xref: /OK3568_Linux_fs/kernel/fs/erofs/zdata.c (revision 4882a59341e53eb6f0b4789bf948001014eff981)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2018 HUAWEI, Inc.
4  *             https://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  * since pclustersize is variable for big pcluster feature, introduce slab
15  * pools implementation for different pcluster sizes.
16  */
17 struct z_erofs_pcluster_slab {
18 	struct kmem_cache *slab;
19 	unsigned int maxpages;
20 	char name[48];
21 };
22 
23 #define _PCLP(n) { .maxpages = n }
24 
25 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
26 	_PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
27 	_PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
28 };
29 
z_erofs_destroy_pcluster_pool(void)30 static void z_erofs_destroy_pcluster_pool(void)
31 {
32 	int i;
33 
34 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
35 		if (!pcluster_pool[i].slab)
36 			continue;
37 		kmem_cache_destroy(pcluster_pool[i].slab);
38 		pcluster_pool[i].slab = NULL;
39 	}
40 }
41 
z_erofs_create_pcluster_pool(void)42 static int z_erofs_create_pcluster_pool(void)
43 {
44 	struct z_erofs_pcluster_slab *pcs;
45 	struct z_erofs_pcluster *a;
46 	unsigned int size;
47 
48 	for (pcs = pcluster_pool;
49 	     pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
50 		size = struct_size(a, compressed_pages, pcs->maxpages);
51 
52 		sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
53 		pcs->slab = kmem_cache_create(pcs->name, size, 0,
54 					      SLAB_RECLAIM_ACCOUNT, NULL);
55 		if (pcs->slab)
56 			continue;
57 
58 		z_erofs_destroy_pcluster_pool();
59 		return -ENOMEM;
60 	}
61 	return 0;
62 }
63 
z_erofs_alloc_pcluster(unsigned int nrpages)64 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
65 {
66 	int i;
67 
68 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
69 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
70 		struct z_erofs_pcluster *pcl;
71 
72 		if (nrpages > pcs->maxpages)
73 			continue;
74 
75 		pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
76 		if (!pcl)
77 			return ERR_PTR(-ENOMEM);
78 		pcl->pclusterpages = nrpages;
79 		return pcl;
80 	}
81 	return ERR_PTR(-EINVAL);
82 }
83 
z_erofs_free_pcluster(struct z_erofs_pcluster * pcl)84 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
85 {
86 	int i;
87 
88 	for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
89 		struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
90 
91 		if (pcl->pclusterpages > pcs->maxpages)
92 			continue;
93 
94 		kmem_cache_free(pcs->slab, pcl);
95 		return;
96 	}
97 	DBG_BUGON(1);
98 }
99 
100 /*
101  * a compressed_pages[] placeholder in order to avoid
102  * being filled with file pages for in-place decompression.
103  */
104 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
105 
106 /* how to allocate cached pages for a pcluster */
107 enum z_erofs_cache_alloctype {
108 	DONTALLOC,	/* don't allocate any cached pages */
109 	DELAYEDALLOC,	/* delayed allocation (at the time of submitting io) */
110 	/*
111 	 * try to use cached I/O if page allocation succeeds or fallback
112 	 * to in-place I/O instead to avoid any direct reclaim.
113 	 */
114 	TRYALLOC,
115 };
116 
117 /*
118  * tagged pointer with 1-bit tag for all compressed pages
119  * tag 0 - the page is just found with an extra page reference
120  */
121 typedef tagptr1_t compressed_page_t;
122 
123 #define tag_compressed_page_justfound(page) \
124 	tagptr_fold(compressed_page_t, page, 1)
125 
126 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
127 
z_erofs_exit_zip_subsystem(void)128 void z_erofs_exit_zip_subsystem(void)
129 {
130 	destroy_workqueue(z_erofs_workqueue);
131 	z_erofs_destroy_pcluster_pool();
132 }
133 
z_erofs_init_workqueue(void)134 static inline int z_erofs_init_workqueue(void)
135 {
136 	const unsigned int onlinecpus = num_possible_cpus();
137 
138 	/*
139 	 * no need to spawn too many threads, limiting threads could minimum
140 	 * scheduling overhead, perhaps per-CPU threads should be better?
141 	 */
142 	z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
143 					    WQ_UNBOUND | WQ_HIGHPRI,
144 					    onlinecpus + onlinecpus / 4);
145 	return z_erofs_workqueue ? 0 : -ENOMEM;
146 }
147 
z_erofs_init_zip_subsystem(void)148 int __init z_erofs_init_zip_subsystem(void)
149 {
150 	int err = z_erofs_create_pcluster_pool();
151 
152 	if (err)
153 		return err;
154 	err = z_erofs_init_workqueue();
155 	if (err)
156 		z_erofs_destroy_pcluster_pool();
157 	return err;
158 }
159 
160 enum z_erofs_collectmode {
161 	COLLECT_SECONDARY,
162 	COLLECT_PRIMARY,
163 	/*
164 	 * The current collection was the tail of an exist chain, in addition
165 	 * that the previous processed chained collections are all decided to
166 	 * be hooked up to it.
167 	 * A new chain will be created for the remaining collections which are
168 	 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
169 	 * the next collection cannot reuse the whole page safely in
170 	 * the following scenario:
171 	 *  ________________________________________________________________
172 	 * |      tail (partial) page     |       head (partial) page       |
173 	 * |   (belongs to the next cl)   |   (belongs to the current cl)   |
174 	 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
175 	 */
176 	COLLECT_PRIMARY_HOOKED,
177 	/*
178 	 * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it
179 	 * could be dispatched into bypass queue later due to uptodated managed
180 	 * pages. All related online pages cannot be reused for inplace I/O (or
181 	 * pagevec) since it can be directly decoded without I/O submission.
182 	 */
183 	COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
184 	/*
185 	 * The current collection has been linked with the owned chain, and
186 	 * could also be linked with the remaining collections, which means
187 	 * if the processing page is the tail page of the collection, thus
188 	 * the current collection can safely use the whole page (since
189 	 * the previous collection is under control) for in-place I/O, as
190 	 * illustrated below:
191 	 *  ________________________________________________________________
192 	 * |  tail (partial) page |          head (partial) page           |
193 	 * |  (of the current cl) |      (of the previous collection)      |
194 	 * |  PRIMARY_FOLLOWED or |                                        |
195 	 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
196 	 *
197 	 * [  (*) the above page can be used as inplace I/O.               ]
198 	 */
199 	COLLECT_PRIMARY_FOLLOWED,
200 };
201 
202 struct z_erofs_collector {
203 	struct z_erofs_pagevec_ctor vector;
204 
205 	struct z_erofs_pcluster *pcl, *tailpcl;
206 	struct z_erofs_collection *cl;
207 	/* a pointer used to pick up inplace I/O pages */
208 	struct page **icpage_ptr;
209 	z_erofs_next_pcluster_t owned_head;
210 
211 	enum z_erofs_collectmode mode;
212 };
213 
214 struct z_erofs_decompress_frontend {
215 	struct inode *const inode;
216 
217 	struct z_erofs_collector clt;
218 	struct erofs_map_blocks map;
219 
220 	bool readahead;
221 	/* used for applying cache strategy on the fly */
222 	bool backmost;
223 	erofs_off_t headoffset;
224 };
225 
226 #define COLLECTOR_INIT() { \
227 	.owned_head = Z_EROFS_PCLUSTER_TAIL, \
228 	.mode = COLLECT_PRIMARY_FOLLOWED }
229 
230 #define DECOMPRESS_FRONTEND_INIT(__i) { \
231 	.inode = __i, .clt = COLLECTOR_INIT(), \
232 	.backmost = true, }
233 
234 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
235 static DEFINE_MUTEX(z_pagemap_global_lock);
236 
preload_compressed_pages(struct z_erofs_collector * clt,struct address_space * mc,enum z_erofs_cache_alloctype type,struct list_head * pagepool)237 static void preload_compressed_pages(struct z_erofs_collector *clt,
238 				     struct address_space *mc,
239 				     enum z_erofs_cache_alloctype type,
240 				     struct list_head *pagepool)
241 {
242 	struct z_erofs_pcluster *pcl = clt->pcl;
243 	bool standalone = true;
244 	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
245 			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
246 	struct page **pages;
247 	pgoff_t index;
248 
249 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
250 		return;
251 
252 	pages = pcl->compressed_pages;
253 	index = pcl->obj.index;
254 	for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) {
255 		struct page *page;
256 		compressed_page_t t;
257 		struct page *newpage = NULL;
258 
259 		/* the compressed page was loaded before */
260 		if (READ_ONCE(*pages))
261 			continue;
262 
263 		page = find_get_page(mc, index);
264 
265 		if (page) {
266 			t = tag_compressed_page_justfound(page);
267 		} else {
268 			/* I/O is needed, no possible to decompress directly */
269 			standalone = false;
270 			switch (type) {
271 			case DELAYEDALLOC:
272 				t = tagptr_init(compressed_page_t,
273 						PAGE_UNALLOCATED);
274 				break;
275 			case TRYALLOC:
276 				newpage = erofs_allocpage(pagepool, gfp);
277 				if (!newpage)
278 					continue;
279 				set_page_private(newpage,
280 						 Z_EROFS_PREALLOCATED_PAGE);
281 				t = tag_compressed_page_justfound(newpage);
282 				break;
283 			default:        /* DONTALLOC */
284 				continue;
285 			}
286 		}
287 
288 		if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
289 			continue;
290 
291 		if (page) {
292 			put_page(page);
293 		} else if (newpage) {
294 			set_page_private(newpage, 0);
295 			list_add(&newpage->lru, pagepool);
296 		}
297 	}
298 
299 	/*
300 	 * don't do inplace I/O if all compressed pages are available in
301 	 * managed cache since it can be moved to the bypass queue instead.
302 	 */
303 	if (standalone)
304 		clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
305 }
306 
307 /* 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)308 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
309 				       struct erofs_workgroup *grp)
310 {
311 	struct z_erofs_pcluster *const pcl =
312 		container_of(grp, struct z_erofs_pcluster, obj);
313 	struct address_space *const mapping = MNGD_MAPPING(sbi);
314 	int i;
315 
316 	/*
317 	 * refcount of workgroup is now freezed as 1,
318 	 * therefore no need to worry about available decompression users.
319 	 */
320 	for (i = 0; i < pcl->pclusterpages; ++i) {
321 		struct page *page = pcl->compressed_pages[i];
322 
323 		if (!page)
324 			continue;
325 
326 		/* block other users from reclaiming or migrating the page */
327 		if (!trylock_page(page))
328 			return -EBUSY;
329 
330 		if (page->mapping != mapping)
331 			continue;
332 
333 		/* barrier is implied in the following 'unlock_page' */
334 		WRITE_ONCE(pcl->compressed_pages[i], NULL);
335 		detach_page_private(page);
336 		unlock_page(page);
337 	}
338 	return 0;
339 }
340 
erofs_try_to_free_cached_page(struct address_space * mapping,struct page * page)341 int erofs_try_to_free_cached_page(struct address_space *mapping,
342 				  struct page *page)
343 {
344 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
345 	int ret = 0;	/* 0 - busy */
346 
347 	if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
348 		unsigned int i;
349 
350 		for (i = 0; i < pcl->pclusterpages; ++i) {
351 			if (pcl->compressed_pages[i] == page) {
352 				WRITE_ONCE(pcl->compressed_pages[i], NULL);
353 				ret = 1;
354 				break;
355 			}
356 		}
357 		erofs_workgroup_unfreeze(&pcl->obj, 1);
358 
359 		if (ret)
360 			detach_page_private(page);
361 	}
362 	return ret;
363 }
364 
365 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
z_erofs_try_inplace_io(struct z_erofs_collector * clt,struct page * page)366 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
367 				   struct page *page)
368 {
369 	struct z_erofs_pcluster *const pcl = clt->pcl;
370 
371 	while (clt->icpage_ptr > pcl->compressed_pages)
372 		if (!cmpxchg(--clt->icpage_ptr, NULL, page))
373 			return true;
374 	return false;
375 }
376 
377 /* 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,bool pvec_safereuse)378 static int z_erofs_attach_page(struct z_erofs_collector *clt,
379 			       struct page *page, enum z_erofs_page_type type,
380 			       bool pvec_safereuse)
381 {
382 	int ret;
383 
384 	/* give priority for inplaceio */
385 	if (clt->mode >= COLLECT_PRIMARY &&
386 	    type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
387 	    z_erofs_try_inplace_io(clt, page))
388 		return 0;
389 
390 	ret = z_erofs_pagevec_enqueue(&clt->vector, page, type,
391 				      pvec_safereuse);
392 	clt->cl->vcnt += (unsigned int)ret;
393 	return ret ? 0 : -EAGAIN;
394 }
395 
396 static enum z_erofs_collectmode
try_to_claim_pcluster(struct z_erofs_pcluster * pcl,z_erofs_next_pcluster_t * owned_head)397 try_to_claim_pcluster(struct z_erofs_pcluster *pcl,
398 		      z_erofs_next_pcluster_t *owned_head)
399 {
400 	/* let's claim these following types of pclusters */
401 retry:
402 	if (pcl->next == Z_EROFS_PCLUSTER_NIL) {
403 		/* type 1, nil pcluster */
404 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
405 			    *owned_head) != Z_EROFS_PCLUSTER_NIL)
406 			goto retry;
407 
408 		*owned_head = &pcl->next;
409 		/* lucky, I am the followee :) */
410 		return COLLECT_PRIMARY_FOLLOWED;
411 	} else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) {
412 		/*
413 		 * type 2, link to the end of a existing open chain,
414 		 * be careful that its submission itself is governed
415 		 * by the original owned chain.
416 		 */
417 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
418 			    *owned_head) != Z_EROFS_PCLUSTER_TAIL)
419 			goto retry;
420 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
421 		return COLLECT_PRIMARY_HOOKED;
422 	}
423 	return COLLECT_PRIMARY;	/* :( better luck next time */
424 }
425 
z_erofs_lookup_collection(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)426 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
427 				     struct inode *inode,
428 				     struct erofs_map_blocks *map)
429 {
430 	struct z_erofs_pcluster *pcl = clt->pcl;
431 	struct z_erofs_collection *cl;
432 	unsigned int length;
433 
434 	/* to avoid unexpected loop formed by corrupted images */
435 	if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
436 		DBG_BUGON(1);
437 		return -EFSCORRUPTED;
438 	}
439 
440 	cl = z_erofs_primarycollection(pcl);
441 	if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
442 		DBG_BUGON(1);
443 		return -EFSCORRUPTED;
444 	}
445 
446 	length = READ_ONCE(pcl->length);
447 	if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
448 		if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
449 			DBG_BUGON(1);
450 			return -EFSCORRUPTED;
451 		}
452 	} else {
453 		unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
454 
455 		if (map->m_flags & EROFS_MAP_FULL_MAPPED)
456 			llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
457 
458 		while (llen > length &&
459 		       length != cmpxchg_relaxed(&pcl->length, length, llen)) {
460 			cpu_relax();
461 			length = READ_ONCE(pcl->length);
462 		}
463 	}
464 	mutex_lock(&cl->lock);
465 	/* used to check tail merging loop due to corrupted images */
466 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
467 		clt->tailpcl = pcl;
468 	clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head);
469 	/* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */
470 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
471 		clt->tailpcl = NULL;
472 	clt->cl = cl;
473 	return 0;
474 }
475 
z_erofs_register_collection(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)476 static int z_erofs_register_collection(struct z_erofs_collector *clt,
477 				       struct inode *inode,
478 				       struct erofs_map_blocks *map)
479 {
480 	struct z_erofs_pcluster *pcl;
481 	struct z_erofs_collection *cl;
482 	struct erofs_workgroup *grp;
483 	int err;
484 
485 	/* no available pcluster, let's allocate one */
486 	pcl = z_erofs_alloc_pcluster(map->m_plen >> PAGE_SHIFT);
487 	if (IS_ERR(pcl))
488 		return PTR_ERR(pcl);
489 
490 	atomic_set(&pcl->obj.refcount, 1);
491 	pcl->obj.index = map->m_pa >> PAGE_SHIFT;
492 
493 	pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
494 		(map->m_flags & EROFS_MAP_FULL_MAPPED ?
495 			Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
496 
497 	if (map->m_flags & EROFS_MAP_ZIPPED)
498 		pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
499 	else
500 		pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
501 
502 	/* new pclusters should be claimed as type 1, primary and followed */
503 	pcl->next = clt->owned_head;
504 	clt->mode = COLLECT_PRIMARY_FOLLOWED;
505 
506 	cl = z_erofs_primarycollection(pcl);
507 	cl->pageofs = map->m_la & ~PAGE_MASK;
508 
509 	/*
510 	 * lock all primary followed works before visible to others
511 	 * and mutex_trylock *never* fails for a new pcluster.
512 	 */
513 	mutex_init(&cl->lock);
514 	DBG_BUGON(!mutex_trylock(&cl->lock));
515 
516 	grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
517 	if (IS_ERR(grp)) {
518 		err = PTR_ERR(grp);
519 		goto err_out;
520 	}
521 
522 	if (grp != &pcl->obj) {
523 		clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
524 		err = -EEXIST;
525 		goto err_out;
526 	}
527 	/* used to check tail merging loop due to corrupted images */
528 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
529 		clt->tailpcl = pcl;
530 	clt->owned_head = &pcl->next;
531 	clt->pcl = pcl;
532 	clt->cl = cl;
533 	return 0;
534 
535 err_out:
536 	mutex_unlock(&cl->lock);
537 	z_erofs_free_pcluster(pcl);
538 	return err;
539 }
540 
z_erofs_collector_begin(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)541 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
542 				   struct inode *inode,
543 				   struct erofs_map_blocks *map)
544 {
545 	struct erofs_workgroup *grp;
546 	int ret;
547 
548 	DBG_BUGON(clt->cl);
549 
550 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
551 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
552 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
553 
554 	if (!PAGE_ALIGNED(map->m_pa)) {
555 		DBG_BUGON(1);
556 		return -EINVAL;
557 	}
558 
559 	grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
560 	if (grp) {
561 		clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
562 	} else {
563 		ret = z_erofs_register_collection(clt, inode, map);
564 
565 		if (!ret)
566 			goto out;
567 		if (ret != -EEXIST)
568 			return ret;
569 	}
570 
571 	ret = z_erofs_lookup_collection(clt, inode, map);
572 	if (ret) {
573 		erofs_workgroup_put(&clt->pcl->obj);
574 		return ret;
575 	}
576 
577 out:
578 	z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
579 				  clt->cl->pagevec, clt->cl->vcnt);
580 
581 	/* since file-backed online pages are traversed in reverse order */
582 	clt->icpage_ptr = clt->pcl->compressed_pages + clt->pcl->pclusterpages;
583 	return 0;
584 }
585 
586 /*
587  * keep in mind that no referenced pclusters will be freed
588  * only after a RCU grace period.
589  */
z_erofs_rcu_callback(struct rcu_head * head)590 static void z_erofs_rcu_callback(struct rcu_head *head)
591 {
592 	struct z_erofs_collection *const cl =
593 		container_of(head, struct z_erofs_collection, rcu);
594 
595 	z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster,
596 					   primary_collection));
597 }
598 
erofs_workgroup_free_rcu(struct erofs_workgroup * grp)599 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
600 {
601 	struct z_erofs_pcluster *const pcl =
602 		container_of(grp, struct z_erofs_pcluster, obj);
603 	struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
604 
605 	call_rcu(&cl->rcu, z_erofs_rcu_callback);
606 }
607 
z_erofs_collection_put(struct z_erofs_collection * cl)608 static void z_erofs_collection_put(struct z_erofs_collection *cl)
609 {
610 	struct z_erofs_pcluster *const pcl =
611 		container_of(cl, struct z_erofs_pcluster, primary_collection);
612 
613 	erofs_workgroup_put(&pcl->obj);
614 }
615 
z_erofs_collector_end(struct z_erofs_collector * clt)616 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
617 {
618 	struct z_erofs_collection *cl = clt->cl;
619 
620 	if (!cl)
621 		return false;
622 
623 	z_erofs_pagevec_ctor_exit(&clt->vector, false);
624 	mutex_unlock(&cl->lock);
625 
626 	/*
627 	 * if all pending pages are added, don't hold its reference
628 	 * any longer if the pcluster isn't hosted by ourselves.
629 	 */
630 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
631 		z_erofs_collection_put(cl);
632 
633 	clt->cl = NULL;
634 	return true;
635 }
636 
should_alloc_managed_pages(struct z_erofs_decompress_frontend * fe,unsigned int cachestrategy,erofs_off_t la)637 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
638 				       unsigned int cachestrategy,
639 				       erofs_off_t la)
640 {
641 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
642 		return false;
643 
644 	if (fe->backmost)
645 		return true;
646 
647 	return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
648 		la < fe->headoffset;
649 }
650 
z_erofs_do_read_page(struct z_erofs_decompress_frontend * fe,struct page * page,struct list_head * pagepool)651 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
652 				struct page *page, struct list_head *pagepool)
653 {
654 	struct inode *const inode = fe->inode;
655 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
656 	struct erofs_map_blocks *const map = &fe->map;
657 	struct z_erofs_collector *const clt = &fe->clt;
658 	const loff_t offset = page_offset(page);
659 	bool tight = true;
660 
661 	enum z_erofs_cache_alloctype cache_strategy;
662 	enum z_erofs_page_type page_type;
663 	unsigned int cur, end, spiltted, index;
664 	int err = 0;
665 
666 	/* register locked file pages as online pages in pack */
667 	z_erofs_onlinepage_init(page);
668 
669 	spiltted = 0;
670 	end = PAGE_SIZE;
671 repeat:
672 	cur = end - 1;
673 
674 	/* lucky, within the range of the current map_blocks */
675 	if (offset + cur >= map->m_la &&
676 	    offset + cur < map->m_la + map->m_llen) {
677 		/* didn't get a valid collection previously (very rare) */
678 		if (!clt->cl)
679 			goto restart_now;
680 		goto hitted;
681 	}
682 
683 	/* go ahead the next map_blocks */
684 	erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
685 
686 	if (z_erofs_collector_end(clt))
687 		fe->backmost = false;
688 
689 	map->m_la = offset + cur;
690 	map->m_llen = 0;
691 	err = z_erofs_map_blocks_iter(inode, map, 0);
692 	if (err)
693 		goto err_out;
694 
695 restart_now:
696 	if (!(map->m_flags & EROFS_MAP_MAPPED))
697 		goto hitted;
698 
699 	err = z_erofs_collector_begin(clt, inode, map);
700 	if (err)
701 		goto err_out;
702 
703 	/* preload all compressed pages (maybe downgrade role if necessary) */
704 	if (should_alloc_managed_pages(fe, sbi->ctx.cache_strategy, map->m_la))
705 		cache_strategy = TRYALLOC;
706 	else
707 		cache_strategy = DONTALLOC;
708 
709 	preload_compressed_pages(clt, MNGD_MAPPING(sbi),
710 				 cache_strategy, pagepool);
711 
712 hitted:
713 	/*
714 	 * Ensure the current partial page belongs to this submit chain rather
715 	 * than other concurrent submit chains or the noio(bypass) chain since
716 	 * those chains are handled asynchronously thus the page cannot be used
717 	 * for inplace I/O or pagevec (should be processed in strict order.)
718 	 */
719 	tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
720 		  clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
721 
722 	cur = end - min_t(unsigned int, offset + end - map->m_la, end);
723 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
724 		zero_user_segment(page, cur, end);
725 		goto next_part;
726 	}
727 
728 	/* let's derive page type */
729 	page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
730 		(!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
731 			(tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
732 				Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
733 
734 	if (cur)
735 		tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
736 
737 retry:
738 	err = z_erofs_attach_page(clt, page, page_type,
739 				  clt->mode >= COLLECT_PRIMARY_FOLLOWED);
740 	/* should allocate an additional staging page for pagevec */
741 	if (err == -EAGAIN) {
742 		struct page *const newpage =
743 				alloc_page(GFP_NOFS | __GFP_NOFAIL);
744 
745 		set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE);
746 		err = z_erofs_attach_page(clt, newpage,
747 					  Z_EROFS_PAGE_TYPE_EXCLUSIVE, true);
748 		if (!err)
749 			goto retry;
750 	}
751 
752 	if (err)
753 		goto err_out;
754 
755 	index = page->index - (map->m_la >> PAGE_SHIFT);
756 
757 	z_erofs_onlinepage_fixup(page, index, true);
758 
759 	/* bump up the number of spiltted parts of a page */
760 	++spiltted;
761 	/* also update nr_pages */
762 	clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
763 next_part:
764 	/* can be used for verification */
765 	map->m_llen = offset + cur - map->m_la;
766 
767 	end = cur;
768 	if (end > 0)
769 		goto repeat;
770 
771 out:
772 	z_erofs_onlinepage_endio(page);
773 
774 	erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
775 		  __func__, page, spiltted, map->m_llen);
776 	return err;
777 
778 	/* if some error occurred while processing this page */
779 err_out:
780 	SetPageError(page);
781 	goto out;
782 }
783 
784 static void z_erofs_decompressqueue_work(struct work_struct *work);
z_erofs_decompress_kickoff(struct z_erofs_decompressqueue * io,bool sync,int bios)785 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
786 				       bool sync, int bios)
787 {
788 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
789 
790 	/* wake up the caller thread for sync decompression */
791 	if (sync) {
792 		if (!atomic_add_return(bios, &io->pending_bios))
793 			complete(&io->u.done);
794 
795 		return;
796 	}
797 
798 	if (atomic_add_return(bios, &io->pending_bios))
799 		return;
800 	/* Use workqueue and sync decompression for atomic contexts only */
801 	if (in_atomic() || irqs_disabled()) {
802 		queue_work(z_erofs_workqueue, &io->u.work);
803 		sbi->ctx.readahead_sync_decompress = true;
804 		return;
805 	}
806 	z_erofs_decompressqueue_work(&io->u.work);
807 }
808 
z_erofs_page_is_invalidated(struct page * page)809 static bool z_erofs_page_is_invalidated(struct page *page)
810 {
811 	return !page->mapping && !z_erofs_is_shortlived_page(page);
812 }
813 
z_erofs_decompressqueue_endio(struct bio * bio)814 static void z_erofs_decompressqueue_endio(struct bio *bio)
815 {
816 	tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
817 	struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
818 	blk_status_t err = bio->bi_status;
819 	struct bio_vec *bvec;
820 	struct bvec_iter_all iter_all;
821 
822 	bio_for_each_segment_all(bvec, bio, iter_all) {
823 		struct page *page = bvec->bv_page;
824 
825 		DBG_BUGON(PageUptodate(page));
826 		DBG_BUGON(z_erofs_page_is_invalidated(page));
827 
828 		if (err)
829 			SetPageError(page);
830 
831 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
832 			if (!err)
833 				SetPageUptodate(page);
834 			unlock_page(page);
835 		}
836 	}
837 	z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
838 	bio_put(bio);
839 }
840 
z_erofs_decompress_pcluster(struct super_block * sb,struct z_erofs_pcluster * pcl,struct list_head * pagepool)841 static int z_erofs_decompress_pcluster(struct super_block *sb,
842 				       struct z_erofs_pcluster *pcl,
843 				       struct list_head *pagepool)
844 {
845 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
846 	struct z_erofs_pagevec_ctor ctor;
847 	unsigned int i, inputsize, outputsize, llen, nr_pages;
848 	struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
849 	struct page **pages, **compressed_pages, *page;
850 
851 	enum z_erofs_page_type page_type;
852 	bool overlapped, partial;
853 	struct z_erofs_collection *cl;
854 	int err;
855 
856 	might_sleep();
857 	cl = z_erofs_primarycollection(pcl);
858 	DBG_BUGON(!READ_ONCE(cl->nr_pages));
859 
860 	mutex_lock(&cl->lock);
861 	nr_pages = cl->nr_pages;
862 
863 	if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
864 		pages = pages_onstack;
865 	} else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
866 		   mutex_trylock(&z_pagemap_global_lock)) {
867 		pages = z_pagemap_global;
868 	} else {
869 		gfp_t gfp_flags = GFP_KERNEL;
870 
871 		if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
872 			gfp_flags |= __GFP_NOFAIL;
873 
874 		pages = kvmalloc_array(nr_pages, sizeof(struct page *),
875 				       gfp_flags);
876 
877 		/* fallback to global pagemap for the lowmem scenario */
878 		if (!pages) {
879 			mutex_lock(&z_pagemap_global_lock);
880 			pages = z_pagemap_global;
881 		}
882 	}
883 
884 	for (i = 0; i < nr_pages; ++i)
885 		pages[i] = NULL;
886 
887 	err = 0;
888 	z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
889 				  cl->pagevec, 0);
890 
891 	for (i = 0; i < cl->vcnt; ++i) {
892 		unsigned int pagenr;
893 
894 		page = z_erofs_pagevec_dequeue(&ctor, &page_type);
895 
896 		/* all pages in pagevec ought to be valid */
897 		DBG_BUGON(!page);
898 		DBG_BUGON(z_erofs_page_is_invalidated(page));
899 
900 		if (z_erofs_put_shortlivedpage(pagepool, page))
901 			continue;
902 
903 		if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
904 			pagenr = 0;
905 		else
906 			pagenr = z_erofs_onlinepage_index(page);
907 
908 		DBG_BUGON(pagenr >= nr_pages);
909 
910 		/*
911 		 * currently EROFS doesn't support multiref(dedup),
912 		 * so here erroring out one multiref page.
913 		 */
914 		if (pages[pagenr]) {
915 			DBG_BUGON(1);
916 			SetPageError(pages[pagenr]);
917 			z_erofs_onlinepage_endio(pages[pagenr]);
918 			err = -EFSCORRUPTED;
919 		}
920 		pages[pagenr] = page;
921 	}
922 	z_erofs_pagevec_ctor_exit(&ctor, true);
923 
924 	overlapped = false;
925 	compressed_pages = pcl->compressed_pages;
926 
927 	for (i = 0; i < pcl->pclusterpages; ++i) {
928 		unsigned int pagenr;
929 
930 		page = compressed_pages[i];
931 
932 		/* all compressed pages ought to be valid */
933 		DBG_BUGON(!page);
934 		DBG_BUGON(z_erofs_page_is_invalidated(page));
935 
936 		if (!z_erofs_is_shortlived_page(page)) {
937 			if (erofs_page_is_managed(sbi, page)) {
938 				if (!PageUptodate(page))
939 					err = -EIO;
940 				continue;
941 			}
942 
943 			/*
944 			 * only if non-head page can be selected
945 			 * for inplace decompression
946 			 */
947 			pagenr = z_erofs_onlinepage_index(page);
948 
949 			DBG_BUGON(pagenr >= nr_pages);
950 			if (pages[pagenr]) {
951 				DBG_BUGON(1);
952 				SetPageError(pages[pagenr]);
953 				z_erofs_onlinepage_endio(pages[pagenr]);
954 				err = -EFSCORRUPTED;
955 			}
956 			pages[pagenr] = page;
957 
958 			overlapped = true;
959 		}
960 
961 		/* PG_error needs checking for all non-managed pages */
962 		if (PageError(page)) {
963 			DBG_BUGON(PageUptodate(page));
964 			err = -EIO;
965 		}
966 	}
967 
968 	if (err)
969 		goto out;
970 
971 	llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
972 	if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
973 		outputsize = llen;
974 		partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
975 	} else {
976 		outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
977 		partial = true;
978 	}
979 
980 	inputsize = pcl->pclusterpages * PAGE_SIZE;
981 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
982 					.sb = sb,
983 					.in = compressed_pages,
984 					.out = pages,
985 					.pageofs_out = cl->pageofs,
986 					.inputsize = inputsize,
987 					.outputsize = outputsize,
988 					.alg = pcl->algorithmformat,
989 					.inplace_io = overlapped,
990 					.partial_decoding = partial
991 				 }, pagepool);
992 
993 out:
994 	/* must handle all compressed pages before ending pages */
995 	for (i = 0; i < pcl->pclusterpages; ++i) {
996 		page = compressed_pages[i];
997 
998 		if (erofs_page_is_managed(sbi, page))
999 			continue;
1000 
1001 		/* recycle all individual short-lived pages */
1002 		(void)z_erofs_put_shortlivedpage(pagepool, page);
1003 
1004 		WRITE_ONCE(compressed_pages[i], NULL);
1005 	}
1006 
1007 	for (i = 0; i < nr_pages; ++i) {
1008 		page = pages[i];
1009 		if (!page)
1010 			continue;
1011 
1012 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1013 
1014 		/* recycle all individual short-lived pages */
1015 		if (z_erofs_put_shortlivedpage(pagepool, page))
1016 			continue;
1017 
1018 		if (err < 0)
1019 			SetPageError(page);
1020 
1021 		z_erofs_onlinepage_endio(page);
1022 	}
1023 
1024 	if (pages == z_pagemap_global)
1025 		mutex_unlock(&z_pagemap_global_lock);
1026 	else if (pages != pages_onstack)
1027 		kvfree(pages);
1028 
1029 	cl->nr_pages = 0;
1030 	cl->vcnt = 0;
1031 
1032 	/* all cl locks MUST be taken before the following line */
1033 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1034 
1035 	/* all cl locks SHOULD be released right now */
1036 	mutex_unlock(&cl->lock);
1037 
1038 	z_erofs_collection_put(cl);
1039 	return err;
1040 }
1041 
z_erofs_decompress_queue(const struct z_erofs_decompressqueue * io,struct list_head * pagepool)1042 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1043 				     struct list_head *pagepool)
1044 {
1045 	z_erofs_next_pcluster_t owned = io->head;
1046 
1047 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1048 		struct z_erofs_pcluster *pcl;
1049 
1050 		/* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1051 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1052 
1053 		/* no possible that 'owned' equals NULL */
1054 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1055 
1056 		pcl = container_of(owned, struct z_erofs_pcluster, next);
1057 		owned = READ_ONCE(pcl->next);
1058 
1059 		z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
1060 	}
1061 }
1062 
z_erofs_decompressqueue_work(struct work_struct * work)1063 static void z_erofs_decompressqueue_work(struct work_struct *work)
1064 {
1065 	struct z_erofs_decompressqueue *bgq =
1066 		container_of(work, struct z_erofs_decompressqueue, u.work);
1067 	LIST_HEAD(pagepool);
1068 
1069 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1070 	z_erofs_decompress_queue(bgq, &pagepool);
1071 
1072 	put_pages_list(&pagepool);
1073 	kvfree(bgq);
1074 }
1075 
pickup_page_for_submission(struct z_erofs_pcluster * pcl,unsigned int nr,struct list_head * pagepool,struct address_space * mc,gfp_t gfp)1076 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1077 					       unsigned int nr,
1078 					       struct list_head *pagepool,
1079 					       struct address_space *mc,
1080 					       gfp_t gfp)
1081 {
1082 	const pgoff_t index = pcl->obj.index;
1083 	bool tocache = false;
1084 
1085 	struct address_space *mapping;
1086 	struct page *oldpage, *page;
1087 
1088 	compressed_page_t t;
1089 	int justfound;
1090 
1091 repeat:
1092 	page = READ_ONCE(pcl->compressed_pages[nr]);
1093 	oldpage = page;
1094 
1095 	if (!page)
1096 		goto out_allocpage;
1097 
1098 	/*
1099 	 * the cached page has not been allocated and
1100 	 * an placeholder is out there, prepare it now.
1101 	 */
1102 	if (page == PAGE_UNALLOCATED) {
1103 		tocache = true;
1104 		goto out_allocpage;
1105 	}
1106 
1107 	/* process the target tagged pointer */
1108 	t = tagptr_init(compressed_page_t, page);
1109 	justfound = tagptr_unfold_tags(t);
1110 	page = tagptr_unfold_ptr(t);
1111 
1112 	/*
1113 	 * preallocated cached pages, which is used to avoid direct reclaim
1114 	 * otherwise, it will go inplace I/O path instead.
1115 	 */
1116 	if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1117 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1118 		set_page_private(page, 0);
1119 		tocache = true;
1120 		goto out_tocache;
1121 	}
1122 	mapping = READ_ONCE(page->mapping);
1123 
1124 	/*
1125 	 * file-backed online pages in plcuster are all locked steady,
1126 	 * therefore it is impossible for `mapping' to be NULL.
1127 	 */
1128 	if (mapping && mapping != mc)
1129 		/* ought to be unmanaged pages */
1130 		goto out;
1131 
1132 	/* directly return for shortlived page as well */
1133 	if (z_erofs_is_shortlived_page(page))
1134 		goto out;
1135 
1136 	lock_page(page);
1137 
1138 	/* only true if page reclaim goes wrong, should never happen */
1139 	DBG_BUGON(justfound && PagePrivate(page));
1140 
1141 	/* the page is still in manage cache */
1142 	if (page->mapping == mc) {
1143 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1144 
1145 		ClearPageError(page);
1146 		if (!PagePrivate(page)) {
1147 			/*
1148 			 * impossible to be !PagePrivate(page) for
1149 			 * the current restriction as well if
1150 			 * the page is already in compressed_pages[].
1151 			 */
1152 			DBG_BUGON(!justfound);
1153 
1154 			justfound = 0;
1155 			set_page_private(page, (unsigned long)pcl);
1156 			SetPagePrivate(page);
1157 		}
1158 
1159 		/* no need to submit io if it is already up-to-date */
1160 		if (PageUptodate(page)) {
1161 			unlock_page(page);
1162 			page = NULL;
1163 		}
1164 		goto out;
1165 	}
1166 
1167 	/*
1168 	 * the managed page has been truncated, it's unsafe to
1169 	 * reuse this one, let's allocate a new cache-managed page.
1170 	 */
1171 	DBG_BUGON(page->mapping);
1172 	DBG_BUGON(!justfound);
1173 
1174 	tocache = true;
1175 	unlock_page(page);
1176 	put_page(page);
1177 out_allocpage:
1178 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1179 	if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1180 		list_add(&page->lru, pagepool);
1181 		cond_resched();
1182 		goto repeat;
1183 	}
1184 out_tocache:
1185 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1186 		/* turn into temporary page if fails (1 ref) */
1187 		set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1188 		goto out;
1189 	}
1190 	attach_page_private(page, pcl);
1191 	/* drop a refcount added by allocpage (then we have 2 refs here) */
1192 	put_page(page);
1193 
1194 out:	/* the only exit (for tracing and debugging) */
1195 	return page;
1196 }
1197 
1198 static struct z_erofs_decompressqueue *
jobqueue_init(struct super_block * sb,struct z_erofs_decompressqueue * fgq,bool * fg)1199 jobqueue_init(struct super_block *sb,
1200 	      struct z_erofs_decompressqueue *fgq, bool *fg)
1201 {
1202 	struct z_erofs_decompressqueue *q;
1203 
1204 	if (fg && !*fg) {
1205 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1206 		if (!q) {
1207 			*fg = true;
1208 			goto fg_out;
1209 		}
1210 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1211 	} else {
1212 fg_out:
1213 		q = fgq;
1214 		init_completion(&fgq->u.done);
1215 		atomic_set(&fgq->pending_bios, 0);
1216 	}
1217 	q->sb = sb;
1218 	q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1219 	return q;
1220 }
1221 
1222 /* define decompression jobqueue types */
1223 enum {
1224 	JQ_BYPASS,
1225 	JQ_SUBMIT,
1226 	NR_JOBQUEUES,
1227 };
1228 
jobqueueset_init(struct super_block * sb,struct z_erofs_decompressqueue * q[],struct z_erofs_decompressqueue * fgq,bool * fg)1229 static void *jobqueueset_init(struct super_block *sb,
1230 			      struct z_erofs_decompressqueue *q[],
1231 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1232 {
1233 	/*
1234 	 * if managed cache is enabled, bypass jobqueue is needed,
1235 	 * no need to read from device for all pclusters in this queue.
1236 	 */
1237 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1238 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1239 
1240 	return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1241 }
1242 
move_to_bypass_jobqueue(struct z_erofs_pcluster * pcl,z_erofs_next_pcluster_t qtail[],z_erofs_next_pcluster_t owned_head)1243 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1244 				    z_erofs_next_pcluster_t qtail[],
1245 				    z_erofs_next_pcluster_t owned_head)
1246 {
1247 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1248 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1249 
1250 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1251 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1252 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1253 
1254 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1255 
1256 	WRITE_ONCE(*submit_qtail, owned_head);
1257 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1258 
1259 	qtail[JQ_BYPASS] = &pcl->next;
1260 }
1261 
z_erofs_submit_queue(struct super_block * sb,struct z_erofs_decompress_frontend * f,struct list_head * pagepool,struct z_erofs_decompressqueue * fgq,bool * force_fg)1262 static void z_erofs_submit_queue(struct super_block *sb,
1263 				 struct z_erofs_decompress_frontend *f,
1264 				 struct list_head *pagepool,
1265 				 struct z_erofs_decompressqueue *fgq,
1266 				 bool *force_fg)
1267 {
1268 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
1269 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1270 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1271 	void *bi_private;
1272 	z_erofs_next_pcluster_t owned_head = f->clt.owned_head;
1273 	/* since bio will be NULL, no need to initialize last_index */
1274 	pgoff_t last_index;
1275 	unsigned int nr_bios = 0;
1276 	struct bio *bio = NULL;
1277 
1278 	bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1279 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1280 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1281 
1282 	/* by default, all need io submission */
1283 	q[JQ_SUBMIT]->head = owned_head;
1284 
1285 	do {
1286 		struct z_erofs_pcluster *pcl;
1287 		pgoff_t cur, end;
1288 		unsigned int i = 0;
1289 		bool bypass = true;
1290 
1291 		/* no possible 'owned_head' equals the following */
1292 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1293 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1294 
1295 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1296 
1297 		cur = pcl->obj.index;
1298 		end = cur + pcl->pclusterpages;
1299 
1300 		/* close the main owned chain at first */
1301 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1302 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1303 
1304 		do {
1305 			struct page *page;
1306 
1307 			page = pickup_page_for_submission(pcl, i++, pagepool,
1308 							  MNGD_MAPPING(sbi),
1309 							  GFP_NOFS);
1310 			if (!page)
1311 				continue;
1312 
1313 			if (bio && cur != last_index + 1) {
1314 submit_bio_retry:
1315 				submit_bio(bio);
1316 				bio = NULL;
1317 			}
1318 
1319 			if (!bio) {
1320 				bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1321 
1322 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1323 				bio_set_dev(bio, sb->s_bdev);
1324 				bio->bi_iter.bi_sector = (sector_t)cur <<
1325 					LOG_SECTORS_PER_BLOCK;
1326 				bio->bi_private = bi_private;
1327 				bio->bi_opf = REQ_OP_READ;
1328 				if (f->readahead)
1329 					bio->bi_opf |= REQ_RAHEAD;
1330 				++nr_bios;
1331 			}
1332 
1333 			if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1334 				goto submit_bio_retry;
1335 
1336 			last_index = cur;
1337 			bypass = false;
1338 		} while (++cur < end);
1339 
1340 		if (!bypass)
1341 			qtail[JQ_SUBMIT] = &pcl->next;
1342 		else
1343 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1344 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1345 
1346 	if (bio)
1347 		submit_bio(bio);
1348 
1349 	/*
1350 	 * although background is preferred, no one is pending for submission.
1351 	 * don't issue workqueue for decompression but drop it directly instead.
1352 	 */
1353 	if (!*force_fg && !nr_bios) {
1354 		kvfree(q[JQ_SUBMIT]);
1355 		return;
1356 	}
1357 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1358 }
1359 
z_erofs_runqueue(struct super_block * sb,struct z_erofs_decompress_frontend * f,struct list_head * pagepool,bool force_fg)1360 static void z_erofs_runqueue(struct super_block *sb,
1361 			     struct z_erofs_decompress_frontend *f,
1362 			     struct list_head *pagepool, bool force_fg)
1363 {
1364 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1365 
1366 	if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL)
1367 		return;
1368 	z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1369 
1370 	/* handle bypass queue (no i/o pclusters) immediately */
1371 	z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1372 
1373 	if (!force_fg)
1374 		return;
1375 
1376 	/* wait until all bios are completed */
1377 	wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1378 
1379 	/* handle synchronous decompress queue in the caller context */
1380 	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1381 }
1382 
z_erofs_readpage(struct file * file,struct page * page)1383 static int z_erofs_readpage(struct file *file, struct page *page)
1384 {
1385 	struct inode *const inode = page->mapping->host;
1386 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1387 	int err;
1388 	LIST_HEAD(pagepool);
1389 
1390 	trace_erofs_readpage(page, false);
1391 
1392 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1393 
1394 	err = z_erofs_do_read_page(&f, page, &pagepool);
1395 	(void)z_erofs_collector_end(&f.clt);
1396 
1397 	/* if some compressed cluster ready, need submit them anyway */
1398 	z_erofs_runqueue(inode->i_sb, &f, &pagepool, true);
1399 
1400 	if (err)
1401 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1402 
1403 	if (f.map.mpage)
1404 		put_page(f.map.mpage);
1405 
1406 	/* clean up the remaining free pages */
1407 	put_pages_list(&pagepool);
1408 	return err;
1409 }
1410 
z_erofs_readahead(struct readahead_control * rac)1411 static void z_erofs_readahead(struct readahead_control *rac)
1412 {
1413 	struct inode *const inode = rac->mapping->host;
1414 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1415 
1416 	unsigned int nr_pages = readahead_count(rac);
1417 	bool sync = (sbi->ctx.readahead_sync_decompress &&
1418 			nr_pages <= sbi->ctx.max_sync_decompress_pages);
1419 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1420 	struct page *page, *head = NULL;
1421 	LIST_HEAD(pagepool);
1422 
1423 	trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1424 
1425 	f.readahead = true;
1426 	f.headoffset = readahead_pos(rac);
1427 
1428 	while ((page = readahead_page(rac))) {
1429 		prefetchw(&page->flags);
1430 
1431 		/*
1432 		 * A pure asynchronous readahead is indicated if
1433 		 * a PG_readahead marked page is hitted at first.
1434 		 * Let's also do asynchronous decompression for this case.
1435 		 */
1436 		sync &= !(PageReadahead(page) && !head);
1437 
1438 		set_page_private(page, (unsigned long)head);
1439 		head = page;
1440 	}
1441 
1442 	while (head) {
1443 		struct page *page = head;
1444 		int err;
1445 
1446 		/* traversal in reverse order */
1447 		head = (void *)page_private(page);
1448 
1449 		err = z_erofs_do_read_page(&f, page, &pagepool);
1450 		if (err)
1451 			erofs_err(inode->i_sb,
1452 				  "readahead error at page %lu @ nid %llu",
1453 				  page->index, EROFS_I(inode)->nid);
1454 		put_page(page);
1455 	}
1456 
1457 	(void)z_erofs_collector_end(&f.clt);
1458 
1459 	z_erofs_runqueue(inode->i_sb, &f, &pagepool, sync);
1460 
1461 	if (f.map.mpage)
1462 		put_page(f.map.mpage);
1463 
1464 	/* clean up the remaining free pages */
1465 	put_pages_list(&pagepool);
1466 }
1467 
1468 const struct address_space_operations z_erofs_aops = {
1469 	.readpage = z_erofs_readpage,
1470 	.readahead = z_erofs_readahead,
1471 };
1472 
1473