Whamcloud - gitweb
e0c9901f4dc5fc758a505e9e25518be8aa40455e
[fs/lustre-release.git] / lustre / osc / osc_cache.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  *
28  */
29 /*
30  * This file is part of Lustre, http://www.lustre.org/
31  * Lustre is a trademark of Sun Microsystems, Inc.
32  *
33  * osc cache management.
34  *
35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_OSC
39
40 #include <lustre_osc.h>
41 #include <lustre_dlm.h>
42
43 #include "osc_internal.h"
44
45 static int extent_debug; /* set it to be true for more debug */
46
47 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
48 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
49                            enum osc_extent_state state);
50 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
51                               struct osc_async_page *oap, int sent, int rc);
52 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
53                           int cmd);
54 static int osc_refresh_count(const struct lu_env *env,
55                              struct osc_async_page *oap, int cmd);
56 static int osc_io_unplug_async(const struct lu_env *env,
57                                struct client_obd *cli, struct osc_object *osc);
58 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
59                            unsigned int lost_grant, unsigned int dirty_grant);
60
61 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
62                                   const char *func, int line);
63 #define osc_extent_tree_dump(mask, obj) \
64         osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
65
66 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
67                                 unsigned int unused);
68
69 /** \addtogroup osc
70  *  @{
71  */
72
73 /* ------------------ osc extent ------------------ */
74 static inline char *ext_flags(struct osc_extent *ext, char *flags)
75 {
76         char *buf = flags;
77         *buf++ = ext->oe_rw ? 'r' : 'w';
78         if (!RB_EMPTY_NODE(&ext->oe_node))
79                 *buf++ = 'i';
80         if (ext->oe_sync)
81                 *buf++ = 'S';
82         if (ext->oe_srvlock)
83                 *buf++ = 's';
84         if (ext->oe_hp)
85                 *buf++ = 'h';
86         if (ext->oe_urgent)
87                 *buf++ = 'u';
88         if (ext->oe_memalloc)
89                 *buf++ = 'm';
90         if (ext->oe_trunc_pending)
91                 *buf++ = 't';
92         if (ext->oe_fsync_wait)
93                 *buf++ = 'Y';
94         *buf = 0;
95         return flags;
96 }
97
98 #define EXTSTR       "[%lu -> %lu/%lu]"
99 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
100 static const char *oes_strings[] = {
101         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
102
103 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
104         static struct cfs_debug_limit_state cdls;                             \
105         struct osc_extent *__ext = (extent);                                  \
106         char __buf[16];                                                       \
107                                                                               \
108         __CDEBUG_WITH_LOC(file, func, line, mask, &cdls,                      \
109                 "extent %p@{" EXTSTR ", "                                     \
110                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
111                 /* ----- extent part 0 ----- */                               \
112                 __ext, EXTPARA(__ext),                                        \
113                 /* ----- part 1 ----- */                                      \
114                 kref_read(&__ext->oe_refc),                                   \
115                 atomic_read(&__ext->oe_users),                                \
116                 list_empty_marker(&__ext->oe_link),                           \
117                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
118                 __ext->oe_obj,                                                \
119                 /* ----- part 2 ----- */                                      \
120                 __ext->oe_grants, __ext->oe_nr_pages,                         \
121                 list_empty_marker(&__ext->oe_pages),                          \
122                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
123                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
124                 /* ----- part 4 ----- */                                      \
125                 ## __VA_ARGS__);                                              \
126         if (mask == D_ERROR && __ext->oe_dlmlock != NULL)                     \
127                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
128         else                                                                  \
129                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
130 } while (0)
131
132 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...)                            \
133         OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,          \
134                                  mask, ext, fmt, ## __VA_ARGS__)
135
136 #undef EASSERTF
137 #define EASSERTF(expr, ext, fmt, args...) do {                          \
138         if (!(expr)) {                                                  \
139                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
140                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
141                 LASSERT(expr);                                          \
142         }                                                               \
143 } while (0)
144
145 #undef EASSERT
146 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
147
148 static inline struct osc_extent *rb_extent(struct rb_node *n)
149 {
150         return rb_entry_safe(n, struct osc_extent, oe_node);
151 }
152
153 static inline struct osc_extent *next_extent(struct osc_extent *ext)
154 {
155         if (ext == NULL)
156                 return NULL;
157
158         LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
159         return rb_extent(rb_next(&ext->oe_node));
160 }
161
162 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
163 {
164         if (ext == NULL)
165                 return NULL;
166
167         LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
168         return rb_extent(rb_prev(&ext->oe_node));
169 }
170
171 static inline struct osc_extent *first_extent(struct osc_object *obj)
172 {
173         return rb_extent(rb_first(&obj->oo_root));
174 }
175
176 /* object must be locked by caller. */
177 static int osc_extent_sanity_check0(struct osc_extent *ext,
178                                     const char *func, const int line)
179 {
180         struct osc_object *obj = ext->oe_obj;
181         struct osc_async_page *oap;
182         size_t page_count;
183         int rc = 0;
184
185         assert_osc_object_is_locked(obj);
186
187         if (ext->oe_state >= OES_STATE_MAX)
188                 GOTO(out, rc = 10);
189
190         if (kref_read(&ext->oe_refc) <= 0)
191                 GOTO(out, rc = 20);
192
193         if (kref_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
194                 GOTO(out, rc = 30);
195
196         switch (ext->oe_state) {
197         case OES_INV:
198                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
199                         GOTO(out, rc = 35);
200                 GOTO(out, rc = 0);
201                 break;
202         case OES_ACTIVE:
203                 if (atomic_read(&ext->oe_users) == 0)
204                         GOTO(out, rc = 40);
205                 if (ext->oe_hp)
206                         GOTO(out, rc = 50);
207                 if (ext->oe_fsync_wait && !ext->oe_urgent)
208                         GOTO(out, rc = 55);
209                 break;
210         case OES_CACHE:
211                 if (ext->oe_grants == 0)
212                         GOTO(out, rc = 60);
213                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
214                         GOTO(out, rc = 65);
215                 /* fallthrough */
216         default:
217                 if (atomic_read(&ext->oe_users) > 0)
218                         GOTO(out, rc = 70);
219         }
220
221         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
222                 GOTO(out, rc = 80);
223
224         if (ext->oe_sync && ext->oe_grants > 0)
225                 GOTO(out, rc = 90);
226
227         if (ext->oe_dlmlock != NULL &&
228             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
229             !ldlm_is_failed(ext->oe_dlmlock)) {
230                 struct ldlm_extent *extent;
231
232                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
233                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
234                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
235                         GOTO(out, rc = 100);
236
237                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
238                         GOTO(out, rc = 102);
239         }
240
241         if (ext->oe_nr_pages > ext->oe_mppr)
242                 GOTO(out, rc = 105);
243
244         /* Do not verify page list if extent is in RPC. This is because an
245          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
246         if (ext->oe_state > OES_CACHE)
247                 GOTO(out, rc = 0);
248
249         if (!extent_debug)
250                 GOTO(out, rc = 0);
251
252         page_count = 0;
253         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
254                 pgoff_t index = osc_index(oap2osc(oap));
255                 ++page_count;
256                 if (index > ext->oe_end || index < ext->oe_start)
257                         GOTO(out, rc = 110);
258         }
259         if (page_count != ext->oe_nr_pages)
260                 GOTO(out, rc = 120);
261
262 out:
263         if (rc != 0)
264                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
265                                          "sanity check %p failed: rc = %d\n",
266                                          ext, rc);
267         return rc;
268 }
269
270 #define sanity_check_nolock(ext) \
271         osc_extent_sanity_check0(ext, __func__, __LINE__)
272
273 #define sanity_check(ext) ({                                                   \
274         int __res;                                                             \
275         osc_object_lock((ext)->oe_obj);                                        \
276         __res = sanity_check_nolock(ext);                                      \
277         osc_object_unlock((ext)->oe_obj);                                      \
278         __res;                                                                 \
279 })
280
281 static inline bool
282 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
283 {
284         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
285 }
286
287 /**
288  * sanity check - to make sure there is no overlapped extent in the tree.
289  */
290 static int osc_extent_is_overlapped(struct osc_object *obj,
291                                     struct osc_extent *ext)
292 {
293         struct osc_extent *tmp;
294
295         assert_osc_object_is_locked(obj);
296
297         if (!extent_debug)
298                 return 0;
299
300         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
301                 if (tmp == ext)
302                         continue;
303                 if (overlapped(tmp, ext))
304                         return 1;
305         }
306         return 0;
307 }
308
309 static void osc_extent_state_set(struct osc_extent *ext, int state)
310 {
311         assert_osc_object_is_locked(ext->oe_obj);
312         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
313
314         /* Never try to sanity check a state changing extent :-) */
315         /* LASSERT(sanity_check_nolock(ext) == 0); */
316
317         /* TODO: validate the state machine */
318         smp_store_release(&ext->oe_state, state);
319         wake_up(&ext->oe_waitq);
320 }
321
322 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
323 {
324         struct osc_extent *ext;
325
326         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
327         if (ext == NULL)
328                 return NULL;
329
330         RB_CLEAR_NODE(&ext->oe_node);
331         ext->oe_obj = obj;
332         cl_object_get(osc2cl(obj));
333         kref_init(&ext->oe_refc);
334         atomic_set(&ext->oe_users, 0);
335         INIT_LIST_HEAD(&ext->oe_link);
336         ext->oe_state = OES_INV;
337         INIT_LIST_HEAD(&ext->oe_pages);
338         init_waitqueue_head(&ext->oe_waitq);
339         ext->oe_dlmlock = NULL;
340
341         return ext;
342 }
343
344 static void osc_extent_free(struct kref *kref)
345 {
346         struct osc_extent *ext = container_of(kref, struct osc_extent,
347                                               oe_refc);
348
349         LASSERT(list_empty(&ext->oe_link));
350         LASSERT(atomic_read(&ext->oe_users) == 0);
351         LASSERT(ext->oe_state == OES_INV);
352         LASSERT(RB_EMPTY_NODE(&ext->oe_node));
353
354         if (ext->oe_dlmlock) {
355                 lu_ref_del(&ext->oe_dlmlock->l_reference,
356                            "osc_extent", ext);
357                 LDLM_LOCK_PUT(ext->oe_dlmlock);
358                 ext->oe_dlmlock = NULL;
359         }
360 #if 0
361         /* If/When cl_object_put drops the need for 'env',
362          * this code can be enabled, and matching code in
363          * osc_extent_put removed.
364          */
365         cl_object_put(osc2cl(ext->oe_obj));
366
367         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
368 #endif
369 }
370
371 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
372 {
373         LASSERT(kref_read(&ext->oe_refc) >= 0);
374         kref_get(&ext->oe_refc);
375         return ext;
376 }
377
378 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
379 {
380         LASSERT(kref_read(&ext->oe_refc) > 0);
381         if (kref_put(&ext->oe_refc, osc_extent_free)) {
382                 /* This should be in osc_extent_free(), but
383                  * while we need to pass 'env' it cannot be.
384                  */
385                 cl_object_put(env, osc2cl(ext->oe_obj));
386
387                 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
388         }
389 }
390
391 /**
392  * osc_extent_put_trust() is a special version of osc_extent_put() when
393  * it's known that the caller is not the last user. This is to address the
394  * problem of lacking of lu_env ;-).
395  */
396 static void osc_extent_put_trust(struct osc_extent *ext)
397 {
398         LASSERT(kref_read(&ext->oe_refc) > 1);
399         assert_osc_object_is_locked(ext->oe_obj);
400         osc_extent_put(NULL, ext);
401 }
402
403 /**
404  * Return the extent which includes pgoff @index, or return the greatest
405  * previous extent in the tree.
406  */
407 static struct osc_extent *osc_extent_search(struct osc_object *obj,
408                                             pgoff_t index)
409 {
410         struct rb_node    *n = obj->oo_root.rb_node;
411         struct osc_extent *tmp, *p = NULL;
412
413         assert_osc_object_is_locked(obj);
414         while (n != NULL) {
415                 tmp = rb_extent(n);
416                 if (index < tmp->oe_start) {
417                         n = n->rb_left;
418                 } else if (index > tmp->oe_end) {
419                         p = rb_extent(n);
420                         n = n->rb_right;
421                 } else {
422                         return tmp;
423                 }
424         }
425         return p;
426 }
427
428 /*
429  * Return the extent covering @index, otherwise return NULL.
430  * caller must have held object lock.
431  */
432 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
433                                             pgoff_t index)
434 {
435         struct osc_extent *ext;
436
437         ext = osc_extent_search(obj, index);
438         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
439                 return osc_extent_get(ext);
440         return NULL;
441 }
442
443 /* caller must have held object lock. */
444 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
445 {
446         struct rb_node   **n      = &obj->oo_root.rb_node;
447         struct rb_node    *parent = NULL;
448         struct osc_extent *tmp;
449
450         LASSERT(RB_EMPTY_NODE(&ext->oe_node));
451         LASSERT(ext->oe_obj == obj);
452         assert_osc_object_is_locked(obj);
453         while (*n != NULL) {
454                 tmp = rb_extent(*n);
455                 parent = *n;
456
457                 if (ext->oe_end < tmp->oe_start)
458                         n = &(*n)->rb_left;
459                 else if (ext->oe_start > tmp->oe_end)
460                         n = &(*n)->rb_right;
461                 else
462                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
463         }
464         rb_link_node(&ext->oe_node, parent, n);
465         rb_insert_color(&ext->oe_node, &obj->oo_root);
466         osc_extent_get(ext);
467 }
468
469 /* caller must have held object lock. */
470 static void osc_extent_erase(struct osc_extent *ext)
471 {
472         struct osc_object *obj = ext->oe_obj;
473         assert_osc_object_is_locked(obj);
474         if (!RB_EMPTY_NODE(&ext->oe_node)) {
475                 rb_erase(&ext->oe_node, &obj->oo_root);
476                 RB_CLEAR_NODE(&ext->oe_node);
477                 /* rbtree held a refcount */
478                 osc_extent_put_trust(ext);
479         }
480 }
481
482 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
483 {
484         struct osc_object *obj = ext->oe_obj;
485
486         assert_osc_object_is_locked(obj);
487         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
488         if (ext->oe_state == OES_CACHE) {
489                 osc_extent_state_set(ext, OES_ACTIVE);
490                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
491         }
492         atomic_inc(&ext->oe_users);
493         list_del_init(&ext->oe_link);
494         return osc_extent_get(ext);
495 }
496
497 static void __osc_extent_remove(struct osc_extent *ext)
498 {
499         assert_osc_object_is_locked(ext->oe_obj);
500         LASSERT(list_empty(&ext->oe_pages));
501         osc_extent_erase(ext);
502         list_del_init(&ext->oe_link);
503         osc_extent_state_set(ext, OES_INV);
504         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
505 }
506
507 static void osc_extent_remove(struct osc_extent *ext)
508 {
509         struct osc_object *obj = ext->oe_obj;
510
511         osc_object_lock(obj);
512         __osc_extent_remove(ext);
513         osc_object_unlock(obj);
514 }
515
516 /**
517  * This function is used to merge extents to get better performance. It checks
518  * if @cur and @victim are contiguous at block level.
519  */
520 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
521                             struct osc_extent *victim)
522 {
523         struct osc_object       *obj = cur->oe_obj;
524         struct client_obd       *cli = osc_cli(obj);
525         pgoff_t                  chunk_start;
526         pgoff_t                  chunk_end;
527         int                      ppc_bits;
528
529         LASSERT(cur->oe_state == OES_CACHE);
530         assert_osc_object_is_locked(obj);
531         if (victim == NULL)
532                 return -EINVAL;
533
534         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
535                 return -EBUSY;
536
537         if (cur->oe_max_end != victim->oe_max_end)
538                 return -ERANGE;
539
540         /*
541          * In the rare case max_pages_per_rpc (mppr) is changed, don't
542          * merge extents until after old ones have been sent, or the
543          * "extents are aligned to RPCs" checks are unhappy.
544          */
545         if (cur->oe_mppr != victim->oe_mppr)
546                 return -ERANGE;
547
548         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
549         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
550         chunk_start = cur->oe_start >> ppc_bits;
551         chunk_end   = cur->oe_end   >> ppc_bits;
552         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
553             chunk_end + 1 != victim->oe_start >> ppc_bits)
554                 return -ERANGE;
555
556         /* overall extent size should not exceed the max supported limit
557          * reported by the server */
558         if (cur->oe_end - cur->oe_start + 1 +
559             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
560                 return -ERANGE;
561
562         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
563
564         cur->oe_start     = min(cur->oe_start, victim->oe_start);
565         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
566         /* per-extent tax should be accounted only once for the whole extent */
567         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
568         cur->oe_nr_pages += victim->oe_nr_pages;
569         /* only the following bits are needed to merge */
570         cur->oe_urgent   |= victim->oe_urgent;
571         cur->oe_memalloc |= victim->oe_memalloc;
572         list_splice_init(&victim->oe_pages, &cur->oe_pages);
573         victim->oe_nr_pages = 0;
574
575         osc_extent_get(victim);
576         __osc_extent_remove(victim);
577         osc_extent_put(env, victim);
578
579         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
580         return 0;
581 }
582
583 /**
584  * Drop user count of osc_extent, and unplug IO asynchronously.
585  */
586 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
587 {
588         struct osc_object *obj = ext->oe_obj;
589         struct client_obd *cli = osc_cli(obj);
590         int rc = 0;
591         ENTRY;
592
593         LASSERT(atomic_read(&ext->oe_users) > 0);
594         LASSERT(sanity_check(ext) == 0);
595         LASSERT(ext->oe_grants > 0);
596
597         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
598                 LASSERT(ext->oe_state == OES_ACTIVE);
599                 if (ext->oe_trunc_pending) {
600                         /* a truncate process is waiting for this extent.
601                          * This may happen due to a race, check
602                          * osc_cache_truncate_start(). */
603                         osc_extent_state_set(ext, OES_TRUNC);
604                         ext->oe_trunc_pending = 0;
605                         osc_object_unlock(obj);
606                 } else {
607                         int grant = 0;
608
609                         osc_extent_state_set(ext, OES_CACHE);
610                         osc_update_pending(obj, OBD_BRW_WRITE,
611                                            ext->oe_nr_pages);
612
613                         /* try to merge the previous and next extent. */
614                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
615                                 grant += cli->cl_grant_extent_tax;
616                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
617                                 grant += cli->cl_grant_extent_tax;
618
619                         if (ext->oe_hp)
620                                 list_move_tail(&ext->oe_link,
621                                                &obj->oo_hp_exts);
622                         else if (ext->oe_urgent)
623                                 list_move_tail(&ext->oe_link,
624                                                &obj->oo_urgent_exts);
625                         else if (ext->oe_nr_pages == ext->oe_mppr) {
626                                 list_move_tail(&ext->oe_link,
627                                                &obj->oo_full_exts);
628                         }
629                         osc_object_unlock(obj);
630                         if (grant > 0)
631                                 osc_unreserve_grant(cli, 0, grant);
632                 }
633
634                 osc_io_unplug_async(env, cli, obj);
635         }
636         osc_extent_put(env, ext);
637         RETURN(rc);
638 }
639
640 /**
641  * Find or create an extent which includes @index, core function to manage
642  * extent tree.
643  */
644 static struct osc_extent *osc_extent_find(const struct lu_env *env,
645                                           struct osc_object *obj, pgoff_t index,
646                                           unsigned int *grants)
647 {
648         struct client_obd *cli = osc_cli(obj);
649         struct osc_lock   *olck;
650         struct cl_lock_descr *descr;
651         struct osc_extent *cur;
652         struct osc_extent *ext;
653         struct osc_extent *conflict = NULL;
654         struct osc_extent *found = NULL;
655         pgoff_t    chunk;
656         pgoff_t    max_end;
657         unsigned int max_pages; /* max_pages_per_rpc */
658         unsigned int chunksize;
659         int        ppc_bits; /* pages per chunk bits */
660         pgoff_t    chunk_mask;
661         int        rc;
662         ENTRY;
663
664         cur = osc_extent_alloc(obj);
665         if (cur == NULL)
666                 RETURN(ERR_PTR(-ENOMEM));
667
668         olck = osc_env_io(env)->oi_write_osclock;
669         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
670         LASSERT(olck->ols_state == OLS_GRANTED);
671
672         descr = &olck->ols_cl.cls_lock->cll_descr;
673         LASSERT(descr->cld_mode >= CLM_WRITE);
674
675         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
676                  "chunkbits: %u\n", cli->cl_chunkbits);
677         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
678         chunk_mask = ~((1 << ppc_bits) - 1);
679         chunksize  = 1 << cli->cl_chunkbits;
680         chunk      = index >> ppc_bits;
681
682         /* align end to RPC edge. */
683         max_pages = cli->cl_max_pages_per_rpc;
684         if ((max_pages & ~chunk_mask) != 0) {
685                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
686                        max_pages, cli->cl_chunkbits, chunk_mask);
687                 RETURN(ERR_PTR(-EINVAL));
688         }
689         max_end = index - (index % max_pages) + max_pages - 1;
690         max_end = min_t(pgoff_t, max_end, descr->cld_end);
691
692         /* initialize new extent by parameters so far */
693         cur->oe_max_end = max_end;
694         cur->oe_start   = index & chunk_mask;
695         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
696         if (cur->oe_start < descr->cld_start)
697                 cur->oe_start = descr->cld_start;
698         if (cur->oe_end > max_end)
699                 cur->oe_end = max_end;
700         cur->oe_grants  = chunksize + cli->cl_grant_extent_tax;
701         cur->oe_mppr    = max_pages;
702         if (olck->ols_dlmlock != NULL) {
703                 LASSERT(olck->ols_hold);
704                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
705                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
706         }
707
708         /* grants has been allocated by caller */
709         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
710                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
711         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
712                  EXTPARA(cur));
713
714 restart:
715         osc_object_lock(obj);
716         ext = osc_extent_search(obj, cur->oe_start);
717         if (!ext)
718                 ext = first_extent(obj);
719         for (; ext; ext = next_extent(ext)) {
720                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
721                 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
722
723                 LASSERT(sanity_check_nolock(ext) == 0);
724                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
725                         break;
726
727                 /* if covering by different locks, no chance to match */
728                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
729                         EASSERTF(!overlapped(ext, cur), ext,
730                                  EXTSTR"\n", EXTPARA(cur));
731
732                         continue;
733                 }
734
735                 /* discontiguous chunks? */
736                 if (chunk + 1 < ext_chk_start)
737                         continue;
738
739                 /* ok, from now on, ext and cur have these attrs:
740                  * 1. covered by the same lock
741                  * 2. contiguous at chunk level or overlapping. */
742
743                 if (overlapped(ext, cur)) {
744                         /* cur is the minimum unit, so overlapping means
745                          * full contain. */
746                         EASSERTF((ext->oe_start <= cur->oe_start &&
747                                   ext->oe_end >= cur->oe_end),
748                                  ext, EXTSTR"\n", EXTPARA(cur));
749
750                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
751                                 /* for simplicity, we wait for this extent to
752                                  * finish before going forward. */
753                                 conflict = osc_extent_get(ext);
754                                 break;
755                         }
756
757                         found = osc_extent_hold(ext);
758                         break;
759                 }
760
761                 /* non-overlapped extent */
762                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
763                         /* we can't do anything for a non OES_CACHE extent, or
764                          * if there is someone waiting for this extent to be
765                          * flushed, try next one. */
766                         continue;
767
768                 if (osc_extent_merge(env, ext, cur) == 0) {
769                         LASSERT(*grants >= chunksize);
770                         *grants -= chunksize;
771                         found = osc_extent_hold(ext);
772
773                         /*
774                          * Try to merge with the next one too because we
775                          * might have just filled in a gap.
776                          */
777                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
778                                 /* we can save extent tax from next extent */
779                                 *grants += cli->cl_grant_extent_tax;
780
781                         break;
782                 }
783         }
784
785         osc_extent_tree_dump(D_CACHE, obj);
786         if (found != NULL) {
787                 LASSERT(conflict == NULL);
788                 if (!IS_ERR(found)) {
789                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
790                         OSC_EXTENT_DUMP(D_CACHE, found,
791                                         "found caching ext for %lu.\n", index);
792                 }
793         } else if (conflict == NULL) {
794                 /* create a new extent */
795                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
796                 LASSERT(*grants >= cur->oe_grants);
797                 *grants -= cur->oe_grants;
798
799                 cur->oe_state = OES_CACHE;
800                 found = osc_extent_hold(cur);
801                 osc_extent_insert(obj, cur);
802                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
803                                 index, descr->cld_end);
804         }
805         osc_object_unlock(obj);
806
807         if (conflict != NULL) {
808                 LASSERT(found == NULL);
809
810                 /* waiting for IO to finish. Please notice that it's impossible
811                  * to be an OES_TRUNC extent. */
812                 rc = osc_extent_wait(env, conflict, OES_INV);
813                 osc_extent_put(env, conflict);
814                 conflict = NULL;
815                 if (rc < 0)
816                         GOTO(out, found = ERR_PTR(rc));
817
818                 goto restart;
819         }
820         EXIT;
821
822 out:
823         osc_extent_put(env, cur);
824         return found;
825 }
826
827 /**
828  * Called when IO is finished to an extent.
829  */
830 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
831                       int sent, int rc)
832 {
833         struct client_obd *cli = osc_cli(ext->oe_obj);
834         struct osc_async_page *oap;
835         struct osc_async_page *tmp;
836         int nr_pages = ext->oe_nr_pages;
837         int lost_grant = 0;
838         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
839         loff_t last_off = 0;
840         int last_count = -1;
841         ENTRY;
842
843         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
844
845         ext->oe_rc = rc ?: ext->oe_nr_pages;
846         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
847
848         osc_lru_add_batch(cli, &ext->oe_pages);
849         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
850                                      oap_pending_item) {
851                 list_del_init(&oap->oap_rpc_item);
852                 list_del_init(&oap->oap_pending_item);
853                 if (last_off <= oap->oap_obj_off) {
854                         last_off = oap->oap_obj_off;
855                         last_count = oap->oap_count;
856                 }
857
858                 --ext->oe_nr_pages;
859                 osc_ap_completion(env, cli, oap, sent, rc);
860         }
861         EASSERT(ext->oe_nr_pages == 0, ext);
862
863         if (!sent) {
864                 lost_grant = ext->oe_grants;
865         } else if (blocksize < PAGE_SIZE &&
866                    last_count != PAGE_SIZE) {
867                 /* For short writes we shouldn't count parts of pages that
868                  * span a whole chunk on the OST side, or our accounting goes
869                  * wrong.  Should match the code in filter_grant_check. */
870                 int offset = last_off & ~PAGE_MASK;
871                 int count = last_count + (offset & (blocksize - 1));
872                 int end = (offset + last_count) & (blocksize - 1);
873                 if (end)
874                         count += blocksize - end;
875
876                 lost_grant = PAGE_SIZE - count;
877         }
878         if (ext->oe_grants > 0)
879                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
880
881         osc_extent_remove(ext);
882         /* put the refcount for RPC */
883         osc_extent_put(env, ext);
884         RETURN(0);
885 }
886
887 /**
888  * Wait for the extent's state to become @state.
889  */
890 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
891                            enum osc_extent_state state)
892 {
893         struct osc_object *obj = ext->oe_obj;
894         int rc = 0;
895         ENTRY;
896
897         osc_object_lock(obj);
898         LASSERT(sanity_check_nolock(ext) == 0);
899         /* `Kick' this extent only if the caller is waiting for it to be
900          * written out. */
901         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
902                 if (ext->oe_state == OES_ACTIVE) {
903                         ext->oe_urgent = 1;
904                 } else if (ext->oe_state == OES_CACHE) {
905                         ext->oe_urgent = 1;
906                         osc_extent_hold(ext);
907                         rc = 1;
908                 }
909         }
910         osc_object_unlock(obj);
911         if (rc == 1)
912                 osc_extent_release(env, ext);
913
914         /* wait for the extent until its state becomes @state */
915         rc = wait_event_idle_timeout(ext->oe_waitq,
916                                      smp_load_acquire(&ext->oe_state) == state,
917                                      cfs_time_seconds(600));
918         if (rc == 0) {
919                 OSC_EXTENT_DUMP(D_ERROR, ext,
920                         "%s: wait ext to %u timedout, recovery in progress?\n",
921                         cli_name(osc_cli(obj)), state);
922
923                 wait_event_idle(ext->oe_waitq,
924                                 smp_load_acquire(&ext->oe_state) == state);
925         }
926         if (ext->oe_rc < 0)
927                 rc = ext->oe_rc;
928         else
929                 rc = 0;
930         RETURN(rc);
931 }
932
933 /**
934  * Discard pages with index greater than @size. If @ext is overlapped with
935  * @size, then partial truncate happens.
936  */
937 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
938                                 bool partial)
939 {
940         struct lu_env         *env;
941         struct cl_io          *io;
942         struct osc_object     *obj = ext->oe_obj;
943         struct client_obd     *cli = osc_cli(obj);
944         struct osc_async_page *oap;
945         struct osc_async_page *tmp;
946         struct pagevec        *pvec;
947         int                    pages_in_chunk = 0;
948         int                    ppc_bits    = cli->cl_chunkbits -
949                                              PAGE_SHIFT;
950         __u64                  trunc_chunk = trunc_index >> ppc_bits;
951         int                    grants   = 0;
952         int                    nr_pages = 0;
953         int                    rc       = 0;
954         __u16                  refcheck;
955         ENTRY;
956
957         LASSERT(sanity_check(ext) == 0);
958         LASSERT(ext->oe_state == OES_TRUNC);
959         LASSERT(!ext->oe_urgent);
960
961         /* Request new lu_env.
962          * We can't use that env from osc_cache_truncate_start() because
963          * it's from lov_io_sub and not fully initialized. */
964         env = cl_env_get(&refcheck);
965         if (IS_ERR(env))
966                 RETURN(PTR_ERR(env));
967
968         io  = osc_env_thread_io(env);
969         io->ci_obj = cl_object_top(osc2cl(obj));
970         io->ci_ignore_layout = 1;
971         pvec = &osc_env_info(env)->oti_pagevec;
972         ll_pagevec_init(pvec, 0);
973         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
974         if (rc < 0)
975                 GOTO(out, rc);
976
977         /* discard all pages with index greater than trunc_index */
978         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
979                                      oap_pending_item) {
980                 pgoff_t index = osc_index(oap2osc(oap));
981                 struct cl_page  *page = oap2cl_page(oap);
982
983                 LASSERT(list_empty(&oap->oap_rpc_item));
984
985                 /* only discard the pages with their index greater than
986                  * trunc_index, and ... */
987                 if (index < trunc_index ||
988                     (index == trunc_index && partial)) {
989                         /* accounting how many pages remaining in the chunk
990                          * so that we can calculate grants correctly. */
991                         if (index >> ppc_bits == trunc_chunk)
992                                 ++pages_in_chunk;
993                         continue;
994                 }
995
996                 list_del_init(&oap->oap_pending_item);
997
998                 cl_page_get(page);
999                 lu_ref_add(&page->cp_reference, "truncate", current);
1000
1001                 if (cl_page_own(env, io, page) == 0) {
1002                         cl_page_discard(env, io, page);
1003                         cl_page_disown(env, io, page);
1004                 } else {
1005                         LASSERT(page->cp_state == CPS_FREEING);
1006                         LASSERT(0);
1007                 }
1008
1009                 lu_ref_del(&page->cp_reference, "truncate", current);
1010                 cl_pagevec_put(env, page, pvec);
1011
1012                 --ext->oe_nr_pages;
1013                 ++nr_pages;
1014         }
1015         pagevec_release(pvec);
1016
1017         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1018                       ext->oe_nr_pages == 0),
1019                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1020
1021         osc_object_lock(obj);
1022         if (ext->oe_nr_pages == 0) {
1023                 LASSERT(pages_in_chunk == 0);
1024                 grants = ext->oe_grants;
1025                 ext->oe_grants = 0;
1026         } else { /* calculate how many grants we can free */
1027                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1028                 pgoff_t last_index;
1029
1030
1031                 /* if there is no pages in this chunk, we can also free grants
1032                  * for the last chunk */
1033                 if (pages_in_chunk == 0) {
1034                         /* if this is the 1st chunk and no pages in this chunk,
1035                          * ext->oe_nr_pages must be zero, so we should be in
1036                          * the other if-clause. */
1037                         LASSERT(trunc_chunk > 0);
1038                         --trunc_chunk;
1039                         ++chunks;
1040                 }
1041
1042                 /* this is what we can free from this extent */
1043                 grants          = chunks << cli->cl_chunkbits;
1044                 ext->oe_grants -= grants;
1045                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1046                 ext->oe_end     = min(last_index, ext->oe_max_end);
1047                 LASSERT(ext->oe_end >= ext->oe_start);
1048                 LASSERT(ext->oe_grants > 0);
1049         }
1050         osc_object_unlock(obj);
1051
1052         if (grants > 0 || nr_pages > 0)
1053                 osc_free_grant(cli, nr_pages, grants, grants);
1054
1055 out:
1056         cl_io_fini(env, io);
1057         cl_env_put(env, &refcheck);
1058         RETURN(rc);
1059 }
1060
1061 /**
1062  * This function is used to make the extent prepared for transfer.
1063  * A race with flusing page - ll_writepage() has to be handled cautiously.
1064  */
1065 static int osc_extent_make_ready(const struct lu_env *env,
1066                                  struct osc_extent *ext)
1067 {
1068         struct osc_async_page *oap;
1069         struct osc_async_page *last = NULL;
1070         struct osc_object *obj = ext->oe_obj;
1071         unsigned int page_count = 0;
1072         int rc;
1073         ENTRY;
1074
1075         /* we're going to grab page lock, so object lock must not be taken. */
1076         LASSERT(sanity_check(ext) == 0);
1077         /* in locking state, any process should not touch this extent. */
1078         EASSERT(ext->oe_state == OES_LOCKING, ext);
1079         EASSERT(ext->oe_owner != NULL, ext);
1080
1081         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1082
1083         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1084                 ++page_count;
1085                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1086                         last = oap;
1087
1088                 /* checking ASYNC_READY is race safe */
1089                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1090                         continue;
1091
1092                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1093                 switch (rc) {
1094                 case 0:
1095                         spin_lock(&oap->oap_lock);
1096                         oap->oap_async_flags |= ASYNC_READY;
1097                         spin_unlock(&oap->oap_lock);
1098                         break;
1099                 case -EALREADY:
1100                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1101                         break;
1102                 default:
1103                         LASSERTF(0, "unknown return code: %d\n", rc);
1104                 }
1105         }
1106
1107         LASSERT(page_count == ext->oe_nr_pages);
1108         LASSERT(last != NULL);
1109         /* the last page is the only one we need to refresh its count by
1110          * the size of file. */
1111         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1112                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1113                 LASSERT(last_oap_count > 0);
1114                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1115                 last->oap_count = last_oap_count;
1116                 spin_lock(&last->oap_lock);
1117                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1118                 spin_unlock(&last->oap_lock);
1119         }
1120
1121         /* for the rest of pages, we don't need to call osf_refresh_count()
1122          * because it's known they are not the last page */
1123         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1124                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1125                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1126                         spin_lock(&oap->oap_lock);
1127                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1128                         spin_unlock(&oap->oap_lock);
1129                 }
1130         }
1131
1132         osc_object_lock(obj);
1133         osc_extent_state_set(ext, OES_RPC);
1134         osc_object_unlock(obj);
1135         /* get a refcount for RPC. */
1136         osc_extent_get(ext);
1137
1138         RETURN(0);
1139 }
1140
1141 /**
1142  * Quick and simple version of osc_extent_find(). This function is frequently
1143  * called to expand the extent for the same IO. To expand the extent, the
1144  * page index must be in the same or next chunk of ext->oe_end.
1145  */
1146 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1147                              unsigned int *grants)
1148 {
1149         struct osc_object *obj = ext->oe_obj;
1150         struct client_obd *cli = osc_cli(obj);
1151         struct osc_extent *next;
1152         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1153         pgoff_t chunk = index >> ppc_bits;
1154         pgoff_t end_chunk;
1155         pgoff_t end_index;
1156         unsigned int chunksize = 1 << cli->cl_chunkbits;
1157         int rc = 0;
1158         ENTRY;
1159
1160         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1161         osc_object_lock(obj);
1162         LASSERT(sanity_check_nolock(ext) == 0);
1163         end_chunk = ext->oe_end >> ppc_bits;
1164         if (chunk > end_chunk + 1)
1165                 GOTO(out, rc = -ERANGE);
1166
1167         if (end_chunk >= chunk)
1168                 GOTO(out, rc = 0);
1169
1170         LASSERT(end_chunk + 1 == chunk);
1171
1172         /* try to expand this extent to cover @index */
1173         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1174
1175         /* don't go over the maximum extent size reported by server */
1176         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1177                 GOTO(out, rc = -ERANGE);
1178
1179         next = next_extent(ext);
1180         if (next != NULL && next->oe_start <= end_index)
1181                 /* complex mode - overlapped with the next extent,
1182                  * this case will be handled by osc_extent_find() */
1183                 GOTO(out, rc = -EAGAIN);
1184
1185         ext->oe_end = end_index;
1186         ext->oe_grants += chunksize;
1187         LASSERT(*grants >= chunksize);
1188         *grants -= chunksize;
1189         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1190                  "overlapped after expanding for %lu.\n", index);
1191         EXIT;
1192
1193 out:
1194         osc_object_unlock(obj);
1195         RETURN(rc);
1196 }
1197
1198 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1199                                   const char *func, int line)
1200 {
1201         struct osc_extent *ext;
1202         int cnt;
1203
1204         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1205                 return;
1206
1207         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1208                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1209
1210         /* osc_object_lock(obj); */
1211         cnt = 1;
1212         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1213                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1214
1215         cnt = 1;
1216         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1217                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1218
1219         cnt = 1;
1220         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1221                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1222
1223         cnt = 1;
1224         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1225                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1226         /* osc_object_unlock(obj); */
1227 }
1228
1229 /* ------------------ osc extent end ------------------ */
1230
1231 static inline int osc_is_ready(struct osc_object *osc)
1232 {
1233         return !list_empty(&osc->oo_ready_item) ||
1234                !list_empty(&osc->oo_hp_ready_item);
1235 }
1236
1237 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1238         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1239                (OSC), osc_is_ready(OSC),                                       \
1240                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1241                list_empty_marker(&(OSC)->oo_ready_item),                       \
1242                atomic_read(&(OSC)->oo_nr_writes),                              \
1243                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1244                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1245                atomic_read(&(OSC)->oo_nr_reads),                               \
1246                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1247                ##args)
1248
1249 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1250                           int cmd)
1251 {
1252         struct osc_page *opg  = oap2osc_page(oap);
1253         struct cl_page  *page = oap2cl_page(oap);
1254         int result;
1255
1256         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1257
1258         ENTRY;
1259         result = cl_page_make_ready(env, page, CRT_WRITE);
1260         if (result == 0)
1261                 opg->ops_submit_time = ktime_get();
1262         RETURN(result);
1263 }
1264
1265 static int osc_refresh_count(const struct lu_env *env,
1266                              struct osc_async_page *oap, int cmd)
1267 {
1268         struct osc_page  *opg = oap2osc_page(oap);
1269         pgoff_t index = osc_index(oap2osc(oap));
1270         struct cl_object *obj;
1271         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1272         int result;
1273         loff_t kms;
1274
1275         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1276         LASSERT(!(cmd & OBD_BRW_READ));
1277         LASSERT(opg != NULL);
1278         obj = opg->ops_cl.cpl_obj;
1279
1280         cl_object_attr_lock(obj);
1281         result = cl_object_attr_get(env, obj, attr);
1282         cl_object_attr_unlock(obj);
1283         if (result < 0)
1284                 return result;
1285         kms = attr->cat_kms;
1286         if (cl_offset(obj, index) >= kms)
1287                 /* catch race with truncate */
1288                 return 0;
1289         else if (cl_offset(obj, index + 1) > kms)
1290                 /* catch sub-page write at end of file */
1291                 return kms & ~PAGE_MASK;
1292         else
1293                 return PAGE_SIZE;
1294 }
1295
1296 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1297                           int cmd, int rc)
1298 {
1299         struct osc_page   *opg  = oap2osc_page(oap);
1300         struct cl_page    *page = oap2cl_page(oap);
1301         enum cl_req_type   crt;
1302         int srvlock;
1303
1304         ENTRY;
1305
1306         cmd &= ~OBD_BRW_NOQUOTA;
1307         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1308                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1309         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1310                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1311         LASSERT(opg->ops_transfer_pinned);
1312
1313         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1314         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1315         opg->ops_transfer_pinned = 0;
1316
1317         opg->ops_submit_time = ktime_set(0, 0);
1318         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1319
1320         /* statistic */
1321         if (rc == 0 && srvlock) {
1322                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1323                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1324                 size_t bytes = oap->oap_count;
1325
1326                 if (crt == CRT_READ)
1327                         stats->os_lockless_reads += bytes;
1328                 else
1329                         stats->os_lockless_writes += bytes;
1330         }
1331
1332         /*
1333          * This has to be the last operation with the page, as locks are
1334          * released in cl_page_completion() and nothing except for the
1335          * reference counter protects page from concurrent reclaim.
1336          */
1337         lu_ref_del(&page->cp_reference, "transfer", page);
1338
1339         cl_page_completion(env, page, crt, rc);
1340         cl_page_put(env, page);
1341
1342         RETURN(0);
1343 }
1344
1345 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1346         struct client_obd *__tmp = (cli);                               \
1347         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1348                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1349                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1350                "left: %ld, waiters: %d }" fmt "\n",                     \
1351                cli_name(__tmp),                                         \
1352                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1353                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1354                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1355                __tmp->cl_dirty_grant,                                   \
1356                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1357                atomic_long_read(&__tmp->cl_lru_in_list),                \
1358                atomic_long_read(&__tmp->cl_lru_busy),                   \
1359                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1360 } while (0)
1361
1362 /* caller must hold loi_list_lock */
1363 static void osc_consume_write_grant(struct client_obd *cli,
1364                                     struct brw_page *pga)
1365 {
1366         assert_spin_locked(&cli->cl_loi_list_lock);
1367         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1368         cli->cl_dirty_pages++;
1369         pga->flag |= OBD_BRW_FROM_GRANT;
1370         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1371                PAGE_SIZE, pga, pga->pg);
1372         osc_update_next_shrink(cli);
1373 }
1374
1375 /* the companion to osc_consume_write_grant, called when a brw has completed.
1376  * must be called with the loi lock held. */
1377 static void osc_release_write_grant(struct client_obd *cli,
1378                                     struct brw_page *pga)
1379 {
1380         ENTRY;
1381
1382         assert_spin_locked(&cli->cl_loi_list_lock);
1383         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1384                 EXIT;
1385                 return;
1386         }
1387
1388         pga->flag &= ~OBD_BRW_FROM_GRANT;
1389         atomic_long_dec(&obd_dirty_pages);
1390         cli->cl_dirty_pages--;
1391         EXIT;
1392 }
1393
1394 /**
1395  * To avoid sleeping with object lock held, it's good for us allocate enough
1396  * grants before entering into critical section.
1397  *
1398  * client_obd_list_lock held by caller
1399  */
1400 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1401 {
1402         int rc = -EDQUOT;
1403
1404         if (cli->cl_avail_grant >= bytes) {
1405                 cli->cl_avail_grant    -= bytes;
1406                 cli->cl_reserved_grant += bytes;
1407                 rc = 0;
1408         }
1409         return rc;
1410 }
1411
1412 static void __osc_unreserve_grant(struct client_obd *cli,
1413                                   unsigned int reserved, unsigned int unused)
1414 {
1415         /* it's quite normal for us to get more grant than reserved.
1416          * Thinking about a case that two extents merged by adding a new
1417          * chunk, we can save one extent tax. If extent tax is greater than
1418          * one chunk, we can save more grant by adding a new chunk */
1419         cli->cl_reserved_grant -= reserved;
1420         if (unused > reserved) {
1421                 cli->cl_avail_grant += reserved;
1422                 cli->cl_lost_grant  += unused - reserved;
1423                 cli->cl_dirty_grant -= unused - reserved;
1424         } else {
1425                 cli->cl_avail_grant += unused;
1426                 cli->cl_dirty_grant += reserved - unused;
1427         }
1428 }
1429
1430 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1431                                        unsigned int reserved,
1432                                        unsigned int unused)
1433 {
1434         __osc_unreserve_grant(cli, reserved, unused);
1435         if (unused > 0)
1436                 osc_wake_cache_waiters(cli);
1437 }
1438
1439 static void osc_unreserve_grant(struct client_obd *cli,
1440                                 unsigned int reserved, unsigned int unused)
1441 {
1442         spin_lock(&cli->cl_loi_list_lock);
1443         osc_unreserve_grant_nolock(cli, reserved, unused);
1444         spin_unlock(&cli->cl_loi_list_lock);
1445 }
1446
1447 /**
1448  * Free grant after IO is finished or canceled.
1449  *
1450  * @lost_grant is used to remember how many grants we have allocated but not
1451  * used, we should return these grants to OST. There're two cases where grants
1452  * can be lost:
1453  * 1. truncate;
1454  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1455  *    written. In this case OST may use less chunks to serve this partial
1456  *    write. OSTs don't actually know the page size on the client side. so
1457  *    clients have to calculate lost grant by the blocksize on the OST.
1458  *    See filter_grant_check() for details.
1459  */
1460 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1461                            unsigned int lost_grant, unsigned int dirty_grant)
1462 {
1463         unsigned long grant;
1464
1465         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1466
1467         spin_lock(&cli->cl_loi_list_lock);
1468         atomic_long_sub(nr_pages, &obd_dirty_pages);
1469         cli->cl_dirty_pages -= nr_pages;
1470         cli->cl_lost_grant += lost_grant;
1471         cli->cl_dirty_grant -= dirty_grant;
1472         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1473                 /* borrow some grant from truncate to avoid the case that
1474                  * truncate uses up all avail grant */
1475                 cli->cl_lost_grant -= grant;
1476                 cli->cl_avail_grant += grant;
1477         }
1478         osc_wake_cache_waiters(cli);
1479         spin_unlock(&cli->cl_loi_list_lock);
1480         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1481                lost_grant, cli->cl_lost_grant,
1482                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1483                cli->cl_dirty_grant);
1484 }
1485
1486 /**
1487  * The companion to osc_enter_cache(), called when @oap is no longer part of
1488  * the dirty accounting due to error.
1489  */
1490 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1491 {
1492         spin_lock(&cli->cl_loi_list_lock);
1493         osc_release_write_grant(cli, &oap->oap_brw_page);
1494         spin_unlock(&cli->cl_loi_list_lock);
1495 }
1496
1497 /**
1498  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1499  * is available.
1500  */
1501 static int osc_enter_cache_try(struct client_obd *cli,
1502                                struct osc_async_page *oap,
1503                                int bytes)
1504 {
1505         int rc;
1506
1507         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1508
1509         rc = osc_reserve_grant(cli, bytes);
1510         if (rc < 0)
1511                 return 0;
1512
1513         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1514                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1515                     obd_max_dirty_pages) {
1516                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1517                         rc = 1;
1518                         goto out;
1519                 } else
1520                         atomic_long_dec(&obd_dirty_pages);
1521         }
1522         __osc_unreserve_grant(cli, bytes, bytes);
1523
1524 out:
1525         return rc;
1526 }
1527
1528 /* Following two inlines exist to pass code fragments
1529  * to wait_event_idle_exclusive_timeout_cmd().  Passing
1530  * code fragments as macro args can look confusing, so
1531  * we provide inlines to encapsulate them.
1532  */
1533 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1534                                          struct client_obd *cli,
1535                                          struct osc_async_page *oap)
1536 {
1537         spin_unlock(&cli->cl_loi_list_lock);
1538         osc_io_unplug_async(env, cli, NULL);
1539         CDEBUG(D_CACHE,
1540                "%s: sleeping for cache space for %p\n",
1541                cli_name(cli), oap);
1542 }
1543
1544 static inline void cli_lock_after_unplug(struct client_obd *cli)
1545 {
1546         spin_lock(&cli->cl_loi_list_lock);
1547 }
1548 /**
1549  * The main entry to reserve dirty page accounting. Usually the grant reserved
1550  * in this function will be freed in bulk in osc_free_grant() unless it fails
1551  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1552  *
1553  * The process will be put into sleep if it's already run out of grant.
1554  */
1555 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1556                            struct osc_async_page *oap, int bytes)
1557 {
1558         struct osc_object *osc = oap->oap_obj;
1559         struct lov_oinfo *loi = osc->oo_oinfo;
1560         int rc = -EDQUOT;
1561         int remain;
1562         bool entered = false;
1563         /* We cannot wait for a long time here since we are holding ldlm lock
1564          * across the actual IO. If no requests complete fast (e.g. due to
1565          * overloaded OST that takes a long time to process everything, we'd
1566          * get evicted if we wait for a normal obd_timeout or some such.
1567          * So we try to wait half the time it would take the client to be
1568          * evicted by server which is half obd_timeout when AT is off
1569          * or at least ldlm_enqueue_min with AT on.
1570          * See LU-13131 */
1571         unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1572                                                           ldlm_enqueue_min / 2);
1573
1574         ENTRY;
1575
1576         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1577
1578         spin_lock(&cli->cl_loi_list_lock);
1579
1580         /* force the caller to try sync io.  this can jump the list
1581          * of queued writes and create a discontiguous rpc stream */
1582         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1583             cli->cl_dirty_max_pages == 0 ||
1584             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1585                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1586                 GOTO(out, rc = -EDQUOT);
1587         }
1588
1589         /*
1590          * We can wait here for two reasons: too many dirty pages in cache, or
1591          * run out of grants. In both cases we should write dirty pages out.
1592          * Adding a cache waiter will trigger urgent write-out no matter what
1593          * RPC size will be.
1594          * The exiting condition (other than success) is no avail grants
1595          * and no dirty pages caching, that really means there is no space
1596          * on the OST.
1597          */
1598         remain = wait_event_idle_exclusive_timeout_cmd(
1599                 cli->cl_cache_waiters,
1600                 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1601                 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1602                 timeout,
1603                 cli_unlock_and_unplug(env, cli, oap),
1604                 cli_lock_after_unplug(cli));
1605
1606         if (entered) {
1607                 if (remain == timeout)
1608                         OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1609                 else
1610                         OSC_DUMP_GRANT(D_CACHE, cli,
1611                                        "finally got grant space\n");
1612                 wake_up(&cli->cl_cache_waiters);
1613                 rc = 0;
1614         } else if (remain == 0) {
1615                 OSC_DUMP_GRANT(D_CACHE, cli,
1616                                "timeout, fall back to sync i/o\n");
1617                 osc_extent_tree_dump(D_CACHE, osc);
1618                 /* fall back to synchronous I/O */
1619         } else {
1620                 OSC_DUMP_GRANT(D_CACHE, cli,
1621                                "no grant space, fall back to sync i/o\n");
1622                 wake_up_all(&cli->cl_cache_waiters);
1623         }
1624         EXIT;
1625 out:
1626         spin_unlock(&cli->cl_loi_list_lock);
1627         RETURN(rc);
1628 }
1629
1630 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1631 {
1632         int hprpc = !!list_empty(&osc->oo_hp_exts);
1633         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1634 }
1635
1636 /* This maintains the lists of pending pages to read/write for a given object
1637  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1638  * to quickly find objects that are ready to send an RPC. */
1639 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1640                          int cmd)
1641 {
1642         int invalid_import = 0;
1643         ENTRY;
1644
1645         /* if we have an invalid import we want to drain the queued pages
1646          * by forcing them through rpcs that immediately fail and complete
1647          * the pages.  recovery relies on this to empty the queued pages
1648          * before canceling the locks and evicting down the llite pages */
1649         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1650                 invalid_import = 1;
1651
1652         if (cmd & OBD_BRW_WRITE) {
1653                 if (atomic_read(&osc->oo_nr_writes) == 0)
1654                         RETURN(0);
1655                 if (invalid_import) {
1656                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1657                         RETURN(1);
1658                 }
1659                 if (!list_empty(&osc->oo_hp_exts)) {
1660                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1661                         RETURN(1);
1662                 }
1663                 if (!list_empty(&osc->oo_urgent_exts)) {
1664                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1665                         RETURN(1);
1666                 }
1667                 /* trigger a write rpc stream as long as there are dirtiers
1668                  * waiting for space.  as they're waiting, they're not going to
1669                  * create more pages to coalesce with what's waiting..
1670                  */
1671                 if (waitqueue_active(&cli->cl_cache_waiters)) {
1672                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1673                         RETURN(1);
1674                 }
1675                 if (!list_empty(&osc->oo_full_exts)) {
1676                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1677                         RETURN(1);
1678                 }
1679         } else {
1680                 if (atomic_read(&osc->oo_nr_reads) == 0)
1681                         RETURN(0);
1682                 if (invalid_import) {
1683                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1684                         RETURN(1);
1685                 }
1686                 /* all read are urgent. */
1687                 if (!list_empty(&osc->oo_reading_exts))
1688                         RETURN(1);
1689         }
1690
1691         RETURN(0);
1692 }
1693
1694 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1695 {
1696         struct client_obd *cli = osc_cli(obj);
1697         if (cmd & OBD_BRW_WRITE) {
1698                 atomic_add(delta, &obj->oo_nr_writes);
1699                 atomic_add(delta, &cli->cl_pending_w_pages);
1700                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1701         } else {
1702                 atomic_add(delta, &obj->oo_nr_reads);
1703                 atomic_add(delta, &cli->cl_pending_r_pages);
1704                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1705         }
1706         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1707 }
1708
1709 static int osc_makes_hprpc(struct osc_object *obj)
1710 {
1711         return !list_empty(&obj->oo_hp_exts);
1712 }
1713
1714 static void on_list(struct list_head *item, struct list_head *list,
1715                     int should_be_on)
1716 {
1717         if (list_empty(item) && should_be_on)
1718                 list_add_tail(item, list);
1719         else if (!list_empty(item) && !should_be_on)
1720                 list_del_init(item);
1721 }
1722
1723 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1724  * can find pages to build into rpcs quickly */
1725 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1726 {
1727         if (osc_makes_hprpc(osc)) {
1728                 /* HP rpc */
1729                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1730                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1731         } else {
1732                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1733                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1734                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1735                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1736         }
1737
1738         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1739                 atomic_read(&osc->oo_nr_writes) > 0);
1740
1741         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1742                 atomic_read(&osc->oo_nr_reads) > 0);
1743
1744         return osc_is_ready(osc);
1745 }
1746
1747 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1748 {
1749         int is_ready;
1750
1751         spin_lock(&cli->cl_loi_list_lock);
1752         is_ready = __osc_list_maint(cli, osc);
1753         spin_unlock(&cli->cl_loi_list_lock);
1754
1755         return is_ready;
1756 }
1757
1758 /* this is trying to propogate async writeback errors back up to the
1759  * application.  As an async write fails we record the error code for later if
1760  * the app does an fsync.  As long as errors persist we force future rpcs to be
1761  * sync so that the app can get a sync error and break the cycle of queueing
1762  * pages for which writeback will fail. */
1763 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1764                            int rc)
1765 {
1766         if (rc) {
1767                 if (!ar->ar_rc)
1768                         ar->ar_rc = rc;
1769
1770                 ar->ar_force_sync = 1;
1771                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1772                 return;
1773
1774         }
1775
1776         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1777                 ar->ar_force_sync = 0;
1778 }
1779
1780 /* this must be called holding the loi list lock to give coverage to exit_cache,
1781  * async_flag maintenance, and oap_request */
1782 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1783                               struct osc_async_page *oap, int sent, int rc)
1784 {
1785         struct osc_object *osc = oap->oap_obj;
1786         struct lov_oinfo  *loi = osc->oo_oinfo;
1787         __u64 xid = 0;
1788
1789         ENTRY;
1790         if (oap->oap_request != NULL) {
1791                 xid = ptlrpc_req_xid(oap->oap_request);
1792                 ptlrpc_req_finished(oap->oap_request);
1793                 oap->oap_request = NULL;
1794         }
1795
1796         /* As the transfer for this page is being done, clear the flags */
1797         spin_lock(&oap->oap_lock);
1798         oap->oap_async_flags = 0;
1799         spin_unlock(&oap->oap_lock);
1800
1801         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1802                 spin_lock(&cli->cl_loi_list_lock);
1803                 osc_process_ar(&cli->cl_ar, xid, rc);
1804                 osc_process_ar(&loi->loi_ar, xid, rc);
1805                 spin_unlock(&cli->cl_loi_list_lock);
1806         }
1807
1808         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1809         if (rc)
1810                 CERROR("completion on oap %p obj %p returns %d.\n",
1811                        oap, osc, rc);
1812
1813         EXIT;
1814 }
1815
1816 struct extent_rpc_data {
1817         struct list_head        *erd_rpc_list;
1818         unsigned int            erd_page_count;
1819         unsigned int            erd_max_pages;
1820         unsigned int            erd_max_chunks;
1821         unsigned int            erd_max_extents;
1822 };
1823
1824 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1825 {
1826         struct client_obd *cli = osc_cli(ext->oe_obj);
1827         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1828
1829         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1830 }
1831
1832 static inline bool
1833 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1834 {
1835         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1836                 return false;
1837
1838         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1839                 return false;
1840
1841         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1842                 return false;
1843
1844         if (!ext->oe_grants != !in_rpc->oe_grants)
1845                 return false;
1846
1847         if (ext->oe_dio != in_rpc->oe_dio)
1848                 return false;
1849
1850         /* It's possible to have overlap on DIO */
1851         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1852                 return false;
1853
1854         return true;
1855 }
1856
1857 /**
1858  * Try to add extent to one RPC. We need to think about the following things:
1859  * - # of pages must not be over max_pages_per_rpc
1860  * - extent must be compatible with previous ones
1861  */
1862 static int try_to_add_extent_for_io(struct client_obd *cli,
1863                                     struct osc_extent *ext,
1864                                     struct extent_rpc_data *data)
1865 {
1866         struct osc_extent *tmp;
1867         unsigned int chunk_count;
1868         ENTRY;
1869
1870         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1871                 ext);
1872         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1873
1874         if (data->erd_max_extents == 0)
1875                 RETURN(0);
1876
1877         chunk_count = osc_extent_chunks(ext);
1878         EASSERTF(data->erd_page_count != 0 ||
1879                  chunk_count <= data->erd_max_chunks, ext,
1880                  "The first extent to be fit in a RPC contains %u chunks, "
1881                  "which is over the limit %u.\n", chunk_count,
1882                  data->erd_max_chunks);
1883         if (chunk_count > data->erd_max_chunks)
1884                 RETURN(0);
1885
1886         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1887         EASSERTF(data->erd_page_count != 0 ||
1888                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1889                 "The first extent to be fit in a RPC contains %u pages, "
1890                 "which is over the limit %u.\n", ext->oe_nr_pages,
1891                 data->erd_max_pages);
1892         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1893                 RETURN(0);
1894
1895         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1896                 EASSERT(tmp->oe_owner == current, tmp);
1897
1898                 if (!can_merge(ext, tmp))
1899                         RETURN(0);
1900         }
1901
1902         data->erd_max_extents--;
1903         data->erd_max_chunks -= chunk_count;
1904         data->erd_page_count += ext->oe_nr_pages;
1905         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1906         ext->oe_owner = current;
1907         RETURN(1);
1908 }
1909
1910 /**
1911  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1912  * get_write_extent() takes all appropriate extents in atomic.
1913  *
1914  * The following policy is used to collect extents for IO:
1915  * 1. Add as many HP extents as possible;
1916  * 2. Add the first urgent extent in urgent extent list and take it out of
1917  *    urgent list;
1918  * 3. Add subsequent extents of this urgent extent;
1919  * 4. If urgent list is not empty, goto 2;
1920  * 5. Traverse the extent tree from the 1st extent;
1921  * 6. Above steps exit if there is no space in this RPC.
1922  */
1923 static unsigned int get_write_extents(struct osc_object *obj,
1924                                       struct list_head *rpclist)
1925 {
1926         struct client_obd *cli = osc_cli(obj);
1927         struct osc_extent *ext;
1928         struct extent_rpc_data data = {
1929                 .erd_rpc_list   = rpclist,
1930                 .erd_page_count = 0,
1931                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
1932                 .erd_max_chunks = osc_max_write_chunks(cli),
1933                 .erd_max_extents = 256,
1934         };
1935
1936         assert_osc_object_is_locked(obj);
1937         while (!list_empty(&obj->oo_hp_exts)) {
1938                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1939                                  oe_link);
1940                 if (!try_to_add_extent_for_io(cli, ext, &data))
1941                         return data.erd_page_count;
1942                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1943         }
1944         if (data.erd_page_count == data.erd_max_pages)
1945                 return data.erd_page_count;
1946
1947         while (!list_empty(&obj->oo_urgent_exts)) {
1948                 ext = list_entry(obj->oo_urgent_exts.next,
1949                                  struct osc_extent, oe_link);
1950                 if (!try_to_add_extent_for_io(cli, ext, &data))
1951                         return data.erd_page_count;
1952         }
1953         if (data.erd_page_count == data.erd_max_pages)
1954                 return data.erd_page_count;
1955
1956         /* One key difference between full extents and other extents: full
1957          * extents can usually only be added if the rpclist was empty, so if we
1958          * can't add one, we continue on to trying to add normal extents.  This
1959          * is so we don't miss adding extra extents to an RPC containing high
1960          * priority or urgent extents. */
1961         while (!list_empty(&obj->oo_full_exts)) {
1962                 ext = list_entry(obj->oo_full_exts.next,
1963                                  struct osc_extent, oe_link);
1964                 if (!try_to_add_extent_for_io(cli, ext, &data))
1965                         break;
1966         }
1967         if (data.erd_page_count == data.erd_max_pages)
1968                 return data.erd_page_count;
1969
1970         for (ext = first_extent(obj);
1971              ext;
1972              ext = next_extent(ext)) {
1973                 if ((ext->oe_state != OES_CACHE) ||
1974                     /* this extent may be already in current rpclist */
1975                     (!list_empty(&ext->oe_link) && ext->oe_owner))
1976                         continue;
1977
1978                 if (!try_to_add_extent_for_io(cli, ext, &data))
1979                         return data.erd_page_count;
1980         }
1981         return data.erd_page_count;
1982 }
1983
1984 static int
1985 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1986                    struct osc_object *osc)
1987 __must_hold(osc)
1988 {
1989         LIST_HEAD(rpclist);
1990         struct osc_extent *ext;
1991         struct osc_extent *tmp;
1992         struct osc_extent *first = NULL;
1993         unsigned int page_count = 0;
1994         int srvlock = 0;
1995         int rc = 0;
1996         ENTRY;
1997
1998         assert_osc_object_is_locked(osc);
1999
2000         page_count = get_write_extents(osc, &rpclist);
2001         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2002
2003         if (list_empty(&rpclist))
2004                 RETURN(0);
2005
2006         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2007
2008         list_for_each_entry(ext, &rpclist, oe_link) {
2009                 LASSERT(ext->oe_state == OES_CACHE ||
2010                         ext->oe_state == OES_LOCK_DONE);
2011                 if (ext->oe_state == OES_CACHE)
2012                         osc_extent_state_set(ext, OES_LOCKING);
2013                 else
2014                         osc_extent_state_set(ext, OES_RPC);
2015         }
2016
2017         /* we're going to grab page lock, so release object lock because
2018          * lock order is page lock -> object lock. */
2019         osc_object_unlock(osc);
2020
2021         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2022                 if (ext->oe_state == OES_LOCKING) {
2023                         rc = osc_extent_make_ready(env, ext);
2024                         if (unlikely(rc < 0)) {
2025                                 list_del_init(&ext->oe_link);
2026                                 osc_extent_finish(env, ext, 0, rc);
2027                                 continue;
2028                         }
2029                 }
2030                 if (first == NULL) {
2031                         first = ext;
2032                         srvlock = ext->oe_srvlock;
2033                 } else {
2034                         LASSERT(srvlock == ext->oe_srvlock);
2035                 }
2036         }
2037
2038         if (!list_empty(&rpclist)) {
2039                 LASSERT(page_count > 0);
2040                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2041                 LASSERT(list_empty(&rpclist));
2042         }
2043
2044         osc_object_lock(osc);
2045         RETURN(rc);
2046 }
2047
2048 /**
2049  * prepare pages for ASYNC io and put pages in send queue.
2050  *
2051  * \param cmd OBD_BRW_* macroses
2052  * \param lop pending pages
2053  *
2054  * \return zero if no page added to send queue.
2055  * \return 1 if pages successfully added to send queue.
2056  * \return negative on errors.
2057  */
2058 static int
2059 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2060                   struct osc_object *osc)
2061 __must_hold(osc)
2062 {
2063         struct osc_extent *ext;
2064         struct osc_extent *next;
2065         LIST_HEAD(rpclist);
2066         struct extent_rpc_data data = {
2067                 .erd_rpc_list   = &rpclist,
2068                 .erd_page_count = 0,
2069                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2070                 .erd_max_chunks = UINT_MAX,
2071                 .erd_max_extents = UINT_MAX,
2072         };
2073         int rc = 0;
2074         ENTRY;
2075
2076         assert_osc_object_is_locked(osc);
2077         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2078                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2079                 if (!try_to_add_extent_for_io(cli, ext, &data))
2080                         break;
2081                 osc_extent_state_set(ext, OES_RPC);
2082                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2083         }
2084         LASSERT(data.erd_page_count <= data.erd_max_pages);
2085
2086         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2087
2088         if (!list_empty(&rpclist)) {
2089                 osc_object_unlock(osc);
2090
2091                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2092                 LASSERT(list_empty(&rpclist));
2093
2094                 osc_object_lock(osc);
2095         }
2096         RETURN(rc);
2097 }
2098
2099 #define list_to_obj(list, item) ({                                            \
2100         struct list_head *__tmp = (list)->next;                               \
2101         list_del_init(__tmp);                                         \
2102         list_entry(__tmp, struct osc_object, oo_##item);                      \
2103 })
2104
2105 /* This is called by osc_check_rpcs() to find which objects have pages that
2106  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2107 static struct osc_object *osc_next_obj(struct client_obd *cli)
2108 {
2109         ENTRY;
2110
2111         /* First return objects that have blocked locks so that they
2112          * will be flushed quickly and other clients can get the lock,
2113          * then objects which have pages ready to be stuffed into RPCs */
2114         if (!list_empty(&cli->cl_loi_hp_ready_list))
2115                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2116         if (!list_empty(&cli->cl_loi_ready_list))
2117                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2118
2119         /* then if we have cache waiters, return all objects with queued
2120          * writes.  This is especially important when many small files
2121          * have filled up the cache and not been fired into rpcs because
2122          * they don't pass the nr_pending/object threshhold
2123          */
2124         if (waitqueue_active(&cli->cl_cache_waiters) &&
2125             !list_empty(&cli->cl_loi_write_list))
2126                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2127
2128         /* then return all queued objects when we have an invalid import
2129          * so that they get flushed */
2130         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2131                 if (!list_empty(&cli->cl_loi_write_list))
2132                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2133                                            write_item));
2134                 if (!list_empty(&cli->cl_loi_read_list))
2135                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2136                                            read_item));
2137         }
2138         RETURN(NULL);
2139 }
2140
2141 /* called with the loi list lock held */
2142 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2143 __must_hold(&cli->cl_loi_list_lock)
2144 {
2145         struct osc_object *osc;
2146         int rc = 0;
2147         ENTRY;
2148
2149         while ((osc = osc_next_obj(cli)) != NULL) {
2150                 struct cl_object *obj = osc2cl(osc);
2151                 struct lu_ref_link link;
2152
2153                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2154
2155                 /* even if we have reached our max in flight RPCs, we still
2156                  * allow all high-priority RPCs through to prevent their
2157                  * starvation and leading to server evicting us for not
2158                  * writing out pages in a timely manner LU-13131 */
2159                 if (osc_max_rpc_in_flight(cli, osc) &&
2160                     list_empty(&osc->oo_hp_exts)) {
2161                         __osc_list_maint(cli, osc);
2162                         break;
2163                 }
2164
2165                 cl_object_get(obj);
2166                 spin_unlock(&cli->cl_loi_list_lock);
2167                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2168
2169                 /* attempt some read/write balancing by alternating between
2170                  * reads and writes in an object.  The makes_rpc checks here
2171                  * would be redundant if we were getting read/write work items
2172                  * instead of objects.  we don't want send_oap_rpc to drain a
2173                  * partial read pending queue when we're given this object to
2174                  * do io on writes while there are cache waiters */
2175                 osc_object_lock(osc);
2176                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2177                         rc = osc_send_write_rpc(env, cli, osc);
2178                         if (rc < 0) {
2179                                 CERROR("Write request failed with %d\n", rc);
2180
2181                                 /* osc_send_write_rpc failed, mostly because of
2182                                  * memory pressure.
2183                                  *
2184                                  * It can't break here, because if:
2185                                  *  - a page was submitted by osc_io_submit, so
2186                                  *    page locked;
2187                                  *  - no request in flight
2188                                  *  - no subsequent request
2189                                  * The system will be in live-lock state,
2190                                  * because there is no chance to call
2191                                  * osc_io_unplug() and osc_check_rpcs() any
2192                                  * more. pdflush can't help in this case,
2193                                  * because it might be blocked at grabbing
2194                                  * the page lock as we mentioned.
2195                                  *
2196                                  * Anyway, continue to drain pages. */
2197                                 /* break; */
2198                         }
2199                 }
2200                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2201                         rc = osc_send_read_rpc(env, cli, osc);
2202                         if (rc < 0)
2203                                 CERROR("Read request failed with %d\n", rc);
2204                 }
2205                 osc_object_unlock(osc);
2206
2207                 osc_list_maint(cli, osc);
2208                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2209                 cl_object_put(env, obj);
2210
2211                 spin_lock(&cli->cl_loi_list_lock);
2212         }
2213 }
2214
2215 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2216                    struct osc_object *osc, int async)
2217 {
2218         int rc = 0;
2219
2220         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2221                 return 0;
2222
2223         if (!async) {
2224                 spin_lock(&cli->cl_loi_list_lock);
2225                 osc_check_rpcs(env, cli);
2226                 spin_unlock(&cli->cl_loi_list_lock);
2227         } else {
2228                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2229                 LASSERT(cli->cl_writeback_work != NULL);
2230                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2231         }
2232         return rc;
2233 }
2234 EXPORT_SYMBOL(osc_io_unplug0);
2235
2236 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2237                         struct page *page, loff_t offset)
2238 {
2239         struct obd_export     *exp = osc_export(osc);
2240         struct osc_async_page *oap = &ops->ops_oap;
2241         ENTRY;
2242
2243         if (!page)
2244                 return cfs_size_round(sizeof(*oap));
2245
2246         oap->oap_magic = OAP_MAGIC;
2247         oap->oap_cli = &exp->exp_obd->u.cli;
2248         oap->oap_obj = osc;
2249
2250         oap->oap_page = page;
2251         oap->oap_obj_off = offset;
2252         LASSERT(!(offset & ~PAGE_MASK));
2253
2254         INIT_LIST_HEAD(&oap->oap_pending_item);
2255         INIT_LIST_HEAD(&oap->oap_rpc_item);
2256
2257         spin_lock_init(&oap->oap_lock);
2258         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2259                oap, page, oap->oap_obj_off);
2260         RETURN(0);
2261 }
2262 EXPORT_SYMBOL(osc_prep_async_page);
2263
2264 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2265                        struct osc_page *ops, cl_commit_cbt cb)
2266 {
2267         struct osc_io *oio = osc_env_io(env);
2268         struct osc_extent     *ext = NULL;
2269         struct osc_async_page *oap = &ops->ops_oap;
2270         struct client_obd     *cli = oap->oap_cli;
2271         struct osc_object     *osc = oap->oap_obj;
2272         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2273         pgoff_t index;
2274         unsigned int tmp;
2275         unsigned int grants = 0;
2276         u32    brw_flags = OBD_BRW_ASYNC;
2277         int    cmd = OBD_BRW_WRITE;
2278         int    need_release = 0;
2279         int    rc = 0;
2280         ENTRY;
2281
2282         if (oap->oap_magic != OAP_MAGIC)
2283                 RETURN(-EINVAL);
2284
2285         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2286                 RETURN(-EIO);
2287
2288         if (!list_empty(&oap->oap_pending_item) ||
2289             !list_empty(&oap->oap_rpc_item))
2290                 RETURN(-EBUSY);
2291
2292         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2293         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2294         if (oio->oi_cap_sys_resource || io->ci_noquota) {
2295                 brw_flags |= OBD_BRW_NOQUOTA;
2296                 cmd |= OBD_BRW_NOQUOTA;
2297         }
2298
2299         /* check if the file's owner/group is over quota */
2300         if (!(cmd & OBD_BRW_NOQUOTA)) {
2301                 struct cl_object *obj;
2302                 struct cl_attr   *attr;
2303                 unsigned int qid[LL_MAXQUOTAS];
2304
2305                 obj = cl_object_top(&osc->oo_cl);
2306                 attr = &osc_env_info(env)->oti_attr;
2307
2308                 cl_object_attr_lock(obj);
2309                 rc = cl_object_attr_get(env, obj, attr);
2310                 cl_object_attr_unlock(obj);
2311
2312                 qid[USRQUOTA] = attr->cat_uid;
2313                 qid[GRPQUOTA] = attr->cat_gid;
2314                 qid[PRJQUOTA] = attr->cat_projid;
2315                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2316                         rc = -EDQUOT;
2317                 if (rc)
2318                         RETURN(rc);
2319         }
2320
2321         oap->oap_cmd = cmd;
2322         oap->oap_page_off = ops->ops_from;
2323         oap->oap_count = ops->ops_to - ops->ops_from;
2324         /* No need to hold a lock here,
2325          * since this page is not in any list yet. */
2326         oap->oap_async_flags = 0;
2327         oap->oap_brw_flags = brw_flags;
2328
2329         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2330                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2331
2332         index = osc_index(oap2osc(oap));
2333
2334         /* Add this page into extent by the following steps:
2335          * 1. if there exists an active extent for this IO, mostly this page
2336          *    can be added to the active extent and sometimes we need to
2337          *    expand extent to accomodate this page;
2338          * 2. otherwise, a new extent will be allocated. */
2339
2340         ext = oio->oi_active;
2341         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2342                 /* one chunk plus extent overhead must be enough to write this
2343                  * page */
2344                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2345                 if (ext->oe_end >= index)
2346                         grants = 0;
2347
2348                 /* it doesn't need any grant to dirty this page */
2349                 spin_lock(&cli->cl_loi_list_lock);
2350                 rc = osc_enter_cache_try(cli, oap, grants);
2351                 if (rc == 0) { /* try failed */
2352                         grants = 0;
2353                         need_release = 1;
2354                 } else if (ext->oe_end < index) {
2355                         tmp = grants;
2356                         /* try to expand this extent */
2357                         rc = osc_extent_expand(ext, index, &tmp);
2358                         if (rc < 0) {
2359                                 need_release = 1;
2360                                 /* don't free reserved grant */
2361                         } else {
2362                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2363                                                 "expanded for %lu.\n", index);
2364                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2365                                 grants = 0;
2366                         }
2367                 }
2368                 spin_unlock(&cli->cl_loi_list_lock);
2369                 rc = 0;
2370         } else if (ext != NULL) {
2371                 /* index is located outside of active extent */
2372                 need_release = 1;
2373         }
2374         if (need_release) {
2375                 osc_extent_release(env, ext);
2376                 oio->oi_active = NULL;
2377                 ext = NULL;
2378         }
2379
2380         if (ext == NULL) {
2381                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2382
2383                 /* try to find new extent to cover this page */
2384                 LASSERT(oio->oi_active == NULL);
2385                 /* we may have allocated grant for this page if we failed
2386                  * to expand the previous active extent. */
2387                 LASSERT(ergo(grants > 0, grants >= tmp));
2388
2389                 rc = 0;
2390                 if (grants == 0) {
2391                         /* We haven't allocated grant for this page, and we
2392                          * must not hold a page lock while we do enter_cache,
2393                          * so we must mark dirty & unlock any pages in the
2394                          * write commit pagevec. */
2395                         if (pagevec_count(pvec)) {
2396                                 cb(env, io, pvec);
2397                                 pagevec_reinit(pvec);
2398                         }
2399                         rc = osc_enter_cache(env, cli, oap, tmp);
2400                         if (rc == 0)
2401                                 grants = tmp;
2402                 }
2403
2404                 tmp = grants;
2405                 if (rc == 0) {
2406                         ext = osc_extent_find(env, osc, index, &tmp);
2407                         if (IS_ERR(ext)) {
2408                                 LASSERT(tmp == grants);
2409                                 osc_exit_cache(cli, oap);
2410                                 rc = PTR_ERR(ext);
2411                                 ext = NULL;
2412                         } else {
2413                                 oio->oi_active = ext;
2414                         }
2415                 }
2416                 if (grants > 0)
2417                         osc_unreserve_grant(cli, grants, tmp);
2418         }
2419
2420         LASSERT(ergo(rc == 0, ext != NULL));
2421         if (ext != NULL) {
2422                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2423                          ext, "index = %lu.\n", index);
2424                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2425
2426                 osc_object_lock(osc);
2427                 if (ext->oe_nr_pages == 0)
2428                         ext->oe_srvlock = ops->ops_srvlock;
2429                 else
2430                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2431                 ++ext->oe_nr_pages;
2432                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2433                 osc_object_unlock(osc);
2434
2435                 if (!ext->oe_layout_version)
2436                         ext->oe_layout_version = io->ci_layout_version;
2437         }
2438
2439         RETURN(rc);
2440 }
2441
2442 int osc_teardown_async_page(const struct lu_env *env,
2443                             struct osc_object *obj, struct osc_page *ops)
2444 {
2445         struct osc_async_page *oap = &ops->ops_oap;
2446         int rc = 0;
2447         ENTRY;
2448
2449         LASSERT(oap->oap_magic == OAP_MAGIC);
2450
2451         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2452                oap, ops, osc_index(oap2osc(oap)));
2453
2454         if (!list_empty(&oap->oap_rpc_item)) {
2455                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2456                 rc = -EBUSY;
2457         } else if (!list_empty(&oap->oap_pending_item)) {
2458                 struct osc_extent *ext = NULL;
2459
2460                 osc_object_lock(obj);
2461                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2462                 osc_object_unlock(obj);
2463                 /* only truncated pages are allowed to be taken out.
2464                  * See osc_extent_truncate() and osc_cache_truncate_start()
2465                  * for details. */
2466                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2467                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2468                                         osc_index(oap2osc(oap)));
2469                         rc = -EBUSY;
2470                 }
2471                 if (ext != NULL)
2472                         osc_extent_put(env, ext);
2473         }
2474         RETURN(rc);
2475 }
2476
2477 /**
2478  * This is called when a page is picked up by kernel to write out.
2479  *
2480  * We should find out the corresponding extent and add the whole extent
2481  * into urgent list. The extent may be being truncated or used, handle it
2482  * carefully.
2483  */
2484 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2485                          struct osc_page *ops)
2486 {
2487         struct osc_extent *ext   = NULL;
2488         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2489         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2490         pgoff_t            index = osc_index(ops);
2491         struct osc_async_page *oap = &ops->ops_oap;
2492         bool unplug = false;
2493         int rc = 0;
2494         ENTRY;
2495
2496         osc_object_lock(obj);
2497         ext = osc_extent_lookup(obj, index);
2498         if (ext == NULL) {
2499                 osc_extent_tree_dump(D_ERROR, obj);
2500                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2501         }
2502
2503         switch (ext->oe_state) {
2504         case OES_RPC:
2505         case OES_LOCK_DONE:
2506                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2507                 LASSERT(0);
2508                 break;
2509         case OES_LOCKING:
2510                 /* If we know this extent is being written out, we should abort
2511                  * so that the writer can make this page ready. Otherwise, there
2512                  * exists a deadlock problem because other process can wait for
2513                  * page writeback bit holding page lock; and meanwhile in
2514                  * vvp_page_make_ready(), we need to grab page lock before
2515                  * really sending the RPC. */
2516         case OES_TRUNC:
2517                 /* race with truncate, page will be redirtied */
2518         case OES_ACTIVE:
2519                 /* The extent is active so we need to abort and let the caller
2520                  * re-dirty the page. If we continued on here, and we were the
2521                  * one making the extent active, we could deadlock waiting for
2522                  * the page writeback to clear but it won't because the extent
2523                  * is active and won't be written out. */
2524                 GOTO(out, rc = -EAGAIN);
2525         default:
2526                 break;
2527         }
2528
2529         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2530         if (rc)
2531                 GOTO(out, rc);
2532
2533         spin_lock(&oap->oap_lock);
2534         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2535         spin_unlock(&oap->oap_lock);
2536
2537         if (current->flags & PF_MEMALLOC)
2538                 ext->oe_memalloc = 1;
2539
2540         ext->oe_urgent = 1;
2541         if (ext->oe_state == OES_CACHE) {
2542                 OSC_EXTENT_DUMP(D_CACHE, ext,
2543                                 "flush page %p make it urgent.\n", oap);
2544                 if (list_empty(&ext->oe_link))
2545                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2546                 unplug = true;
2547         }
2548         rc = 0;
2549         EXIT;
2550
2551 out:
2552         osc_object_unlock(obj);
2553         osc_extent_put(env, ext);
2554         if (unplug)
2555                 osc_io_unplug_async(env, osc_cli(obj), obj);
2556         return rc;
2557 }
2558
2559 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2560                          struct osc_object *obj, struct list_head *list,
2561                          int brw_flags)
2562 {
2563         struct client_obd     *cli = osc_cli(obj);
2564         struct osc_extent     *ext;
2565         struct osc_async_page *oap;
2566         int     page_count = 0;
2567         int     mppr       = cli->cl_max_pages_per_rpc;
2568         bool    can_merge   = true;
2569         pgoff_t start      = CL_PAGE_EOF;
2570         pgoff_t end        = 0;
2571         ENTRY;
2572
2573         list_for_each_entry(oap, list, oap_pending_item) {
2574                 struct osc_page *opg = oap2osc_page(oap);
2575                 pgoff_t index = osc_index(opg);
2576
2577                 if (index > end)
2578                         end = index;
2579                 if (index < start)
2580                         start = index;
2581                 ++page_count;
2582                 mppr <<= (page_count > mppr);
2583
2584                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2585                         can_merge = false;
2586         }
2587
2588         ext = osc_extent_alloc(obj);
2589         if (ext == NULL) {
2590                 struct osc_async_page *tmp;
2591
2592                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2593                         list_del_init(&oap->oap_pending_item);
2594                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2595                 }
2596                 RETURN(-ENOMEM);
2597         }
2598
2599         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2600         ext->oe_sync = 1;
2601         ext->oe_no_merge = !can_merge;
2602         ext->oe_urgent = 1;
2603         ext->oe_start = start;
2604         ext->oe_end = ext->oe_max_end = end;
2605         ext->oe_obj = obj;
2606         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2607         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2608         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2609         ext->oe_nr_pages = page_count;
2610         ext->oe_mppr = mppr;
2611         list_splice_init(list, &ext->oe_pages);
2612         ext->oe_layout_version = io->ci_layout_version;
2613
2614         osc_object_lock(obj);
2615         /* Reuse the initial refcount for RPC, don't drop it */
2616         osc_extent_state_set(ext, OES_LOCK_DONE);
2617         if (!ext->oe_rw) { /* write */
2618                 if (!ext->oe_srvlock && !ext->oe_dio) {
2619                         /* The most likely case here is from lack of grants
2620                          * so we are either out of quota or out of space.
2621                          * Since this means we are holding locks across
2622                          * potentially multi-striped IO, we must send out
2623                          * everything out instantly to avoid prolonged
2624                          * waits resulting in lock eviction (likely since
2625                          * the extended wait in osc_cache_enter() did not
2626                          * yield any additional grant due to a timeout.
2627                          * LU-13131 */
2628                         ext->oe_hp = 1;
2629                         list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2630                 } else {
2631                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2632                 }
2633                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2634         } else {
2635                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2636                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2637         }
2638         osc_object_unlock(obj);
2639
2640         osc_io_unplug_async(env, cli, obj);
2641         RETURN(0);
2642 }
2643
2644 /**
2645  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2646  */
2647 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2648                              __u64 size, struct osc_extent **extp)
2649 {
2650         struct client_obd *cli = osc_cli(obj);
2651         struct osc_extent *ext;
2652         struct osc_extent *waiting = NULL;
2653         pgoff_t index;
2654         LIST_HEAD(list);
2655         int result = 0;
2656         bool partial;
2657         ENTRY;
2658
2659         /* pages with index greater or equal to index will be truncated. */
2660         index = cl_index(osc2cl(obj), size);
2661         partial = size > cl_offset(osc2cl(obj), index);
2662
2663 again:
2664         osc_object_lock(obj);
2665         ext = osc_extent_search(obj, index);
2666         if (ext == NULL)
2667                 ext = first_extent(obj);
2668         else if (ext->oe_end < index)
2669                 ext = next_extent(ext);
2670         while (ext != NULL) {
2671                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2672
2673                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2674                         /* if ext is in urgent state, it means there must exist
2675                          * a page already having been flushed by write_page().
2676                          * We have to wait for this extent because we can't
2677                          * truncate that page. */
2678                         OSC_EXTENT_DUMP(D_CACHE, ext,
2679                                         "waiting for busy extent\n");
2680                         waiting = osc_extent_get(ext);
2681                         break;
2682                 }
2683
2684                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2685
2686                 osc_extent_get(ext);
2687                 if (ext->oe_state == OES_ACTIVE) {
2688                         /* though we grab inode mutex for write path, but we
2689                          * release it before releasing extent(in osc_io_end()),
2690                          * so there is a race window that an extent is still
2691                          * in OES_ACTIVE when truncate starts. */
2692                         LASSERT(!ext->oe_trunc_pending);
2693                         ext->oe_trunc_pending = 1;
2694                 } else {
2695                         EASSERT(ext->oe_state == OES_CACHE, ext);
2696                         osc_extent_state_set(ext, OES_TRUNC);
2697                         osc_update_pending(obj, OBD_BRW_WRITE,
2698                                            -ext->oe_nr_pages);
2699                 }
2700                 /* This extent could be on the full extents list, that's OK */
2701                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2702                 if (!list_empty(&ext->oe_link))
2703                         list_move_tail(&ext->oe_link, &list);
2704                 else
2705                         list_add_tail(&ext->oe_link, &list);
2706
2707                 ext = next_extent(ext);
2708         }
2709         osc_object_unlock(obj);
2710
2711         osc_list_maint(cli, obj);
2712
2713         while (!list_empty(&list)) {
2714                 int rc;
2715
2716                 ext = list_entry(list.next, struct osc_extent, oe_link);
2717                 list_del_init(&ext->oe_link);
2718
2719                 /* extent may be in OES_ACTIVE state because inode mutex
2720                  * is released before osc_io_end() in file write case */
2721                 if (ext->oe_state != OES_TRUNC)
2722                         osc_extent_wait(env, ext, OES_TRUNC);
2723
2724                 rc = osc_extent_truncate(ext, index, partial);
2725                 if (rc < 0) {
2726                         if (result == 0)
2727                                 result = rc;
2728
2729                         OSC_EXTENT_DUMP(D_ERROR, ext,
2730                                         "truncate error %d\n", rc);
2731                 } else if (ext->oe_nr_pages == 0) {
2732                         osc_extent_remove(ext);
2733                 } else {
2734                         /* this must be an overlapped extent which means only
2735                          * part of pages in this extent have been truncated.
2736                          */
2737                         EASSERTF(ext->oe_start <= index, ext,
2738                                  "trunc index = %lu/%d.\n", index, partial);
2739                         /* fix index to skip this partially truncated extent */
2740                         index = ext->oe_end + 1;
2741                         partial = false;
2742
2743                         /* we need to hold this extent in OES_TRUNC state so
2744                          * that no writeback will happen. This is to avoid
2745                          * BUG 17397.
2746                          * Only partial truncate can reach here, if @size is
2747                          * not zero, the caller should provide a valid @extp. */
2748                         LASSERT(*extp == NULL);
2749                         *extp = osc_extent_get(ext);
2750                         OSC_EXTENT_DUMP(D_CACHE, ext,
2751                                         "trunc at %llu\n", size);
2752                 }
2753                 osc_extent_put(env, ext);
2754         }
2755         if (waiting != NULL) {
2756                 int rc;
2757
2758                 /* ignore the result of osc_extent_wait the write initiator
2759                  * should take care of it. */
2760                 rc = osc_extent_wait(env, waiting, OES_INV);
2761                 if (rc < 0)
2762                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2763
2764                 osc_extent_put(env, waiting);
2765                 waiting = NULL;
2766                 goto again;
2767         }
2768         RETURN(result);
2769 }
2770 EXPORT_SYMBOL(osc_cache_truncate_start);
2771
2772 /**
2773  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2774  */
2775 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2776 {
2777         if (ext != NULL) {
2778                 struct osc_object *obj = ext->oe_obj;
2779                 bool unplug = false;
2780
2781                 EASSERT(ext->oe_nr_pages > 0, ext);
2782                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2783                 EASSERT(!ext->oe_urgent, ext);
2784
2785                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2786                 osc_object_lock(obj);
2787                 osc_extent_state_set(ext, OES_CACHE);
2788                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2789                         ext->oe_urgent = 1;
2790                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2791                         unplug = true;
2792                 }
2793                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2794                 osc_object_unlock(obj);
2795                 osc_extent_put(env, ext);
2796
2797                 if (unplug)
2798                         osc_io_unplug_async(env, osc_cli(obj), obj);
2799         }
2800 }
2801
2802 /**
2803  * Wait for extents in a specific range to be written out.
2804  * The caller must have called osc_cache_writeback_range() to issue IO
2805  * otherwise it will take a long time for this function to finish.
2806  *
2807  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2808  * nobody else can dirty this range of file while we're waiting for
2809  * extents to be written.
2810  */
2811 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2812                          pgoff_t start, pgoff_t end)
2813 {
2814         struct osc_extent *ext;
2815         pgoff_t index = start;
2816         int     result = 0;
2817         ENTRY;
2818
2819 again:
2820         osc_object_lock(obj);
2821         ext = osc_extent_search(obj, index);
2822         if (ext == NULL)
2823                 ext = first_extent(obj);
2824         else if (ext->oe_end < index)
2825                 ext = next_extent(ext);
2826         while (ext != NULL) {
2827                 int rc;
2828
2829                 if (ext->oe_start > end)
2830                         break;
2831
2832                 if (!ext->oe_fsync_wait) {
2833                         ext = next_extent(ext);
2834                         continue;
2835                 }
2836
2837                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2838                              ext->oe_hp || ext->oe_urgent), ext);
2839                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2840                              !ext->oe_hp && ext->oe_urgent), ext);
2841
2842                 index = ext->oe_end + 1;
2843                 osc_extent_get(ext);
2844                 osc_object_unlock(obj);
2845
2846                 rc = osc_extent_wait(env, ext, OES_INV);
2847                 if (result == 0)
2848                         result = rc;
2849                 osc_extent_put(env, ext);
2850                 goto again;
2851         }
2852         osc_object_unlock(obj);
2853
2854         OSC_IO_DEBUG(obj, "sync file range.\n");
2855         RETURN(result);
2856 }
2857 EXPORT_SYMBOL(osc_cache_wait_range);
2858
2859 /**
2860  * Called to write out a range of osc object.
2861  *
2862  * @hp     : should be set this is caused by lock cancel;
2863  * @discard: is set if dirty pages should be dropped - file will be deleted or
2864  *         truncated, this implies there is no partially discarding extents.
2865  *
2866  * Return how many pages will be issued, or error code if error occurred.
2867  */
2868 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2869                               pgoff_t start, pgoff_t end, int hp, int discard)
2870 {
2871         struct osc_extent *ext;
2872         LIST_HEAD(discard_list);
2873         bool unplug = false;
2874         int result = 0;
2875         ENTRY;
2876
2877         osc_object_lock(obj);
2878         ext = osc_extent_search(obj, start);
2879         if (ext == NULL)
2880                 ext = first_extent(obj);
2881         else if (ext->oe_end < start)
2882                 ext = next_extent(ext);
2883         while (ext != NULL) {
2884                 if (ext->oe_start > end)
2885                         break;
2886
2887                 ext->oe_fsync_wait = 1;
2888                 switch (ext->oe_state) {
2889                 case OES_CACHE:
2890                         result += ext->oe_nr_pages;
2891                         if (!discard) {
2892                                 struct list_head *list = NULL;
2893                                 if (hp) {
2894                                         EASSERT(!ext->oe_hp, ext);
2895                                         ext->oe_hp = 1;
2896                                         list = &obj->oo_hp_exts;
2897                                 } else if (!ext->oe_urgent && !ext->oe_hp) {
2898                                         ext->oe_urgent = 1;
2899                                         list = &obj->oo_urgent_exts;
2900                                 }
2901                                 if (list != NULL)
2902                                         list_move_tail(&ext->oe_link, list);
2903                                 unplug = true;
2904                         } else {
2905                                 struct client_obd *cli = osc_cli(obj);
2906                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2907                                 pgoff_t align_by = (1 << pcc_bits);
2908                                 pgoff_t a_start = round_down(start, align_by);
2909                                 pgoff_t a_end = round_up(end, align_by);
2910
2911                                 /* overflow case */
2912                                 if (end && !a_end)
2913                                         a_end = CL_PAGE_EOF;
2914                                 /* the only discarder is lock cancelling, so
2915                                  * [start, end], aligned by chunk size, must
2916                                  * contain this extent */
2917                                 LASSERTF(ext->oe_start >= a_start &&
2918                                          ext->oe_end <= a_end,
2919                                          "ext [%lu, %lu] reg [%lu, %lu] "
2920                                          "orig [%lu %lu] align %lu bits "
2921                                          "%d\n", ext->oe_start, ext->oe_end,
2922                                          a_start, a_end, start, end,
2923                                          align_by, pcc_bits);
2924                                 osc_extent_state_set(ext, OES_LOCKING);
2925                                 ext->oe_owner = current;
2926                                 list_move_tail(&ext->oe_link,
2927                                                    &discard_list);
2928                                 osc_update_pending(obj, OBD_BRW_WRITE,
2929                                                    -ext->oe_nr_pages);
2930                         }
2931                         break;
2932                 case OES_ACTIVE:
2933                         /* It's pretty bad to wait for ACTIVE extents, because
2934                          * we don't know how long we will wait for it to be
2935                          * flushed since it may be blocked at awaiting more
2936                          * grants. We do this for the correctness of fsync. */
2937                         LASSERT(hp == 0 && discard == 0);
2938                         ext->oe_urgent = 1;
2939                         break;
2940                 case OES_TRUNC:
2941                         /* this extent is being truncated, can't do anything
2942                          * for it now. it will be set to urgent after truncate
2943                          * is finished in osc_cache_truncate_end(). */
2944                 default:
2945                         break;
2946                 }
2947                 ext = next_extent(ext);
2948         }
2949         osc_object_unlock(obj);
2950
2951         LASSERT(ergo(!discard, list_empty(&discard_list)));
2952         if (!list_empty(&discard_list)) {
2953                 struct osc_extent *tmp;
2954                 int rc;
2955
2956                 osc_list_maint(osc_cli(obj), obj);
2957                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2958                         list_del_init(&ext->oe_link);
2959                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2960
2961                         /* Discard caching pages. We don't actually write this
2962                          * extent out but we complete it as if we did. */
2963                         rc = osc_extent_make_ready(env, ext);
2964                         if (unlikely(rc < 0)) {
2965                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2966                                                 "make_ready returned %d\n", rc);
2967                                 if (result >= 0)
2968                                         result = rc;
2969                         }
2970
2971                         /* finish the extent as if the pages were sent */
2972                         osc_extent_finish(env, ext, 0, 0);
2973                 }
2974         }
2975
2976         if (unplug)
2977                 osc_io_unplug(env, osc_cli(obj), obj);
2978
2979         if (hp || discard) {
2980                 int rc;
2981                 rc = osc_cache_wait_range(env, obj, start, end);
2982                 if (result >= 0 && rc < 0)
2983                         result = rc;
2984         }
2985
2986         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2987         RETURN(result);
2988 }
2989 EXPORT_SYMBOL(osc_cache_writeback_range);
2990
2991 /**
2992  * Returns a list of pages by a given [start, end] of \a obj.
2993  *
2994  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
2995  * crucial in the face of [offset, EOF] locks.
2996  *
2997  * Return at least one page in @queue unless there is no covered page.
2998  */
2999 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3000                           struct osc_object *osc, pgoff_t start, pgoff_t end,
3001                           osc_page_gang_cbt cb, void *cbdata)
3002 {
3003         struct osc_page *ops;
3004         struct pagevec  *pagevec;
3005         void            **pvec;
3006         pgoff_t         idx;
3007         unsigned int    nr;
3008         unsigned int    i;
3009         unsigned int    j;
3010         bool            res = true;
3011         bool            tree_lock = true;
3012         ENTRY;
3013
3014         idx = start;
3015         pvec = osc_env_info(env)->oti_pvec;
3016         pagevec = &osc_env_info(env)->oti_pagevec;
3017         ll_pagevec_init(pagevec, 0);
3018         spin_lock(&osc->oo_tree_lock);
3019         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3020                                             idx, OTI_PVEC_SIZE)) > 0) {
3021                 struct cl_page *page;
3022                 bool end_of_region = false;
3023
3024                 for (i = 0, j = 0; i < nr; ++i) {
3025                         ops = pvec[i];
3026                         pvec[i] = NULL;
3027
3028                         idx = osc_index(ops);
3029                         if (idx > end) {
3030                                 end_of_region = true;
3031                                 break;
3032                         }
3033
3034                         page = ops->ops_cl.cpl_page;
3035                         LASSERT(page->cp_type == CPT_CACHEABLE);
3036                         if (page->cp_state == CPS_FREEING)
3037                                 continue;
3038
3039                         cl_page_get(page);
3040                         lu_ref_add_atomic(&page->cp_reference,
3041                                           "gang_lookup", current);
3042                         pvec[j++] = ops;
3043                 }
3044                 ++idx;
3045
3046                 /*
3047                  * Here a delicate locking dance is performed. Current thread
3048                  * holds a reference to a page, but has to own it before it
3049                  * can be placed into queue. Owning implies waiting, so
3050                  * radix-tree lock is to be released. After a wait one has to
3051                  * check that pages weren't truncated (cl_page_own() returns
3052                  * error in the latter case).
3053                  */
3054                 spin_unlock(&osc->oo_tree_lock);
3055                 tree_lock = false;
3056
3057                 for (i = 0; i < j; ++i) {
3058                         ops = pvec[i];
3059                         if (res)
3060                                 res = (*cb)(env, io, ops, cbdata);
3061
3062                         page = ops->ops_cl.cpl_page;
3063                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3064                         cl_pagevec_put(env, page, pagevec);
3065                 }
3066                 pagevec_release(pagevec);
3067
3068                 if (nr < OTI_PVEC_SIZE || end_of_region)
3069                         break;
3070
3071                 if (!res)
3072                         break;
3073                 if (need_resched())
3074                         cond_resched();
3075
3076                 spin_lock(&osc->oo_tree_lock);
3077                 tree_lock = true;
3078         }
3079         if (tree_lock)
3080                 spin_unlock(&osc->oo_tree_lock);
3081         RETURN(res);
3082 }
3083 EXPORT_SYMBOL(osc_page_gang_lookup);
3084
3085 /**
3086  * Check if page @page is covered by an extra lock or discard it.
3087  */
3088 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3089                                 struct osc_page *ops, void *cbdata)
3090 {
3091         struct osc_thread_info *info = osc_env_info(env);
3092         struct osc_object *osc = cbdata;
3093         pgoff_t index;
3094
3095         index = osc_index(ops);
3096         if (index >= info->oti_fn_index) {
3097                 struct ldlm_lock *tmp;
3098                 struct cl_page *page = ops->ops_cl.cpl_page;
3099
3100                 /* refresh non-overlapped index */
3101                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3102                                            OSC_DAP_FL_TEST_LOCK);
3103                 if (tmp != NULL) {
3104                         __u64 end = tmp->l_policy_data.l_extent.end;
3105                         /* Cache the first-non-overlapped index so as to skip
3106                          * all pages within [index, oti_fn_index). This is safe
3107                          * because if tmp lock is canceled, it will discard
3108                          * these pages. */
3109                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3110                         if (end == OBD_OBJECT_EOF)
3111                                 info->oti_fn_index = CL_PAGE_EOF;
3112                         LDLM_LOCK_PUT(tmp);
3113                 } else if (cl_page_own(env, io, page) == 0) {
3114                         /* discard the page */
3115                         cl_page_discard(env, io, page);
3116                         cl_page_disown(env, io, page);
3117                 } else {
3118                         LASSERT(page->cp_state == CPS_FREEING);
3119                 }
3120         }
3121
3122         info->oti_next_index = index + 1;
3123         return true;
3124 }
3125
3126 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3127                    struct osc_page *ops, void *cbdata)
3128 {
3129         struct osc_thread_info *info = osc_env_info(env);
3130         struct cl_page *page = ops->ops_cl.cpl_page;
3131
3132         /* page is top page. */
3133         info->oti_next_index = osc_index(ops) + 1;
3134         if (cl_page_own(env, io, page) == 0) {
3135                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3136                           !PageDirty(cl_page_vmpage(page))))
3137                         CL_PAGE_DEBUG(D_ERROR, env, page,
3138                                         "discard dirty page?\n");
3139
3140                 /* discard the page */
3141                 cl_page_discard(env, io, page);
3142                 cl_page_disown(env, io, page);
3143         } else {
3144                 LASSERT(page->cp_state == CPS_FREEING);
3145         }
3146
3147         return true;
3148 }
3149 EXPORT_SYMBOL(osc_discard_cb);
3150
3151 /**
3152  * Discard pages protected by the given lock. This function traverses radix
3153  * tree to find all covering pages and discard them. If a page is being covered
3154  * by other locks, it should remain in cache.
3155  *
3156  * If error happens on any step, the process continues anyway (the reasoning
3157  * behind this being that lock cancellation cannot be delayed indefinitely).
3158  */
3159 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3160                            pgoff_t start, pgoff_t end, bool discard)
3161 {
3162         struct osc_thread_info *info = osc_env_info(env);
3163         struct cl_io *io = osc_env_thread_io(env);
3164         osc_page_gang_cbt cb;
3165         int result;
3166
3167         ENTRY;
3168
3169         io->ci_obj = cl_object_top(osc2cl(osc));
3170         io->ci_ignore_layout = 1;
3171         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);