• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <linux/cpuhotplug.h>
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 
128 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
129 static struct kthread_worker __rcu **z_erofs_pcpu_workers;
130 
erofs_destroy_percpu_workers(void)131 static void erofs_destroy_percpu_workers(void)
132 {
133 	struct kthread_worker *worker;
134 	unsigned int cpu;
135 
136 	for_each_possible_cpu(cpu) {
137 		worker = rcu_dereference_protected(
138 					z_erofs_pcpu_workers[cpu], 1);
139 		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
140 		if (worker)
141 			kthread_destroy_worker(worker);
142 	}
143 	kfree(z_erofs_pcpu_workers);
144 }
145 
erofs_init_percpu_worker(int cpu)146 static struct kthread_worker *erofs_init_percpu_worker(int cpu)
147 {
148 	struct kthread_worker *worker =
149 		kthread_create_worker_on_cpu(cpu, 0, "erofs_worker/%u", cpu);
150 
151 	if (IS_ERR(worker))
152 		return worker;
153 	if (IS_ENABLED(CONFIG_EROFS_FS_PCPU_KTHREAD_HIPRI))
154 		sched_set_fifo_low(worker->task);
155 	else
156 		sched_set_normal(worker->task, 0);
157 	return worker;
158 }
159 
erofs_init_percpu_workers(void)160 static int erofs_init_percpu_workers(void)
161 {
162 	struct kthread_worker *worker;
163 	unsigned int cpu;
164 
165 	z_erofs_pcpu_workers = kcalloc(num_possible_cpus(),
166 			sizeof(struct kthread_worker *), GFP_ATOMIC);
167 	if (!z_erofs_pcpu_workers)
168 		return -ENOMEM;
169 
170 	for_each_online_cpu(cpu) {	/* could miss cpu{off,on}line? */
171 		worker = erofs_init_percpu_worker(cpu);
172 		if (!IS_ERR(worker))
173 			rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
174 	}
175 	return 0;
176 }
177 #else
erofs_destroy_percpu_workers(void)178 static inline void erofs_destroy_percpu_workers(void) {}
erofs_init_percpu_workers(void)179 static inline int erofs_init_percpu_workers(void) { return 0; }
180 #endif
181 
182 #if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_EROFS_FS_PCPU_KTHREAD)
183 static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock);
184 static enum cpuhp_state erofs_cpuhp_state;
185 
erofs_cpu_online(unsigned int cpu)186 static int erofs_cpu_online(unsigned int cpu)
187 {
188 	struct kthread_worker *worker, *old;
189 
190 	worker = erofs_init_percpu_worker(cpu);
191 	if (IS_ERR(worker))
192 		return PTR_ERR(worker);
193 
194 	spin_lock(&z_erofs_pcpu_worker_lock);
195 	old = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
196 			lockdep_is_held(&z_erofs_pcpu_worker_lock));
197 	if (!old)
198 		rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
199 	spin_unlock(&z_erofs_pcpu_worker_lock);
200 	if (old)
201 		kthread_destroy_worker(worker);
202 	return 0;
203 }
204 
erofs_cpu_offline(unsigned int cpu)205 static int erofs_cpu_offline(unsigned int cpu)
206 {
207 	struct kthread_worker *worker;
208 
209 	spin_lock(&z_erofs_pcpu_worker_lock);
210 	worker = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
211 			lockdep_is_held(&z_erofs_pcpu_worker_lock));
212 	rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
213 	spin_unlock(&z_erofs_pcpu_worker_lock);
214 
215 	synchronize_rcu();
216 	if (worker)
217 		kthread_destroy_worker(worker);
218 	return 0;
219 }
220 
erofs_cpu_hotplug_init(void)221 static int erofs_cpu_hotplug_init(void)
222 {
223 	int state;
224 
225 	state = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
226 			"fs/erofs:online", erofs_cpu_online, erofs_cpu_offline);
227 	if (state < 0)
228 		return state;
229 
230 	erofs_cpuhp_state = state;
231 	return 0;
232 }
233 
erofs_cpu_hotplug_destroy(void)234 static void erofs_cpu_hotplug_destroy(void)
235 {
236 	if (erofs_cpuhp_state)
237 		cpuhp_remove_state_nocalls(erofs_cpuhp_state);
238 }
239 #else /* !CONFIG_HOTPLUG_CPU || !CONFIG_EROFS_FS_PCPU_KTHREAD */
erofs_cpu_hotplug_init(void)240 static inline int erofs_cpu_hotplug_init(void) { return 0; }
erofs_cpu_hotplug_destroy(void)241 static inline void erofs_cpu_hotplug_destroy(void) {}
242 #endif
243 
z_erofs_exit_zip_subsystem(void)244 void z_erofs_exit_zip_subsystem(void)
245 {
246 	erofs_cpu_hotplug_destroy();
247 	erofs_destroy_percpu_workers();
248 	destroy_workqueue(z_erofs_workqueue);
249 	z_erofs_destroy_pcluster_pool();
250 }
251 
z_erofs_init_zip_subsystem(void)252 int __init z_erofs_init_zip_subsystem(void)
253 {
254 	int err = z_erofs_create_pcluster_pool();
255 
256 	if (err)
257 		goto out_error_pcluster_pool;
258 
259 	z_erofs_workqueue = alloc_workqueue("erofs_worker",
260 			WQ_UNBOUND | WQ_HIGHPRI, num_possible_cpus());
261 	if (!z_erofs_workqueue) {
262 		err = -ENOMEM;
263 		goto out_error_workqueue_init;
264 	}
265 
266 	err = erofs_init_percpu_workers();
267 	if (err)
268 		goto out_error_pcpu_worker;
269 
270 	err = erofs_cpu_hotplug_init();
271 	if (err < 0)
272 		goto out_error_cpuhp_init;
273 	return err;
274 
275 out_error_cpuhp_init:
276 	erofs_destroy_percpu_workers();
277 out_error_pcpu_worker:
278 	destroy_workqueue(z_erofs_workqueue);
279 out_error_workqueue_init:
280 	z_erofs_destroy_pcluster_pool();
281 out_error_pcluster_pool:
282 	return err;
283 }
284 
285 enum z_erofs_collectmode {
286 	COLLECT_SECONDARY,
287 	COLLECT_PRIMARY,
288 	/*
289 	 * The current collection was the tail of an exist chain, in addition
290 	 * that the previous processed chained collections are all decided to
291 	 * be hooked up to it.
292 	 * A new chain will be created for the remaining collections which are
293 	 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
294 	 * the next collection cannot reuse the whole page safely in
295 	 * the following scenario:
296 	 *  ________________________________________________________________
297 	 * |      tail (partial) page     |       head (partial) page       |
298 	 * |   (belongs to the next cl)   |   (belongs to the current cl)   |
299 	 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
300 	 */
301 	COLLECT_PRIMARY_HOOKED,
302 	/*
303 	 * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it
304 	 * could be dispatched into bypass queue later due to uptodated managed
305 	 * pages. All related online pages cannot be reused for inplace I/O (or
306 	 * pagevec) since it can be directly decoded without I/O submission.
307 	 */
308 	COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
309 	/*
310 	 * The current collection has been linked with the owned chain, and
311 	 * could also be linked with the remaining collections, which means
312 	 * if the processing page is the tail page of the collection, thus
313 	 * the current collection can safely use the whole page (since
314 	 * the previous collection is under control) for in-place I/O, as
315 	 * illustrated below:
316 	 *  ________________________________________________________________
317 	 * |  tail (partial) page |          head (partial) page           |
318 	 * |  (of the current cl) |      (of the previous collection)      |
319 	 * |  PRIMARY_FOLLOWED or |                                        |
320 	 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
321 	 *
322 	 * [  (*) the above page can be used as inplace I/O.               ]
323 	 */
324 	COLLECT_PRIMARY_FOLLOWED,
325 };
326 
327 struct z_erofs_collector {
328 	struct z_erofs_pagevec_ctor vector;
329 
330 	struct z_erofs_pcluster *pcl, *tailpcl;
331 	struct z_erofs_collection *cl;
332 	/* a pointer used to pick up inplace I/O pages */
333 	struct page **icpage_ptr;
334 	z_erofs_next_pcluster_t owned_head;
335 
336 	enum z_erofs_collectmode mode;
337 };
338 
339 struct z_erofs_decompress_frontend {
340 	struct inode *const inode;
341 
342 	struct z_erofs_collector clt;
343 	struct erofs_map_blocks map;
344 
345 	bool readahead;
346 	/* used for applying cache strategy on the fly */
347 	bool backmost;
348 	erofs_off_t headoffset;
349 };
350 
351 #define COLLECTOR_INIT() { \
352 	.owned_head = Z_EROFS_PCLUSTER_TAIL, \
353 	.mode = COLLECT_PRIMARY_FOLLOWED }
354 
355 #define DECOMPRESS_FRONTEND_INIT(__i) { \
356 	.inode = __i, .clt = COLLECTOR_INIT(), \
357 	.backmost = true, }
358 
359 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
360 static DEFINE_MUTEX(z_pagemap_global_lock);
361 
preload_compressed_pages(struct z_erofs_collector * clt,struct address_space * mc,enum z_erofs_cache_alloctype type,struct list_head * pagepool)362 static void preload_compressed_pages(struct z_erofs_collector *clt,
363 				     struct address_space *mc,
364 				     enum z_erofs_cache_alloctype type,
365 				     struct list_head *pagepool)
366 {
367 	struct z_erofs_pcluster *pcl = clt->pcl;
368 	bool standalone = true;
369 	gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
370 			__GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
371 	struct page **pages;
372 	pgoff_t index;
373 
374 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
375 		return;
376 
377 	pages = pcl->compressed_pages;
378 	index = pcl->obj.index;
379 	for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) {
380 		struct page *page;
381 		compressed_page_t t;
382 		struct page *newpage = NULL;
383 
384 		/* the compressed page was loaded before */
385 		if (READ_ONCE(*pages))
386 			continue;
387 
388 		page = find_get_page(mc, index);
389 
390 		if (page) {
391 			t = tag_compressed_page_justfound(page);
392 		} else {
393 			/* I/O is needed, no possible to decompress directly */
394 			standalone = false;
395 			switch (type) {
396 			case DELAYEDALLOC:
397 				t = tagptr_init(compressed_page_t,
398 						PAGE_UNALLOCATED);
399 				break;
400 			case TRYALLOC:
401 				newpage = erofs_allocpage(pagepool, gfp);
402 				if (!newpage)
403 					continue;
404 				set_page_private(newpage,
405 						 Z_EROFS_PREALLOCATED_PAGE);
406 				t = tag_compressed_page_justfound(newpage);
407 				break;
408 			default:        /* DONTALLOC */
409 				continue;
410 			}
411 		}
412 
413 		if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
414 			continue;
415 
416 		if (page) {
417 			put_page(page);
418 		} else if (newpage) {
419 			set_page_private(newpage, 0);
420 			list_add(&newpage->lru, pagepool);
421 		}
422 	}
423 
424 	/*
425 	 * don't do inplace I/O if all compressed pages are available in
426 	 * managed cache since it can be moved to the bypass queue instead.
427 	 */
428 	if (standalone)
429 		clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
430 }
431 
432 /* 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)433 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
434 				       struct erofs_workgroup *grp)
435 {
436 	struct z_erofs_pcluster *const pcl =
437 		container_of(grp, struct z_erofs_pcluster, obj);
438 	struct address_space *const mapping = MNGD_MAPPING(sbi);
439 	int i;
440 
441 	/*
442 	 * refcount of workgroup is now freezed as 1,
443 	 * therefore no need to worry about available decompression users.
444 	 */
445 	for (i = 0; i < pcl->pclusterpages; ++i) {
446 		struct page *page = pcl->compressed_pages[i];
447 
448 		if (!page)
449 			continue;
450 
451 		/* block other users from reclaiming or migrating the page */
452 		if (!trylock_page(page))
453 			return -EBUSY;
454 
455 		if (page->mapping != mapping)
456 			continue;
457 
458 		/* barrier is implied in the following 'unlock_page' */
459 		WRITE_ONCE(pcl->compressed_pages[i], NULL);
460 		detach_page_private(page);
461 		unlock_page(page);
462 	}
463 	return 0;
464 }
465 
erofs_try_to_free_cached_page(struct address_space * mapping,struct page * page)466 int erofs_try_to_free_cached_page(struct address_space *mapping,
467 				  struct page *page)
468 {
469 	struct z_erofs_pcluster *const pcl = (void *)page_private(page);
470 	int ret = 0;	/* 0 - busy */
471 
472 	if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
473 		unsigned int i;
474 
475 		for (i = 0; i < pcl->pclusterpages; ++i) {
476 			if (pcl->compressed_pages[i] == page) {
477 				WRITE_ONCE(pcl->compressed_pages[i], NULL);
478 				ret = 1;
479 				break;
480 			}
481 		}
482 		erofs_workgroup_unfreeze(&pcl->obj, 1);
483 
484 		if (ret)
485 			detach_page_private(page);
486 	}
487 	return ret;
488 }
489 
490 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
z_erofs_try_inplace_io(struct z_erofs_collector * clt,struct page * page)491 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
492 				   struct page *page)
493 {
494 	struct z_erofs_pcluster *const pcl = clt->pcl;
495 
496 	while (clt->icpage_ptr > pcl->compressed_pages)
497 		if (!cmpxchg(--clt->icpage_ptr, NULL, page))
498 			return true;
499 	return false;
500 }
501 
502 /* 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)503 static int z_erofs_attach_page(struct z_erofs_collector *clt,
504 			       struct page *page, enum z_erofs_page_type type,
505 			       bool pvec_safereuse)
506 {
507 	int ret;
508 
509 	/* give priority for inplaceio */
510 	if (clt->mode >= COLLECT_PRIMARY &&
511 	    type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
512 	    z_erofs_try_inplace_io(clt, page))
513 		return 0;
514 
515 	ret = z_erofs_pagevec_enqueue(&clt->vector, page, type,
516 				      pvec_safereuse);
517 	clt->cl->vcnt += (unsigned int)ret;
518 	return ret ? 0 : -EAGAIN;
519 }
520 
521 static enum z_erofs_collectmode
try_to_claim_pcluster(struct z_erofs_pcluster * pcl,z_erofs_next_pcluster_t * owned_head)522 try_to_claim_pcluster(struct z_erofs_pcluster *pcl,
523 		      z_erofs_next_pcluster_t *owned_head)
524 {
525 	/* let's claim these following types of pclusters */
526 retry:
527 	if (pcl->next == Z_EROFS_PCLUSTER_NIL) {
528 		/* type 1, nil pcluster */
529 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
530 			    *owned_head) != Z_EROFS_PCLUSTER_NIL)
531 			goto retry;
532 
533 		*owned_head = &pcl->next;
534 		/* lucky, I am the followee :) */
535 		return COLLECT_PRIMARY_FOLLOWED;
536 	} else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) {
537 		/*
538 		 * type 2, link to the end of a existing open chain,
539 		 * be careful that its submission itself is governed
540 		 * by the original owned chain.
541 		 */
542 		if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
543 			    *owned_head) != Z_EROFS_PCLUSTER_TAIL)
544 			goto retry;
545 		*owned_head = Z_EROFS_PCLUSTER_TAIL;
546 		return COLLECT_PRIMARY_HOOKED;
547 	}
548 	return COLLECT_PRIMARY;	/* :( better luck next time */
549 }
550 
z_erofs_lookup_collection(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)551 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
552 				     struct inode *inode,
553 				     struct erofs_map_blocks *map)
554 {
555 	struct z_erofs_pcluster *pcl = clt->pcl;
556 	struct z_erofs_collection *cl;
557 	unsigned int length;
558 
559 	/* to avoid unexpected loop formed by corrupted images */
560 	if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
561 		DBG_BUGON(1);
562 		return -EFSCORRUPTED;
563 	}
564 
565 	cl = z_erofs_primarycollection(pcl);
566 	if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
567 		DBG_BUGON(1);
568 		return -EFSCORRUPTED;
569 	}
570 
571 	length = READ_ONCE(pcl->length);
572 	if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
573 		if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
574 			DBG_BUGON(1);
575 			return -EFSCORRUPTED;
576 		}
577 	} else {
578 		unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
579 
580 		if (map->m_flags & EROFS_MAP_FULL_MAPPED)
581 			llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
582 
583 		while (llen > length &&
584 		       length != cmpxchg_relaxed(&pcl->length, length, llen)) {
585 			cpu_relax();
586 			length = READ_ONCE(pcl->length);
587 		}
588 	}
589 	mutex_lock(&cl->lock);
590 	/* used to check tail merging loop due to corrupted images */
591 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
592 		clt->tailpcl = pcl;
593 	clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head);
594 	/* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */
595 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
596 		clt->tailpcl = NULL;
597 	clt->cl = cl;
598 	return 0;
599 }
600 
z_erofs_register_collection(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)601 static int z_erofs_register_collection(struct z_erofs_collector *clt,
602 				       struct inode *inode,
603 				       struct erofs_map_blocks *map)
604 {
605 	struct z_erofs_pcluster *pcl;
606 	struct z_erofs_collection *cl;
607 	struct erofs_workgroup *grp;
608 	int err;
609 
610 	/* no available pcluster, let's allocate one */
611 	pcl = z_erofs_alloc_pcluster(map->m_plen >> PAGE_SHIFT);
612 	if (IS_ERR(pcl))
613 		return PTR_ERR(pcl);
614 
615 	atomic_set(&pcl->obj.refcount, 1);
616 	pcl->obj.index = map->m_pa >> PAGE_SHIFT;
617 
618 	pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
619 		(map->m_flags & EROFS_MAP_FULL_MAPPED ?
620 			Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
621 
622 	if (map->m_flags & EROFS_MAP_ZIPPED)
623 		pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
624 	else
625 		pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
626 
627 	/* new pclusters should be claimed as type 1, primary and followed */
628 	pcl->next = clt->owned_head;
629 	clt->mode = COLLECT_PRIMARY_FOLLOWED;
630 
631 	cl = z_erofs_primarycollection(pcl);
632 	cl->pageofs = map->m_la & ~PAGE_MASK;
633 
634 	/*
635 	 * lock all primary followed works before visible to others
636 	 * and mutex_trylock *never* fails for a new pcluster.
637 	 */
638 	mutex_init(&cl->lock);
639 	DBG_BUGON(!mutex_trylock(&cl->lock));
640 
641 	grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
642 	if (IS_ERR(grp)) {
643 		err = PTR_ERR(grp);
644 		goto err_out;
645 	}
646 
647 	if (grp != &pcl->obj) {
648 		clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
649 		err = -EEXIST;
650 		goto err_out;
651 	}
652 	/* used to check tail merging loop due to corrupted images */
653 	if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
654 		clt->tailpcl = pcl;
655 	clt->owned_head = &pcl->next;
656 	clt->pcl = pcl;
657 	clt->cl = cl;
658 	return 0;
659 
660 err_out:
661 	mutex_unlock(&cl->lock);
662 	z_erofs_free_pcluster(pcl);
663 	return err;
664 }
665 
z_erofs_collector_begin(struct z_erofs_collector * clt,struct inode * inode,struct erofs_map_blocks * map)666 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
667 				   struct inode *inode,
668 				   struct erofs_map_blocks *map)
669 {
670 	struct erofs_workgroup *grp;
671 	int ret;
672 
673 	DBG_BUGON(clt->cl);
674 
675 	/* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
676 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
677 	DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
678 
679 	if (!PAGE_ALIGNED(map->m_pa)) {
680 		DBG_BUGON(1);
681 		return -EINVAL;
682 	}
683 
684 	grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
685 	if (grp) {
686 		clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
687 	} else {
688 		ret = z_erofs_register_collection(clt, inode, map);
689 
690 		if (!ret)
691 			goto out;
692 		if (ret != -EEXIST)
693 			return ret;
694 	}
695 
696 	ret = z_erofs_lookup_collection(clt, inode, map);
697 	if (ret) {
698 		erofs_workgroup_put(&clt->pcl->obj);
699 		return ret;
700 	}
701 
702 out:
703 	z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
704 				  clt->cl->pagevec, clt->cl->vcnt);
705 
706 	/* since file-backed online pages are traversed in reverse order */
707 	clt->icpage_ptr = clt->pcl->compressed_pages + clt->pcl->pclusterpages;
708 	return 0;
709 }
710 
711 /*
712  * keep in mind that no referenced pclusters will be freed
713  * only after a RCU grace period.
714  */
z_erofs_rcu_callback(struct rcu_head * head)715 static void z_erofs_rcu_callback(struct rcu_head *head)
716 {
717 	struct z_erofs_collection *const cl =
718 		container_of(head, struct z_erofs_collection, rcu);
719 
720 	z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster,
721 					   primary_collection));
722 }
723 
erofs_workgroup_free_rcu(struct erofs_workgroup * grp)724 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
725 {
726 	struct z_erofs_pcluster *const pcl =
727 		container_of(grp, struct z_erofs_pcluster, obj);
728 	struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
729 
730 	call_rcu(&cl->rcu, z_erofs_rcu_callback);
731 }
732 
z_erofs_collection_put(struct z_erofs_collection * cl)733 static void z_erofs_collection_put(struct z_erofs_collection *cl)
734 {
735 	struct z_erofs_pcluster *const pcl =
736 		container_of(cl, struct z_erofs_pcluster, primary_collection);
737 
738 	erofs_workgroup_put(&pcl->obj);
739 }
740 
z_erofs_collector_end(struct z_erofs_collector * clt)741 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
742 {
743 	struct z_erofs_collection *cl = clt->cl;
744 
745 	if (!cl)
746 		return false;
747 
748 	z_erofs_pagevec_ctor_exit(&clt->vector, false);
749 	mutex_unlock(&cl->lock);
750 
751 	/*
752 	 * if all pending pages are added, don't hold its reference
753 	 * any longer if the pcluster isn't hosted by ourselves.
754 	 */
755 	if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
756 		z_erofs_collection_put(cl);
757 
758 	clt->cl = NULL;
759 	return true;
760 }
761 
should_alloc_managed_pages(struct z_erofs_decompress_frontend * fe,unsigned int cachestrategy,erofs_off_t la)762 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
763 				       unsigned int cachestrategy,
764 				       erofs_off_t la)
765 {
766 	if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
767 		return false;
768 
769 	if (fe->backmost)
770 		return true;
771 
772 	return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
773 		la < fe->headoffset;
774 }
775 
z_erofs_do_read_page(struct z_erofs_decompress_frontend * fe,struct page * page,struct list_head * pagepool)776 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
777 				struct page *page, struct list_head *pagepool)
778 {
779 	struct inode *const inode = fe->inode;
780 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
781 	struct erofs_map_blocks *const map = &fe->map;
782 	struct z_erofs_collector *const clt = &fe->clt;
783 	const loff_t offset = page_offset(page);
784 	bool tight = true;
785 
786 	enum z_erofs_cache_alloctype cache_strategy;
787 	enum z_erofs_page_type page_type;
788 	unsigned int cur, end, spiltted, index;
789 	int err = 0;
790 
791 	/* register locked file pages as online pages in pack */
792 	z_erofs_onlinepage_init(page);
793 
794 	spiltted = 0;
795 	end = PAGE_SIZE;
796 repeat:
797 	cur = end - 1;
798 
799 	/* lucky, within the range of the current map_blocks */
800 	if (offset + cur >= map->m_la &&
801 	    offset + cur < map->m_la + map->m_llen) {
802 		/* didn't get a valid collection previously (very rare) */
803 		if (!clt->cl)
804 			goto restart_now;
805 		goto hitted;
806 	}
807 
808 	/* go ahead the next map_blocks */
809 	erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
810 
811 	if (z_erofs_collector_end(clt))
812 		fe->backmost = false;
813 
814 	map->m_la = offset + cur;
815 	map->m_llen = 0;
816 	err = z_erofs_map_blocks_iter(inode, map, 0);
817 	if (err)
818 		goto err_out;
819 
820 restart_now:
821 	if (!(map->m_flags & EROFS_MAP_MAPPED))
822 		goto hitted;
823 
824 	err = z_erofs_collector_begin(clt, inode, map);
825 	if (err)
826 		goto err_out;
827 
828 	/* preload all compressed pages (maybe downgrade role if necessary) */
829 	if (should_alloc_managed_pages(fe, sbi->ctx.cache_strategy, map->m_la))
830 		cache_strategy = TRYALLOC;
831 	else
832 		cache_strategy = DONTALLOC;
833 
834 	preload_compressed_pages(clt, MNGD_MAPPING(sbi),
835 				 cache_strategy, pagepool);
836 
837 hitted:
838 	/*
839 	 * Ensure the current partial page belongs to this submit chain rather
840 	 * than other concurrent submit chains or the noio(bypass) chain since
841 	 * those chains are handled asynchronously thus the page cannot be used
842 	 * for inplace I/O or pagevec (should be processed in strict order.)
843 	 */
844 	tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
845 		  clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
846 
847 	cur = end - min_t(erofs_off_t, offset + end - map->m_la, end);
848 	if (!(map->m_flags & EROFS_MAP_MAPPED)) {
849 		zero_user_segment(page, cur, end);
850 		++spiltted;
851 		tight = false;
852 		goto next_part;
853 	}
854 
855 	/* let's derive page type */
856 	page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
857 		(!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
858 			(tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
859 				Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
860 
861 	if (cur)
862 		tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
863 
864 retry:
865 	err = z_erofs_attach_page(clt, page, page_type,
866 				  clt->mode >= COLLECT_PRIMARY_FOLLOWED);
867 	/* should allocate an additional staging page for pagevec */
868 	if (err == -EAGAIN) {
869 		struct page *const newpage =
870 				alloc_page(GFP_NOFS | __GFP_NOFAIL);
871 
872 		set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE);
873 		err = z_erofs_attach_page(clt, newpage,
874 					  Z_EROFS_PAGE_TYPE_EXCLUSIVE, true);
875 		if (!err)
876 			goto retry;
877 	}
878 
879 	if (err)
880 		goto err_out;
881 
882 	index = page->index - (map->m_la >> PAGE_SHIFT);
883 
884 	z_erofs_onlinepage_fixup(page, index, true);
885 
886 	/* bump up the number of spiltted parts of a page */
887 	++spiltted;
888 	/* also update nr_pages */
889 	clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
890 next_part:
891 	/* can be used for verification */
892 	map->m_llen = offset + cur - map->m_la;
893 
894 	end = cur;
895 	if (end > 0)
896 		goto repeat;
897 
898 out:
899 	z_erofs_onlinepage_endio(page);
900 
901 	erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
902 		  __func__, page, spiltted, map->m_llen);
903 	return err;
904 
905 	/* if some error occurred while processing this page */
906 err_out:
907 	SetPageError(page);
908 	goto out;
909 }
910 
911 static void z_erofs_decompressqueue_work(struct work_struct *work);
912 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
z_erofs_decompressqueue_kthread_work(struct kthread_work * work)913 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
914 {
915 	z_erofs_decompressqueue_work((struct work_struct *)work);
916 }
917 #endif
z_erofs_decompress_kickoff(struct z_erofs_decompressqueue * io,bool sync,int bios)918 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
919 				       bool sync, int bios)
920 {
921 	struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
922 
923 	/* wake up the caller thread for sync decompression */
924 	if (sync) {
925 		if (!atomic_add_return(bios, &io->pending_bios))
926 			complete(&io->u.done);
927 
928 		return;
929 	}
930 
931 	if (atomic_add_return(bios, &io->pending_bios))
932 		return;
933 	/* Use workqueue and sync decompression for atomic contexts only */
934 	if (in_atomic() || irqs_disabled()) {
935 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
936 		struct kthread_worker *worker;
937 
938 		rcu_read_lock();
939 		worker = rcu_dereference(
940 				z_erofs_pcpu_workers[raw_smp_processor_id()]);
941 		if (!worker) {
942 			INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);
943 			queue_work(z_erofs_workqueue, &io->u.work);
944 		} else {
945 			kthread_queue_work(worker, &io->u.kthread_work);
946 		}
947 		rcu_read_unlock();
948 #else
949 		queue_work(z_erofs_workqueue, &io->u.work);
950 #endif
951 		sbi->ctx.readahead_sync_decompress = true;
952 		return;
953 	}
954 	z_erofs_decompressqueue_work(&io->u.work);
955 }
956 
z_erofs_page_is_invalidated(struct page * page)957 static bool z_erofs_page_is_invalidated(struct page *page)
958 {
959 	return !page->mapping && !z_erofs_is_shortlived_page(page);
960 }
961 
z_erofs_decompressqueue_endio(struct bio * bio)962 static void z_erofs_decompressqueue_endio(struct bio *bio)
963 {
964 	tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
965 	struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
966 	blk_status_t err = bio->bi_status;
967 	struct bio_vec *bvec;
968 	struct bvec_iter_all iter_all;
969 
970 	bio_for_each_segment_all(bvec, bio, iter_all) {
971 		struct page *page = bvec->bv_page;
972 
973 		DBG_BUGON(PageUptodate(page));
974 		DBG_BUGON(z_erofs_page_is_invalidated(page));
975 
976 		if (err)
977 			SetPageError(page);
978 
979 		if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
980 			if (!err)
981 				SetPageUptodate(page);
982 			unlock_page(page);
983 		}
984 	}
985 	z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
986 	bio_put(bio);
987 }
988 
z_erofs_decompress_pcluster(struct super_block * sb,struct z_erofs_pcluster * pcl,struct list_head * pagepool)989 static int z_erofs_decompress_pcluster(struct super_block *sb,
990 				       struct z_erofs_pcluster *pcl,
991 				       struct list_head *pagepool)
992 {
993 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
994 	struct z_erofs_pagevec_ctor ctor;
995 	unsigned int i, inputsize, outputsize, llen, nr_pages;
996 	struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
997 	struct page **pages, **compressed_pages, *page;
998 
999 	enum z_erofs_page_type page_type;
1000 	bool overlapped, partial;
1001 	struct z_erofs_collection *cl;
1002 	int err;
1003 
1004 	might_sleep();
1005 	cl = z_erofs_primarycollection(pcl);
1006 	DBG_BUGON(!READ_ONCE(cl->nr_pages));
1007 
1008 	mutex_lock(&cl->lock);
1009 	nr_pages = cl->nr_pages;
1010 
1011 	if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
1012 		pages = pages_onstack;
1013 	} else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
1014 		   mutex_trylock(&z_pagemap_global_lock)) {
1015 		pages = z_pagemap_global;
1016 	} else {
1017 		gfp_t gfp_flags = GFP_KERNEL;
1018 
1019 		if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
1020 			gfp_flags |= __GFP_NOFAIL;
1021 
1022 		pages = kvmalloc_array(nr_pages, sizeof(struct page *),
1023 				       gfp_flags);
1024 
1025 		/* fallback to global pagemap for the lowmem scenario */
1026 		if (!pages) {
1027 			mutex_lock(&z_pagemap_global_lock);
1028 			pages = z_pagemap_global;
1029 		}
1030 	}
1031 
1032 	for (i = 0; i < nr_pages; ++i)
1033 		pages[i] = NULL;
1034 
1035 	err = 0;
1036 	z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
1037 				  cl->pagevec, 0);
1038 
1039 	for (i = 0; i < cl->vcnt; ++i) {
1040 		unsigned int pagenr;
1041 
1042 		page = z_erofs_pagevec_dequeue(&ctor, &page_type);
1043 
1044 		/* all pages in pagevec ought to be valid */
1045 		DBG_BUGON(!page);
1046 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1047 
1048 		if (z_erofs_put_shortlivedpage(pagepool, page))
1049 			continue;
1050 
1051 		if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
1052 			pagenr = 0;
1053 		else
1054 			pagenr = z_erofs_onlinepage_index(page);
1055 
1056 		DBG_BUGON(pagenr >= nr_pages);
1057 
1058 		/*
1059 		 * currently EROFS doesn't support multiref(dedup),
1060 		 * so here erroring out one multiref page.
1061 		 */
1062 		if (pages[pagenr]) {
1063 			DBG_BUGON(1);
1064 			SetPageError(pages[pagenr]);
1065 			z_erofs_onlinepage_endio(pages[pagenr]);
1066 			err = -EFSCORRUPTED;
1067 		}
1068 		pages[pagenr] = page;
1069 	}
1070 	z_erofs_pagevec_ctor_exit(&ctor, true);
1071 
1072 	overlapped = false;
1073 	compressed_pages = pcl->compressed_pages;
1074 
1075 	for (i = 0; i < pcl->pclusterpages; ++i) {
1076 		unsigned int pagenr;
1077 
1078 		page = compressed_pages[i];
1079 
1080 		/* all compressed pages ought to be valid */
1081 		DBG_BUGON(!page);
1082 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1083 
1084 		if (!z_erofs_is_shortlived_page(page)) {
1085 			if (erofs_page_is_managed(sbi, page)) {
1086 				if (!PageUptodate(page))
1087 					err = -EIO;
1088 				continue;
1089 			}
1090 
1091 			/*
1092 			 * only if non-head page can be selected
1093 			 * for inplace decompression
1094 			 */
1095 			pagenr = z_erofs_onlinepage_index(page);
1096 
1097 			DBG_BUGON(pagenr >= nr_pages);
1098 			if (pages[pagenr]) {
1099 				DBG_BUGON(1);
1100 				SetPageError(pages[pagenr]);
1101 				z_erofs_onlinepage_endio(pages[pagenr]);
1102 				err = -EFSCORRUPTED;
1103 			}
1104 			pages[pagenr] = page;
1105 
1106 			overlapped = true;
1107 		}
1108 
1109 		/* PG_error needs checking for all non-managed pages */
1110 		if (PageError(page)) {
1111 			DBG_BUGON(PageUptodate(page));
1112 			err = -EIO;
1113 		}
1114 	}
1115 
1116 	if (err)
1117 		goto out;
1118 
1119 	llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
1120 	if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
1121 		outputsize = llen;
1122 		partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
1123 	} else {
1124 		outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
1125 		partial = true;
1126 	}
1127 
1128 	inputsize = pcl->pclusterpages * PAGE_SIZE;
1129 	err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
1130 					.sb = sb,
1131 					.in = compressed_pages,
1132 					.out = pages,
1133 					.pageofs_out = cl->pageofs,
1134 					.inputsize = inputsize,
1135 					.outputsize = outputsize,
1136 					.alg = pcl->algorithmformat,
1137 					.inplace_io = overlapped,
1138 					.partial_decoding = partial
1139 				 }, pagepool);
1140 
1141 out:
1142 	/* must handle all compressed pages before ending pages */
1143 	for (i = 0; i < pcl->pclusterpages; ++i) {
1144 		page = compressed_pages[i];
1145 
1146 		if (erofs_page_is_managed(sbi, page))
1147 			continue;
1148 
1149 		/* recycle all individual short-lived pages */
1150 		(void)z_erofs_put_shortlivedpage(pagepool, page);
1151 
1152 		WRITE_ONCE(compressed_pages[i], NULL);
1153 	}
1154 
1155 	for (i = 0; i < nr_pages; ++i) {
1156 		page = pages[i];
1157 		if (!page)
1158 			continue;
1159 
1160 		DBG_BUGON(z_erofs_page_is_invalidated(page));
1161 
1162 		/* recycle all individual short-lived pages */
1163 		if (z_erofs_put_shortlivedpage(pagepool, page))
1164 			continue;
1165 
1166 		if (err < 0)
1167 			SetPageError(page);
1168 
1169 		z_erofs_onlinepage_endio(page);
1170 	}
1171 
1172 	if (pages == z_pagemap_global)
1173 		mutex_unlock(&z_pagemap_global_lock);
1174 	else if (pages != pages_onstack)
1175 		kvfree(pages);
1176 
1177 	cl->nr_pages = 0;
1178 	cl->vcnt = 0;
1179 
1180 	/* all cl locks MUST be taken before the following line */
1181 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1182 
1183 	/* all cl locks SHOULD be released right now */
1184 	mutex_unlock(&cl->lock);
1185 
1186 	z_erofs_collection_put(cl);
1187 	return err;
1188 }
1189 
z_erofs_decompress_queue(const struct z_erofs_decompressqueue * io,struct list_head * pagepool)1190 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1191 				     struct list_head *pagepool)
1192 {
1193 	z_erofs_next_pcluster_t owned = io->head;
1194 
1195 	while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1196 		struct z_erofs_pcluster *pcl;
1197 
1198 		/* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1199 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1200 
1201 		/* no possible that 'owned' equals NULL */
1202 		DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1203 
1204 		pcl = container_of(owned, struct z_erofs_pcluster, next);
1205 		owned = READ_ONCE(pcl->next);
1206 
1207 		z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
1208 	}
1209 }
1210 
z_erofs_decompressqueue_work(struct work_struct * work)1211 static void z_erofs_decompressqueue_work(struct work_struct *work)
1212 {
1213 	struct z_erofs_decompressqueue *bgq =
1214 		container_of(work, struct z_erofs_decompressqueue, u.work);
1215 	LIST_HEAD(pagepool);
1216 
1217 	DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1218 	z_erofs_decompress_queue(bgq, &pagepool);
1219 
1220 	put_pages_list(&pagepool);
1221 	kvfree(bgq);
1222 }
1223 
pickup_page_for_submission(struct z_erofs_pcluster * pcl,unsigned int nr,struct list_head * pagepool,struct address_space * mc,gfp_t gfp)1224 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1225 					       unsigned int nr,
1226 					       struct list_head *pagepool,
1227 					       struct address_space *mc,
1228 					       gfp_t gfp)
1229 {
1230 	const pgoff_t index = pcl->obj.index;
1231 	bool tocache = false;
1232 
1233 	struct address_space *mapping;
1234 	struct page *oldpage, *page;
1235 
1236 	compressed_page_t t;
1237 	int justfound;
1238 
1239 repeat:
1240 	page = READ_ONCE(pcl->compressed_pages[nr]);
1241 	oldpage = page;
1242 
1243 	if (!page)
1244 		goto out_allocpage;
1245 
1246 	/*
1247 	 * the cached page has not been allocated and
1248 	 * an placeholder is out there, prepare it now.
1249 	 */
1250 	if (page == PAGE_UNALLOCATED) {
1251 		tocache = true;
1252 		goto out_allocpage;
1253 	}
1254 
1255 	/* process the target tagged pointer */
1256 	t = tagptr_init(compressed_page_t, page);
1257 	justfound = tagptr_unfold_tags(t);
1258 	page = tagptr_unfold_ptr(t);
1259 
1260 	/*
1261 	 * preallocated cached pages, which is used to avoid direct reclaim
1262 	 * otherwise, it will go inplace I/O path instead.
1263 	 */
1264 	if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1265 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1266 		set_page_private(page, 0);
1267 		tocache = true;
1268 		goto out_tocache;
1269 	}
1270 	mapping = READ_ONCE(page->mapping);
1271 
1272 	/*
1273 	 * file-backed online pages in plcuster are all locked steady,
1274 	 * therefore it is impossible for `mapping' to be NULL.
1275 	 */
1276 	if (mapping && mapping != mc)
1277 		/* ought to be unmanaged pages */
1278 		goto out;
1279 
1280 	/* directly return for shortlived page as well */
1281 	if (z_erofs_is_shortlived_page(page))
1282 		goto out;
1283 
1284 	lock_page(page);
1285 
1286 	/* only true if page reclaim goes wrong, should never happen */
1287 	DBG_BUGON(justfound && PagePrivate(page));
1288 
1289 	/* the page is still in manage cache */
1290 	if (page->mapping == mc) {
1291 		WRITE_ONCE(pcl->compressed_pages[nr], page);
1292 
1293 		ClearPageError(page);
1294 		if (!PagePrivate(page)) {
1295 			/*
1296 			 * impossible to be !PagePrivate(page) for
1297 			 * the current restriction as well if
1298 			 * the page is already in compressed_pages[].
1299 			 */
1300 			DBG_BUGON(!justfound);
1301 
1302 			justfound = 0;
1303 			set_page_private(page, (unsigned long)pcl);
1304 			SetPagePrivate(page);
1305 		}
1306 
1307 		/* no need to submit io if it is already up-to-date */
1308 		if (PageUptodate(page)) {
1309 			unlock_page(page);
1310 			page = NULL;
1311 		}
1312 		goto out;
1313 	}
1314 
1315 	/*
1316 	 * the managed page has been truncated, it's unsafe to
1317 	 * reuse this one, let's allocate a new cache-managed page.
1318 	 */
1319 	DBG_BUGON(page->mapping);
1320 	DBG_BUGON(!justfound);
1321 
1322 	tocache = true;
1323 	unlock_page(page);
1324 	put_page(page);
1325 out_allocpage:
1326 	page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1327 	if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1328 		list_add(&page->lru, pagepool);
1329 		cond_resched();
1330 		goto repeat;
1331 	}
1332 out_tocache:
1333 	if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1334 		/* turn into temporary page if fails (1 ref) */
1335 		set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1336 		goto out;
1337 	}
1338 	attach_page_private(page, pcl);
1339 	/* drop a refcount added by allocpage (then we have 2 refs here) */
1340 	put_page(page);
1341 
1342 out:	/* the only exit (for tracing and debugging) */
1343 	return page;
1344 }
1345 
1346 static struct z_erofs_decompressqueue *
jobqueue_init(struct super_block * sb,struct z_erofs_decompressqueue * fgq,bool * fg)1347 jobqueue_init(struct super_block *sb,
1348 	      struct z_erofs_decompressqueue *fgq, bool *fg)
1349 {
1350 	struct z_erofs_decompressqueue *q;
1351 
1352 	if (fg && !*fg) {
1353 		q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1354 		if (!q) {
1355 			*fg = true;
1356 			goto fg_out;
1357 		}
1358 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1359 		kthread_init_work(&q->u.kthread_work,
1360 				  z_erofs_decompressqueue_kthread_work);
1361 #else
1362 		INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1363 #endif
1364 	} else {
1365 fg_out:
1366 		q = fgq;
1367 		init_completion(&fgq->u.done);
1368 		atomic_set(&fgq->pending_bios, 0);
1369 	}
1370 	q->sb = sb;
1371 	q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1372 	return q;
1373 }
1374 
1375 /* define decompression jobqueue types */
1376 enum {
1377 	JQ_BYPASS,
1378 	JQ_SUBMIT,
1379 	NR_JOBQUEUES,
1380 };
1381 
jobqueueset_init(struct super_block * sb,struct z_erofs_decompressqueue * q[],struct z_erofs_decompressqueue * fgq,bool * fg)1382 static void *jobqueueset_init(struct super_block *sb,
1383 			      struct z_erofs_decompressqueue *q[],
1384 			      struct z_erofs_decompressqueue *fgq, bool *fg)
1385 {
1386 	/*
1387 	 * if managed cache is enabled, bypass jobqueue is needed,
1388 	 * no need to read from device for all pclusters in this queue.
1389 	 */
1390 	q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1391 	q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1392 
1393 	return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1394 }
1395 
move_to_bypass_jobqueue(struct z_erofs_pcluster * pcl,z_erofs_next_pcluster_t qtail[],z_erofs_next_pcluster_t owned_head)1396 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1397 				    z_erofs_next_pcluster_t qtail[],
1398 				    z_erofs_next_pcluster_t owned_head)
1399 {
1400 	z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1401 	z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1402 
1403 	DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1404 	if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1405 		owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1406 
1407 	WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1408 
1409 	WRITE_ONCE(*submit_qtail, owned_head);
1410 	WRITE_ONCE(*bypass_qtail, &pcl->next);
1411 
1412 	qtail[JQ_BYPASS] = &pcl->next;
1413 }
1414 
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)1415 static void z_erofs_submit_queue(struct super_block *sb,
1416 				 struct z_erofs_decompress_frontend *f,
1417 				 struct list_head *pagepool,
1418 				 struct z_erofs_decompressqueue *fgq,
1419 				 bool *force_fg)
1420 {
1421 	struct erofs_sb_info *const sbi = EROFS_SB(sb);
1422 	z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1423 	struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1424 	void *bi_private;
1425 	z_erofs_next_pcluster_t owned_head = f->clt.owned_head;
1426 	/* since bio will be NULL, no need to initialize last_index */
1427 	pgoff_t last_index;
1428 	unsigned int nr_bios = 0;
1429 	struct bio *bio = NULL;
1430 
1431 	bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1432 	qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1433 	qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1434 
1435 	/* by default, all need io submission */
1436 	q[JQ_SUBMIT]->head = owned_head;
1437 
1438 	do {
1439 		struct z_erofs_pcluster *pcl;
1440 		pgoff_t cur, end;
1441 		unsigned int i = 0;
1442 		bool bypass = true;
1443 
1444 		/* no possible 'owned_head' equals the following */
1445 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1446 		DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1447 
1448 		pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1449 
1450 		cur = pcl->obj.index;
1451 		end = cur + pcl->pclusterpages;
1452 
1453 		/* close the main owned chain at first */
1454 		owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1455 				     Z_EROFS_PCLUSTER_TAIL_CLOSED);
1456 
1457 		do {
1458 			struct page *page;
1459 
1460 			page = pickup_page_for_submission(pcl, i++, pagepool,
1461 							  MNGD_MAPPING(sbi),
1462 							  GFP_NOFS);
1463 			if (!page)
1464 				continue;
1465 
1466 			if (bio && cur != last_index + 1) {
1467 submit_bio_retry:
1468 				submit_bio(bio);
1469 				bio = NULL;
1470 			}
1471 
1472 			if (!bio) {
1473 				bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1474 
1475 				bio->bi_end_io = z_erofs_decompressqueue_endio;
1476 				bio_set_dev(bio, sb->s_bdev);
1477 				bio->bi_iter.bi_sector = (sector_t)cur <<
1478 					LOG_SECTORS_PER_BLOCK;
1479 				bio->bi_private = bi_private;
1480 				bio->bi_opf = REQ_OP_READ;
1481 				if (f->readahead)
1482 					bio->bi_opf |= REQ_RAHEAD;
1483 				++nr_bios;
1484 			}
1485 
1486 			if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1487 				goto submit_bio_retry;
1488 
1489 			last_index = cur;
1490 			bypass = false;
1491 		} while (++cur < end);
1492 
1493 		if (!bypass)
1494 			qtail[JQ_SUBMIT] = &pcl->next;
1495 		else
1496 			move_to_bypass_jobqueue(pcl, qtail, owned_head);
1497 	} while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1498 
1499 	if (bio)
1500 		submit_bio(bio);
1501 
1502 	/*
1503 	 * although background is preferred, no one is pending for submission.
1504 	 * don't issue decompression but drop it directly instead.
1505 	 */
1506 	if (!*force_fg && !nr_bios) {
1507 		kvfree(q[JQ_SUBMIT]);
1508 		return;
1509 	}
1510 	z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1511 }
1512 
z_erofs_runqueue(struct super_block * sb,struct z_erofs_decompress_frontend * f,struct list_head * pagepool,bool force_fg)1513 static void z_erofs_runqueue(struct super_block *sb,
1514 			     struct z_erofs_decompress_frontend *f,
1515 			     struct list_head *pagepool, bool force_fg)
1516 {
1517 	struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1518 
1519 	if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL)
1520 		return;
1521 	z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1522 
1523 	/* handle bypass queue (no i/o pclusters) immediately */
1524 	z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1525 
1526 	if (!force_fg)
1527 		return;
1528 
1529 	/* wait until all bios are completed */
1530 	wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1531 
1532 	/* handle synchronous decompress queue in the caller context */
1533 	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1534 }
1535 
z_erofs_readpage(struct file * file,struct page * page)1536 static int z_erofs_readpage(struct file *file, struct page *page)
1537 {
1538 	struct inode *const inode = page->mapping->host;
1539 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1540 	int err;
1541 	LIST_HEAD(pagepool);
1542 
1543 	trace_erofs_readpage(page, false);
1544 
1545 	f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1546 
1547 	err = z_erofs_do_read_page(&f, page, &pagepool);
1548 	(void)z_erofs_collector_end(&f.clt);
1549 
1550 	/* if some compressed cluster ready, need submit them anyway */
1551 	z_erofs_runqueue(inode->i_sb, &f, &pagepool, true);
1552 
1553 	if (err)
1554 		erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1555 
1556 	if (f.map.mpage)
1557 		put_page(f.map.mpage);
1558 
1559 	/* clean up the remaining free pages */
1560 	put_pages_list(&pagepool);
1561 	return err;
1562 }
1563 
z_erofs_readahead(struct readahead_control * rac)1564 static void z_erofs_readahead(struct readahead_control *rac)
1565 {
1566 	struct inode *const inode = rac->mapping->host;
1567 	struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1568 
1569 	unsigned int nr_pages = readahead_count(rac);
1570 	bool sync = (sbi->ctx.readahead_sync_decompress &&
1571 			nr_pages <= sbi->ctx.max_sync_decompress_pages);
1572 	struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1573 	struct page *page, *head = NULL;
1574 	LIST_HEAD(pagepool);
1575 
1576 	trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1577 
1578 	f.readahead = true;
1579 	f.headoffset = readahead_pos(rac);
1580 
1581 	while ((page = readahead_page(rac))) {
1582 		prefetchw(&page->flags);
1583 
1584 		/*
1585 		 * A pure asynchronous readahead is indicated if
1586 		 * a PG_readahead marked page is hitted at first.
1587 		 * Let's also do asynchronous decompression for this case.
1588 		 */
1589 		sync &= !(PageReadahead(page) && !head);
1590 
1591 		set_page_private(page, (unsigned long)head);
1592 		head = page;
1593 	}
1594 
1595 	while (head) {
1596 		struct page *page = head;
1597 		int err;
1598 
1599 		/* traversal in reverse order */
1600 		head = (void *)page_private(page);
1601 
1602 		err = z_erofs_do_read_page(&f, page, &pagepool);
1603 		if (err)
1604 			erofs_err(inode->i_sb,
1605 				  "readahead error at page %lu @ nid %llu",
1606 				  page->index, EROFS_I(inode)->nid);
1607 		put_page(page);
1608 	}
1609 
1610 	(void)z_erofs_collector_end(&f.clt);
1611 
1612 	z_erofs_runqueue(inode->i_sb, &f, &pagepool, sync);
1613 
1614 	if (f.map.mpage)
1615 		put_page(f.map.mpage);
1616 
1617 	/* clean up the remaining free pages */
1618 	put_pages_list(&pagepool);
1619 }
1620 
1621 const struct address_space_operations z_erofs_aops = {
1622 	.readpage = z_erofs_readpage,
1623 	.readahead = z_erofs_readahead,
1624 };
1625 
1626