xref: /linux/fs/erofs/zdata.c (revision e1c4c5436b4ad579762fbe78bfabc8aef59bd5b1)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2018 HUAWEI, Inc.
4  *             https://www.huawei.com/
5  * Copyright (C) 2022 Alibaba Cloud
6  */
7 #include "zdata.h"
8 #include "compress.h"
9 #include <linux/prefetch.h>
10 #include <linux/psi.h>
11 
12 #include <trace/events/erofs.h>
13 
14 /*
15  * since pclustersize is variable for big pcluster feature, introduce slab
16  * pools implementation for different pcluster sizes.
17  */
18 struct z_erofs_pcluster_slab {
19 	struct kmem_cache *slab;
20 	unsigned int maxpages;
21 	char name[48];
22 };
23 
24 #define _PCLP(n) { .maxpages = n }
25 
26 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
27 	_PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
28 	_PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
29 };
30 
31 struct z_erofs_bvec_iter {
32 	struct page *bvpage;
33 	struct z_erofs_bvset *bvset;
34 	unsigned int nr, cur;
35 };
36 
37 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter)
38 {
39 	if (iter->bvpage)
40 		kunmap_local(iter->bvset);
41 	return iter->bvpage;
42 }
43 
44 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter)
45 {
46 	unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec;
47 	/* have to access nextpage in advance, otherwise it will be unmapped */
48 	struct page *nextpage = iter->bvset->nextpage;
49 	struct page *oldpage;
50 
51 	DBG_BUGON(!nextpage);
52 	oldpage = z_erofs_bvec_iter_end(iter);
53 	iter->bvpage = nextpage;
54 	iter->bvset = kmap_local_page(nextpage);
55 	iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec);
56 	iter->cur = 0;
57 	return oldpage;
58 }
59 
60 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter,
61 				    struct z_erofs_bvset_inline *bvset,
62 				    unsigned int bootstrap_nr,
63 				    unsigned int cur)
64 {
65 	*iter = (struct z_erofs_bvec_iter) {
66 		.nr = bootstrap_nr,
67 		.bvset = (struct z_erofs_bvset *)bvset,
68 	};
69 
70 	while (cur > iter->nr) {
71 		cur -= iter->nr;
72 		z_erofs_bvset_flip(iter);
73 	}
74 	iter->cur = cur;
75 }
76 
77 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter,
78 				struct z_erofs_bvec *bvec,
79 				struct page **candidate_bvpage)
80 {
81 	if (iter->cur == iter->nr) {
82 		if (!*candidate_bvpage)
83 			return -EAGAIN;
84 
85 		DBG_BUGON(iter->bvset->nextpage);
86 		iter->bvset->nextpage = *candidate_bvpage;
87 		z_erofs_bvset_flip(iter);
88 
89 		iter->bvset->nextpage = NULL;
90 		*candidate_bvpage = NULL;
91 	}
92 	iter->bvset->bvec[iter->cur++] = *bvec;
93 	return 0;
94 }
95 
96 static void z_erofs_bvec_dequeue(struct z_erofs_bvec_iter *iter,
97 				 struct z_erofs_bvec *bvec,
98 				 struct page **old_bvpage)
99 {
100 	if (iter->cur == iter->nr)
101 		*old_bvpage = z_erofs_bvset_flip(iter);
102 	else
103 		*old_bvpage = NULL;
104 	*bvec = iter->bvset->bvec[iter->cur++];
105 }
106 
107 static void z_erofs_destroy_pcluster_pool(void)
108 {
109 	int i;
110 
111 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
112 		if (!pcluster_pool[i].slab)
113 			continue;
114 		kmem_cache_destroy(pcluster_pool[i].slab);
115 		pcluster_pool[i].slab = NULL;
116 	}
117 }
118 
119 static int z_erofs_create_pcluster_pool(void)
120 {
121 	struct z_erofs_pcluster_slab *pcs;
122 	struct z_erofs_pcluster *a;
123 	unsigned int size;
124 
125 	for (pcs = pcluster_pool;
126 	     pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
127 		size = struct_size(a, compressed_bvecs, pcs->maxpages);
128 
129 		sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
130 		pcs->slab = kmem_cache_create(pcs->name, size, 0,
131 					      SLAB_RECLAIM_ACCOUNT, NULL);
132 		if (pcs->slab)
133 			continue;
134 
135 		z_erofs_destroy_pcluster_pool();
136 		return -ENOMEM;
137 	}
138 	return 0;
139 }
140 
141 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
142 {
143 	int i;
144 
145 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
146 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
147 		struct z_erofs_pcluster *pcl;
148 
149 		if (nrpages > pcs->maxpages)
150 			continue;
151 
152 		pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
153 		if (!pcl)
154 			return ERR_PTR(-ENOMEM);
155 		pcl->pclusterpages = nrpages;
156 		return pcl;
157 	}
158 	return ERR_PTR(-EINVAL);
159 }
160 
161 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
162 {
163 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
164 	int i;
165 
166 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
167 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
168 
169 		if (pclusterpages > pcs->maxpages)
170 			continue;
171 
172 		kmem_cache_free(pcs->slab, pcl);
173 		return;
174 	}
175 	DBG_BUGON(1);
176 }
177 
178 /* how to allocate cached pages for a pcluster */
179 enum z_erofs_cache_alloctype {
180 	DONTALLOC,	/* don't allocate any cached pages */
181 	/*
182 	 * try to use cached I/O if page allocation succeeds or fallback
183 	 * to in-place I/O instead to avoid any direct reclaim.
184 	 */
185 	TRYALLOC,
186 };
187 
188 /*
189  * tagged pointer with 1-bit tag for all compressed pages
190  * tag 0 - the page is just found with an extra page reference
191  */
192 typedef tagptr1_t compressed_page_t;
193 
194 #define tag_compressed_page_justfound(page) \
195 	tagptr_fold(compressed_page_t, page, 1)
196 
197 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
198 
199 void z_erofs_exit_zip_subsystem(void)
200 {
201 	destroy_workqueue(z_erofs_workqueue);
202 	z_erofs_destroy_pcluster_pool();
203 }
204 
205 static inline int z_erofs_init_workqueue(void)
206 {
207 	const unsigned int onlinecpus = num_possible_cpus();
208 
209 	/*
210 	 * no need to spawn too many threads, limiting threads could minimum
211 	 * scheduling overhead, perhaps per-CPU threads should be better?
212 	 */
213 	z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
214 					    WQ_UNBOUND | WQ_HIGHPRI,
215 					    onlinecpus + onlinecpus / 4);
216 	return z_erofs_workqueue ? 0 : -ENOMEM;
217 }
218 
219 int __init z_erofs_init_zip_subsystem(void)
220 {
221 	int err = z_erofs_create_pcluster_pool();
222 
223 	if (err)
224 		return err;
225 	err = z_erofs_init_workqueue();
226 	if (err)
227 		z_erofs_destroy_pcluster_pool();
228 	return err;
229 }
230 
231 enum z_erofs_pclustermode {
232 	Z_EROFS_PCLUSTER_INFLIGHT,
233 	/*
234 	 * The current pclusters was the tail of an exist chain, in addition
235 	 * that the previous processed chained pclusters are all decided to
236 	 * be hooked up to it.
237 	 * A new chain will be created for the remaining pclusters which are
238 	 * not processed yet, so different from Z_EROFS_PCLUSTER_FOLLOWED,
239 	 * the next pcluster cannot reuse the whole page safely for inplace I/O
240 	 * in the following scenario:
241 	 *  ________________________________________________________________
242 	 * |      tail (partial) page     |       head (partial) page       |
243 	 * |   (belongs to the next pcl)  |   (belongs to the current pcl)  |
244 	 * |_______PCLUSTER_FOLLOWED______|________PCLUSTER_HOOKED__________|
245 	 */
246 	Z_EROFS_PCLUSTER_HOOKED,
247 	/*
248 	 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it
249 	 * could be dispatched into bypass queue later due to uptodated managed
250 	 * pages. All related online pages cannot be reused for inplace I/O (or
251 	 * bvpage) since it can be directly decoded without I/O submission.
252 	 */
253 	Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE,
254 	/*
255 	 * The current collection has been linked with the owned chain, and
256 	 * could also be linked with the remaining collections, which means
257 	 * if the processing page is the tail page of the collection, thus
258 	 * the current collection can safely use the whole page (since
259 	 * the previous collection is under control) for in-place I/O, as
260 	 * illustrated below:
261 	 *  ________________________________________________________________
262 	 * |  tail (partial) page |          head (partial) page           |
263 	 * |  (of the current cl) |      (of the previous collection)      |
264 	 * | PCLUSTER_FOLLOWED or |                                        |
265 	 * |_____PCLUSTER_HOOKED__|___________PCLUSTER_FOLLOWED____________|
266 	 *
267 	 * [  (*) the above page can be used as inplace I/O.               ]
268 	 */
269 	Z_EROFS_PCLUSTER_FOLLOWED,
270 };
271 
272 struct z_erofs_decompress_frontend {
273 	struct inode *const inode;
274 	struct erofs_map_blocks map;
275 	struct z_erofs_bvec_iter biter;
276 
277 	struct page *candidate_bvpage;
278 	struct z_erofs_pcluster *pcl, *tailpcl;
279 	z_erofs_next_pcluster_t owned_head;
280 	enum z_erofs_pclustermode mode;
281 
282 	bool readahead;
283 	/* used for applying cache strategy on the fly */
284 	bool backmost;
285 	erofs_off_t headoffset;
286 
287 	/* a pointer used to pick up inplace I/O pages */
288 	unsigned int icur;
289 };
290 
291 #define DECOMPRESS_FRONTEND_INIT(__i) { \
292 	.inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \
293 	.mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true }
294 
295 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe,
296 			       enum z_erofs_cache_alloctype type,
297 			       struct page **pagepool)
298 {
299 	struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));
300 	struct z_erofs_pcluster *pcl = fe->pcl;
301 	bool standalone = true;
302 	/*
303 	 * optimistic allocation without direct reclaim since inplace I/O
304 	 * can be used if low memory otherwise.
305 	 */
306 	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
307 			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
308 	unsigned int i;
309 
310 	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED)
311 		return;
312 
313 	for (i = 0; i < pcl->pclusterpages; ++i) {
314 		struct page *page;
315 		compressed_page_t t;
316 		struct page *newpage = NULL;
317 
318 		/* the compressed page was loaded before */
319 		if (READ_ONCE(pcl->compressed_bvecs[i].page))
320 			continue;
321 
322 		page = find_get_page(mc, pcl->obj.index + i);
323 
324 		if (page) {
325 			t = tag_compressed_page_justfound(page);
326 		} else {
327 			/* I/O is needed, no possible to decompress directly */
328 			standalone = false;
329 			switch (type) {
330 			case TRYALLOC:
331 				newpage = erofs_allocpage(pagepool, gfp);
332 				if (!newpage)
333 					continue;
334 				set_page_private(newpage,
335 						 Z_EROFS_PREALLOCATED_PAGE);
336 				t = tag_compressed_page_justfound(newpage);
337 				break;
338 			default:        /* DONTALLOC */
339 				continue;
340 			}
341 		}
342 
343 		if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL,
344 				     tagptr_cast_ptr(t)))
345 			continue;
346 
347 		if (page)
348 			put_page(page);
349 		else if (newpage)
350 			erofs_pagepool_add(pagepool, newpage);
351 	}
352 
353 	/*
354 	 * don't do inplace I/O if all compressed pages are available in
355 	 * managed cache since it can be moved to the bypass queue instead.
356 	 */
357 	if (standalone)
358 		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
359 }
360 
361 /* called by erofs_shrinker to get rid of all compressed_pages */
362 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
363 				       struct erofs_workgroup *grp)
364 {
365 	struct z_erofs_pcluster *const pcl =
366 		container_of(grp, struct z_erofs_pcluster, obj);
367 	int i;
368 
369 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
370 	/*
371 	 * refcount of workgroup is now freezed as 1,
372 	 * therefore no need to worry about available decompression users.
373 	 */
374 	for (i = 0; i < pcl->pclusterpages; ++i) {
375 		struct page *page = pcl->compressed_bvecs[i].page;
376 
377 		if (!page)
378 			continue;
379 
380 		/* block other users from reclaiming or migrating the page */
381 		if (!trylock_page(page))
382 			return -EBUSY;
383 
384 		if (!erofs_page_is_managed(sbi, page))
385 			continue;
386 
387 		/* barrier is implied in the following 'unlock_page' */
388 		WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
389 		detach_page_private(page);
390 		unlock_page(page);
391 	}
392 	return 0;
393 }
394 
395 int erofs_try_to_free_cached_page(struct page *page)
396 {
397 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
398 	int ret, i;
399 
400 	if (!erofs_workgroup_try_to_freeze(&pcl->obj, 1))
401 		return 0;
402 
403 	ret = 0;
404 	DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
405 	for (i = 0; i < pcl->pclusterpages; ++i) {
406 		if (pcl->compressed_bvecs[i].page == page) {
407 			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
408 			ret = 1;
409 			break;
410 		}
411 	}
412 	erofs_workgroup_unfreeze(&pcl->obj, 1);
413 	if (ret)
414 		detach_page_private(page);
415 	return ret;
416 }
417 
418 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
419 				   struct z_erofs_bvec *bvec)
420 {
421 	struct z_erofs_pcluster *const pcl = fe->pcl;
422 
423 	while (fe->icur > 0) {
424 		if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
425 			     NULL, bvec->page)) {
426 			pcl->compressed_bvecs[fe->icur] = *bvec;
427 			return true;
428 		}
429 	}
430 	return false;
431 }
432 
433 /* callers must be with pcluster lock held */
434 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,
435 			       struct z_erofs_bvec *bvec, bool exclusive)
436 {
437 	int ret;
438 
439 	if (exclusive) {
440 		/* give priority for inplaceio to use file pages first */
441 		if (z_erofs_try_inplace_io(fe, bvec))
442 			return 0;
443 		/* otherwise, check if it can be used as a bvpage */
444 		if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED &&
445 		    !fe->candidate_bvpage)
446 			fe->candidate_bvpage = bvec->page;
447 	}
448 	ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage);
449 	fe->pcl->vcnt += (ret >= 0);
450 	return ret;
451 }
452 
453 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
454 {
455 	struct z_erofs_pcluster *pcl = f->pcl;
456 	z_erofs_next_pcluster_t *owned_head = &f->owned_head;
457 
458 	/* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
459 	if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
460 		    *owned_head) == Z_EROFS_PCLUSTER_NIL) {
461 		*owned_head = &pcl->next;
462 		/* so we can attach this pcluster to our submission chain. */
463 		f->mode = Z_EROFS_PCLUSTER_FOLLOWED;
464 		return;
465 	}
466 
467 	/*
468 	 * type 2, link to the end of an existing open chain, be careful
469 	 * that its submission is controlled by the original attached chain.
470 	 */
471 	if (*owned_head != &pcl->next && pcl != f->tailpcl &&
472 	    cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
473 		    *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
474 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
475 		f->mode = Z_EROFS_PCLUSTER_HOOKED;
476 		f->tailpcl = NULL;
477 		return;
478 	}
479 	/* type 3, it belongs to a chain, but it isn't the end of the chain */
480 	f->mode = Z_EROFS_PCLUSTER_INFLIGHT;
481 }
482 
483 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)
484 {
485 	struct erofs_map_blocks *map = &fe->map;
486 	bool ztailpacking = map->m_flags & EROFS_MAP_META;
487 	struct z_erofs_pcluster *pcl;
488 	struct erofs_workgroup *grp;
489 	int err;
490 
491 	if (!(map->m_flags & EROFS_MAP_ENCODED)) {
492 		DBG_BUGON(1);
493 		return -EFSCORRUPTED;
494 	}
495 
496 	/* no available pcluster, let's allocate one */
497 	pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
498 				     map->m_plen >> PAGE_SHIFT);
499 	if (IS_ERR(pcl))
500 		return PTR_ERR(pcl);
501 
502 	atomic_set(&pcl->obj.refcount, 1);
503 	pcl->algorithmformat = map->m_algorithmformat;
504 	pcl->length = 0;
505 	pcl->partial = true;
506 
507 	/* new pclusters should be claimed as type 1, primary and followed */
508 	pcl->next = fe->owned_head;
509 	pcl->pageofs_out = map->m_la & ~PAGE_MASK;
510 	fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;
511 
512 	/*
513 	 * lock all primary followed works before visible to others
514 	 * and mutex_trylock *never* fails for a new pcluster.
515 	 */
516 	mutex_init(&pcl->lock);
517 	DBG_BUGON(!mutex_trylock(&pcl->lock));
518 
519 	if (ztailpacking) {
520 		pcl->obj.index = 0;	/* which indicates ztailpacking */
521 		pcl->pageofs_in = erofs_blkoff(map->m_pa);
522 		pcl->tailpacking_size = map->m_plen;
523 	} else {
524 		pcl->obj.index = map->m_pa >> PAGE_SHIFT;
525 
526 		grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);
527 		if (IS_ERR(grp)) {
528 			err = PTR_ERR(grp);
529 			goto err_out;
530 		}
531 
532 		if (grp != &pcl->obj) {
533 			fe->pcl = container_of(grp,
534 					struct z_erofs_pcluster, obj);
535 			err = -EEXIST;
536 			goto err_out;
537 		}
538 	}
539 	/* used to check tail merging loop due to corrupted images */
540 	if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
541 		fe->tailpcl = pcl;
542 	fe->owned_head = &pcl->next;
543 	fe->pcl = pcl;
544 	return 0;
545 
546 err_out:
547 	mutex_unlock(&pcl->lock);
548 	z_erofs_free_pcluster(pcl);
549 	return err;
550 }
551 
552 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
553 {
554 	struct erofs_map_blocks *map = &fe->map;
555 	struct erofs_workgroup *grp = NULL;
556 	int ret;
557 
558 	DBG_BUGON(fe->pcl);
559 
560 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */
561 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
562 	DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
563 
564 	if (!(map->m_flags & EROFS_MAP_META)) {
565 		grp = erofs_find_workgroup(fe->inode->i_sb,
566 					   map->m_pa >> PAGE_SHIFT);
567 	} else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
568 		DBG_BUGON(1);
569 		return -EFSCORRUPTED;
570 	}
571 
572 	if (grp) {
573 		fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
574 		ret = -EEXIST;
575 	} else {
576 		ret = z_erofs_register_pcluster(fe);
577 	}
578 
579 	if (ret == -EEXIST) {
580 		mutex_lock(&fe->pcl->lock);
581 		/* used to check tail merging loop due to corrupted images */
582 		if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
583 			fe->tailpcl = fe->pcl;
584 
585 		z_erofs_try_to_claim_pcluster(fe);
586 	} else if (ret) {
587 		return ret;
588 	}
589 	z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
590 				Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
591 	/* since file-backed online pages are traversed in reverse order */
592 	fe->icur = z_erofs_pclusterpages(fe->pcl);
593 	return 0;
594 }
595 
596 /*
597  * keep in mind that no referenced pclusters will be freed
598  * only after a RCU grace period.
599  */
600 static void z_erofs_rcu_callback(struct rcu_head *head)
601 {
602 	z_erofs_free_pcluster(container_of(head,
603 			struct z_erofs_pcluster, rcu));
604 }
605 
606 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
607 {
608 	struct z_erofs_pcluster *const pcl =
609 		container_of(grp, struct z_erofs_pcluster, obj);
610 
611 	call_rcu(&pcl->rcu, z_erofs_rcu_callback);
612 }
613 
614 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe)
615 {
616 	struct z_erofs_pcluster *pcl = fe->pcl;
617 
618 	if (!pcl)
619 		return false;
620 
621 	z_erofs_bvec_iter_end(&fe->biter);
622 	mutex_unlock(&pcl->lock);
623 
624 	if (fe->candidate_bvpage) {
625 		DBG_BUGON(z_erofs_is_shortlived_page(fe->candidate_bvpage));
626 		fe->candidate_bvpage = NULL;
627 	}
628 
629 	/*
630 	 * if all pending pages are added, don't hold its reference
631 	 * any longer if the pcluster isn't hosted by ourselves.
632 	 */
633 	if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE)
634 		erofs_workgroup_put(&pcl->obj);
635 
636 	fe->pcl = NULL;
637 	return true;
638 }
639 
640 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
641 				       unsigned int cachestrategy,
642 				       erofs_off_t la)
643 {
644 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
645 		return false;
646 
647 	if (fe->backmost)
648 		return true;
649 
650 	return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
651 		la < fe->headoffset;
652 }
653 
654 static int z_erofs_read_fragment(struct inode *inode, erofs_off_t pos,
655 				 struct page *page, unsigned int pageofs,
656 				 unsigned int len)
657 {
658 	struct inode *packed_inode = EROFS_I_SB(inode)->packed_inode;
659 	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
660 	u8 *src, *dst;
661 	unsigned int i, cnt;
662 
663 	pos += EROFS_I(inode)->z_fragmentoff;
664 	for (i = 0; i < len; i += cnt) {
665 		cnt = min_t(unsigned int, len - i,
666 			    EROFS_BLKSIZ - erofs_blkoff(pos));
667 		src = erofs_bread(&buf, packed_inode,
668 				  erofs_blknr(pos), EROFS_KMAP);
669 		if (IS_ERR(src)) {
670 			erofs_put_metabuf(&buf);
671 			return PTR_ERR(src);
672 		}
673 
674 		dst = kmap_local_page(page);
675 		memcpy(dst + pageofs + i, src + erofs_blkoff(pos), cnt);
676 		kunmap_local(dst);
677 		pos += cnt;
678 	}
679 	erofs_put_metabuf(&buf);
680 	return 0;
681 }
682 
683 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
684 				struct page *page, struct page **pagepool)
685 {
686 	struct inode *const inode = fe->inode;
687 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
688 	struct erofs_map_blocks *const map = &fe->map;
689 	const loff_t offset = page_offset(page);
690 	bool tight = true, exclusive;
691 
692 	enum z_erofs_cache_alloctype cache_strategy;
693 	unsigned int cur, end, spiltted;
694 	int err = 0;
695 
696 	/* register locked file pages as online pages in pack */
697 	z_erofs_onlinepage_init(page);
698 
699 	spiltted = 0;
700 	end = PAGE_SIZE;
701 repeat:
702 	cur = end - 1;
703 
704 	if (offset + cur < map->m_la ||
705 	    offset + cur >= map->m_la + map->m_llen) {
706 		erofs_dbg("out-of-range map @ pos %llu", offset + cur);
707 
708 		if (z_erofs_collector_end(fe))
709 			fe->backmost = false;
710 		map->m_la = offset + cur;
711 		map->m_llen = 0;
712 		err = z_erofs_map_blocks_iter(inode, map, 0);
713 		if (err)
714 			goto out;
715 	} else {
716 		if (fe->pcl)
717 			goto hitted;
718 		/* didn't get a valid pcluster previously (very rare) */
719 	}
720 
721 	if (!(map->m_flags & EROFS_MAP_MAPPED) ||
722 	    map->m_flags & EROFS_MAP_FRAGMENT)
723 		goto hitted;
724 
725 	err = z_erofs_collector_begin(fe);
726 	if (err)
727 		goto out;
728 
729 	if (z_erofs_is_inline_pcluster(fe->pcl)) {
730 		void *mp;
731 
732 		mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
733 					erofs_blknr(map->m_pa), EROFS_NO_KMAP);
734 		if (IS_ERR(mp)) {
735 			err = PTR_ERR(mp);
736 			erofs_err(inode->i_sb,
737 				  "failed to get inline page, err %d", err);
738 			goto out;
739 		}
740 		get_page(fe->map.buf.page);
741 		WRITE_ONCE(fe->pcl->compressed_bvecs[0].page,
742 			   fe->map.buf.page);
743 		fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
744 	} else {
745 		/* bind cache first when cached decompression is preferred */
746 		if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy,
747 					       map->m_la))
748 			cache_strategy = TRYALLOC;
749 		else
750 			cache_strategy = DONTALLOC;
751 
752 		z_erofs_bind_cache(fe, cache_strategy, pagepool);
753 	}
754 hitted:
755 	/*
756 	 * Ensure the current partial page belongs to this submit chain rather
757 	 * than other concurrent submit chains or the noio(bypass) chain since
758 	 * those chains are handled asynchronously thus the page cannot be used
759 	 * for inplace I/O or bvpage (should be processed in a strict order.)
760 	 */
761 	tight &= (fe->mode >= Z_EROFS_PCLUSTER_HOOKED &&
762 		  fe->mode != Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE);
763 
764 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
765 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
766 		zero_user_segment(page, cur, end);
767 		goto next_part;
768 	}
769 	if (map->m_flags & EROFS_MAP_FRAGMENT) {
770 		unsigned int pageofs, skip, len;
771 
772 		if (offset > map->m_la) {
773 			pageofs = 0;
774 			skip = offset - map->m_la;
775 		} else {
776 			pageofs = map->m_la & ~PAGE_MASK;
777 			skip = 0;
778 		}
779 		len = min_t(unsigned int, map->m_llen - skip, end - cur);
780 		err = z_erofs_read_fragment(inode, skip, page, pageofs, len);
781 		if (err)
782 			goto out;
783 		++spiltted;
784 		tight = false;
785 		goto next_part;
786 	}
787 
788 	exclusive = (!cur && (!spiltted || tight));
789 	if (cur)
790 		tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED);
791 
792 retry:
793 	err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) {
794 					.page = page,
795 					.offset = offset - map->m_la,
796 					.end = end,
797 				  }), exclusive);
798 	/* should allocate an additional short-lived page for bvset */
799 	if (err == -EAGAIN && !fe->candidate_bvpage) {
800 		fe->candidate_bvpage = alloc_page(GFP_NOFS | __GFP_NOFAIL);
801 		set_page_private(fe->candidate_bvpage,
802 				 Z_EROFS_SHORTLIVED_PAGE);
803 		goto retry;
804 	}
805 
806 	if (err) {
807 		DBG_BUGON(err == -EAGAIN && fe->candidate_bvpage);
808 		goto out;
809 	}
810 
811 	z_erofs_onlinepage_split(page);
812 	/* bump up the number of spiltted parts of a page */
813 	++spiltted;
814 	if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))
815 		fe->pcl->multibases = true;
816 
817 	if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&
818 	    !(map->m_flags & EROFS_MAP_PARTIAL_REF) &&
819 	    fe->pcl->length == map->m_llen)
820 		fe->pcl->partial = false;
821 	if (fe->pcl->length < offset + end - map->m_la) {
822 		fe->pcl->length = offset + end - map->m_la;
823 		fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK;
824 	}
825 next_part:
826 	/* shorten the remaining extent to update progress */
827 	map->m_llen = offset + cur - map->m_la;
828 	map->m_flags &= ~EROFS_MAP_FULL_MAPPED;
829 
830 	end = cur;
831 	if (end > 0)
832 		goto repeat;
833 
834 out:
835 	if (err)
836 		z_erofs_page_mark_eio(page);
837 	z_erofs_onlinepage_endio(page);
838 
839 	erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
840 		  __func__, page, spiltted, map->m_llen);
841 	return err;
842 }
843 
844 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi,
845 				       unsigned int readahead_pages)
846 {
847 	/* auto: enable for read_folio, disable for readahead */
848 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
849 	    !readahead_pages)
850 		return true;
851 
852 	if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
853 	    (readahead_pages <= sbi->opt.max_sync_decompress_pages))
854 		return true;
855 
856 	return false;
857 }
858 
859 static bool z_erofs_page_is_invalidated(struct page *page)
860 {
861 	return !page->mapping && !z_erofs_is_shortlived_page(page);
862 }
863 
864 struct z_erofs_decompress_backend {
865 	struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];
866 	struct super_block *sb;
867 	struct z_erofs_pcluster *pcl;
868 
869 	/* pages with the longest decompressed length for deduplication */
870 	struct page **decompressed_pages;
871 	/* pages to keep the compressed data */
872 	struct page **compressed_pages;
873 
874 	struct list_head decompressed_secondary_bvecs;
875 	struct page **pagepool;
876 	unsigned int onstack_used, nr_pages;
877 };
878 
879 struct z_erofs_bvec_item {
880 	struct z_erofs_bvec bvec;
881 	struct list_head list;
882 };
883 
884 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,
885 					 struct z_erofs_bvec *bvec)
886 {
887 	struct z_erofs_bvec_item *item;
888 
889 	if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK)) {
890 		unsigned int pgnr;
891 		struct page *oldpage;
892 
893 		pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;
894 		DBG_BUGON(pgnr >= be->nr_pages);
895 		oldpage = be->decompressed_pages[pgnr];
896 		be->decompressed_pages[pgnr] = bvec->page;
897 
898 		if (!oldpage)
899 			return;
900 	}
901 
902 	/* (cold path) one pcluster is requested multiple times */
903 	item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);
904 	item->bvec = *bvec;
905 	list_add(&item->list, &be->decompressed_secondary_bvecs);
906 }
907 
908 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,
909 				      int err)
910 {
911 	unsigned int off0 = be->pcl->pageofs_out;
912 	struct list_head *p, *n;
913 
914 	list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {
915 		struct z_erofs_bvec_item *bvi;
916 		unsigned int end, cur;
917 		void *dst, *src;
918 
919 		bvi = container_of(p, struct z_erofs_bvec_item, list);
920 		cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;
921 		end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,
922 			    bvi->bvec.end);
923 		dst = kmap_local_page(bvi->bvec.page);
924 		while (cur < end) {
925 			unsigned int pgnr, scur, len;
926 
927 			pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;
928 			DBG_BUGON(pgnr >= be->nr_pages);
929 
930 			scur = bvi->bvec.offset + cur -
931 					((pgnr << PAGE_SHIFT) - off0);
932 			len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);
933 			if (!be->decompressed_pages[pgnr]) {
934 				err = -EFSCORRUPTED;
935 				cur += len;
936 				continue;
937 			}
938 			src = kmap_local_page(be->decompressed_pages[pgnr]);
939 			memcpy(dst + cur, src + scur, len);
940 			kunmap_local(src);
941 			cur += len;
942 		}
943 		kunmap_local(dst);
944 		if (err)
945 			z_erofs_page_mark_eio(bvi->bvec.page);
946 		z_erofs_onlinepage_endio(bvi->bvec.page);
947 		list_del(p);
948 		kfree(bvi);
949 	}
950 }
951 
952 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)
953 {
954 	struct z_erofs_pcluster *pcl = be->pcl;
955 	struct z_erofs_bvec_iter biter;
956 	struct page *old_bvpage;
957 	int i;
958 
959 	z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);
960 	for (i = 0; i < pcl->vcnt; ++i) {
961 		struct z_erofs_bvec bvec;
962 
963 		z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);
964 
965 		if (old_bvpage)
966 			z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
967 
968 		DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));
969 		z_erofs_do_decompressed_bvec(be, &bvec);
970 	}
971 
972 	old_bvpage = z_erofs_bvec_iter_end(&biter);
973 	if (old_bvpage)
974 		z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
975 }
976 
977 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,
978 				  bool *overlapped)
979 {
980 	struct z_erofs_pcluster *pcl = be->pcl;
981 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
982 	int i, err = 0;
983 
984 	*overlapped = false;
985 	for (i = 0; i < pclusterpages; ++i) {
986 		struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];
987 		struct page *page = bvec->page;
988 
989 		/* compressed pages ought to be present before decompressing */
990 		if (!page) {
991 			DBG_BUGON(1);
992 			continue;
993 		}
994 		be->compressed_pages[i] = page;
995 
996 		if (z_erofs_is_inline_pcluster(pcl)) {
997 			if (!PageUptodate(page))
998 				err = -EIO;
999 			continue;
1000 		}
1001 
1002 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1003 		if (!z_erofs_is_shortlived_page(page)) {
1004 			if (erofs_page_is_managed(EROFS_SB(be->sb), page)) {
1005 				if (!PageUptodate(page))
1006 					err = -EIO;
1007 				continue;
1008 			}
1009 			z_erofs_do_decompressed_bvec(be, bvec);
1010 			*overlapped = true;
1011 		}
1012 	}
1013 
1014 	if (err)
1015 		return err;
1016 	return 0;
1017 }
1018 
1019 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,
1020 				       int err)
1021 {
1022 	struct erofs_sb_info *const sbi = EROFS_SB(be->sb);
1023 	struct z_erofs_pcluster *pcl = be->pcl;
1024 	unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1025 	unsigned int i, inputsize;
1026 	int err2;
1027 	struct page *page;
1028 	bool overlapped;
1029 
1030 	mutex_lock(&pcl->lock);
1031 	be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;
1032 
1033 	/* allocate (de)compressed page arrays if cannot be kept on stack */
1034 	be->decompressed_pages = NULL;
1035 	be->compressed_pages = NULL;
1036 	be->onstack_used = 0;
1037 	if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {
1038 		be->decompressed_pages = be->onstack_pages;
1039 		be->onstack_used = be->nr_pages;
1040 		memset(be->decompressed_pages, 0,
1041 		       sizeof(struct page *) * be->nr_pages);
1042 	}
1043 
1044 	if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)
1045 		be->compressed_pages = be->onstack_pages + be->onstack_used;
1046 
1047 	if (!be->decompressed_pages)
1048 		be->decompressed_pages =
1049 			kvcalloc(be->nr_pages, sizeof(struct page *),
1050 				 GFP_KERNEL | __GFP_NOFAIL);
1051 	if (!be->compressed_pages)
1052 		be->compressed_pages =
1053 			kvcalloc(pclusterpages, sizeof(struct page *),
1054 				 GFP_KERNEL | __GFP_NOFAIL);
1055 
1056 	z_erofs_parse_out_bvecs(be);
1057 	err2 = z_erofs_parse_in_bvecs(be, &overlapped);
1058 	if (err2)
1059 		err = err2;
1060 	if (err)
1061 		goto out;
1062 
1063 	if (z_erofs_is_inline_pcluster(pcl))
1064 		inputsize = pcl->tailpacking_size;
1065 	else
1066 		inputsize = pclusterpages * PAGE_SIZE;
1067 
1068 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
1069 					.sb = be->sb,
1070 					.in = be->compressed_pages,
1071 					.out = be->decompressed_pages,
1072 					.pageofs_in = pcl->pageofs_in,
1073 					.pageofs_out = pcl->pageofs_out,
1074 					.inputsize = inputsize,
1075 					.outputsize = pcl->length,
1076 					.alg = pcl->algorithmformat,
1077 					.inplace_io = overlapped,
1078 					.partial_decoding = pcl->partial,
1079 					.fillgaps = pcl->multibases,
1080 				 }, be->pagepool);
1081 
1082 out:
1083 	/* must handle all compressed pages before actual file pages */
1084 	if (z_erofs_is_inline_pcluster(pcl)) {
1085 		page = pcl->compressed_bvecs[0].page;
1086 		WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);
1087 		put_page(page);
1088 	} else {
1089 		for (i = 0; i < pclusterpages; ++i) {
1090 			page = pcl->compressed_bvecs[i].page;
1091 
1092 			if (erofs_page_is_managed(sbi, page))
1093 				continue;
1094 
1095 			/* recycle all individual short-lived pages */
1096 			(void)z_erofs_put_shortlivedpage(be->pagepool, page);
1097 			WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
1098 		}
1099 	}
1100 	if (be->compressed_pages < be->onstack_pages ||
1101 	    be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)
1102 		kvfree(be->compressed_pages);
1103 	z_erofs_fill_other_copies(be, err);
1104 
1105 	for (i = 0; i < be->nr_pages; ++i) {
1106 		page = be->decompressed_pages[i];
1107 		if (!page)
1108 			continue;
1109 
1110 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1111 
1112 		/* recycle all individual short-lived pages */
1113 		if (z_erofs_put_shortlivedpage(be->pagepool, page))
1114 			continue;
1115 		if (err)
1116 			z_erofs_page_mark_eio(page);
1117 		z_erofs_onlinepage_endio(page);
1118 	}
1119 
1120 	if (be->decompressed_pages != be->onstack_pages)
1121 		kvfree(be->decompressed_pages);
1122 
1123 	pcl->length = 0;
1124 	pcl->partial = true;
1125 	pcl->multibases = false;
1126 	pcl->bvset.nextpage = NULL;
1127 	pcl->vcnt = 0;
1128 
1129 	/* pcluster lock MUST be taken before the following line */
1130 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1131 	mutex_unlock(&pcl->lock);
1132 	return err;
1133 }
1134 
1135 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1136 				     struct page **pagepool)
1137 {
1138 	struct z_erofs_decompress_backend be = {
1139 		.sb = io->sb,
1140 		.pagepool = pagepool,
1141 		.decompressed_secondary_bvecs =
1142 			LIST_HEAD_INIT(be.decompressed_secondary_bvecs),
1143 	};
1144 	z_erofs_next_pcluster_t owned = io->head;
1145 
1146 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1147 		/* impossible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1148 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1149 		/* impossible that 'owned' equals Z_EROFS_PCLUSTER_NIL */
1150 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1151 
1152 		be.pcl = container_of(owned, struct z_erofs_pcluster, next);
1153 		owned = READ_ONCE(be.pcl->next);
1154 
1155 		z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0);
1156 		erofs_workgroup_put(&be.pcl->obj);
1157 	}
1158 }
1159 
1160 static void z_erofs_decompressqueue_work(struct work_struct *work)
1161 {
1162 	struct z_erofs_decompressqueue *bgq =
1163 		container_of(work, struct z_erofs_decompressqueue, u.work);
1164 	struct page *pagepool = NULL;
1165 
1166 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1167 	z_erofs_decompress_queue(bgq, &pagepool);
1168 
1169 	erofs_release_pages(&pagepool);
1170 	kvfree(bgq);
1171 }
1172 
1173 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1174 				       bool sync, int bios)
1175 {
1176 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1177 
1178 	/* wake up the caller thread for sync decompression */
1179 	if (sync) {
1180 		if (!atomic_add_return(bios, &io->pending_bios))
1181 			complete(&io->u.done);
1182 		return;
1183 	}
1184 
1185 	if (atomic_add_return(bios, &io->pending_bios))
1186 		return;
1187 	/* Use workqueue and sync decompression for atomic contexts only */
1188 	if (in_atomic() || irqs_disabled()) {
1189 		queue_work(z_erofs_workqueue, &io->u.work);
1190 		/* enable sync decompression for readahead */
1191 		if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1192 			sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1193 		return;
1194 	}
1195 	z_erofs_decompressqueue_work(&io->u.work);
1196 }
1197 
1198 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1199 					       unsigned int nr,
1200 					       struct page **pagepool,
1201 					       struct address_space *mc)
1202 {
1203 	const pgoff_t index = pcl->obj.index;
1204 	gfp_t gfp = mapping_gfp_mask(mc);
1205 	bool tocache = false;
1206 
1207 	struct address_space *mapping;
1208 	struct page *oldpage, *page;
1209 
1210 	compressed_page_t t;
1211 	int justfound;
1212 
1213 repeat:
1214 	page = READ_ONCE(pcl->compressed_bvecs[nr].page);
1215 	oldpage = page;
1216 
1217 	if (!page)
1218 		goto out_allocpage;
1219 
1220 	/* process the target tagged pointer */
1221 	t = tagptr_init(compressed_page_t, page);
1222 	justfound = tagptr_unfold_tags(t);
1223 	page = tagptr_unfold_ptr(t);
1224 
1225 	/*
1226 	 * preallocated cached pages, which is used to avoid direct reclaim
1227 	 * otherwise, it will go inplace I/O path instead.
1228 	 */
1229 	if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1230 		WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1231 		set_page_private(page, 0);
1232 		tocache = true;
1233 		goto out_tocache;
1234 	}
1235 	mapping = READ_ONCE(page->mapping);
1236 
1237 	/*
1238 	 * file-backed online pages in plcuster are all locked steady,
1239 	 * therefore it is impossible for `mapping' to be NULL.
1240 	 */
1241 	if (mapping && mapping != mc)
1242 		/* ought to be unmanaged pages */
1243 		goto out;
1244 
1245 	/* directly return for shortlived page as well */
1246 	if (z_erofs_is_shortlived_page(page))
1247 		goto out;
1248 
1249 	lock_page(page);
1250 
1251 	/* only true if page reclaim goes wrong, should never happen */
1252 	DBG_BUGON(justfound && PagePrivate(page));
1253 
1254 	/* the page is still in manage cache */
1255 	if (page->mapping == mc) {
1256 		WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1257 
1258 		if (!PagePrivate(page)) {
1259 			/*
1260 			 * impossible to be !PagePrivate(page) for
1261 			 * the current restriction as well if
1262 			 * the page is already in compressed_bvecs[].
1263 			 */
1264 			DBG_BUGON(!justfound);
1265 
1266 			justfound = 0;
1267 			set_page_private(page, (unsigned long)pcl);
1268 			SetPagePrivate(page);
1269 		}
1270 
1271 		/* no need to submit io if it is already up-to-date */
1272 		if (PageUptodate(page)) {
1273 			unlock_page(page);
1274 			page = NULL;
1275 		}
1276 		goto out;
1277 	}
1278 
1279 	/*
1280 	 * the managed page has been truncated, it's unsafe to
1281 	 * reuse this one, let's allocate a new cache-managed page.
1282 	 */
1283 	DBG_BUGON(page->mapping);
1284 	DBG_BUGON(!justfound);
1285 
1286 	tocache = true;
1287 	unlock_page(page);
1288 	put_page(page);
1289 out_allocpage:
1290 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1291 	if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page,
1292 			       oldpage, page)) {
1293 		erofs_pagepool_add(pagepool, page);
1294 		cond_resched();
1295 		goto repeat;
1296 	}
1297 out_tocache:
1298 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1299 		/* turn into temporary page if fails (1 ref) */
1300 		set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1301 		goto out;
1302 	}
1303 	attach_page_private(page, pcl);
1304 	/* drop a refcount added by allocpage (then we have 2 refs here) */
1305 	put_page(page);
1306 
1307 out:	/* the only exit (for tracing and debugging) */
1308 	return page;
1309 }
1310 
1311 static struct z_erofs_decompressqueue *
1312 jobqueue_init(struct super_block *sb,
1313 	      struct z_erofs_decompressqueue *fgq, bool *fg)
1314 {
1315 	struct z_erofs_decompressqueue *q;
1316 
1317 	if (fg && !*fg) {
1318 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1319 		if (!q) {
1320 			*fg = true;
1321 			goto fg_out;
1322 		}
1323 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1324 	} else {
1325 fg_out:
1326 		q = fgq;
1327 		init_completion(&fgq->u.done);
1328 		atomic_set(&fgq->pending_bios, 0);
1329 		q->eio = false;
1330 	}
1331 	q->sb = sb;
1332 	q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1333 	return q;
1334 }
1335 
1336 /* define decompression jobqueue types */
1337 enum {
1338 	JQ_BYPASS,
1339 	JQ_SUBMIT,
1340 	NR_JOBQUEUES,
1341 };
1342 
1343 static void *jobqueueset_init(struct super_block *sb,
1344 			      struct z_erofs_decompressqueue *q[],
1345 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1346 {
1347 	/*
1348 	 * if managed cache is enabled, bypass jobqueue is needed,
1349 	 * no need to read from device for all pclusters in this queue.
1350 	 */
1351 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1352 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1353 
1354 	return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1355 }
1356 
1357 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1358 				    z_erofs_next_pcluster_t qtail[],
1359 				    z_erofs_next_pcluster_t owned_head)
1360 {
1361 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1362 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1363 
1364 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1365 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1366 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1367 
1368 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1369 
1370 	WRITE_ONCE(*submit_qtail, owned_head);
1371 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1372 
1373 	qtail[JQ_BYPASS] = &pcl->next;
1374 }
1375 
1376 static void z_erofs_decompressqueue_endio(struct bio *bio)
1377 {
1378 	tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
1379 	struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
1380 	blk_status_t err = bio->bi_status;
1381 	struct bio_vec *bvec;
1382 	struct bvec_iter_all iter_all;
1383 
1384 	bio_for_each_segment_all(bvec, bio, iter_all) {
1385 		struct page *page = bvec->bv_page;
1386 
1387 		DBG_BUGON(PageUptodate(page));
1388 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1389 
1390 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1391 			if (!err)
1392 				SetPageUptodate(page);
1393 			unlock_page(page);
1394 		}
1395 	}
1396 	if (err)
1397 		q->eio = true;
1398 	z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
1399 	bio_put(bio);
1400 }
1401 
1402 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
1403 				 struct page **pagepool,
1404 				 struct z_erofs_decompressqueue *fgq,
1405 				 bool *force_fg)
1406 {
1407 	struct super_block *sb = f->inode->i_sb;
1408 	struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));
1409 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1410 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1411 	void *bi_private;
1412 	z_erofs_next_pcluster_t owned_head = f->owned_head;
1413 	/* bio is NULL initially, so no need to initialize last_{index,bdev} */
1414 	pgoff_t last_index;
1415 	struct block_device *last_bdev;
1416 	unsigned int nr_bios = 0;
1417 	struct bio *bio = NULL;
1418 	/* initialize to 1 to make skip psi_memstall_leave unless needed */
1419 	unsigned long pflags = 1;
1420 
1421 	bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1422 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1423 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1424 
1425 	/* by default, all need io submission */
1426 	q[JQ_SUBMIT]->head = owned_head;
1427 
1428 	do {
1429 		struct erofs_map_dev mdev;
1430 		struct z_erofs_pcluster *pcl;
1431 		pgoff_t cur, end;
1432 		unsigned int i = 0;
1433 		bool bypass = true;
1434 
1435 		/* no possible 'owned_head' equals the following */
1436 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1437 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1438 
1439 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1440 
1441 		/* close the main owned chain at first */
1442 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1443 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1444 		if (z_erofs_is_inline_pcluster(pcl)) {
1445 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1446 			continue;
1447 		}
1448 
1449 		/* no device id here, thus it will always succeed */
1450 		mdev = (struct erofs_map_dev) {
1451 			.m_pa = blknr_to_addr(pcl->obj.index),
1452 		};
1453 		(void)erofs_map_dev(sb, &mdev);
1454 
1455 		cur = erofs_blknr(mdev.m_pa);
1456 		end = cur + pcl->pclusterpages;
1457 
1458 		do {
1459 			struct page *page;
1460 
1461 			page = pickup_page_for_submission(pcl, i++, pagepool,
1462 							  mc);
1463 			if (!page)
1464 				continue;
1465 
1466 			if (bio && (cur != last_index + 1 ||
1467 				    last_bdev != mdev.m_bdev)) {
1468 submit_bio_retry:
1469 				if (!pflags)
1470 					psi_memstall_leave(&pflags);
1471 				submit_bio(bio);
1472 				bio = NULL;
1473 			}
1474 
1475 			if (unlikely(PageWorkingset(page)))
1476 				psi_memstall_enter(&pflags);
1477 
1478 			if (!bio) {
1479 				bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1480 						REQ_OP_READ, GFP_NOIO);
1481 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1482 
1483 				last_bdev = mdev.m_bdev;
1484 				bio->bi_iter.bi_sector = (sector_t)cur <<
1485 					LOG_SECTORS_PER_BLOCK;
1486 				bio->bi_private = bi_private;
1487 				if (f->readahead)
1488 					bio->bi_opf |= REQ_RAHEAD;
1489 				++nr_bios;
1490 			}
1491 
1492 			if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1493 				goto submit_bio_retry;
1494 
1495 			last_index = cur;
1496 			bypass = false;
1497 		} while (++cur < end);
1498 
1499 		if (!bypass)
1500 			qtail[JQ_SUBMIT] = &pcl->next;
1501 		else
1502 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1503 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1504 
1505 	if (bio) {
1506 		if (!pflags)
1507 			psi_memstall_leave(&pflags);
1508 		submit_bio(bio);
1509 	}
1510 
1511 	/*
1512 	 * although background is preferred, no one is pending for submission.
1513 	 * don't issue workqueue for decompression but drop it directly instead.
1514 	 */
1515 	if (!*force_fg && !nr_bios) {
1516 		kvfree(q[JQ_SUBMIT]);
1517 		return;
1518 	}
1519 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1520 }
1521 
1522 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f,
1523 			     struct page **pagepool, bool force_fg)
1524 {
1525 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1526 
1527 	if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1528 		return;
1529 	z_erofs_submit_queue(f, pagepool, io, &force_fg);
1530 
1531 	/* handle bypass queue (no i/o pclusters) immediately */
1532 	z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1533 
1534 	if (!force_fg)
1535 		return;
1536 
1537 	/* wait until all bios are completed */
1538 	wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1539 
1540 	/* handle synchronous decompress queue in the caller context */
1541 	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1542 }
1543 
1544 /*
1545  * Since partial uptodate is still unimplemented for now, we have to use
1546  * approximate readmore strategies as a start.
1547  */
1548 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1549 				      struct readahead_control *rac,
1550 				      erofs_off_t end,
1551 				      struct page **pagepool,
1552 				      bool backmost)
1553 {
1554 	struct inode *inode = f->inode;
1555 	struct erofs_map_blocks *map = &f->map;
1556 	erofs_off_t cur;
1557 	int err;
1558 
1559 	if (backmost) {
1560 		map->m_la = end;
1561 		err = z_erofs_map_blocks_iter(inode, map,
1562 					      EROFS_GET_BLOCKS_READMORE);
1563 		if (err)
1564 			return;
1565 
1566 		/* expend ra for the trailing edge if readahead */
1567 		if (rac) {
1568 			loff_t newstart = readahead_pos(rac);
1569 
1570 			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1571 			readahead_expand(rac, newstart, cur - newstart);
1572 			return;
1573 		}
1574 		end = round_up(end, PAGE_SIZE);
1575 	} else {
1576 		end = round_up(map->m_la, PAGE_SIZE);
1577 
1578 		if (!map->m_llen)
1579 			return;
1580 	}
1581 
1582 	cur = map->m_la + map->m_llen - 1;
1583 	while (cur >= end) {
1584 		pgoff_t index = cur >> PAGE_SHIFT;
1585 		struct page *page;
1586 
1587 		page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1588 		if (page) {
1589 			if (PageUptodate(page)) {
1590 				unlock_page(page);
1591 			} else {
1592 				err = z_erofs_do_read_page(f, page, pagepool);
1593 				if (err)
1594 					erofs_err(inode->i_sb,
1595 						  "readmore error at page %lu @ nid %llu",
1596 						  index, EROFS_I(inode)->nid);
1597 			}
1598 			put_page(page);
1599 		}
1600 
1601 		if (cur < PAGE_SIZE)
1602 			break;
1603 		cur = (index << PAGE_SHIFT) - 1;
1604 	}
1605 }
1606 
1607 static int z_erofs_read_folio(struct file *file, struct folio *folio)
1608 {
1609 	struct page *page = &folio->page;
1610 	struct inode *const inode = page->mapping->host;
1611 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1612 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1613 	struct page *pagepool = NULL;
1614 	int err;
1615 
1616 	trace_erofs_readpage(page, false);
1617 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1618 
1619 	z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1620 				  &pagepool, true);
1621 	err = z_erofs_do_read_page(&f, page, &pagepool);
1622 	z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1623 
1624 	(void)z_erofs_collector_end(&f);
1625 
1626 	/* if some compressed cluster ready, need submit them anyway */
1627 	z_erofs_runqueue(&f, &pagepool,
1628 			 z_erofs_get_sync_decompress_policy(sbi, 0));
1629 
1630 	if (err)
1631 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1632 
1633 	erofs_put_metabuf(&f.map.buf);
1634 	erofs_release_pages(&pagepool);
1635 	return err;
1636 }
1637 
1638 static void z_erofs_readahead(struct readahead_control *rac)
1639 {
1640 	struct inode *const inode = rac->mapping->host;
1641 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1642 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1643 	struct page *pagepool = NULL, *head = NULL, *page;
1644 	unsigned int nr_pages;
1645 
1646 	f.readahead = true;
1647 	f.headoffset = readahead_pos(rac);
1648 
1649 	z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1650 				  readahead_length(rac) - 1, &pagepool, true);
1651 	nr_pages = readahead_count(rac);
1652 	trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1653 
1654 	while ((page = readahead_page(rac))) {
1655 		set_page_private(page, (unsigned long)head);
1656 		head = page;
1657 	}
1658 
1659 	while (head) {
1660 		struct page *page = head;
1661 		int err;
1662 
1663 		/* traversal in reverse order */
1664 		head = (void *)page_private(page);
1665 
1666 		err = z_erofs_do_read_page(&f, page, &pagepool);
1667 		if (err)
1668 			erofs_err(inode->i_sb,
1669 				  "readahead error at page %lu @ nid %llu",
1670 				  page->index, EROFS_I(inode)->nid);
1671 		put_page(page);
1672 	}
1673 	z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1674 	(void)z_erofs_collector_end(&f);
1675 
1676 	z_erofs_runqueue(&f, &pagepool,
1677 			 z_erofs_get_sync_decompress_policy(sbi, nr_pages));
1678 	erofs_put_metabuf(&f.map.buf);
1679 	erofs_release_pages(&pagepool);
1680 }
1681 
1682 const struct address_space_operations z_erofs_aops = {
1683 	.read_folio = z_erofs_read_folio,
1684 	.readahead = z_erofs_readahead,
1685 };
1686