• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2018 HUAWEI, Inc.
4  *             http://www.huawei.com/
5  * Created by Gao Xiang <gaoxiang25@huawei.com>
6  */
7 #include "zdata.h"
8 #include "compress.h"
9 #include <linux/prefetch.h>
10 
11 #include <trace/events/erofs.h>
12 
13 /*
14  * a compressed_pages[] placeholder in order to avoid
15  * being filled with file pages for in-place decompression.
16  */
17 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
18 
19 /* how to allocate cached pages for a pcluster */
20 enum z_erofs_cache_alloctype {
21 	DONTALLOC,	/* don't allocate any cached pages */
22 	DELAYEDALLOC,	/* delayed allocation (at the time of submitting io) */
23 };
24 
25 /*
26  * tagged pointer with 1-bit tag for all compressed pages
27  * tag 0 - the page is just found with an extra page reference
28  */
29 typedef tagptr1_t compressed_page_t;
30 
31 #define tag_compressed_page_justfound(page) \
32 	tagptr_fold(compressed_page_t, page, 1)
33 
34 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
35 static struct kmem_cache *pcluster_cachep __read_mostly;
36 
z_erofs_exit_zip_subsystem(void)37 void z_erofs_exit_zip_subsystem(void)
38 {
39 	destroy_workqueue(z_erofs_workqueue);
40 	kmem_cache_destroy(pcluster_cachep);
41 }
42 
z_erofs_init_workqueue(void)43 static inline int z_erofs_init_workqueue(void)
44 {
45 	const unsigned int onlinecpus = num_possible_cpus();
46 	const unsigned int flags = WQ_UNBOUND | WQ_HIGHPRI | WQ_CPU_INTENSIVE;
47 
48 	/*
49 	 * no need to spawn too many threads, limiting threads could minimum
50 	 * scheduling overhead, perhaps per-CPU threads should be better?
51 	 */
52 	z_erofs_workqueue = alloc_workqueue("erofs_unzipd", flags,
53 					    onlinecpus + onlinecpus / 4);
54 	return z_erofs_workqueue ? 0 : -ENOMEM;
55 }
56 
z_erofs_pcluster_init_once(void * ptr)57 static void z_erofs_pcluster_init_once(void *ptr)
58 {
59 	struct z_erofs_pcluster *pcl = ptr;
60 	struct z_erofs_collection *cl = z_erofs_primarycollection(pcl);
61 	unsigned int i;
62 
63 	mutex_init(&cl->lock);
64 	cl->nr_pages = 0;
65 	cl->vcnt = 0;
66 	for (i = 0; i < Z_EROFS_CLUSTER_MAX_PAGES; ++i)
67 		pcl->compressed_pages[i] = NULL;
68 }
69 
z_erofs_pcluster_init_always(struct z_erofs_pcluster * pcl)70 static void z_erofs_pcluster_init_always(struct z_erofs_pcluster *pcl)
71 {
72 	struct z_erofs_collection *cl = z_erofs_primarycollection(pcl);
73 
74 	atomic_set(&pcl->obj.refcount, 1);
75 
76 	DBG_BUGON(cl->nr_pages);
77 	DBG_BUGON(cl->vcnt);
78 }
79 
z_erofs_init_zip_subsystem(void)80 int __init z_erofs_init_zip_subsystem(void)
81 {
82 	pcluster_cachep = kmem_cache_create("erofs_compress",
83 					    Z_EROFS_WORKGROUP_SIZE, 0,
84 					    SLAB_RECLAIM_ACCOUNT,
85 					    z_erofs_pcluster_init_once);
86 	if (pcluster_cachep) {
87 		if (!z_erofs_init_workqueue())
88 			return 0;
89 
90 		kmem_cache_destroy(pcluster_cachep);
91 	}
92 	return -ENOMEM;
93 }
94 
95 enum z_erofs_collectmode {
96 	COLLECT_SECONDARY,
97 	COLLECT_PRIMARY,
98 	/*
99 	 * The current collection was the tail of an exist chain, in addition
100 	 * that the previous processed chained collections are all decided to
101 	 * be hooked up to it.
102 	 * A new chain will be created for the remaining collections which are
103 	 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
104 	 * the next collection cannot reuse the whole page safely in
105 	 * the following scenario:
106 	 *  ________________________________________________________________
107 	 * |      tail (partial) page     |       head (partial) page       |
108 	 * |   (belongs to the next cl)   |   (belongs to the current cl)   |
109 	 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
110 	 */
111 	COLLECT_PRIMARY_HOOKED,
112 	COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
113 	/*
114 	 * The current collection has been linked with the owned chain, and
115 	 * could also be linked with the remaining collections, which means
116 	 * if the processing page is the tail page of the collection, thus
117 	 * the current collection can safely use the whole page (since
118 	 * the previous collection is under control) for in-place I/O, as
119 	 * illustrated below:
120 	 *  ________________________________________________________________
121 	 * |  tail (partial) page |          head (partial) page           |
122 	 * |  (of the current cl) |      (of the previous collection)      |
123 	 * |  PRIMARY_FOLLOWED or |                                        |
124 	 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
125 	 *
126 	 * [  (*) the above page can be used as inplace I/O.               ]
127 	 */
128 	COLLECT_PRIMARY_FOLLOWED,
129 };
130 
131 struct z_erofs_collector {
132 	struct z_erofs_pagevec_ctor vector;
133 
134 	struct z_erofs_pcluster *pcl, *tailpcl;
135 	struct z_erofs_collection *cl;
136 	struct page **compressedpages;
137 	z_erofs_next_pcluster_t owned_head;
138 
139 	enum z_erofs_collectmode mode;
140 };
141 
142 struct z_erofs_decompress_frontend {
143 	struct inode *const inode;
144 
145 	struct z_erofs_collector clt;
146 	struct erofs_map_blocks map;
147 
148 	/* used for applying cache strategy on the fly */
149 	bool backmost;
150 	erofs_off_t headoffset;
151 };
152 
153 #define COLLECTOR_INIT() { \
154 	.owned_head = Z_EROFS_PCLUSTER_TAIL, \
155 	.mode = COLLECT_PRIMARY_FOLLOWED }
156 
157 #define DECOMPRESS_FRONTEND_INIT(__i) { \
158 	.inode = __i, .clt = COLLECTOR_INIT(), \
159 	.backmost = true, }
160 
161 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
162 static DEFINE_MUTEX(z_pagemap_global_lock);
163 
preload_compressed_pages(struct z_erofs_collector * clt,struct address_space * mc,enum z_erofs_cache_alloctype type,struct list_head * pagepool)164 static void preload_compressed_pages(struct z_erofs_collector *clt,
165 				     struct address_space *mc,
166 				     enum z_erofs_cache_alloctype type,
167 				     struct list_head *pagepool)
168 {
169 	const struct z_erofs_pcluster *pcl = clt->pcl;
170 	const unsigned int clusterpages = BIT(pcl->clusterbits);
171 	struct page **pages = clt->compressedpages;
172 	pgoff_t index = pcl->obj.index + (pages - pcl->compressed_pages);
173 	bool standalone = true;
174 
175 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
176 		return;
177 
178 	for (; pages < pcl->compressed_pages + clusterpages; ++pages) {
179 		struct page *page;
180 		compressed_page_t t;
181 
182 		/* the compressed page was loaded before */
183 		if (READ_ONCE(*pages))
184 			continue;
185 
186 		page = find_get_page(mc, index);
187 
188 		if (page) {
189 			t = tag_compressed_page_justfound(page);
190 		} else if (type == DELAYEDALLOC) {
191 			t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED);
192 		} else {	/* DONTALLOC */
193 			if (standalone)
194 				clt->compressedpages = pages;
195 			standalone = false;
196 			continue;
197 		}
198 
199 		if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
200 			continue;
201 
202 		if (page)
203 			put_page(page);
204 	}
205 
206 	if (standalone)		/* downgrade to PRIMARY_FOLLOWED_NOINPLACE */
207 		clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
208 }
209 
210 /* called by erofs_shrinker to get rid of all compressed_pages */
erofs_try_to_free_all_cached_pages(struct erofs_sb_info * sbi,struct erofs_workgroup * grp)211 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
212 				       struct erofs_workgroup *grp)
213 {
214 	struct z_erofs_pcluster *const pcl =
215 		container_of(grp, struct z_erofs_pcluster, obj);
216 	struct address_space *const mapping = MNGD_MAPPING(sbi);
217 	const unsigned int clusterpages = BIT(pcl->clusterbits);
218 	int i;
219 
220 	/*
221 	 * refcount of workgroup is now freezed as 1,
222 	 * therefore no need to worry about available decompression users.
223 	 */
224 	for (i = 0; i < clusterpages; ++i) {
225 		struct page *page = pcl->compressed_pages[i];
226 
227 		if (!page)
228 			continue;
229 
230 		/* block other users from reclaiming or migrating the page */
231 		if (!trylock_page(page))
232 			return -EBUSY;
233 
234 		if (page->mapping != mapping)
235 			continue;
236 
237 		/* barrier is implied in the following 'unlock_page' */
238 		WRITE_ONCE(pcl->compressed_pages[i], NULL);
239 		set_page_private(page, 0);
240 		ClearPagePrivate(page);
241 
242 		unlock_page(page);
243 		put_page(page);
244 	}
245 	return 0;
246 }
247 
erofs_try_to_free_cached_page(struct address_space * mapping,struct page * page)248 int erofs_try_to_free_cached_page(struct address_space *mapping,
249 				  struct page *page)
250 {
251 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
252 	const unsigned int clusterpages = BIT(pcl->clusterbits);
253 	int ret = 0;	/* 0 - busy */
254 
255 	if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
256 		unsigned int i;
257 
258 		for (i = 0; i < clusterpages; ++i) {
259 			if (pcl->compressed_pages[i] == page) {
260 				WRITE_ONCE(pcl->compressed_pages[i], NULL);
261 				ret = 1;
262 				break;
263 			}
264 		}
265 		erofs_workgroup_unfreeze(&pcl->obj, 1);
266 
267 		if (ret) {
268 			ClearPagePrivate(page);
269 			put_page(page);
270 		}
271 	}
272 	return ret;
273 }
274 
275 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
z_erofs_try_inplace_io(struct z_erofs_collector * clt,struct page * page)276 static inline bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
277 					  struct page *page)
278 {
279 	struct z_erofs_pcluster *const pcl = clt->pcl;
280 	const unsigned int clusterpages = BIT(pcl->clusterbits);
281 
282 	while (clt->compressedpages < pcl->compressed_pages + clusterpages) {
283 		if (!cmpxchg(clt->compressedpages++, NULL, page))
284 			return true;
285 	}
286 	return false;
287 }
288 
289 /* callers must be with collection lock held */
z_erofs_attach_page(struct z_erofs_collector * clt,struct page * page,enum z_erofs_page_type type)290 static int z_erofs_attach_page(struct z_erofs_collector *clt,
291 			       struct page *page,
292 			       enum z_erofs_page_type type)
293 {
294 	int ret;
295 	bool occupied;
296 
297 	/* give priority for inplaceio */
298 	if (clt->mode >= COLLECT_PRIMARY &&
299 	    type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
300 	    z_erofs_try_inplace_io(clt, page))
301 		return 0;
302 
303 	ret = z_erofs_pagevec_enqueue(&clt->vector,
304 				      page, type, &occupied);
305 	clt->cl->vcnt += (unsigned int)ret;
306 
307 	return ret ? 0 : -EAGAIN;
308 }
309 
310 static enum z_erofs_collectmode
try_to_claim_pcluster(struct z_erofs_pcluster * pcl,z_erofs_next_pcluster_t * owned_head)311 try_to_claim_pcluster(struct z_erofs_pcluster *pcl,
312 		      z_erofs_next_pcluster_t *owned_head)
313 {
314 	/* let's claim these following types of pclusters */
315 retry:
316 	if (pcl->next == Z_EROFS_PCLUSTER_NIL) {
317 		/* type 1, nil pcluster */
318 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
319 			    *owned_head) != Z_EROFS_PCLUSTER_NIL)
320 			goto retry;
321 
322 		*owned_head = &pcl->next;
323 		/* lucky, I am the followee :) */
324 		return COLLECT_PRIMARY_FOLLOWED;
325 	} else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) {
326 		/*
327 		 * type 2, link to the end of a existing open chain,
328 		 * be careful that its submission itself is governed
329 		 * by the original owned chain.
330 		 */
331 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
332 			    *owned_head) != Z_EROFS_PCLUSTER_TAIL)
333 			goto retry;
334 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
335 		return COLLECT_PRIMARY_HOOKED;
336 	}
337 	return COLLECT_PRIMARY;	/* :( better luck next time */
338 }
339 
cllookup(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)340 static struct z_erofs_collection *cllookup(struct z_erofs_collector *clt,
341 					   struct inode *inode,
342 					   struct erofs_map_blocks *map)
343 {
344 	struct erofs_workgroup *grp;
345 	struct z_erofs_pcluster *pcl;
346 	struct z_erofs_collection *cl;
347 	unsigned int length;
348 	bool tag;
349 
350 	grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT, &tag);
351 	if (!grp)
352 		return NULL;
353 
354 	pcl = container_of(grp, struct z_erofs_pcluster, obj);
355 	if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
356 		DBG_BUGON(1);
357 		erofs_workgroup_put(grp);
358 		return ERR_PTR(-EFSCORRUPTED);
359 	}
360 
361 	cl = z_erofs_primarycollection(pcl);
362 	if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
363 		DBG_BUGON(1);
364 		erofs_workgroup_put(grp);
365 		return ERR_PTR(-EFSCORRUPTED);
366 	}
367 
368 	length = READ_ONCE(pcl->length);
369 	if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
370 		if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
371 			DBG_BUGON(1);
372 			erofs_workgroup_put(grp);
373 			return ERR_PTR(-EFSCORRUPTED);
374 		}
375 	} else {
376 		unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
377 
378 		if (map->m_flags & EROFS_MAP_FULL_MAPPED)
379 			llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
380 
381 		while (llen > length &&
382 		       length != cmpxchg_relaxed(&pcl->length, length, llen)) {
383 			cpu_relax();
384 			length = READ_ONCE(pcl->length);
385 		}
386 	}
387 	mutex_lock(&cl->lock);
388 	/* used to check tail merging loop due to corrupted images */
389 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
390 		clt->tailpcl = pcl;
391 	clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head);
392 	/* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */
393 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
394 		clt->tailpcl = NULL;
395 	clt->pcl = pcl;
396 	clt->cl = cl;
397 	return cl;
398 }
399 
clregister(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)400 static struct z_erofs_collection *clregister(struct z_erofs_collector *clt,
401 					     struct inode *inode,
402 					     struct erofs_map_blocks *map)
403 {
404 	struct z_erofs_pcluster *pcl;
405 	struct z_erofs_collection *cl;
406 	int err;
407 
408 	/* no available workgroup, let's allocate one */
409 	pcl = kmem_cache_alloc(pcluster_cachep, GFP_NOFS);
410 	if (!pcl)
411 		return ERR_PTR(-ENOMEM);
412 
413 	z_erofs_pcluster_init_always(pcl);
414 	pcl->obj.index = map->m_pa >> PAGE_SHIFT;
415 
416 	pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
417 		(map->m_flags & EROFS_MAP_FULL_MAPPED ?
418 			Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
419 
420 	if (map->m_flags & EROFS_MAP_ZIPPED)
421 		pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
422 	else
423 		pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
424 
425 	pcl->clusterbits = EROFS_I(inode)->z_physical_clusterbits[0];
426 	pcl->clusterbits -= PAGE_SHIFT;
427 
428 	/* new pclusters should be claimed as type 1, primary and followed */
429 	pcl->next = clt->owned_head;
430 	clt->mode = COLLECT_PRIMARY_FOLLOWED;
431 
432 	cl = z_erofs_primarycollection(pcl);
433 	cl->pageofs = map->m_la & ~PAGE_MASK;
434 
435 	/*
436 	 * lock all primary followed works before visible to others
437 	 * and mutex_trylock *never* fails for a new pcluster.
438 	 */
439 	mutex_trylock(&cl->lock);
440 
441 	err = erofs_register_workgroup(inode->i_sb, &pcl->obj, 0);
442 	if (err) {
443 		mutex_unlock(&cl->lock);
444 		kmem_cache_free(pcluster_cachep, pcl);
445 		return ERR_PTR(-EAGAIN);
446 	}
447 	/* used to check tail merging loop due to corrupted images */
448 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
449 		clt->tailpcl = pcl;
450 	clt->owned_head = &pcl->next;
451 	clt->pcl = pcl;
452 	clt->cl = cl;
453 	return cl;
454 }
455 
z_erofs_collector_begin(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)456 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
457 				   struct inode *inode,
458 				   struct erofs_map_blocks *map)
459 {
460 	struct z_erofs_collection *cl;
461 
462 	DBG_BUGON(clt->cl);
463 
464 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
465 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
466 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
467 
468 	if (!PAGE_ALIGNED(map->m_pa)) {
469 		DBG_BUGON(1);
470 		return -EINVAL;
471 	}
472 
473 repeat:
474 	cl = cllookup(clt, inode, map);
475 	if (!cl) {
476 		cl = clregister(clt, inode, map);
477 
478 		if (cl == ERR_PTR(-EAGAIN))
479 			goto repeat;
480 	}
481 
482 	if (IS_ERR(cl))
483 		return PTR_ERR(cl);
484 
485 	z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
486 				  cl->pagevec, cl->vcnt);
487 
488 	clt->compressedpages = clt->pcl->compressed_pages;
489 	if (clt->mode <= COLLECT_PRIMARY) /* cannot do in-place I/O */
490 		clt->compressedpages += Z_EROFS_CLUSTER_MAX_PAGES;
491 	return 0;
492 }
493 
494 /*
495  * keep in mind that no referenced pclusters will be freed
496  * only after a RCU grace period.
497  */
z_erofs_rcu_callback(struct rcu_head * head)498 static void z_erofs_rcu_callback(struct rcu_head *head)
499 {
500 	struct z_erofs_collection *const cl =
501 		container_of(head, struct z_erofs_collection, rcu);
502 
503 	kmem_cache_free(pcluster_cachep,
504 			container_of(cl, struct z_erofs_pcluster,
505 				     primary_collection));
506 }
507 
erofs_workgroup_free_rcu(struct erofs_workgroup * grp)508 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
509 {
510 	struct z_erofs_pcluster *const pcl =
511 		container_of(grp, struct z_erofs_pcluster, obj);
512 	struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
513 
514 	call_rcu(&cl->rcu, z_erofs_rcu_callback);
515 }
516 
z_erofs_collection_put(struct z_erofs_collection * cl)517 static void z_erofs_collection_put(struct z_erofs_collection *cl)
518 {
519 	struct z_erofs_pcluster *const pcl =
520 		container_of(cl, struct z_erofs_pcluster, primary_collection);
521 
522 	erofs_workgroup_put(&pcl->obj);
523 }
524 
z_erofs_collector_end(struct z_erofs_collector * clt)525 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
526 {
527 	struct z_erofs_collection *cl = clt->cl;
528 
529 	if (!cl)
530 		return false;
531 
532 	z_erofs_pagevec_ctor_exit(&clt->vector, false);
533 	mutex_unlock(&cl->lock);
534 
535 	/*
536 	 * if all pending pages are added, don't hold its reference
537 	 * any longer if the pcluster isn't hosted by ourselves.
538 	 */
539 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
540 		z_erofs_collection_put(cl);
541 
542 	clt->cl = NULL;
543 	return true;
544 }
545 
__stagingpage_alloc(struct list_head * pagepool,gfp_t gfp)546 static inline struct page *__stagingpage_alloc(struct list_head *pagepool,
547 					       gfp_t gfp)
548 {
549 	struct page *page = erofs_allocpage(pagepool, gfp, true);
550 
551 	page->mapping = Z_EROFS_MAPPING_STAGING;
552 	return page;
553 }
554 
should_alloc_managed_pages(struct z_erofs_decompress_frontend * fe,unsigned int cachestrategy,erofs_off_t la)555 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
556 				       unsigned int cachestrategy,
557 				       erofs_off_t la)
558 {
559 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
560 		return false;
561 
562 	if (fe->backmost)
563 		return true;
564 
565 	return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
566 		la < fe->headoffset;
567 }
568 
z_erofs_do_read_page(struct z_erofs_decompress_frontend * fe,struct page * page,struct list_head * pagepool)569 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
570 				struct page *page,
571 				struct list_head *pagepool)
572 {
573 	struct inode *const inode = fe->inode;
574 	struct erofs_sb_info *const sbi __maybe_unused = EROFS_I_SB(inode);
575 	struct erofs_map_blocks *const map = &fe->map;
576 	struct z_erofs_collector *const clt = &fe->clt;
577 	const loff_t offset = page_offset(page);
578 	bool tight = true;
579 
580 	enum z_erofs_cache_alloctype cache_strategy;
581 	enum z_erofs_page_type page_type;
582 	unsigned int cur, end, spiltted, index;
583 	int err = 0;
584 
585 	/* register locked file pages as online pages in pack */
586 	z_erofs_onlinepage_init(page);
587 
588 	spiltted = 0;
589 	end = PAGE_SIZE;
590 repeat:
591 	cur = end - 1;
592 
593 	/* lucky, within the range of the current map_blocks */
594 	if (offset + cur >= map->m_la &&
595 	    offset + cur < map->m_la + map->m_llen) {
596 		/* didn't get a valid collection previously (very rare) */
597 		if (!clt->cl)
598 			goto restart_now;
599 		goto hitted;
600 	}
601 
602 	/* go ahead the next map_blocks */
603 	erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
604 
605 	if (z_erofs_collector_end(clt))
606 		fe->backmost = false;
607 
608 	map->m_la = offset + cur;
609 	map->m_llen = 0;
610 	err = z_erofs_map_blocks_iter(inode, map, 0);
611 	if (err)
612 		goto err_out;
613 
614 restart_now:
615 	if (!(map->m_flags & EROFS_MAP_MAPPED))
616 		goto hitted;
617 
618 	err = z_erofs_collector_begin(clt, inode, map);
619 	if (err)
620 		goto err_out;
621 
622 	/* preload all compressed pages (maybe downgrade role if necessary) */
623 	if (should_alloc_managed_pages(fe, sbi->cache_strategy, map->m_la))
624 		cache_strategy = DELAYEDALLOC;
625 	else
626 		cache_strategy = DONTALLOC;
627 
628 	preload_compressed_pages(clt, MNGD_MAPPING(sbi),
629 				 cache_strategy, pagepool);
630 
631 hitted:
632 	/*
633 	 * Ensure the current partial page belongs to this submit chain rather
634 	 * than other concurrent submit chains or the noio(bypass) chain since
635 	 * those chains are handled asynchronously thus the page cannot be used
636 	 * for inplace I/O or pagevec (should be processed in strict order.)
637 	 */
638 	tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
639 		  clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
640 
641 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
642 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
643 		zero_user_segment(page, cur, end);
644 		goto next_part;
645 	}
646 
647 	/* let's derive page type */
648 	page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
649 		(!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
650 			(tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
651 				Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
652 
653 	if (cur)
654 		tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
655 
656 retry:
657 	err = z_erofs_attach_page(clt, page, page_type);
658 	/* should allocate an additional staging page for pagevec */
659 	if (err == -EAGAIN) {
660 		struct page *const newpage =
661 			__stagingpage_alloc(pagepool, GFP_NOFS);
662 
663 		err = z_erofs_attach_page(clt, newpage,
664 					  Z_EROFS_PAGE_TYPE_EXCLUSIVE);
665 		if (!err)
666 			goto retry;
667 	}
668 
669 	if (err)
670 		goto err_out;
671 
672 	index = page->index - (map->m_la >> PAGE_SHIFT);
673 
674 	z_erofs_onlinepage_fixup(page, index, true);
675 
676 	/* bump up the number of spiltted parts of a page */
677 	++spiltted;
678 	/* also update nr_pages */
679 	clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
680 next_part:
681 	/* can be used for verification */
682 	map->m_llen = offset + cur - map->m_la;
683 
684 	end = cur;
685 	if (end > 0)
686 		goto repeat;
687 
688 out:
689 	z_erofs_onlinepage_endio(page);
690 
691 	erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
692 		  __func__, page, spiltted, map->m_llen);
693 	return err;
694 
695 	/* if some error occurred while processing this page */
696 err_out:
697 	SetPageError(page);
698 	goto out;
699 }
700 
z_erofs_vle_unzip_kickoff(void * ptr,int bios)701 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios)
702 {
703 	tagptr1_t t = tagptr_init(tagptr1_t, ptr);
704 	struct z_erofs_unzip_io *io = tagptr_unfold_ptr(t);
705 	bool background = tagptr_unfold_tags(t);
706 
707 	if (!background) {
708 		unsigned long flags;
709 
710 		spin_lock_irqsave(&io->u.wait.lock, flags);
711 		if (!atomic_add_return(bios, &io->pending_bios))
712 			wake_up_locked(&io->u.wait);
713 		spin_unlock_irqrestore(&io->u.wait.lock, flags);
714 		return;
715 	}
716 
717 	if (!atomic_add_return(bios, &io->pending_bios))
718 		queue_work(z_erofs_workqueue, &io->u.work);
719 }
720 
z_erofs_vle_read_endio(struct bio * bio)721 static inline void z_erofs_vle_read_endio(struct bio *bio)
722 {
723 	struct erofs_sb_info *sbi = NULL;
724 	blk_status_t err = bio->bi_status;
725 	struct bio_vec *bvec;
726 	struct bvec_iter_all iter_all;
727 
728 	bio_for_each_segment_all(bvec, bio, iter_all) {
729 		struct page *page = bvec->bv_page;
730 		bool cachemngd = false;
731 
732 		DBG_BUGON(PageUptodate(page));
733 		DBG_BUGON(!page->mapping);
734 
735 		if (!sbi && !z_erofs_page_is_staging(page))
736 			sbi = EROFS_SB(page->mapping->host->i_sb);
737 
738 		/* sbi should already be gotten if the page is managed */
739 		if (sbi)
740 			cachemngd = erofs_page_is_managed(sbi, page);
741 
742 		if (err)
743 			SetPageError(page);
744 		else if (cachemngd)
745 			SetPageUptodate(page);
746 
747 		if (cachemngd)
748 			unlock_page(page);
749 	}
750 
751 	z_erofs_vle_unzip_kickoff(bio->bi_private, -1);
752 	bio_put(bio);
753 }
754 
z_erofs_decompress_pcluster(struct super_block * sb,struct z_erofs_pcluster * pcl,struct list_head * pagepool)755 static int z_erofs_decompress_pcluster(struct super_block *sb,
756 				       struct z_erofs_pcluster *pcl,
757 				       struct list_head *pagepool)
758 {
759 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
760 	const unsigned int clusterpages = BIT(pcl->clusterbits);
761 	struct z_erofs_pagevec_ctor ctor;
762 	unsigned int i, outputsize, llen, nr_pages;
763 	struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
764 	struct page **pages, **compressed_pages, *page;
765 
766 	enum z_erofs_page_type page_type;
767 	bool overlapped, partial;
768 	struct z_erofs_collection *cl;
769 	int err;
770 
771 	might_sleep();
772 	cl = z_erofs_primarycollection(pcl);
773 	DBG_BUGON(!READ_ONCE(cl->nr_pages));
774 
775 	mutex_lock(&cl->lock);
776 	nr_pages = cl->nr_pages;
777 
778 	if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
779 		pages = pages_onstack;
780 	} else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
781 		   mutex_trylock(&z_pagemap_global_lock)) {
782 		pages = z_pagemap_global;
783 	} else {
784 		gfp_t gfp_flags = GFP_KERNEL;
785 
786 		if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
787 			gfp_flags |= __GFP_NOFAIL;
788 
789 		pages = kvmalloc_array(nr_pages, sizeof(struct page *),
790 				       gfp_flags);
791 
792 		/* fallback to global pagemap for the lowmem scenario */
793 		if (!pages) {
794 			mutex_lock(&z_pagemap_global_lock);
795 			pages = z_pagemap_global;
796 		}
797 	}
798 
799 	for (i = 0; i < nr_pages; ++i)
800 		pages[i] = NULL;
801 
802 	err = 0;
803 	z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
804 				  cl->pagevec, 0);
805 
806 	for (i = 0; i < cl->vcnt; ++i) {
807 		unsigned int pagenr;
808 
809 		page = z_erofs_pagevec_dequeue(&ctor, &page_type);
810 
811 		/* all pages in pagevec ought to be valid */
812 		DBG_BUGON(!page);
813 		DBG_BUGON(!page->mapping);
814 
815 		if (z_erofs_put_stagingpage(pagepool, page))
816 			continue;
817 
818 		if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
819 			pagenr = 0;
820 		else
821 			pagenr = z_erofs_onlinepage_index(page);
822 
823 		DBG_BUGON(pagenr >= nr_pages);
824 
825 		/*
826 		 * currently EROFS doesn't support multiref(dedup),
827 		 * so here erroring out one multiref page.
828 		 */
829 		if (pages[pagenr]) {
830 			DBG_BUGON(1);
831 			SetPageError(pages[pagenr]);
832 			z_erofs_onlinepage_endio(pages[pagenr]);
833 			err = -EFSCORRUPTED;
834 		}
835 		pages[pagenr] = page;
836 	}
837 	z_erofs_pagevec_ctor_exit(&ctor, true);
838 
839 	overlapped = false;
840 	compressed_pages = pcl->compressed_pages;
841 
842 	for (i = 0; i < clusterpages; ++i) {
843 		unsigned int pagenr;
844 
845 		page = compressed_pages[i];
846 
847 		/* all compressed pages ought to be valid */
848 		DBG_BUGON(!page);
849 		DBG_BUGON(!page->mapping);
850 
851 		if (!z_erofs_page_is_staging(page)) {
852 			if (erofs_page_is_managed(sbi, page)) {
853 				if (!PageUptodate(page))
854 					err = -EIO;
855 				continue;
856 			}
857 
858 			/*
859 			 * only if non-head page can be selected
860 			 * for inplace decompression
861 			 */
862 			pagenr = z_erofs_onlinepage_index(page);
863 
864 			DBG_BUGON(pagenr >= nr_pages);
865 			if (pages[pagenr]) {
866 				DBG_BUGON(1);
867 				SetPageError(pages[pagenr]);
868 				z_erofs_onlinepage_endio(pages[pagenr]);
869 				err = -EFSCORRUPTED;
870 			}
871 			pages[pagenr] = page;
872 
873 			overlapped = true;
874 		}
875 
876 		/* PG_error needs checking for inplaced and staging pages */
877 		if (PageError(page)) {
878 			DBG_BUGON(PageUptodate(page));
879 			err = -EIO;
880 		}
881 	}
882 
883 	if (err)
884 		goto out;
885 
886 	llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
887 	if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
888 		outputsize = llen;
889 		partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
890 	} else {
891 		outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
892 		partial = true;
893 	}
894 
895 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
896 					.sb = sb,
897 					.in = compressed_pages,
898 					.out = pages,
899 					.pageofs_out = cl->pageofs,
900 					.inputsize = PAGE_SIZE,
901 					.outputsize = outputsize,
902 					.alg = pcl->algorithmformat,
903 					.inplace_io = overlapped,
904 					.partial_decoding = partial
905 				 }, pagepool);
906 
907 out:
908 	/* must handle all compressed pages before endding pages */
909 	for (i = 0; i < clusterpages; ++i) {
910 		page = compressed_pages[i];
911 
912 		if (erofs_page_is_managed(sbi, page))
913 			continue;
914 
915 		/* recycle all individual staging pages */
916 		(void)z_erofs_put_stagingpage(pagepool, page);
917 
918 		WRITE_ONCE(compressed_pages[i], NULL);
919 	}
920 
921 	for (i = 0; i < nr_pages; ++i) {
922 		page = pages[i];
923 		if (!page)
924 			continue;
925 
926 		DBG_BUGON(!page->mapping);
927 
928 		/* recycle all individual staging pages */
929 		if (z_erofs_put_stagingpage(pagepool, page))
930 			continue;
931 
932 		if (err < 0)
933 			SetPageError(page);
934 
935 		z_erofs_onlinepage_endio(page);
936 	}
937 
938 	if (pages == z_pagemap_global)
939 		mutex_unlock(&z_pagemap_global_lock);
940 	else if (pages != pages_onstack)
941 		kvfree(pages);
942 
943 	cl->nr_pages = 0;
944 	cl->vcnt = 0;
945 
946 	/* all cl locks MUST be taken before the following line */
947 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
948 
949 	/* all cl locks SHOULD be released right now */
950 	mutex_unlock(&cl->lock);
951 
952 	z_erofs_collection_put(cl);
953 	return err;
954 }
955 
z_erofs_vle_unzip_all(struct super_block * sb,struct z_erofs_unzip_io * io,struct list_head * pagepool)956 static void z_erofs_vle_unzip_all(struct super_block *sb,
957 				  struct z_erofs_unzip_io *io,
958 				  struct list_head *pagepool)
959 {
960 	z_erofs_next_pcluster_t owned = io->head;
961 
962 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
963 		struct z_erofs_pcluster *pcl;
964 
965 		/* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
966 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
967 
968 		/* no possible that 'owned' equals NULL */
969 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
970 
971 		pcl = container_of(owned, struct z_erofs_pcluster, next);
972 		owned = READ_ONCE(pcl->next);
973 
974 		z_erofs_decompress_pcluster(sb, pcl, pagepool);
975 	}
976 }
977 
z_erofs_vle_unzip_wq(struct work_struct * work)978 static void z_erofs_vle_unzip_wq(struct work_struct *work)
979 {
980 	struct z_erofs_unzip_io_sb *iosb =
981 		container_of(work, struct z_erofs_unzip_io_sb, io.u.work);
982 	LIST_HEAD(pagepool);
983 
984 	DBG_BUGON(iosb->io.head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
985 	z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &pagepool);
986 
987 	put_pages_list(&pagepool);
988 	kvfree(iosb);
989 }
990 
pickup_page_for_submission(struct z_erofs_pcluster * pcl,unsigned int nr,struct list_head * pagepool,struct address_space * mc,gfp_t gfp)991 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
992 					       unsigned int nr,
993 					       struct list_head *pagepool,
994 					       struct address_space *mc,
995 					       gfp_t gfp)
996 {
997 	/* determined at compile time to avoid too many #ifdefs */
998 	const bool nocache = __builtin_constant_p(mc) ? !mc : false;
999 	const pgoff_t index = pcl->obj.index;
1000 	bool tocache = false;
1001 
1002 	struct address_space *mapping;
1003 	struct page *oldpage, *page;
1004 
1005 	compressed_page_t t;
1006 	int justfound;
1007 
1008 repeat:
1009 	page = READ_ONCE(pcl->compressed_pages[nr]);
1010 	oldpage = page;
1011 
1012 	if (!page)
1013 		goto out_allocpage;
1014 
1015 	/*
1016 	 * the cached page has not been allocated and
1017 	 * an placeholder is out there, prepare it now.
1018 	 */
1019 	if (!nocache && page == PAGE_UNALLOCATED) {
1020 		tocache = true;
1021 		goto out_allocpage;
1022 	}
1023 
1024 	/* process the target tagged pointer */
1025 	t = tagptr_init(compressed_page_t, page);
1026 	justfound = tagptr_unfold_tags(t);
1027 	page = tagptr_unfold_ptr(t);
1028 
1029 	mapping = READ_ONCE(page->mapping);
1030 
1031 	/*
1032 	 * if managed cache is disabled, it's no way to
1033 	 * get such a cached-like page.
1034 	 */
1035 	if (nocache) {
1036 		/* if managed cache is disabled, it is impossible `justfound' */
1037 		DBG_BUGON(justfound);
1038 
1039 		/* and it should be locked, not uptodate, and not truncated */
1040 		DBG_BUGON(!PageLocked(page));
1041 		DBG_BUGON(PageUptodate(page));
1042 		DBG_BUGON(!mapping);
1043 		goto out;
1044 	}
1045 
1046 	/*
1047 	 * unmanaged (file) pages are all locked solidly,
1048 	 * therefore it is impossible for `mapping' to be NULL.
1049 	 */
1050 	if (mapping && mapping != mc)
1051 		/* ought to be unmanaged pages */
1052 		goto out;
1053 
1054 	lock_page(page);
1055 
1056 	/* only true if page reclaim goes wrong, should never happen */
1057 	DBG_BUGON(justfound && PagePrivate(page));
1058 
1059 	/* the page is still in manage cache */
1060 	if (page->mapping == mc) {
1061 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1062 
1063 		ClearPageError(page);
1064 		if (!PagePrivate(page)) {
1065 			/*
1066 			 * impossible to be !PagePrivate(page) for
1067 			 * the current restriction as well if
1068 			 * the page is already in compressed_pages[].
1069 			 */
1070 			DBG_BUGON(!justfound);
1071 
1072 			justfound = 0;
1073 			set_page_private(page, (unsigned long)pcl);
1074 			SetPagePrivate(page);
1075 		}
1076 
1077 		/* no need to submit io if it is already up-to-date */
1078 		if (PageUptodate(page)) {
1079 			unlock_page(page);
1080 			page = NULL;
1081 		}
1082 		goto out;
1083 	}
1084 
1085 	/*
1086 	 * the managed page has been truncated, it's unsafe to
1087 	 * reuse this one, let's allocate a new cache-managed page.
1088 	 */
1089 	DBG_BUGON(page->mapping);
1090 	DBG_BUGON(!justfound);
1091 
1092 	tocache = true;
1093 	unlock_page(page);
1094 	put_page(page);
1095 out_allocpage:
1096 	page = __stagingpage_alloc(pagepool, gfp);
1097 	if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1098 		list_add(&page->lru, pagepool);
1099 		cpu_relax();
1100 		goto repeat;
1101 	}
1102 	if (nocache || !tocache)
1103 		goto out;
1104 	if (add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1105 		page->mapping = Z_EROFS_MAPPING_STAGING;
1106 		goto out;
1107 	}
1108 
1109 	set_page_private(page, (unsigned long)pcl);
1110 	SetPagePrivate(page);
1111 out:	/* the only exit (for tracing and debugging) */
1112 	return page;
1113 }
1114 
jobqueue_init(struct super_block * sb,struct z_erofs_unzip_io * io,bool foreground)1115 static struct z_erofs_unzip_io *jobqueue_init(struct super_block *sb,
1116 					      struct z_erofs_unzip_io *io,
1117 					      bool foreground)
1118 {
1119 	struct z_erofs_unzip_io_sb *iosb;
1120 
1121 	if (foreground) {
1122 		/* waitqueue available for foreground io */
1123 		DBG_BUGON(!io);
1124 
1125 		init_waitqueue_head(&io->u.wait);
1126 		atomic_set(&io->pending_bios, 0);
1127 		goto out;
1128 	}
1129 
1130 	iosb = kvzalloc(sizeof(*iosb), GFP_KERNEL | __GFP_NOFAIL);
1131 	DBG_BUGON(!iosb);
1132 
1133 	/* initialize fields in the allocated descriptor */
1134 	io = &iosb->io;
1135 	iosb->sb = sb;
1136 	INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq);
1137 out:
1138 	io->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1139 	return io;
1140 }
1141 
1142 /* define decompression jobqueue types */
1143 enum {
1144 	JQ_BYPASS,
1145 	JQ_SUBMIT,
1146 	NR_JOBQUEUES,
1147 };
1148 
jobqueueset_init(struct super_block * sb,z_erofs_next_pcluster_t qtail[],struct z_erofs_unzip_io * q[],struct z_erofs_unzip_io * fgq,bool forcefg)1149 static void *jobqueueset_init(struct super_block *sb,
1150 			      z_erofs_next_pcluster_t qtail[],
1151 			      struct z_erofs_unzip_io *q[],
1152 			      struct z_erofs_unzip_io *fgq,
1153 			      bool forcefg)
1154 {
1155 	/*
1156 	 * if managed cache is enabled, bypass jobqueue is needed,
1157 	 * no need to read from device for all pclusters in this queue.
1158 	 */
1159 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, true);
1160 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1161 
1162 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, forcefg);
1163 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1164 
1165 	return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], !forcefg));
1166 }
1167 
move_to_bypass_jobqueue(struct z_erofs_pcluster * pcl,z_erofs_next_pcluster_t qtail[],z_erofs_next_pcluster_t owned_head)1168 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1169 				    z_erofs_next_pcluster_t qtail[],
1170 				    z_erofs_next_pcluster_t owned_head)
1171 {
1172 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1173 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1174 
1175 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1176 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1177 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1178 
1179 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1180 
1181 	WRITE_ONCE(*submit_qtail, owned_head);
1182 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1183 
1184 	qtail[JQ_BYPASS] = &pcl->next;
1185 }
1186 
postsubmit_is_all_bypassed(struct z_erofs_unzip_io * q[],unsigned int nr_bios,bool force_fg)1187 static bool postsubmit_is_all_bypassed(struct z_erofs_unzip_io *q[],
1188 				       unsigned int nr_bios,
1189 				       bool force_fg)
1190 {
1191 	/*
1192 	 * although background is preferred, no one is pending for submission.
1193 	 * don't issue workqueue for decompression but drop it directly instead.
1194 	 */
1195 	if (force_fg || nr_bios)
1196 		return false;
1197 
1198 	kvfree(container_of(q[JQ_SUBMIT], struct z_erofs_unzip_io_sb, io));
1199 	return true;
1200 }
1201 
z_erofs_vle_submit_all(struct super_block * sb,z_erofs_next_pcluster_t owned_head,struct list_head * pagepool,struct z_erofs_unzip_io * fgq,bool force_fg)1202 static bool z_erofs_vle_submit_all(struct super_block *sb,
1203 				   z_erofs_next_pcluster_t owned_head,
1204 				   struct list_head *pagepool,
1205 				   struct z_erofs_unzip_io *fgq,
1206 				   bool force_fg)
1207 {
1208 	struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb);
1209 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1210 	struct z_erofs_unzip_io *q[NR_JOBQUEUES];
1211 	struct bio *bio;
1212 	void *bi_private;
1213 	/* since bio will be NULL, no need to initialize last_index */
1214 	pgoff_t uninitialized_var(last_index);
1215 	bool force_submit = false;
1216 	unsigned int nr_bios;
1217 
1218 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1219 		return false;
1220 
1221 	force_submit = false;
1222 	bio = NULL;
1223 	nr_bios = 0;
1224 	bi_private = jobqueueset_init(sb, qtail, q, fgq, force_fg);
1225 
1226 	/* by default, all need io submission */
1227 	q[JQ_SUBMIT]->head = owned_head;
1228 
1229 	do {
1230 		struct z_erofs_pcluster *pcl;
1231 		unsigned int clusterpages;
1232 		pgoff_t first_index;
1233 		struct page *page;
1234 		unsigned int i = 0, bypass = 0;
1235 		int err;
1236 
1237 		/* no possible 'owned_head' equals the following */
1238 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1239 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1240 
1241 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1242 
1243 		clusterpages = BIT(pcl->clusterbits);
1244 
1245 		/* close the main owned chain at first */
1246 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1247 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1248 
1249 		first_index = pcl->obj.index;
1250 		force_submit |= (first_index != last_index + 1);
1251 
1252 repeat:
1253 		page = pickup_page_for_submission(pcl, i, pagepool,
1254 						  MNGD_MAPPING(sbi),
1255 						  GFP_NOFS);
1256 		if (!page) {
1257 			force_submit = true;
1258 			++bypass;
1259 			goto skippage;
1260 		}
1261 
1262 		if (bio && force_submit) {
1263 submit_bio_retry:
1264 			submit_bio(bio);
1265 			bio = NULL;
1266 		}
1267 
1268 		if (!bio) {
1269 			bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1270 
1271 			bio->bi_end_io = z_erofs_vle_read_endio;
1272 			bio_set_dev(bio, sb->s_bdev);
1273 			bio->bi_iter.bi_sector = (sector_t)(first_index + i) <<
1274 				LOG_SECTORS_PER_BLOCK;
1275 			bio->bi_private = bi_private;
1276 			bio->bi_opf = REQ_OP_READ;
1277 
1278 			++nr_bios;
1279 		}
1280 
1281 		err = bio_add_page(bio, page, PAGE_SIZE, 0);
1282 		if (err < PAGE_SIZE)
1283 			goto submit_bio_retry;
1284 
1285 		force_submit = false;
1286 		last_index = first_index + i;
1287 skippage:
1288 		if (++i < clusterpages)
1289 			goto repeat;
1290 
1291 		if (bypass < clusterpages)
1292 			qtail[JQ_SUBMIT] = &pcl->next;
1293 		else
1294 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1295 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1296 
1297 	if (bio)
1298 		submit_bio(bio);
1299 
1300 	if (postsubmit_is_all_bypassed(q, nr_bios, force_fg))
1301 		return true;
1302 
1303 	z_erofs_vle_unzip_kickoff(bi_private, nr_bios);
1304 	return true;
1305 }
1306 
z_erofs_submit_and_unzip(struct super_block * sb,struct z_erofs_collector * clt,struct list_head * pagepool,bool force_fg)1307 static void z_erofs_submit_and_unzip(struct super_block *sb,
1308 				     struct z_erofs_collector *clt,
1309 				     struct list_head *pagepool,
1310 				     bool force_fg)
1311 {
1312 	struct z_erofs_unzip_io io[NR_JOBQUEUES];
1313 
1314 	if (!z_erofs_vle_submit_all(sb, clt->owned_head,
1315 				    pagepool, io, force_fg))
1316 		return;
1317 
1318 	/* decompress no I/O pclusters immediately */
1319 	z_erofs_vle_unzip_all(sb, &io[JQ_BYPASS], pagepool);
1320 
1321 	if (!force_fg)
1322 		return;
1323 
1324 	/* wait until all bios are completed */
1325 	wait_event(io[JQ_SUBMIT].u.wait,
1326 		   !atomic_read(&io[JQ_SUBMIT].pending_bios));
1327 
1328 	/* let's synchronous decompression */
1329 	z_erofs_vle_unzip_all(sb, &io[JQ_SUBMIT], pagepool);
1330 }
1331 
z_erofs_vle_normalaccess_readpage(struct file * file,struct page * page)1332 static int z_erofs_vle_normalaccess_readpage(struct file *file,
1333 					     struct page *page)
1334 {
1335 	struct inode *const inode = page->mapping->host;
1336 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1337 	int err;
1338 	LIST_HEAD(pagepool);
1339 
1340 	trace_erofs_readpage(page, false);
1341 
1342 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1343 
1344 	err = z_erofs_do_read_page(&f, page, &pagepool);
1345 	(void)z_erofs_collector_end(&f.clt);
1346 
1347 	/* if some compressed cluster ready, need submit them anyway */
1348 	z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, true);
1349 
1350 	if (err)
1351 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1352 
1353 	if (f.map.mpage)
1354 		put_page(f.map.mpage);
1355 
1356 	/* clean up the remaining free pages */
1357 	put_pages_list(&pagepool);
1358 	return err;
1359 }
1360 
should_decompress_synchronously(struct erofs_sb_info * sbi,unsigned int nr)1361 static bool should_decompress_synchronously(struct erofs_sb_info *sbi,
1362 					    unsigned int nr)
1363 {
1364 	return nr <= sbi->max_sync_decompress_pages;
1365 }
1366 
z_erofs_vle_normalaccess_readpages(struct file * filp,struct address_space * mapping,struct list_head * pages,unsigned int nr_pages)1367 static int z_erofs_vle_normalaccess_readpages(struct file *filp,
1368 					      struct address_space *mapping,
1369 					      struct list_head *pages,
1370 					      unsigned int nr_pages)
1371 {
1372 	struct inode *const inode = mapping->host;
1373 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1374 
1375 	bool sync = should_decompress_synchronously(sbi, nr_pages);
1376 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1377 	gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1378 	struct page *head = NULL;
1379 	LIST_HEAD(pagepool);
1380 
1381 	trace_erofs_readpages(mapping->host, lru_to_page(pages),
1382 			      nr_pages, false);
1383 
1384 	f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT;
1385 
1386 	for (; nr_pages; --nr_pages) {
1387 		struct page *page = lru_to_page(pages);
1388 
1389 		prefetchw(&page->flags);
1390 		list_del(&page->lru);
1391 
1392 		/*
1393 		 * A pure asynchronous readahead is indicated if
1394 		 * a PG_readahead marked page is hitted at first.
1395 		 * Let's also do asynchronous decompression for this case.
1396 		 */
1397 		sync &= !(PageReadahead(page) && !head);
1398 
1399 		if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1400 			list_add(&page->lru, &pagepool);
1401 			continue;
1402 		}
1403 
1404 		set_page_private(page, (unsigned long)head);
1405 		head = page;
1406 	}
1407 
1408 	while (head) {
1409 		struct page *page = head;
1410 		int err;
1411 
1412 		/* traversal in reverse order */
1413 		head = (void *)page_private(page);
1414 
1415 		err = z_erofs_do_read_page(&f, page, &pagepool);
1416 		if (err)
1417 			erofs_err(inode->i_sb,
1418 				  "readahead error at page %lu @ nid %llu",
1419 				  page->index, EROFS_I(inode)->nid);
1420 		put_page(page);
1421 	}
1422 
1423 	(void)z_erofs_collector_end(&f.clt);
1424 
1425 	z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, sync);
1426 
1427 	if (f.map.mpage)
1428 		put_page(f.map.mpage);
1429 
1430 	/* clean up the remaining free pages */
1431 	put_pages_list(&pagepool);
1432 	return 0;
1433 }
1434 
1435 const struct address_space_operations z_erofs_vle_normalaccess_aops = {
1436 	.readpage = z_erofs_vle_normalaccess_readpage,
1437 	.readpages = z_erofs_vle_normalaccess_readpages,
1438 };
1439 
1440