endio(page);
780 
781 	erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
782 		  __func__, page, spiltted, map->m_llen);
783 	return err;
784 
785 	/* if some error occurred while processing this page */
786 err_out:
787 	SetPageError(page);
788 	goto out;
789 }
790 
791 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi,
792 				       unsigned int readahead_pages)
793 {
794 	/* auto: enable for readpage, disable for readahead */
795 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
796 	    !readahead_pages)
797 		return true;
798 
799 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
800 	    (readahead_pages <= sbi->opt.max_sync_decompress_pages))
801 		return true;
802 
803 	return false;
804 }
805 
806 static bool z_erofs_page_is_invalidated(struct page *page)
807 {
808 	return !page->mapping && !z_erofs_is_shortlived_page(page);
809 }
810 
811 static int z_erofs_decompress_pcluster(struct super_block *sb,
812 				       struct z_erofs_pcluster *pcl,
813 				       struct page **pagepool)
814 {
815 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
816 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
817 	struct z_erofs_pagevec_ctor ctor;
818 	unsigned int i, inputsize, outputsize, llen, nr_pages;
819 	struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
820 	struct page **pages, **compressed_pages, *page;
821 
822 	enum z_erofs_page_type page_type;
823 	bool overlapped, partial;
824 	struct z_erofs_collection *cl;
825 	int err;
826 
827 	might_sleep();
828 	cl = z_erofs_primarycollection(pcl);
829 	DBG_BUGON(!READ_ONCE(cl->nr_pages));
830 
831 	mutex_lock(&cl->lock);
832 	nr_pages = cl->nr_pages;
833 
834 	if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
835 		pages = pages_onstack;
836 	} else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
837 		   mutex_trylock(&z_pagemap_global_lock)) {
838 		pages = z_pagemap_global;
839 	} else {
840 		gfp_t gfp_flags = GFP_KERNEL;
841 
842 		if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
843 			gfp_flags |= __GFP_NOFAIL;
844 
845 		pages = kvmalloc_array(nr_pages, sizeof(struct page *),
846 				       gfp_flags);
847 
848 		/* fallback to global pagemap for the lowmem scenario */
849 		if (!pages) {
850 			mutex_lock(&z_pagemap_global_lock);
851 			pages = z_pagemap_global;
852 		}
853 	}
854 
855 	for (i = 0; i < nr_pages; ++i)
856 		pages[i] = NULL;
857 
858 	err = 0;
859 	z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
860 				  cl->pagevec, 0);
861 
862 	for (i = 0; i < cl->vcnt; ++i) {
863 		unsigned int pagenr;
864 
865 		page = z_erofs_pagevec_dequeue(&ctor, &page_type);
866 
867 		/* all pages in pagevec ought to be valid */
868 		DBG_BUGON(!page);
869 		DBG_BUGON(z_erofs_page_is_invalidated(page));
870 
871 		if (z_erofs_put_shortlivedpage(pagepool, page))
872 			continue;
873 
874 		if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
875 			pagenr = 0;
876 		else
877 			pagenr = z_erofs_onlinepage_index(page);
878 
879 		DBG_BUGON(pagenr >= nr_pages);
880 
881 		/*
882 		 * currently EROFS doesn't support multiref(dedup),
883 		 * so here erroring out one multiref page.
884 		 */
885 		if (pages[pagenr]) {
886 			DBG_BUGON(1);
887 			SetPageError(pages[pagenr]);
888 			z_erofs_onlinepage_endio(pages[pagenr]);
889 			err = -EFSCORRUPTED;
890 		}
891 		pages[pagenr] = page;
892 	}
893 	z_erofs_pagevec_ctor_exit(&ctor, true);
894 
895 	overlapped = false;
896 	compressed_pages = pcl->compressed_pages;
897 
898 	for (i = 0; i < pclusterpages; ++i) {
899 		unsigned int pagenr;
900 
901 		page = compressed_pages[i];
902 		/* all compressed pages ought to be valid */
903 		DBG_BUGON(!page);
904 
905 		if (z_erofs_is_inline_pcluster(pcl)) {
906 			if (!PageUptodate(page))
907 				err = -EIO;
908 			continue;
909 		}
910 
911 		DBG_BUGON(z_erofs_page_is_invalidated(page));
912 		if (!z_erofs_is_shortlived_page(page)) {
913 			if (erofs_page_is_managed(sbi, page)) {
914 				if (!PageUptodate(page))
915 					err = -EIO;
916 				continue;
917 			}
918 
919 			/*
920 			 * only if non-head page can be selected
921 			 * for inplace decompression
922 			 */
923 			pagenr = z_erofs_onlinepage_index(page);
924 
925 			DBG_BUGON(pagenr >= nr_pages);
926 			if (pages[pagenr]) {
927 				DBG_BUGON(1);
928 				SetPageError(pages[pagenr]);
929 				z_erofs_onlinepage_endio(pages[pagenr]);
930 				err = -EFSCORRUPTED;
931 			}
932 			pages[pagenr] = page;
933 
934 			overlapped = true;
935 		}
936 
937 		/* PG_error needs checking for all non-managed pages */
938 		if (PageError(page)) {
939 			DBG_BUGON(PageUptodate(page));
940 			err = -EIO;
941 		}
942 	}
943 
944 	if (err)
945 		goto out;
946 
947 	llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
948 	if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
949 		outputsize = llen;
950 		partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
951 	} else {
952 		outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
953 		partial = true;
954 	}
955 
956 	if (z_erofs_is_inline_pcluster(pcl))
957 		inputsize = pcl->tailpacking_size;
958 	else
959 		inputsize = pclusterpages * PAGE_SIZE;
960 
961 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
962 					.sb = sb,
963 					.in = compressed_pages,
964 					.out = pages,
965 					.pageofs_in = pcl->pageofs_in,
966 					.pageofs_out = cl->pageofs,
967 					.inputsize = inputsize,
968 					.outputsize = outputsize,
969 					.alg = pcl->algorithmformat,
970 					.inplace_io = overlapped,
971 					.partial_decoding = partial
972 				 }, pagepool);
973 
974 out:
975 	/* must handle all compressed pages before actual file pages */
976 	if (z_erofs_is_inline_pcluster(pcl)) {
977 		page = compressed_pages[0];
978 		WRITE_ONCE(compressed_pages[0], NULL);
979 		put_page(page);
980 	} else {
981 		for (i = 0; i < pclusterpages; ++i) {
982 			page = compressed_pages[i];
983 
984 			if (erofs_page_is_managed(sbi, page))
985 				continue;
986 
987 			/* recycle all individual short-lived pages */
988 			(void)z_erofs_put_shortlivedpage(pagepool, page);
989 			WRITE_ONCE(compressed_pages[i], NULL);
990 		}
991 	}
992 
993 	for (i = 0; i < nr_pages; ++i) {
994 		page = pages[i];
995 		if (!page)
996 			continue;
997 
998 		DBG_BUGON(z_erofs_page_is_invalidated(page));
999 
1000 		/* recycle all individual short-lived pages */
1001 		if (z_erofs_put_shortlivedpage(pagepool, page))
1002 			continue;
1003 
1004 		if (err < 0)
1005 			SetPageError(page);
1006 
1007 		z_erofs_onlinepage_endio(page);
1008 	}
1009 
1010 	if (pages == z_pagemap_global)
1011 		mutex_unlock(&z_pagemap_global_lock);
1012 	else if (pages != pages_onstack)
1013 		kvfree(pages);
1014 
1015 	cl->nr_pages = 0;
1016 	cl->vcnt = 0;
1017 
1018 	/* all cl locks MUST be taken before the following line */
1019 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1020 
1021 	/* all cl locks SHOULD be released right now */
1022 	mutex_unlock(&cl->lock);
1023 
1024 	z_erofs_collection_put(cl);
1025 	return err;
1026 }
1027 
1028 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1029 				     struct page **pagepool)
1030 {
1031 	z_erofs_next_pcluster_t owned = io->head;
1032 
1033 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1034 		struct z_erofs_pcluster *pcl;
1035 
1036 		/* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1037 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1038 
1039 		/* no possible that 'owned' equals NULL */
1040 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1041 
1042 		pcl = container_of(owned, struct z_erofs_pcluster, next);
1043 		owned = READ_ONCE(pcl->next);
1044 
1045 		z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
1046 	}
1047 }
1048 
1049 static void z_erofs_decompressqueue_work(struct work_struct *work)
1050 {
1051 	struct z_erofs_decompressqueue *bgq =
1052 		container_of(work, struct z_erofs_decompressqueue, u.work);
1053 	struct page *pagepool = NULL;
1054 
1055 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1056 	z_erofs_decompress_queue(bgq, &pagepool);
1057 
1058 	erofs_release_pages(&pagepool);
1059 	kvfree(bgq);
1060 }
1061 
1062 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1063 				       bool sync, int bios)
1064 {
1065 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1066 
1067 	/* wake up the caller thread for sync decompression */
1068 	if (sync) {
1069 		unsigned long flags;
1070 
1071 		spin_lock_irqsave(&io->u.wait.lock, flags);
1072 		if (!atomic_add_return(bios, &io->pending_bios))
1073 			wake_up_locked(&io->u.wait);
1074 		spin_unlock_irqrestore(&io->u.wait.lock, flags);
1075 		return;
1076 	}
1077 
1078 	if (atomic_add_return(bios, &io->pending_bios))
1079 		return;
1080 	/* Use workqueue and sync decompression for atomic contexts only */
1081 	if (in_atomic() || irqs_disabled()) {
1082 		queue_work(z_erofs_workqueue, &io->u.work);
1083 		/* enable sync decompression for readahead */
1084 		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1085 			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1086 		return;
1087 	}
1088 	z_erofs_decompressqueue_work(&io->u.work);
1089 }
1090 
1091 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1092 					       unsigned int nr,
1093 					       struct page **pagepool,
1094 					       struct address_space *mc)
1095 {
1096 	const pgoff_t index = pcl->obj.index;
1097 	gfp_t gfp = mapping_gfp_mask(mc);
1098 	bool tocache = false;
1099 
1100 	struct address_space *mapping;
1101 	struct page *oldpage, *page;
1102 
1103 	compressed_page_t t;
1104 	int justfound;
1105 
1106 repeat:
1107 	page = READ_ONCE(pcl->compressed_pages[nr]);
1108 	oldpage = page;
1109 
1110 	if (!page)
1111 		goto out_allocpage;
1112 
1113 	/* process the target tagged pointer */
1114 	t = tagptr_init(compressed_page_t, page);
1115 	justfound = tagptr_unfold_tags(t);
1116 	page = tagptr_unfold_ptr(t);
1117 
1118 	/*
1119 	 * preallocated cached pages, which is used to avoid direct reclaim
1120 	 * otherwise, it will go inplace I/O path instead.
1121 	 */
1122 	if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1123 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1124 		set_page_private(page, 0);
1125 		tocache = true;
1126 		goto out_tocache;
1127 	}
1128 	mapping = READ_ONCE(page->mapping);
1129 
1130 	/*
1131 	 * file-backed online pages in plcuster are all locked steady,
1132 	 * therefore it is impossible for `mapping' to be NULL.
1133 	 */
1134 	if (mapping && mapping != mc)
1135 		/* ought to be unmanaged pages */
1136 		goto out;
1137 
1138 	/* directly return for shortlived page as well */
1139 	if (z_erofs_is_shortlived_page(page))
1140 		goto out;
1141 
1142 	lock_page(page);
1143 
1144 	/* only true if page reclaim goes wrong, should never happen */
1145 	DBG_BUGON(justfound && PagePrivate(page));
1146 
1147 	/* the page is still in manage cache */
1148 	if (page->mapping == mc) {
1149 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1150 
1151 		ClearPageError(page);
1152 		if (!PagePrivate(page)) {
1153 			/*
1154 			 * impossible to be !PagePrivate(page) for
1155 			 * the current restriction as well if
1156 			 * the page is already in compressed_pages[].
1157 			 */
1158 			DBG_BUGON(!justfound);
1159 
1160 			justfound = 0;
1161 			set_page_private(page, (unsigned long)pcl);
1162 			SetPagePrivate(page);
1163 		}
1164 
1165 		/* no need to submit io if it is already up-to-date */
1166 		if (PageUptodate(page)) {
1167 			unlock_page(page);
1168 			page = NULL;
1169 		}
1170 		goto out;
1171 	}
1172 
1173 	/*
1174 	 * the managed page has been truncated, it's unsafe to
1175 	 * reuse this one, let's allocate a new cache-managed page.
1176 	 */
1177 	DBG_BUGON(page->mapping);
1178 	DBG_BUGON(!justfound);
1179 
1180 	tocache = true;
1181 	unlock_page(page);
1182 	put_page(page);
1183 out_allocpage:
1184 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1185 	if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1186 		erofs_pagepool_add(pagepool, page);
1187 		cond_resched();
1188 		goto repeat;
1189 	}
1190 out_tocache:
1191 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1192 		/* turn into temporary page if fails (1 ref) */
1193 		set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1194 		goto out;
1195 	}
1196 	attach_page_private(page, pcl);
1197 	/* drop a refcount added by allocpage (then we have 2 refs here) */
1198 	put_page(page);
1199 
1200 out:	/* the only exit (for tracing and debugging) */
1201 	return page;
1202 }
1203 
1204 static struct z_erofs_decompressqueue *
1205 jobqueue_init(struct super_block *sb,
1206 	      struct z_erofs_decompressqueue *fgq, bool *fg)
1207 {
1208 	struct z_erofs_decompressqueue *q;
1209 
1210 	if (fg && !*fg) {
1211 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1212 		if (!q) {
1213 			*fg = true;
1214 			goto fg_out;
1215 		}
1216 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1217 	} else {
1218 fg_out:
1219 		q = fgq;
1220 		init_waitqueue_head(&fgq->u.wait);
1221 		atomic_set(&fgq->pending_bios, 0);
1222 	}
1223 	q->sb = sb;
1224 	q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1225 	return q;
1226 }
1227 
1228 /* define decompression jobqueue types */
1229 enum {
1230 	JQ_BYPASS,
1231 	JQ_SUBMIT,
1232 	NR_JOBQUEUES,
1233 };
1234 
1235 static void *jobqueueset_init(struct super_block *sb,
1236 			      struct z_erofs_decompressqueue *q[],
1237 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1238 {
1239 	/*
1240 	 * if managed cache is enabled, bypass jobqueue is needed,
1241 	 * no need to read from device for all pclusters in this queue.
1242 	 */
1243 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1244 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1245 
1246 	return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1247 }
1248 
1249 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1250 				    z_erofs_next_pcluster_t qtail[],
1251 				    z_erofs_next_pcluster_t owned_head)
1252 {
1253 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1254 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1255 
1256 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1257 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1258 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1259 
1260 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1261 
1262 	WRITE_ONCE(*submit_qtail, owned_head);
1263 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1264 
1265 	qtail[JQ_BYPASS] = &pcl->next;
1266 }
1267 
1268 static void z_erofs_decompressqueue_endio(struct bio *bio)
1269 {
1270 	tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
1271 	struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
1272 	blk_status_t err = bio->bi_status;
1273 	struct bio_vec *bvec;
1274 	struct bvec_iter_all iter_all;
1275 
1276 	bio_for_each_segment_all(bvec, bio, iter_all) {
1277 		struct page *page = bvec->bv_page;
1278 
1279 		DBG_BUGON(PageUptodate(page));
1280 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1281 
1282 		if (err)
1283 			SetPageError(page);
1284 
1285 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1286 			if (!err)
1287 				SetPageUptodate(page);
1288 			unlock_page(page);
1289 		}
1290 	}
1291 	z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
1292 	bio_put(bio);
1293 }
1294 
1295 static void z_erofs_submit_queue(struct super_block *sb,
1296 				 struct z_erofs_decompress_frontend *f,
1297 				 struct page **pagepool,
1298 				 struct z_erofs_decompressqueue *fgq,
1299 				 bool *force_fg)
1300 {
1301 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
1302 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1303 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1304 	void *bi_private;
1305 	z_erofs_next_pcluster_t owned_head = f->owned_head;
1306 	/* bio is NULL initially, so no need to initialize last_{index,bdev} */
1307 	pgoff_t last_index;
1308 	struct block_device *last_bdev;
1309 	unsigned int nr_bios = 0;
1310 	struct bio *bio = NULL;
1311 
1312 	bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1313 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1314 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1315 
1316 	/* by default, all need io submission */
1317 	q[JQ_SUBMIT]->head = owned_head;
1318 
1319 	do {
1320 		struct erofs_map_dev mdev;
1321 		struct z_erofs_pcluster *pcl;
1322 		pgoff_t cur, end;
1323 		unsigned int i = 0;
1324 		bool bypass = true;
1325 
1326 		/* no possible 'owned_head' equals the following */
1327 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1328 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1329 
1330 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1331 
1332 		/* close the main owned chain at first */
1333 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1334 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1335 		if (z_erofs_is_inline_pcluster(pcl)) {
1336 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1337 			continue;
1338 		}
1339 
1340 		/* no device id here, thus it will always succeed */
1341 		mdev = (struct erofs_map_dev) {
1342 			.m_pa = blknr_to_addr(pcl->obj.index),
1343 		};
1344 		(void)erofs_map_dev(sb, &mdev);
1345 
1346 		cur = erofs_blknr(mdev.m_pa);
1347 		end = cur + pcl->pclusterpages;
1348 
1349 		do {
1350 			struct page *page;
1351 
1352 			page = pickup_page_for_submission(pcl, i++, pagepool,
1353 							  MNGD_MAPPING(sbi));
1354 			if (!page)
1355 				continue;
1356 
1357 			if (bio && (cur != last_index + 1 ||
1358 				    last_bdev != mdev.m_bdev)) {
1359 submit_bio_retry:
1360 				submit_bio(bio);
1361 				bio = NULL;
1362 			}
1363 
1364 			if (!bio) {
1365 				bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1366 						REQ_OP_READ, GFP_NOIO);
1367 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1368 
1369 				last_bdev = mdev.m_bdev;
1370 				bio->bi_iter.bi_sector = (sector_t)cur <<
1371 					LOG_SECTORS_PER_BLOCK;
1372 				bio->bi_private = bi_private;
1373 				if (f->readahead)
1374 					bio->bi_opf |= REQ_RAHEAD;
1375 				++nr_bios;
1376 			}
1377 
1378 			if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1379 				goto submit_bio_retry;
1380 
1381 			last_index = cur;
1382 			bypass = false;
1383 		} while (++cur < end);
1384 
1385 		if (!bypass)
1386 			qtail[JQ_SUBMIT] = &pcl->next;
1387 		else
1388 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1389 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1390 
1391 	if (bio)
1392 		submit_bio(bio);
1393 
1394 	/*
1395 	 * although background is preferred, no one is pending for submission.
1396 	 * don't issue workqueue for decompression but drop it directly instead.
1397 	 */
1398 	if (!*force_fg && !nr_bios) {
1399 		kvfree(q[JQ_SUBMIT]);
1400 		return;
1401 	}
1402 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1403 }
1404 
1405 static void z_erofs_runqueue(struct super_block *sb,
1406 			     struct z_erofs_decompress_frontend *f,
1407 			     struct page **pagepool, bool force_fg)
1408 {
1409 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1410 
1411 	if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1412 		return;
1413 	z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1414 
1415 	/* handle bypass queue (no i/o pclusters) immediately */
1416 	z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1417 
1418 	if (!force_fg)
1419 		return;
1420 
1421 	/* wait until all bios are completed */
1422 	io_wait_event(io[JQ_SUBMIT].u.wait,
1423 		      !atomic_read(&io[JQ_SUBMIT].pending_bios));
1424 
1425 	/* handle synchronous decompress queue in the caller context */
1426 	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1427 }
1428 
1429 /*
1430  * Since partial uptodate is still unimplemented for now, we have to use
1431  * approximate readmore strategies as a start.
1432  */
1433 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1434 				      struct readahead_control *rac,
1435 				      erofs_off_t end,
1436 				      struct page **pagepool,
1437 				      bool backmost)
1438 {
1439 	struct inode *inode = f->inode;
1440 	struct erofs_map_blocks *map = &f->map;
1441 	erofs_off_t cur;
1442 	int err;
1443 
1444 	if (backmost) {
1445 		map->m_la = end;
1446 		err = z_erofs_map_blocks_iter(inode, map,
1447 					      EROFS_GET_BLOCKS_READMORE);
1448 		if (err)
1449 			return;
1450 
1451 		/* expend ra for the trailing edge if readahead */
1452 		if (rac) {
1453 			loff_t newstart = readahead_pos(rac);
1454 
1455 			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1456 			readahead_expand(rac, newstart, cur - newstart);
1457 			return;
1458 		}
1459 		end = round_up(end, PAGE_SIZE);
1460 	} else {
1461 		end = round_up(map->m_la, PAGE_SIZE);
1462 
1463 		if (!map->m_llen)
1464 			return;
1465 	}
1466 
1467 	cur = map->m_la + map->m_llen - 1;
1468 	while (cur >= end) {
1469 		pgoff_t index = cur >> PAGE_SHIFT;
1470 		struct page *page;
1471 
1472 		page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1473 		if (!page)
1474 			goto skip;
1475 
1476 		if (PageUptodate(page)) {
1477 			unlock_page(page);
1478 			put_page(page);
1479 			goto skip;
1480 		}
1481 
1482 		err = z_erofs_do_read_page(f, page, pagepool);
1483 		if (err)
1484 			erofs_err(inode->i_sb,
1485 				  "readmore error at page %lu @ nid %llu",
1486 				  index, EROFS_I(inode)->nid);
1487 		put_page(page);
1488 skip:
1489 		if (cur < PAGE_SIZE)
1490 			break;
1491 		cur = (index << PAGE_SHIFT) - 1;
1492 	}
1493 }
1494 
1495 static int z_erofs_readpage(struct file *file, struct page *page)
1496 {
1497 	struct inode *const inode = page->mapping->host;
1498 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1499 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1500 	struct page *pagepool = NULL;
1501 	int err;
1502 
1503 	trace_erofs_readpage(page, false);
1504 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1505 
1506 	z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1507 				  &pagepool, true);
1508 	err = z_erofs_do_read_page(&f, page, &pagepool);
1509 	z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1510 
1511 	(void)z_erofs_collector_end(&f);
1512 
1513 	/* if some compressed cluster ready, need submit them anyway */
1514 	z_erofs_runqueue(inode->i_sb, &f, &pagepool,
1515 			 z_erofs_get_sync_decompress_policy(sbi, 0));
1516 
1517 	if (err)
1518 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1519 
1520 	erofs_put_metabuf(&f.map.buf);
1521 	erofs_release_pages(&pagepool);
1522 	return err;
1523 }
1524 
1525 static void z_erofs_readahead(struct readahead_control *rac)
1526 {
1527 	struct inode *const inode = rac->mapping->host;
1528 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1529 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1530 	struct page *pagepool = NULL, *head = NULL, *page;
1531 	unsigned int nr_pages;
1532 
1533 	f.readahead = true;
1534 	f.headoffset = readahead_pos(rac);
1535 
1536 	z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1537 				  readahead_length(rac) - 1, &pagepool, true);
1538 	nr_pages = readahead_count(rac);
1539 	trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1540 
1541 	while ((page = readahead_page(rac))) {
1542 		set_page_private(page, (unsigned long)head);
1543 		head = page;
1544 	}
1545 
1546 	while (head) {
1547 		struct page *page = head;
1548 		int err;
1549 
1550 		/* traversal in reverse order */
1551 		head = (void *)page_private(page);
1552 
1553 		err = z_erofs_do_read_page(&f, page, &pagepool);
1554 		if (err)
1555 			erofs_err(inode->i_sb,
1556 				  "readahead error at page %lu @ nid %llu",
1557 				  page->index, EROFS_I(inode)->nid);
1558 		put_page(page);
1559 	}
1560 	z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1561 	(void)z_erofs_collector_end(&f);
1562 
1563 	z_erofs_runqueue(inode->i_sb, &f, &pagepool,
1564 			 z_erofs_get_sync_decompress_policy(sbi, nr_pages));
1565 	erofs_put_metabuf(&f.map.buf);
1566 	erofs_release_pages(&pagepool);
1567 }
1568 
1569 const struct address_space_operations z_erofs_aops = {
1570 	.readpage = z_erofs_readpage,
1571 	.readahead = z_erofs_readahead,
1572 };
1573