Whamcloud - gitweb
LU-9679 modules: Use LIST_HEAD for declaring list_heads
[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
42 #include "osc_internal.h"
43
44 static int extent_debug; /* set it to be true for more debug */
45
46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
48                            enum osc_extent_state state);
49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
50                               struct osc_async_page *oap, int sent, int rc);
51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
52                           int cmd);
53 static int osc_refresh_count(const struct lu_env *env,
54                              struct osc_async_page *oap, int cmd);
55 static int osc_io_unplug_async(const struct lu_env *env,
56                                struct client_obd *cli, struct osc_object *osc);
57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
58                            unsigned int lost_grant, unsigned int dirty_grant);
59
60 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
61                                   const char *func, int line);
62 #define osc_extent_tree_dump(mask, obj) \
63         osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
64
65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
66                                 unsigned int unused);
67
68 /** \addtogroup osc
69  *  @{
70  */
71
72 /* ------------------ osc extent ------------------ */
73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
74 {
75         char *buf = flags;
76         *buf++ = ext->oe_rw ? 'r' : 'w';
77         if (ext->oe_intree)
78                 *buf++ = 'i';
79         if (ext->oe_sync)
80                 *buf++ = 'S';
81         if (ext->oe_srvlock)
82                 *buf++ = 's';
83         if (ext->oe_hp)
84                 *buf++ = 'h';
85         if (ext->oe_urgent)
86                 *buf++ = 'u';
87         if (ext->oe_memalloc)
88                 *buf++ = 'm';
89         if (ext->oe_trunc_pending)
90                 *buf++ = 't';
91         if (ext->oe_fsync_wait)
92                 *buf++ = 'Y';
93         *buf = 0;
94         return flags;
95 }
96
97 static inline char list_empty_marker(struct list_head *list)
98 {
99         return list_empty(list) ? '-' : '+';
100 }
101
102 #define EXTSTR       "[%lu -> %lu/%lu]"
103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
104 static const char *oes_strings[] = {
105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
106
107 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
108         static struct cfs_debug_limit_state cdls;                             \
109         struct osc_extent *__ext = (extent);                                  \
110         char __buf[16];                                                       \
111                                                                               \
112         __CDEBUG_WITH_LOC(file, func, line, mask, &cdls,                      \
113                 "extent %p@{" EXTSTR ", "                                     \
114                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
115                 /* ----- extent part 0 ----- */                               \
116                 __ext, EXTPARA(__ext),                                        \
117                 /* ----- part 1 ----- */                                      \
118                 atomic_read(&__ext->oe_refc),                                 \
119                 atomic_read(&__ext->oe_users),                                \
120                 list_empty_marker(&__ext->oe_link),                           \
121                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
122                 __ext->oe_obj,                                                \
123                 /* ----- part 2 ----- */                                      \
124                 __ext->oe_grants, __ext->oe_nr_pages,                         \
125                 list_empty_marker(&__ext->oe_pages),                          \
126                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
127                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
128                 /* ----- part 4 ----- */                                      \
129                 ## __VA_ARGS__);                                              \
130         if (mask == D_ERROR && __ext->oe_dlmlock != NULL)                     \
131                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
132         else                                                                  \
133                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
134 } while (0)
135
136 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...)                            \
137         OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,          \
138                                  mask, ext, fmt, ## __VA_ARGS__)
139
140 #undef EASSERTF
141 #define EASSERTF(expr, ext, fmt, args...) do {                          \
142         if (!(expr)) {                                                  \
143                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
144                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
145                 LASSERT(expr);                                          \
146         }                                                               \
147 } while (0)
148
149 #undef EASSERT
150 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
151
152 static inline struct osc_extent *rb_extent(struct rb_node *n)
153 {
154         if (n == NULL)
155                 return NULL;
156
157         return container_of(n, struct osc_extent, oe_node);
158 }
159
160 static inline struct osc_extent *next_extent(struct osc_extent *ext)
161 {
162         if (ext == NULL)
163                 return NULL;
164
165         LASSERT(ext->oe_intree);
166         return rb_extent(rb_next(&ext->oe_node));
167 }
168
169 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
170 {
171         if (ext == NULL)
172                 return NULL;
173
174         LASSERT(ext->oe_intree);
175         return rb_extent(rb_prev(&ext->oe_node));
176 }
177
178 static inline struct osc_extent *first_extent(struct osc_object *obj)
179 {
180         return rb_extent(rb_first(&obj->oo_root));
181 }
182
183 /* object must be locked by caller. */
184 static int osc_extent_sanity_check0(struct osc_extent *ext,
185                                     const char *func, const int line)
186 {
187         struct osc_object *obj = ext->oe_obj;
188         struct osc_async_page *oap;
189         size_t page_count;
190         int rc = 0;
191
192         if (!osc_object_is_locked(obj))
193                 GOTO(out, rc = 9);
194
195         if (ext->oe_state >= OES_STATE_MAX)
196                 GOTO(out, rc = 10);
197
198         if (atomic_read(&ext->oe_refc) <= 0)
199                 GOTO(out, rc = 20);
200
201         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
202                 GOTO(out, rc = 30);
203
204         switch (ext->oe_state) {
205         case OES_INV:
206                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
207                         GOTO(out, rc = 35);
208                 GOTO(out, rc = 0);
209                 break;
210         case OES_ACTIVE:
211                 if (atomic_read(&ext->oe_users) == 0)
212                         GOTO(out, rc = 40);
213                 if (ext->oe_hp)
214                         GOTO(out, rc = 50);
215                 if (ext->oe_fsync_wait && !ext->oe_urgent)
216                         GOTO(out, rc = 55);
217                 break;
218         case OES_CACHE:
219                 if (ext->oe_grants == 0)
220                         GOTO(out, rc = 60);
221                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
222                         GOTO(out, rc = 65);
223                 /* fallthrough */
224         default:
225                 if (atomic_read(&ext->oe_users) > 0)
226                         GOTO(out, rc = 70);
227         }
228
229         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
230                 GOTO(out, rc = 80);
231
232         if (ext->oe_sync && ext->oe_grants > 0)
233                 GOTO(out, rc = 90);
234
235         if (ext->oe_dlmlock != NULL &&
236             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
237             !ldlm_is_failed(ext->oe_dlmlock)) {
238                 struct ldlm_extent *extent;
239
240                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
241                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
242                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
243                         GOTO(out, rc = 100);
244
245                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
246                         GOTO(out, rc = 102);
247         }
248
249         if (ext->oe_nr_pages > ext->oe_mppr)
250                 GOTO(out, rc = 105);
251
252         /* Do not verify page list if extent is in RPC. This is because an
253          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
254         if (ext->oe_state > OES_CACHE)
255                 GOTO(out, rc = 0);
256
257         if (!extent_debug)
258                 GOTO(out, rc = 0);
259
260         page_count = 0;
261         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
262                 pgoff_t index = osc_index(oap2osc(oap));
263                 ++page_count;
264                 if (index > ext->oe_end || index < ext->oe_start)
265                         GOTO(out, rc = 110);
266         }
267         if (page_count != ext->oe_nr_pages)
268                 GOTO(out, rc = 120);
269
270 out:
271         if (rc != 0)
272                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
273                                          "sanity check %p failed: rc = %d\n",
274                                          ext, rc);
275         return rc;
276 }
277
278 #define sanity_check_nolock(ext) \
279         osc_extent_sanity_check0(ext, __func__, __LINE__)
280
281 #define sanity_check(ext) ({                                                   \
282         int __res;                                                             \
283         osc_object_lock((ext)->oe_obj);                                        \
284         __res = sanity_check_nolock(ext);                                      \
285         osc_object_unlock((ext)->oe_obj);                                      \
286         __res;                                                                 \
287 })
288
289
290 /**
291  * sanity check - to make sure there is no overlapped extent in the tree.
292  */
293 static int osc_extent_is_overlapped(struct osc_object *obj,
294                                     struct osc_extent *ext)
295 {
296         struct osc_extent *tmp;
297
298         LASSERT(osc_object_is_locked(obj));
299
300         if (!extent_debug)
301                 return 0;
302
303         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
304                 if (tmp == ext)
305                         continue;
306                 if (tmp->oe_end >= ext->oe_start &&
307                     tmp->oe_start <= ext->oe_end)
308                         return 1;
309         }
310         return 0;
311 }
312
313 static void osc_extent_state_set(struct osc_extent *ext, int state)
314 {
315         LASSERT(osc_object_is_locked(ext->oe_obj));
316         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
317
318         /* Never try to sanity check a state changing extent :-) */
319         /* LASSERT(sanity_check_nolock(ext) == 0); */
320
321         /* TODO: validate the state machine */
322         ext->oe_state = state;
323         wake_up_all(&ext->oe_waitq);
324 }
325
326 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
327 {
328         struct osc_extent *ext;
329
330         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
331         if (ext == NULL)
332                 return NULL;
333
334         RB_CLEAR_NODE(&ext->oe_node);
335         ext->oe_obj = obj;
336         cl_object_get(osc2cl(obj));
337         atomic_set(&ext->oe_refc, 1);
338         atomic_set(&ext->oe_users, 0);
339         INIT_LIST_HEAD(&ext->oe_link);
340         ext->oe_state = OES_INV;
341         INIT_LIST_HEAD(&ext->oe_pages);
342         init_waitqueue_head(&ext->oe_waitq);
343         ext->oe_dlmlock = NULL;
344
345         return ext;
346 }
347
348 static void osc_extent_free(struct osc_extent *ext)
349 {
350         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
351 }
352
353 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
354 {
355         LASSERT(atomic_read(&ext->oe_refc) >= 0);
356         atomic_inc(&ext->oe_refc);
357         return ext;
358 }
359
360 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
361 {
362         LASSERT(atomic_read(&ext->oe_refc) > 0);
363         if (atomic_dec_and_test(&ext->oe_refc)) {
364                 LASSERT(list_empty(&ext->oe_link));
365                 LASSERT(atomic_read(&ext->oe_users) == 0);
366                 LASSERT(ext->oe_state == OES_INV);
367                 LASSERT(!ext->oe_intree);
368
369                 if (ext->oe_dlmlock != NULL) {
370                         lu_ref_add(&ext->oe_dlmlock->l_reference,
371                                    "osc_extent", ext);
372                         LDLM_LOCK_PUT(ext->oe_dlmlock);
373                         ext->oe_dlmlock = NULL;
374                 }
375                 cl_object_put(env, osc2cl(ext->oe_obj));
376                 osc_extent_free(ext);
377         }
378 }
379
380 /**
381  * osc_extent_put_trust() is a special version of osc_extent_put() when
382  * it's known that the caller is not the last user. This is to address the
383  * problem of lacking of lu_env ;-).
384  */
385 static void osc_extent_put_trust(struct osc_extent *ext)
386 {
387         LASSERT(atomic_read(&ext->oe_refc) > 1);
388         LASSERT(osc_object_is_locked(ext->oe_obj));
389         atomic_dec(&ext->oe_refc);
390 }
391
392 /**
393  * Return the extent which includes pgoff @index, or return the greatest
394  * previous extent in the tree.
395  */
396 static struct osc_extent *osc_extent_search(struct osc_object *obj,
397                                             pgoff_t index)
398 {
399         struct rb_node    *n = obj->oo_root.rb_node;
400         struct osc_extent *tmp, *p = NULL;
401
402         LASSERT(osc_object_is_locked(obj));
403         while (n != NULL) {
404                 tmp = rb_extent(n);
405                 if (index < tmp->oe_start) {
406                         n = n->rb_left;
407                 } else if (index > tmp->oe_end) {
408                         p = rb_extent(n);
409                         n = n->rb_right;
410                 } else {
411                         return tmp;
412                 }
413         }
414         return p;
415 }
416
417 /*
418  * Return the extent covering @index, otherwise return NULL.
419  * caller must have held object lock.
420  */
421 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
422                                             pgoff_t index)
423 {
424         struct osc_extent *ext;
425
426         ext = osc_extent_search(obj, index);
427         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
428                 return osc_extent_get(ext);
429         return NULL;
430 }
431
432 /* caller must have held object lock. */
433 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
434 {
435         struct rb_node   **n      = &obj->oo_root.rb_node;
436         struct rb_node    *parent = NULL;
437         struct osc_extent *tmp;
438
439         LASSERT(ext->oe_intree == 0);
440         LASSERT(ext->oe_obj == obj);
441         LASSERT(osc_object_is_locked(obj));
442         while (*n != NULL) {
443                 tmp = rb_extent(*n);
444                 parent = *n;
445
446                 if (ext->oe_end < tmp->oe_start)
447                         n = &(*n)->rb_left;
448                 else if (ext->oe_start > tmp->oe_end)
449                         n = &(*n)->rb_right;
450                 else
451                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
452         }
453         rb_link_node(&ext->oe_node, parent, n);
454         rb_insert_color(&ext->oe_node, &obj->oo_root);
455         osc_extent_get(ext);
456         ext->oe_intree = 1;
457 }
458
459 /* caller must have held object lock. */
460 static void osc_extent_erase(struct osc_extent *ext)
461 {
462         struct osc_object *obj = ext->oe_obj;
463         LASSERT(osc_object_is_locked(obj));
464         if (ext->oe_intree) {
465                 rb_erase(&ext->oe_node, &obj->oo_root);
466                 ext->oe_intree = 0;
467                 /* rbtree held a refcount */
468                 osc_extent_put_trust(ext);
469         }
470 }
471
472 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
473 {
474         struct osc_object *obj = ext->oe_obj;
475
476         LASSERT(osc_object_is_locked(obj));
477         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
478         if (ext->oe_state == OES_CACHE) {
479                 osc_extent_state_set(ext, OES_ACTIVE);
480                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
481         }
482         atomic_inc(&ext->oe_users);
483         list_del_init(&ext->oe_link);
484         return osc_extent_get(ext);
485 }
486
487 static void __osc_extent_remove(struct osc_extent *ext)
488 {
489         LASSERT(osc_object_is_locked(ext->oe_obj));
490         LASSERT(list_empty(&ext->oe_pages));
491         osc_extent_erase(ext);
492         list_del_init(&ext->oe_link);
493         osc_extent_state_set(ext, OES_INV);
494         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
495 }
496
497 static void osc_extent_remove(struct osc_extent *ext)
498 {
499         struct osc_object *obj = ext->oe_obj;
500
501         osc_object_lock(obj);
502         __osc_extent_remove(ext);
503         osc_object_unlock(obj);
504 }
505
506 /**
507  * This function is used to merge extents to get better performance. It checks
508  * if @cur and @victim are contiguous at block level.
509  */
510 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
511                             struct osc_extent *victim)
512 {
513         struct osc_object       *obj = cur->oe_obj;
514         struct client_obd       *cli = osc_cli(obj);
515         pgoff_t                  chunk_start;
516         pgoff_t                  chunk_end;
517         int                      ppc_bits;
518
519         LASSERT(cur->oe_state == OES_CACHE);
520         LASSERT(osc_object_is_locked(obj));
521         if (victim == NULL)
522                 return -EINVAL;
523
524         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
525                 return -EBUSY;
526
527         if (cur->oe_max_end != victim->oe_max_end)
528                 return -ERANGE;
529
530         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
531         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
532         chunk_start = cur->oe_start >> ppc_bits;
533         chunk_end   = cur->oe_end   >> ppc_bits;
534         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
535             chunk_end + 1 != victim->oe_start >> ppc_bits)
536                 return -ERANGE;
537
538         /* overall extent size should not exceed the max supported limit
539          * reported by the server */
540         if (cur->oe_end - cur->oe_start + 1 +
541             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
542                 return -ERANGE;
543
544         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
545
546         cur->oe_start     = min(cur->oe_start, victim->oe_start);
547         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
548         /* per-extent tax should be accounted only once for the whole extent */
549         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
550         cur->oe_nr_pages += victim->oe_nr_pages;
551         /* only the following bits are needed to merge */
552         cur->oe_urgent   |= victim->oe_urgent;
553         cur->oe_memalloc |= victim->oe_memalloc;
554         list_splice_init(&victim->oe_pages, &cur->oe_pages);
555         list_del_init(&victim->oe_link);
556         victim->oe_nr_pages = 0;
557
558         osc_extent_get(victim);
559         __osc_extent_remove(victim);
560         osc_extent_put(env, victim);
561
562         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
563         return 0;
564 }
565
566 /**
567  * Drop user count of osc_extent, and unplug IO asynchronously.
568  */
569 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
570 {
571         struct osc_object *obj = ext->oe_obj;
572         struct client_obd *cli = osc_cli(obj);
573         int rc = 0;
574         ENTRY;
575
576         LASSERT(atomic_read(&ext->oe_users) > 0);
577         LASSERT(sanity_check(ext) == 0);
578         LASSERT(ext->oe_grants > 0);
579
580         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
581                 LASSERT(ext->oe_state == OES_ACTIVE);
582                 if (ext->oe_trunc_pending) {
583                         /* a truncate process is waiting for this extent.
584                          * This may happen due to a race, check
585                          * osc_cache_truncate_start(). */
586                         osc_extent_state_set(ext, OES_TRUNC);
587                         ext->oe_trunc_pending = 0;
588                         osc_object_unlock(obj);
589                 } else {
590                         int grant = 0;
591
592                         osc_extent_state_set(ext, OES_CACHE);
593                         osc_update_pending(obj, OBD_BRW_WRITE,
594                                            ext->oe_nr_pages);
595
596                         /* try to merge the previous and next extent. */
597                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
598                                 grant += cli->cl_grant_extent_tax;
599                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
600                                 grant += cli->cl_grant_extent_tax;
601
602                         if (ext->oe_urgent)
603                                 list_move_tail(&ext->oe_link,
604                                                &obj->oo_urgent_exts);
605                         else if (ext->oe_nr_pages == ext->oe_mppr) {
606                                 list_move_tail(&ext->oe_link,
607                                                &obj->oo_full_exts);
608                         }
609                         osc_object_unlock(obj);
610                         if (grant > 0)
611                                 osc_unreserve_grant(cli, 0, grant);
612                 }
613
614                 osc_io_unplug_async(env, cli, obj);
615         }
616         osc_extent_put(env, ext);
617         RETURN(rc);
618 }
619
620 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
621 {
622         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
623 }
624
625 /**
626  * Find or create an extent which includes @index, core function to manage
627  * extent tree.
628  */
629 static struct osc_extent *osc_extent_find(const struct lu_env *env,
630                                           struct osc_object *obj, pgoff_t index,
631                                           unsigned int *grants)
632 {
633         struct client_obd *cli = osc_cli(obj);
634         struct osc_lock   *olck;
635         struct cl_lock_descr *descr;
636         struct osc_extent *cur;
637         struct osc_extent *ext;
638         struct osc_extent *conflict = NULL;
639         struct osc_extent *found = NULL;
640         pgoff_t    chunk;
641         pgoff_t    max_end;
642         unsigned int max_pages; /* max_pages_per_rpc */
643         unsigned int chunksize;
644         int        ppc_bits; /* pages per chunk bits */
645         pgoff_t    chunk_mask;
646         int        rc;
647         ENTRY;
648
649         cur = osc_extent_alloc(obj);
650         if (cur == NULL)
651                 RETURN(ERR_PTR(-ENOMEM));
652
653         olck = osc_env_io(env)->oi_write_osclock;
654         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
655         LASSERT(olck->ols_state == OLS_GRANTED);
656
657         descr = &olck->ols_cl.cls_lock->cll_descr;
658         LASSERT(descr->cld_mode >= CLM_WRITE);
659
660         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
661                  "chunkbits: %u\n", cli->cl_chunkbits);
662         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
663         chunk_mask = ~((1 << ppc_bits) - 1);
664         chunksize  = 1 << cli->cl_chunkbits;
665         chunk      = index >> ppc_bits;
666
667         /* align end to RPC edge. */
668         max_pages = cli->cl_max_pages_per_rpc;
669         if ((max_pages & ~chunk_mask) != 0) {
670                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
671                        max_pages, cli->cl_chunkbits, chunk_mask);
672                 RETURN(ERR_PTR(-EINVAL));
673         }
674         max_end = index - (index % max_pages) + max_pages - 1;
675         max_end = min_t(pgoff_t, max_end, descr->cld_end);
676
677         /* initialize new extent by parameters so far */
678         cur->oe_max_end = max_end;
679         cur->oe_start   = index & chunk_mask;
680         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
681         if (cur->oe_start < descr->cld_start)
682                 cur->oe_start = descr->cld_start;
683         if (cur->oe_end > max_end)
684                 cur->oe_end = max_end;
685         cur->oe_grants  = 0;
686         cur->oe_mppr    = max_pages;
687         if (olck->ols_dlmlock != NULL) {
688                 LASSERT(olck->ols_hold);
689                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
690                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
691         }
692
693         /* grants has been allocated by caller */
694         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
695                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
696         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
697                  EXTPARA(cur));
698
699 restart:
700         osc_object_lock(obj);
701         ext = osc_extent_search(obj, cur->oe_start);
702         if (ext == NULL)
703                 ext = first_extent(obj);
704         while (ext != NULL) {
705                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
706                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
707
708                 LASSERT(sanity_check_nolock(ext) == 0);
709                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
710                         break;
711
712                 /* if covering by different locks, no chance to match */
713                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
714                         EASSERTF(!overlapped(ext, cur), ext,
715                                  EXTSTR"\n", EXTPARA(cur));
716
717                         ext = next_extent(ext);
718                         continue;
719                 }
720
721                 /* discontiguous chunks? */
722                 if (chunk + 1 < ext_chk_start) {
723                         ext = next_extent(ext);
724                         continue;
725                 }
726
727                 /* ok, from now on, ext and cur have these attrs:
728                  * 1. covered by the same lock
729                  * 2. contiguous at chunk level or overlapping. */
730
731                 if (overlapped(ext, cur)) {
732                         /* cur is the minimum unit, so overlapping means
733                          * full contain. */
734                         EASSERTF((ext->oe_start <= cur->oe_start &&
735                                   ext->oe_end >= cur->oe_end),
736                                  ext, EXTSTR"\n", EXTPARA(cur));
737
738                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
739                                 /* for simplicity, we wait for this extent to
740                                  * finish before going forward. */
741                                 conflict = osc_extent_get(ext);
742                                 break;
743                         }
744
745                         found = osc_extent_hold(ext);
746                         break;
747                 }
748
749                 /* non-overlapped extent */
750                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
751                         /* we can't do anything for a non OES_CACHE extent, or
752                          * if there is someone waiting for this extent to be
753                          * flushed, try next one. */
754                         ext = next_extent(ext);
755                         continue;
756                 }
757
758                 /* check if they belong to the same rpc slot before trying to
759                  * merge. the extents are not overlapped and contiguous at
760                  * chunk level to get here. */
761                 if (ext->oe_max_end != max_end) {
762                         /* if they don't belong to the same RPC slot or
763                          * max_pages_per_rpc has ever changed, do not merge. */
764                         ext = next_extent(ext);
765                         continue;
766                 }
767
768                 /* check whether maximum extent size will be hit */
769                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
770                     cli->cl_max_extent_pages) {
771                         ext = next_extent(ext);
772                         continue;
773                 }
774
775                 /* it's required that an extent must be contiguous at chunk
776                  * level so that we know the whole extent is covered by grant
777                  * (the pages in the extent are NOT required to be contiguous).
778                  * Otherwise, it will be too much difficult to know which
779                  * chunks have grants allocated. */
780
781                 /* try to do front merge - extend ext's start */
782                 if (chunk + 1 == ext_chk_start) {
783                         /* ext must be chunk size aligned */
784                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
785
786                         /* pull ext's start back to cover cur */
787                         ext->oe_start   = cur->oe_start;
788                         ext->oe_grants += chunksize;
789                         LASSERT(*grants >= chunksize);
790                         *grants -= chunksize;
791
792                         found = osc_extent_hold(ext);
793                 } else if (chunk == ext_chk_end + 1) {
794                         /* rear merge */
795                         ext->oe_end     = cur->oe_end;
796                         ext->oe_grants += chunksize;
797                         LASSERT(*grants >= chunksize);
798                         *grants -= chunksize;
799
800                         /* try to merge with the next one because we just fill
801                          * in a gap */
802                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
803                                 /* we can save extent tax from next extent */
804                                 *grants += cli->cl_grant_extent_tax;
805
806                         found = osc_extent_hold(ext);
807                 }
808                 if (found != NULL)
809                         break;
810
811                 ext = next_extent(ext);
812         }
813
814         osc_extent_tree_dump(D_CACHE, obj);
815         if (found != NULL) {
816                 LASSERT(conflict == NULL);
817                 if (!IS_ERR(found)) {
818                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
819                         OSC_EXTENT_DUMP(D_CACHE, found,
820                                         "found caching ext for %lu.\n", index);
821                 }
822         } else if (conflict == NULL) {
823                 /* create a new extent */
824                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
825                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
826                 LASSERT(*grants >= cur->oe_grants);
827                 *grants -= cur->oe_grants;
828
829                 cur->oe_state = OES_CACHE;
830                 found = osc_extent_hold(cur);
831                 osc_extent_insert(obj, cur);
832                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
833                                 index, descr->cld_end);
834         }
835         osc_object_unlock(obj);
836
837         if (conflict != NULL) {
838                 LASSERT(found == NULL);
839
840                 /* waiting for IO to finish. Please notice that it's impossible
841                  * to be an OES_TRUNC extent. */
842                 rc = osc_extent_wait(env, conflict, OES_INV);
843                 osc_extent_put(env, conflict);
844                 conflict = NULL;
845                 if (rc < 0)
846                         GOTO(out, found = ERR_PTR(rc));
847
848                 goto restart;
849         }
850         EXIT;
851
852 out:
853         osc_extent_put(env, cur);
854         return found;
855 }
856
857 /**
858  * Called when IO is finished to an extent.
859  */
860 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
861                       int sent, int rc)
862 {
863         struct client_obd *cli = osc_cli(ext->oe_obj);
864         struct osc_async_page *oap;
865         struct osc_async_page *tmp;
866         int nr_pages = ext->oe_nr_pages;
867         int lost_grant = 0;
868         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
869         loff_t last_off = 0;
870         int last_count = -1;
871         ENTRY;
872
873         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
874
875         ext->oe_rc = rc ?: ext->oe_nr_pages;
876         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
877
878         osc_lru_add_batch(cli, &ext->oe_pages);
879         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
880                                      oap_pending_item) {
881                 list_del_init(&oap->oap_rpc_item);
882                 list_del_init(&oap->oap_pending_item);
883                 if (last_off <= oap->oap_obj_off) {
884                         last_off = oap->oap_obj_off;
885                         last_count = oap->oap_count;
886                 }
887
888                 --ext->oe_nr_pages;
889                 osc_ap_completion(env, cli, oap, sent, rc);
890         }
891         EASSERT(ext->oe_nr_pages == 0, ext);
892
893         if (!sent) {
894                 lost_grant = ext->oe_grants;
895         } else if (blocksize < PAGE_SIZE &&
896                    last_count != PAGE_SIZE) {
897                 /* For short writes we shouldn't count parts of pages that
898                  * span a whole chunk on the OST side, or our accounting goes
899                  * wrong.  Should match the code in filter_grant_check. */
900                 int offset = last_off & ~PAGE_MASK;
901                 int count = last_count + (offset & (blocksize - 1));
902                 int end = (offset + last_count) & (blocksize - 1);
903                 if (end)
904                         count += blocksize - end;
905
906                 lost_grant = PAGE_SIZE - count;
907         }
908         if (ext->oe_grants > 0)
909                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
910
911         osc_extent_remove(ext);
912         /* put the refcount for RPC */
913         osc_extent_put(env, ext);
914         RETURN(0);
915 }
916
917 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
918 {
919         int ret;
920
921         osc_object_lock(ext->oe_obj);
922         ret = ext->oe_state == state;
923         osc_object_unlock(ext->oe_obj);
924
925         return ret;
926 }
927
928 /**
929  * Wait for the extent's state to become @state.
930  */
931 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
932                            enum osc_extent_state state)
933 {
934         struct osc_object *obj = ext->oe_obj;
935         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
936                                                   LWI_ON_SIGNAL_NOOP, NULL);
937         int rc = 0;
938         ENTRY;
939
940         osc_object_lock(obj);
941         LASSERT(sanity_check_nolock(ext) == 0);
942         /* `Kick' this extent only if the caller is waiting for it to be
943          * written out. */
944         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
945                 if (ext->oe_state == OES_ACTIVE) {
946                         ext->oe_urgent = 1;
947                 } else if (ext->oe_state == OES_CACHE) {
948                         ext->oe_urgent = 1;
949                         osc_extent_hold(ext);
950                         rc = 1;
951                 }
952         }
953         osc_object_unlock(obj);
954         if (rc == 1)
955                 osc_extent_release(env, ext);
956
957         /* wait for the extent until its state becomes @state */
958         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
959         if (rc == -ETIMEDOUT) {
960                 OSC_EXTENT_DUMP(D_ERROR, ext,
961                         "%s: wait ext to %u timedout, recovery in progress?\n",
962                         cli_name(osc_cli(obj)), state);
963
964                 wait_event_idle(ext->oe_waitq, extent_wait_cb(ext, state));
965                 rc = 0;
966         }
967         if (rc == 0 && ext->oe_rc < 0)
968                 rc = ext->oe_rc;
969         RETURN(rc);
970 }
971
972 /**
973  * Discard pages with index greater than @size. If @ext is overlapped with
974  * @size, then partial truncate happens.
975  */
976 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
977                                 bool partial)
978 {
979         struct lu_env         *env;
980         struct cl_io          *io;
981         struct osc_object     *obj = ext->oe_obj;
982         struct client_obd     *cli = osc_cli(obj);
983         struct osc_async_page *oap;
984         struct osc_async_page *tmp;
985         struct pagevec        *pvec;
986         int                    pages_in_chunk = 0;
987         int                    ppc_bits    = cli->cl_chunkbits -
988                                              PAGE_SHIFT;
989         __u64                  trunc_chunk = trunc_index >> ppc_bits;
990         int                    grants   = 0;
991         int                    nr_pages = 0;
992         int                    rc       = 0;
993         __u16                  refcheck;
994         ENTRY;
995
996         LASSERT(sanity_check(ext) == 0);
997         LASSERT(ext->oe_state == OES_TRUNC);
998         LASSERT(!ext->oe_urgent);
999
1000         /* Request new lu_env.
1001          * We can't use that env from osc_cache_truncate_start() because
1002          * it's from lov_io_sub and not fully initialized. */
1003         env = cl_env_get(&refcheck);
1004         if (IS_ERR(env))
1005                 RETURN(PTR_ERR(env));
1006
1007         io  = osc_env_thread_io(env);
1008         io->ci_obj = cl_object_top(osc2cl(obj));
1009         io->ci_ignore_layout = 1;
1010         pvec = &osc_env_info(env)->oti_pagevec;
1011         ll_pagevec_init(pvec, 0);
1012         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1013         if (rc < 0)
1014                 GOTO(out, rc);
1015
1016         /* discard all pages with index greater than trunc_index */
1017         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1018                                      oap_pending_item) {
1019                 pgoff_t index = osc_index(oap2osc(oap));
1020                 struct cl_page  *page = oap2cl_page(oap);
1021
1022                 LASSERT(list_empty(&oap->oap_rpc_item));
1023
1024                 /* only discard the pages with their index greater than
1025                  * trunc_index, and ... */
1026                 if (index < trunc_index ||
1027                     (index == trunc_index && partial)) {
1028                         /* accounting how many pages remaining in the chunk
1029                          * so that we can calculate grants correctly. */
1030                         if (index >> ppc_bits == trunc_chunk)
1031                                 ++pages_in_chunk;
1032                         continue;
1033                 }
1034
1035                 list_del_init(&oap->oap_pending_item);
1036
1037                 cl_page_get(page);
1038                 lu_ref_add(&page->cp_reference, "truncate", current);
1039
1040                 if (cl_page_own(env, io, page) == 0) {
1041                         cl_page_discard(env, io, page);
1042                         cl_page_disown(env, io, page);
1043                 } else {
1044                         LASSERT(page->cp_state == CPS_FREEING);
1045                         LASSERT(0);
1046                 }
1047
1048                 lu_ref_del(&page->cp_reference, "truncate", current);
1049                 cl_pagevec_put(env, page, pvec);
1050
1051                 --ext->oe_nr_pages;
1052                 ++nr_pages;
1053         }
1054         pagevec_release(pvec);
1055
1056         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1057                       ext->oe_nr_pages == 0),
1058                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1059
1060         osc_object_lock(obj);
1061         if (ext->oe_nr_pages == 0) {
1062                 LASSERT(pages_in_chunk == 0);
1063                 grants = ext->oe_grants;
1064                 ext->oe_grants = 0;
1065         } else { /* calculate how many grants we can free */
1066                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1067                 pgoff_t last_index;
1068
1069
1070                 /* if there is no pages in this chunk, we can also free grants
1071                  * for the last chunk */
1072                 if (pages_in_chunk == 0) {
1073                         /* if this is the 1st chunk and no pages in this chunk,
1074                          * ext->oe_nr_pages must be zero, so we should be in
1075                          * the other if-clause. */
1076                         LASSERT(trunc_chunk > 0);
1077                         --trunc_chunk;
1078                         ++chunks;
1079                 }
1080
1081                 /* this is what we can free from this extent */
1082                 grants          = chunks << cli->cl_chunkbits;
1083                 ext->oe_grants -= grants;
1084                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1085                 ext->oe_end     = min(last_index, ext->oe_max_end);
1086                 LASSERT(ext->oe_end >= ext->oe_start);
1087                 LASSERT(ext->oe_grants > 0);
1088         }
1089         osc_object_unlock(obj);
1090
1091         if (grants > 0 || nr_pages > 0)
1092                 osc_free_grant(cli, nr_pages, grants, grants);
1093
1094 out:
1095         cl_io_fini(env, io);
1096         cl_env_put(env, &refcheck);
1097         RETURN(rc);
1098 }
1099
1100 /**
1101  * This function is used to make the extent prepared for transfer.
1102  * A race with flusing page - ll_writepage() has to be handled cautiously.
1103  */
1104 static int osc_extent_make_ready(const struct lu_env *env,
1105                                  struct osc_extent *ext)
1106 {
1107         struct osc_async_page *oap;
1108         struct osc_async_page *last = NULL;
1109         struct osc_object *obj = ext->oe_obj;
1110         unsigned int page_count = 0;
1111         int rc;
1112         ENTRY;
1113
1114         /* we're going to grab page lock, so object lock must not be taken. */
1115         LASSERT(sanity_check(ext) == 0);
1116         /* in locking state, any process should not touch this extent. */
1117         EASSERT(ext->oe_state == OES_LOCKING, ext);
1118         EASSERT(ext->oe_owner != NULL, ext);
1119
1120         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1121
1122         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1123                 ++page_count;
1124                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1125                         last = oap;
1126
1127                 /* checking ASYNC_READY is race safe */
1128                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1129                         continue;
1130
1131                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1132                 switch (rc) {
1133                 case 0:
1134                         spin_lock(&oap->oap_lock);
1135                         oap->oap_async_flags |= ASYNC_READY;
1136                         spin_unlock(&oap->oap_lock);
1137                         break;
1138                 case -EALREADY:
1139                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1140                         break;
1141                 default:
1142                         LASSERTF(0, "unknown return code: %d\n", rc);
1143                 }
1144         }
1145
1146         LASSERT(page_count == ext->oe_nr_pages);
1147         LASSERT(last != NULL);
1148         /* the last page is the only one we need to refresh its count by
1149          * the size of file. */
1150         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1151                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1152                 LASSERT(last_oap_count > 0);
1153                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1154                 last->oap_count = last_oap_count;
1155                 spin_lock(&last->oap_lock);
1156                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1157                 spin_unlock(&last->oap_lock);
1158         }
1159
1160         /* for the rest of pages, we don't need to call osf_refresh_count()
1161          * because it's known they are not the last page */
1162         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1163                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1164                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1165                         spin_lock(&oap->oap_lock);
1166                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1167                         spin_unlock(&oap->oap_lock);
1168                 }
1169         }
1170
1171         osc_object_lock(obj);
1172         osc_extent_state_set(ext, OES_RPC);
1173         osc_object_unlock(obj);
1174         /* get a refcount for RPC. */
1175         osc_extent_get(ext);
1176
1177         RETURN(0);
1178 }
1179
1180 /**
1181  * Quick and simple version of osc_extent_find(). This function is frequently
1182  * called to expand the extent for the same IO. To expand the extent, the
1183  * page index must be in the same or next chunk of ext->oe_end.
1184  */
1185 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1186                              unsigned int *grants)
1187 {
1188         struct osc_object *obj = ext->oe_obj;
1189         struct client_obd *cli = osc_cli(obj);
1190         struct osc_extent *next;
1191         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1192         pgoff_t chunk = index >> ppc_bits;
1193         pgoff_t end_chunk;
1194         pgoff_t end_index;
1195         unsigned int chunksize = 1 << cli->cl_chunkbits;
1196         int rc = 0;
1197         ENTRY;
1198
1199         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1200         osc_object_lock(obj);
1201         LASSERT(sanity_check_nolock(ext) == 0);
1202         end_chunk = ext->oe_end >> ppc_bits;
1203         if (chunk > end_chunk + 1)
1204                 GOTO(out, rc = -ERANGE);
1205
1206         if (end_chunk >= chunk)
1207                 GOTO(out, rc = 0);
1208
1209         LASSERT(end_chunk + 1 == chunk);
1210
1211         /* try to expand this extent to cover @index */
1212         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1213
1214         /* don't go over the maximum extent size reported by server */
1215         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1216                 GOTO(out, rc = -ERANGE);
1217
1218         next = next_extent(ext);
1219         if (next != NULL && next->oe_start <= end_index)
1220                 /* complex mode - overlapped with the next extent,
1221                  * this case will be handled by osc_extent_find() */
1222                 GOTO(out, rc = -EAGAIN);
1223
1224         ext->oe_end = end_index;
1225         ext->oe_grants += chunksize;
1226         LASSERT(*grants >= chunksize);
1227         *grants -= chunksize;
1228         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1229                  "overlapped after expanding for %lu.\n", index);
1230         EXIT;
1231
1232 out:
1233         osc_object_unlock(obj);
1234         RETURN(rc);
1235 }
1236
1237 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1238                                   const char *func, int line)
1239 {
1240         struct osc_extent *ext;
1241         int cnt;
1242
1243         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1244                 return;
1245
1246         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1247                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1248
1249         /* osc_object_lock(obj); */
1250         cnt = 1;
1251         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1252                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1253
1254         cnt = 1;
1255         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1256                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1257
1258         cnt = 1;
1259         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1260                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1261
1262         cnt = 1;
1263         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1264                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1265         /* osc_object_unlock(obj); */
1266 }
1267
1268 /* ------------------ osc extent end ------------------ */
1269
1270 static inline int osc_is_ready(struct osc_object *osc)
1271 {
1272         return !list_empty(&osc->oo_ready_item) ||
1273                !list_empty(&osc->oo_hp_ready_item);
1274 }
1275
1276 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1277         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1278                (OSC), osc_is_ready(OSC),                                       \
1279                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1280                list_empty_marker(&(OSC)->oo_ready_item),                       \
1281                atomic_read(&(OSC)->oo_nr_writes),                              \
1282                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1283                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1284                atomic_read(&(OSC)->oo_nr_reads),                               \
1285                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1286                ##args)
1287
1288 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1289                           int cmd)
1290 {
1291         struct osc_page *opg  = oap2osc_page(oap);
1292         struct cl_page  *page = oap2cl_page(oap);
1293         int result;
1294
1295         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1296
1297         ENTRY;
1298         result = cl_page_make_ready(env, page, CRT_WRITE);
1299         if (result == 0)
1300                 opg->ops_submit_time = ktime_get();
1301         RETURN(result);
1302 }
1303
1304 static int osc_refresh_count(const struct lu_env *env,
1305                              struct osc_async_page *oap, int cmd)
1306 {
1307         struct osc_page  *opg = oap2osc_page(oap);
1308         pgoff_t index = osc_index(oap2osc(oap));
1309         struct cl_object *obj;
1310         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1311         int result;
1312         loff_t kms;
1313
1314         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1315         LASSERT(!(cmd & OBD_BRW_READ));
1316         LASSERT(opg != NULL);
1317         obj = opg->ops_cl.cpl_obj;
1318
1319         cl_object_attr_lock(obj);
1320         result = cl_object_attr_get(env, obj, attr);
1321         cl_object_attr_unlock(obj);
1322         if (result < 0)
1323                 return result;
1324         kms = attr->cat_kms;
1325         if (cl_offset(obj, index) >= kms)
1326                 /* catch race with truncate */
1327                 return 0;
1328         else if (cl_offset(obj, index + 1) > kms)
1329                 /* catch sub-page write at end of file */
1330                 return kms % PAGE_SIZE;
1331         else
1332                 return PAGE_SIZE;
1333 }
1334
1335 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1336                           int cmd, int rc)
1337 {
1338         struct osc_page   *opg  = oap2osc_page(oap);
1339         struct cl_page    *page = oap2cl_page(oap);
1340         enum cl_req_type   crt;
1341         int srvlock;
1342
1343         ENTRY;
1344
1345         cmd &= ~OBD_BRW_NOQUOTA;
1346         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1347                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1348         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1349                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1350         LASSERT(opg->ops_transfer_pinned);
1351
1352         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1353         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1354         opg->ops_transfer_pinned = 0;
1355
1356         opg->ops_submit_time = ktime_set(0, 0);
1357         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1358
1359         /* statistic */
1360         if (rc == 0 && srvlock) {
1361                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1362                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1363                 size_t bytes = oap->oap_count;
1364
1365                 if (crt == CRT_READ)
1366                         stats->os_lockless_reads += bytes;
1367                 else
1368                         stats->os_lockless_writes += bytes;
1369         }
1370
1371         /*
1372          * This has to be the last operation with the page, as locks are
1373          * released in cl_page_completion() and nothing except for the
1374          * reference counter protects page from concurrent reclaim.
1375          */
1376         lu_ref_del(&page->cp_reference, "transfer", page);
1377
1378         cl_page_completion(env, page, crt, rc);
1379         cl_page_put(env, page);
1380
1381         RETURN(0);
1382 }
1383
1384 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1385         struct client_obd *__tmp = (cli);                               \
1386         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1387                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1388                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1389                "left: %ld, waiters: %d }" fmt "\n",                     \
1390                cli_name(__tmp),                                         \
1391                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1392                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1393                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1394                __tmp->cl_dirty_grant,                                   \
1395                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1396                atomic_long_read(&__tmp->cl_lru_in_list),                \
1397                atomic_long_read(&__tmp->cl_lru_busy),                   \
1398                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1399 } while (0)
1400
1401 /* caller must hold loi_list_lock */
1402 static void osc_consume_write_grant(struct client_obd *cli,
1403                                     struct brw_page *pga)
1404 {
1405         assert_spin_locked(&cli->cl_loi_list_lock);
1406         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1407         cli->cl_dirty_pages++;
1408         pga->flag |= OBD_BRW_FROM_GRANT;
1409         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1410                PAGE_SIZE, pga, pga->pg);
1411         osc_update_next_shrink(cli);
1412 }
1413
1414 /* the companion to osc_consume_write_grant, called when a brw has completed.
1415  * must be called with the loi lock held. */
1416 static void osc_release_write_grant(struct client_obd *cli,
1417                                     struct brw_page *pga)
1418 {
1419         ENTRY;
1420
1421         assert_spin_locked(&cli->cl_loi_list_lock);
1422         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1423                 EXIT;
1424                 return;
1425         }
1426
1427         pga->flag &= ~OBD_BRW_FROM_GRANT;
1428         atomic_long_dec(&obd_dirty_pages);
1429         cli->cl_dirty_pages--;
1430         if (pga->flag & OBD_BRW_NOCACHE) {
1431                 pga->flag &= ~OBD_BRW_NOCACHE;
1432                 atomic_long_dec(&obd_dirty_transit_pages);
1433                 cli->cl_dirty_transit--;
1434         }
1435         EXIT;
1436 }
1437
1438 /**
1439  * To avoid sleeping with object lock held, it's good for us allocate enough
1440  * grants before entering into critical section.
1441  *
1442  * client_obd_list_lock held by caller
1443  */
1444 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1445 {
1446         int rc = -EDQUOT;
1447
1448         if (cli->cl_avail_grant >= bytes) {
1449                 cli->cl_avail_grant    -= bytes;
1450                 cli->cl_reserved_grant += bytes;
1451                 rc = 0;
1452         }
1453         return rc;
1454 }
1455
1456 static void __osc_unreserve_grant(struct client_obd *cli,
1457                                   unsigned int reserved, unsigned int unused)
1458 {
1459         /* it's quite normal for us to get more grant than reserved.
1460          * Thinking about a case that two extents merged by adding a new
1461          * chunk, we can save one extent tax. If extent tax is greater than
1462          * one chunk, we can save more grant by adding a new chunk */
1463         cli->cl_reserved_grant -= reserved;
1464         if (unused > reserved) {
1465                 cli->cl_avail_grant += reserved;
1466                 cli->cl_lost_grant  += unused - reserved;
1467                 cli->cl_dirty_grant -= unused - reserved;
1468         } else {
1469                 cli->cl_avail_grant += unused;
1470                 cli->cl_dirty_grant += reserved - unused;
1471         }
1472 }
1473
1474 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1475                                        unsigned int reserved,
1476                                        unsigned int unused)
1477 {
1478         __osc_unreserve_grant(cli, reserved, unused);
1479         if (unused > 0)
1480                 osc_wake_cache_waiters(cli);
1481 }
1482
1483 static void osc_unreserve_grant(struct client_obd *cli,
1484                                 unsigned int reserved, unsigned int unused)
1485 {
1486         spin_lock(&cli->cl_loi_list_lock);
1487         osc_unreserve_grant_nolock(cli, reserved, unused);
1488         spin_unlock(&cli->cl_loi_list_lock);
1489 }
1490
1491 /**
1492  * Free grant after IO is finished or canceled.
1493  *
1494  * @lost_grant is used to remember how many grants we have allocated but not
1495  * used, we should return these grants to OST. There're two cases where grants
1496  * can be lost:
1497  * 1. truncate;
1498  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1499  *    written. In this case OST may use less chunks to serve this partial
1500  *    write. OSTs don't actually know the page size on the client side. so
1501  *    clients have to calculate lost grant by the blocksize on the OST.
1502  *    See filter_grant_check() for details.
1503  */
1504 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1505                            unsigned int lost_grant, unsigned int dirty_grant)
1506 {
1507         unsigned long grant;
1508
1509         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1510
1511         spin_lock(&cli->cl_loi_list_lock);
1512         atomic_long_sub(nr_pages, &obd_dirty_pages);
1513         cli->cl_dirty_pages -= nr_pages;
1514         cli->cl_lost_grant += lost_grant;
1515         cli->cl_dirty_grant -= dirty_grant;
1516         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1517                 /* borrow some grant from truncate to avoid the case that
1518                  * truncate uses up all avail grant */
1519                 cli->cl_lost_grant -= grant;
1520                 cli->cl_avail_grant += grant;
1521         }
1522         osc_wake_cache_waiters(cli);
1523         spin_unlock(&cli->cl_loi_list_lock);
1524         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1525                lost_grant, cli->cl_lost_grant,
1526                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1527                cli->cl_dirty_grant);
1528 }
1529
1530 /**
1531  * The companion to osc_enter_cache(), called when @oap is no longer part of
1532  * the dirty accounting due to error.
1533  */
1534 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1535 {
1536         spin_lock(&cli->cl_loi_list_lock);
1537         osc_release_write_grant(cli, &oap->oap_brw_page);
1538         spin_unlock(&cli->cl_loi_list_lock);
1539 }
1540
1541 /**
1542  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1543  * is available.
1544  */
1545 static int osc_enter_cache_try(struct client_obd *cli,
1546                                struct osc_async_page *oap,
1547                                int bytes, int transient)
1548 {
1549         int rc;
1550
1551         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1552
1553         rc = osc_reserve_grant(cli, bytes);
1554         if (rc < 0)
1555                 return 0;
1556
1557         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1558                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1559                     obd_max_dirty_pages) {
1560                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1561                         if (transient) {
1562                                 cli->cl_dirty_transit++;
1563                                 atomic_long_inc(&obd_dirty_transit_pages);
1564                                 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1565                         }
1566                         rc = 1;
1567                         goto out;
1568                 } else
1569                         atomic_long_dec(&obd_dirty_pages);
1570         }
1571         __osc_unreserve_grant(cli, bytes, bytes);
1572
1573 out:
1574         return rc;
1575 }
1576
1577 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1578 {
1579         int rc;
1580         spin_lock(&cli->cl_loi_list_lock);
1581         rc = list_empty(&ocw->ocw_entry);
1582         spin_unlock(&cli->cl_loi_list_lock);
1583         return rc;
1584 }
1585
1586 /**
1587  * The main entry to reserve dirty page accounting. Usually the grant reserved
1588  * in this function will be freed in bulk in osc_free_grant() unless it fails
1589  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1590  *
1591  * The process will be put into sleep if it's already run out of grant.
1592  */
1593 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1594                            struct osc_async_page *oap, int bytes)
1595 {
1596         struct osc_object       *osc = oap->oap_obj;
1597         struct lov_oinfo        *loi = osc->oo_oinfo;
1598         struct osc_cache_waiter  ocw;
1599         struct l_wait_info       lwi;
1600         int                      rc = -EDQUOT;
1601         ENTRY;
1602
1603         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1604                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1605
1606         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1607
1608         spin_lock(&cli->cl_loi_list_lock);
1609
1610         /* force the caller to try sync io.  this can jump the list
1611          * of queued writes and create a discontiguous rpc stream */
1612         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1613             cli->cl_dirty_max_pages == 0 ||
1614             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1615                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1616                 GOTO(out, rc = -EDQUOT);
1617         }
1618
1619         /* Hopefully normal case - cache space and write credits available */
1620         if (list_empty(&cli->cl_cache_waiters) &&
1621             osc_enter_cache_try(cli, oap, bytes, 0)) {
1622                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1623                 GOTO(out, rc = 0);
1624         }
1625
1626         /* We can get here for two reasons: too many dirty pages in cache, or
1627          * run out of grants. In both cases we should write dirty pages out.
1628          * Adding a cache waiter will trigger urgent write-out no matter what
1629          * RPC size will be.
1630          * The exiting condition is no avail grants and no dirty pages caching,
1631          * that really means there is no space on the OST. */
1632         init_waitqueue_head(&ocw.ocw_waitq);
1633         ocw.ocw_oap   = oap;
1634         ocw.ocw_grant = bytes;
1635         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1636                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1637                 ocw.ocw_rc = 0;
1638                 spin_unlock(&cli->cl_loi_list_lock);
1639
1640                 osc_io_unplug_async(env, cli, NULL);
1641
1642                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1643                        cli_name(cli), &ocw, oap);
1644
1645                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1646
1647                 spin_lock(&cli->cl_loi_list_lock);
1648
1649                 if (rc < 0) {
1650                         /* l_wait_event is interrupted by signal or timed out */
1651                         list_del_init(&ocw.ocw_entry);
1652                         break;
1653                 }
1654                 LASSERT(list_empty(&ocw.ocw_entry));
1655                 rc = ocw.ocw_rc;
1656
1657                 if (rc != -EDQUOT)
1658                         break;
1659                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1660                         rc = 0;
1661                         break;
1662                 }
1663         }
1664
1665         switch (rc) {
1666         case 0:
1667                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1668                 break;
1669         case -ETIMEDOUT:
1670                 OSC_DUMP_GRANT(D_CACHE, cli,
1671                                "timeout, fall back to sync i/o\n");
1672                 osc_extent_tree_dump(D_CACHE, osc);
1673                 /* fall back to synchronous I/O */
1674                 rc = -EDQUOT;
1675                 break;
1676         case -EINTR:
1677                 /* Ensures restartability - LU-3581 */
1678                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1679                 rc = -ERESTARTSYS;
1680                 break;
1681         case -EDQUOT:
1682                 OSC_DUMP_GRANT(D_CACHE, cli,
1683                                "no grant space, fall back to sync i/o\n");
1684                 break;
1685         default:
1686                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1687                        "due to %d, fall back to sync i/o\n",
1688                        cli_name(cli), &ocw, rc);
1689                 break;
1690         }
1691         EXIT;
1692 out:
1693         spin_unlock(&cli->cl_loi_list_lock);
1694         RETURN(rc);
1695 }
1696
1697 /* caller must hold loi_list_lock */
1698 void osc_wake_cache_waiters(struct client_obd *cli)
1699 {
1700         struct list_head *l, *tmp;
1701         struct osc_cache_waiter *ocw;
1702
1703         ENTRY;
1704         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1705                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1706
1707                 ocw->ocw_rc = -EDQUOT;
1708
1709                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1710                         ocw->ocw_rc = 0;
1711
1712                 if (ocw->ocw_rc == 0 ||
1713                     !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1714                         list_del_init(&ocw->ocw_entry);
1715                         CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1716                                "%ld, %d\n", ocw, ocw->ocw_oap,
1717                                cli->cl_avail_grant, ocw->ocw_rc);
1718
1719                         wake_up(&ocw->ocw_waitq);
1720                 }
1721         }
1722
1723         EXIT;
1724 }
1725 EXPORT_SYMBOL(osc_wake_cache_waiters);
1726
1727 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1728 {
1729         int hprpc = !!list_empty(&osc->oo_hp_exts);
1730         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1731 }
1732
1733 /* This maintains the lists of pending pages to read/write for a given object
1734  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1735  * to quickly find objects that are ready to send an RPC. */
1736 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1737                          int cmd)
1738 {
1739         int invalid_import = 0;
1740         ENTRY;
1741
1742         /* if we have an invalid import we want to drain the queued pages
1743          * by forcing them through rpcs that immediately fail and complete
1744          * the pages.  recovery relies on this to empty the queued pages
1745          * before canceling the locks and evicting down the llite pages */
1746         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1747                 invalid_import = 1;
1748
1749         if (cmd & OBD_BRW_WRITE) {
1750                 if (atomic_read(&osc->oo_nr_writes) == 0)
1751                         RETURN(0);
1752                 if (invalid_import) {
1753                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1754                         RETURN(1);
1755                 }
1756                 if (!list_empty(&osc->oo_hp_exts)) {
1757                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1758                         RETURN(1);
1759                 }
1760                 if (!list_empty(&osc->oo_urgent_exts)) {
1761                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1762                         RETURN(1);
1763                 }
1764                 /* trigger a write rpc stream as long as there are dirtiers
1765                  * waiting for space.  as they're waiting, they're not going to
1766                  * create more pages to coalesce with what's waiting.. */
1767                 if (!list_empty(&cli->cl_cache_waiters)) {
1768                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1769                         RETURN(1);
1770                 }
1771                 if (!list_empty(&osc->oo_full_exts)) {
1772                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1773                         RETURN(1);
1774                 }
1775         } else {
1776                 if (atomic_read(&osc->oo_nr_reads) == 0)
1777                         RETURN(0);
1778                 if (invalid_import) {
1779                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1780                         RETURN(1);
1781                 }
1782                 /* all read are urgent. */
1783                 if (!list_empty(&osc->oo_reading_exts))
1784                         RETURN(1);
1785         }
1786
1787         RETURN(0);
1788 }
1789
1790 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1791 {
1792         struct client_obd *cli = osc_cli(obj);
1793         if (cmd & OBD_BRW_WRITE) {
1794                 atomic_add(delta, &obj->oo_nr_writes);
1795                 atomic_add(delta, &cli->cl_pending_w_pages);
1796                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1797         } else {
1798                 atomic_add(delta, &obj->oo_nr_reads);
1799                 atomic_add(delta, &cli->cl_pending_r_pages);
1800                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1801         }
1802         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1803 }
1804
1805 static int osc_makes_hprpc(struct osc_object *obj)
1806 {
1807         return !list_empty(&obj->oo_hp_exts);
1808 }
1809
1810 static void on_list(struct list_head *item, struct list_head *list,
1811                     int should_be_on)
1812 {
1813         if (list_empty(item) && should_be_on)
1814                 list_add_tail(item, list);
1815         else if (!list_empty(item) && !should_be_on)
1816                 list_del_init(item);
1817 }
1818
1819 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1820  * can find pages to build into rpcs quickly */
1821 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1822 {
1823         if (osc_makes_hprpc(osc)) {
1824                 /* HP rpc */
1825                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1826                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1827         } else {
1828                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1829                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1830                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1831                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1832         }
1833
1834         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1835                 atomic_read(&osc->oo_nr_writes) > 0);
1836
1837         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1838                 atomic_read(&osc->oo_nr_reads) > 0);
1839
1840         return osc_is_ready(osc);
1841 }
1842
1843 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1844 {
1845         int is_ready;
1846
1847         spin_lock(&cli->cl_loi_list_lock);
1848         is_ready = __osc_list_maint(cli, osc);
1849         spin_unlock(&cli->cl_loi_list_lock);
1850
1851         return is_ready;
1852 }
1853
1854 /* this is trying to propogate async writeback errors back up to the
1855  * application.  As an async write fails we record the error code for later if
1856  * the app does an fsync.  As long as errors persist we force future rpcs to be
1857  * sync so that the app can get a sync error and break the cycle of queueing
1858  * pages for which writeback will fail. */
1859 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1860                            int rc)
1861 {
1862         if (rc) {
1863                 if (!ar->ar_rc)
1864                         ar->ar_rc = rc;
1865
1866                 ar->ar_force_sync = 1;
1867                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1868                 return;
1869
1870         }
1871
1872         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1873                 ar->ar_force_sync = 0;
1874 }
1875
1876 /* this must be called holding the loi list lock to give coverage to exit_cache,
1877  * async_flag maintenance, and oap_request */
1878 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1879                               struct osc_async_page *oap, int sent, int rc)
1880 {
1881         struct osc_object *osc = oap->oap_obj;
1882         struct lov_oinfo  *loi = osc->oo_oinfo;
1883         __u64 xid = 0;
1884
1885         ENTRY;
1886         if (oap->oap_request != NULL) {
1887                 xid = ptlrpc_req_xid(oap->oap_request);
1888                 ptlrpc_req_finished(oap->oap_request);
1889                 oap->oap_request = NULL;
1890         }
1891
1892         /* As the transfer for this page is being done, clear the flags */
1893         spin_lock(&oap->oap_lock);
1894         oap->oap_async_flags = 0;
1895         spin_unlock(&oap->oap_lock);
1896         oap->oap_interrupted = 0;
1897
1898         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1899                 spin_lock(&cli->cl_loi_list_lock);
1900                 osc_process_ar(&cli->cl_ar, xid, rc);
1901                 osc_process_ar(&loi->loi_ar, xid, rc);
1902                 spin_unlock(&cli->cl_loi_list_lock);
1903         }
1904
1905         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1906         if (rc)
1907                 CERROR("completion on oap %p obj %p returns %d.\n",
1908                        oap, osc, rc);
1909
1910         EXIT;
1911 }
1912
1913 struct extent_rpc_data {
1914         struct list_head        *erd_rpc_list;
1915         unsigned int            erd_page_count;
1916         unsigned int            erd_max_pages;
1917         unsigned int            erd_max_chunks;
1918         unsigned int            erd_max_extents;
1919 };
1920
1921 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1922 {
1923         struct client_obd *cli = osc_cli(ext->oe_obj);
1924         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1925
1926         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1927 }
1928
1929 /**
1930  * Try to add extent to one RPC. We need to think about the following things:
1931  * - # of pages must not be over max_pages_per_rpc
1932  * - extent must be compatible with previous ones
1933  */
1934 static int try_to_add_extent_for_io(struct client_obd *cli,
1935                                     struct osc_extent *ext,
1936                                     struct extent_rpc_data *data)
1937 {
1938         struct osc_extent *tmp;
1939         unsigned int chunk_count;
1940         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1941                                                       struct osc_async_page,
1942                                                       oap_pending_item);
1943         ENTRY;
1944
1945         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1946                 ext);
1947         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1948
1949         if (data->erd_max_extents == 0)
1950                 RETURN(0);
1951
1952         chunk_count = osc_extent_chunks(ext);
1953         EASSERTF(data->erd_page_count != 0 ||
1954                  chunk_count <= data->erd_max_chunks, ext,
1955                  "The first extent to be fit in a RPC contains %u chunks, "
1956                  "which is over the limit %u.\n", chunk_count,
1957                  data->erd_max_chunks);
1958         if (chunk_count > data->erd_max_chunks)
1959                 RETURN(0);
1960
1961         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1962         EASSERTF(data->erd_page_count != 0 ||
1963                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1964                 "The first extent to be fit in a RPC contains %u pages, "
1965                 "which is over the limit %u.\n", ext->oe_nr_pages,
1966                 data->erd_max_pages);
1967         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1968                 RETURN(0);
1969
1970         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1971                 struct osc_async_page *oap2;
1972                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1973                                         oap_pending_item);
1974                 EASSERT(tmp->oe_owner == current, tmp);
1975 #if 0
1976                 if (overlapped(tmp, ext)) {
1977                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1978                         EASSERT(0, ext);
1979                 }
1980 #endif
1981                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1982                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1983                                "in one RPC\n");
1984                         RETURN(0);
1985                 }
1986
1987                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1988                     !tmp->oe_grants != !ext->oe_grants ||
1989                     tmp->oe_ndelay != ext->oe_ndelay ||
1990                     tmp->oe_no_merge || ext->oe_no_merge)
1991                         RETURN(0);
1992
1993                 /* remove break for strict check */
1994                 break;
1995         }
1996
1997         data->erd_max_extents--;
1998         data->erd_max_chunks -= chunk_count;
1999         data->erd_page_count += ext->oe_nr_pages;
2000         list_move_tail(&ext->oe_link, data->erd_rpc_list);
2001         ext->oe_owner = current;
2002         RETURN(1);
2003 }
2004
2005 /**
2006  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
2007  * get_write_extent() takes all appropriate extents in atomic.
2008  *
2009  * The following policy is used to collect extents for IO:
2010  * 1. Add as many HP extents as possible;
2011  * 2. Add the first urgent extent in urgent extent list and take it out of
2012  *    urgent list;
2013  * 3. Add subsequent extents of this urgent extent;
2014  * 4. If urgent list is not empty, goto 2;
2015  * 5. Traverse the extent tree from the 1st extent;
2016  * 6. Above steps exit if there is no space in this RPC.
2017  */
2018 static unsigned int get_write_extents(struct osc_object *obj,
2019                                       struct list_head *rpclist)
2020 {
2021         struct client_obd *cli = osc_cli(obj);
2022         struct osc_extent *ext;
2023         struct extent_rpc_data data = {
2024                 .erd_rpc_list   = rpclist,
2025                 .erd_page_count = 0,
2026                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2027                 .erd_max_chunks = osc_max_write_chunks(cli),
2028                 .erd_max_extents = 256,
2029         };
2030
2031         LASSERT(osc_object_is_locked(obj));
2032         while (!list_empty(&obj->oo_hp_exts)) {
2033                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2034                                  oe_link);
2035                 LASSERT(ext->oe_state == OES_CACHE);
2036                 if (!try_to_add_extent_for_io(cli, ext, &data))
2037                         return data.erd_page_count;
2038                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2039         }
2040         if (data.erd_page_count == data.erd_max_pages)
2041                 return data.erd_page_count;
2042
2043         while (!list_empty(&obj->oo_urgent_exts)) {
2044                 ext = list_entry(obj->oo_urgent_exts.next,
2045                                  struct osc_extent, oe_link);
2046                 if (!try_to_add_extent_for_io(cli, ext, &data))
2047                         return data.erd_page_count;
2048         }
2049         if (data.erd_page_count == data.erd_max_pages)
2050                 return data.erd_page_count;
2051
2052         /* One key difference between full extents and other extents: full
2053          * extents can usually only be added if the rpclist was empty, so if we
2054          * can't add one, we continue on to trying to add normal extents.  This
2055          * is so we don't miss adding extra extents to an RPC containing high
2056          * priority or urgent extents. */
2057         while (!list_empty(&obj->oo_full_exts)) {
2058                 ext = list_entry(obj->oo_full_exts.next,
2059                                  struct osc_extent, oe_link);
2060                 if (!try_to_add_extent_for_io(cli, ext, &data))
2061                         break;
2062         }
2063         if (data.erd_page_count == data.erd_max_pages)
2064                 return data.erd_page_count;
2065
2066         ext = first_extent(obj);
2067         while (ext != NULL) {
2068                 if ((ext->oe_state != OES_CACHE) ||
2069                     /* this extent may be already in current rpclist */
2070                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2071                         ext = next_extent(ext);
2072                         continue;
2073                 }
2074
2075                 if (!try_to_add_extent_for_io(cli, ext, &data))
2076                         return data.erd_page_count;
2077
2078                 ext = next_extent(ext);
2079         }
2080         return data.erd_page_count;
2081 }
2082
2083 static int
2084 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2085                    struct osc_object *osc)
2086 __must_hold(osc)
2087 {
2088         LIST_HEAD(rpclist);
2089         struct osc_extent *ext;
2090         struct osc_extent *tmp;
2091         struct osc_extent *first = NULL;
2092         unsigned int page_count = 0;
2093         int srvlock = 0;
2094         int rc = 0;
2095         ENTRY;
2096
2097         LASSERT(osc_object_is_locked(osc));
2098
2099         page_count = get_write_extents(osc, &rpclist);
2100         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2101
2102         if (list_empty(&rpclist))
2103                 RETURN(0);
2104
2105         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2106
2107         list_for_each_entry(ext, &rpclist, oe_link) {
2108                 LASSERT(ext->oe_state == OES_CACHE ||
2109                         ext->oe_state == OES_LOCK_DONE);
2110                 if (ext->oe_state == OES_CACHE)
2111                         osc_extent_state_set(ext, OES_LOCKING);
2112                 else
2113                         osc_extent_state_set(ext, OES_RPC);
2114         }
2115
2116         /* we're going to grab page lock, so release object lock because
2117          * lock order is page lock -> object lock. */
2118         osc_object_unlock(osc);
2119
2120         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2121                 if (ext->oe_state == OES_LOCKING) {
2122                         rc = osc_extent_make_ready(env, ext);
2123                         if (unlikely(rc < 0)) {
2124                                 list_del_init(&ext->oe_link);
2125                                 osc_extent_finish(env, ext, 0, rc);
2126                                 continue;
2127                         }
2128                 }
2129                 if (first == NULL) {
2130                         first = ext;
2131                         srvlock = ext->oe_srvlock;
2132                 } else {
2133                         LASSERT(srvlock == ext->oe_srvlock);
2134                 }
2135         }
2136
2137         if (!list_empty(&rpclist)) {
2138                 LASSERT(page_count > 0);
2139                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2140                 LASSERT(list_empty(&rpclist));
2141         }
2142
2143         osc_object_lock(osc);
2144         RETURN(rc);
2145 }
2146
2147 /**
2148  * prepare pages for ASYNC io and put pages in send queue.
2149  *
2150  * \param cmd OBD_BRW_* macroses
2151  * \param lop pending pages
2152  *
2153  * \return zero if no page added to send queue.
2154  * \return 1 if pages successfully added to send queue.
2155  * \return negative on errors.
2156  */
2157 static int
2158 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2159                   struct osc_object *osc)
2160 __must_hold(osc)
2161 {
2162         struct osc_extent *ext;
2163         struct osc_extent *next;
2164         LIST_HEAD(rpclist);
2165         struct extent_rpc_data data = {
2166                 .erd_rpc_list   = &rpclist,
2167                 .erd_page_count = 0,
2168                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2169                 .erd_max_chunks = UINT_MAX,
2170                 .erd_max_extents = UINT_MAX,
2171         };
2172         int rc = 0;
2173         ENTRY;
2174
2175         LASSERT(osc_object_is_locked(osc));
2176         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2177                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2178                 if (!try_to_add_extent_for_io(cli, ext, &data))
2179                         break;
2180                 osc_extent_state_set(ext, OES_RPC);
2181                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2182         }
2183         LASSERT(data.erd_page_count <= data.erd_max_pages);
2184
2185         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2186
2187         if (!list_empty(&rpclist)) {
2188                 osc_object_unlock(osc);
2189
2190                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2191                 LASSERT(list_empty(&rpclist));
2192
2193                 osc_object_lock(osc);
2194         }
2195         RETURN(rc);
2196 }
2197
2198 #define list_to_obj(list, item) ({                                            \
2199         struct list_head *__tmp = (list)->next;                               \
2200         list_del_init(__tmp);                                         \
2201         list_entry(__tmp, struct osc_object, oo_##item);                      \
2202 })
2203
2204 /* This is called by osc_check_rpcs() to find which objects have pages that
2205  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2206 static struct osc_object *osc_next_obj(struct client_obd *cli)
2207 {
2208         ENTRY;
2209
2210         /* First return objects that have blocked locks so that they
2211          * will be flushed quickly and other clients can get the lock,
2212          * then objects which have pages ready to be stuffed into RPCs */
2213         if (!list_empty(&cli->cl_loi_hp_ready_list))
2214                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2215         if (!list_empty(&cli->cl_loi_ready_list))
2216                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2217
2218         /* then if we have cache waiters, return all objects with queued
2219          * writes.  This is especially important when many small files
2220          * have filled up the cache and not been fired into rpcs because
2221          * they don't pass the nr_pending/object threshhold */
2222         if (!list_empty(&cli->cl_cache_waiters) &&
2223             !list_empty(&cli->cl_loi_write_list))
2224                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2225
2226         /* then return all queued objects when we have an invalid import
2227          * so that they get flushed */
2228         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2229                 if (!list_empty(&cli->cl_loi_write_list))
2230                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2231                                            write_item));
2232                 if (!list_empty(&cli->cl_loi_read_list))
2233                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2234                                            read_item));
2235         }
2236         RETURN(NULL);
2237 }
2238
2239 /* called with the loi list lock held */
2240 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2241 __must_hold(&cli->cl_loi_list_lock)
2242 {
2243         struct osc_object *osc;
2244         int rc = 0;
2245         ENTRY;
2246
2247         while ((osc = osc_next_obj(cli)) != NULL) {
2248                 struct cl_object *obj = osc2cl(osc);
2249                 struct lu_ref_link link;
2250
2251                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2252
2253                 if (osc_max_rpc_in_flight(cli, osc)) {
2254                         __osc_list_maint(cli, osc);
2255                         break;
2256                 }
2257
2258                 cl_object_get(obj);
2259                 spin_unlock(&cli->cl_loi_list_lock);
2260                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2261
2262                 /* attempt some read/write balancing by alternating between
2263                  * reads and writes in an object.  The makes_rpc checks here
2264                  * would be redundant if we were getting read/write work items
2265                  * instead of objects.  we don't want send_oap_rpc to drain a
2266                  * partial read pending queue when we're given this object to
2267                  * do io on writes while there are cache waiters */
2268                 osc_object_lock(osc);
2269                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2270                         rc = osc_send_write_rpc(env, cli, osc);
2271                         if (rc < 0) {
2272                                 CERROR("Write request failed with %d\n", rc);
2273
2274                                 /* osc_send_write_rpc failed, mostly because of
2275                                  * memory pressure.
2276                                  *
2277                                  * It can't break here, because if:
2278                                  *  - a page was submitted by osc_io_submit, so
2279                                  *    page locked;
2280                                  *  - no request in flight
2281                                  *  - no subsequent request
2282                                  * The system will be in live-lock state,
2283                                  * because there is no chance to call
2284                                  * osc_io_unplug() and osc_check_rpcs() any
2285                                  * more. pdflush can't help in this case,
2286                                  * because it might be blocked at grabbing
2287                                  * the page lock as we mentioned.
2288                                  *
2289                                  * Anyway, continue to drain pages. */
2290                                 /* break; */
2291                         }
2292                 }
2293                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2294                         rc = osc_send_read_rpc(env, cli, osc);
2295                         if (rc < 0)
2296                                 CERROR("Read request failed with %d\n", rc);
2297                 }
2298                 osc_object_unlock(osc);
2299
2300                 osc_list_maint(cli, osc);
2301                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2302                 cl_object_put(env, obj);
2303
2304                 spin_lock(&cli->cl_loi_list_lock);
2305         }
2306 }
2307
2308 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2309                    struct osc_object *osc, int async)
2310 {
2311         int rc = 0;
2312
2313         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2314                 return 0;
2315
2316         if (!async) {
2317                 spin_lock(&cli->cl_loi_list_lock);
2318                 osc_check_rpcs(env, cli);
2319                 spin_unlock(&cli->cl_loi_list_lock);
2320         } else {
2321                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2322                 LASSERT(cli->cl_writeback_work != NULL);
2323                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2324         }
2325         return rc;
2326 }
2327 EXPORT_SYMBOL(osc_io_unplug0);
2328
2329 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2330                         struct page *page, loff_t offset)
2331 {
2332         struct obd_export     *exp = osc_export(osc);
2333         struct osc_async_page *oap = &ops->ops_oap;
2334         ENTRY;
2335
2336         if (!page)
2337                 return cfs_size_round(sizeof(*oap));
2338
2339         oap->oap_magic = OAP_MAGIC;
2340         oap->oap_cli = &exp->exp_obd->u.cli;
2341         oap->oap_obj = osc;
2342
2343         oap->oap_page = page;
2344         oap->oap_obj_off = offset;
2345         LASSERT(!(offset & ~PAGE_MASK));
2346
2347         INIT_LIST_HEAD(&oap->oap_pending_item);
2348         INIT_LIST_HEAD(&oap->oap_rpc_item);
2349
2350         spin_lock_init(&oap->oap_lock);
2351         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2352                oap, page, oap->oap_obj_off);
2353         RETURN(0);
2354 }
2355 EXPORT_SYMBOL(osc_prep_async_page);
2356
2357 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2358                        struct osc_page *ops, cl_commit_cbt cb)
2359 {
2360         struct osc_io *oio = osc_env_io(env);
2361         struct osc_extent     *ext = NULL;
2362         struct osc_async_page *oap = &ops->ops_oap;
2363         struct client_obd     *cli = oap->oap_cli;
2364         struct osc_object     *osc = oap->oap_obj;
2365         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2366         pgoff_t index;
2367         unsigned int tmp;
2368         unsigned int grants = 0;
2369         u32    brw_flags = OBD_BRW_ASYNC;
2370         int    cmd = OBD_BRW_WRITE;
2371         int    need_release = 0;
2372         int    rc = 0;
2373         ENTRY;
2374
2375         if (oap->oap_magic != OAP_MAGIC)
2376                 RETURN(-EINVAL);
2377
2378         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2379                 RETURN(-EIO);
2380
2381         if (!list_empty(&oap->oap_pending_item) ||
2382             !list_empty(&oap->oap_rpc_item))
2383                 RETURN(-EBUSY);
2384
2385         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2386         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2387         if (oio->oi_cap_sys_resource) {
2388                 brw_flags |= OBD_BRW_NOQUOTA;
2389                 cmd |= OBD_BRW_NOQUOTA;
2390         }
2391
2392         /* check if the file's owner/group is over quota */
2393         if (!(cmd & OBD_BRW_NOQUOTA)) {
2394                 struct cl_object *obj;
2395                 struct cl_attr   *attr;
2396                 unsigned int qid[LL_MAXQUOTAS];
2397
2398                 obj = cl_object_top(&osc->oo_cl);
2399                 attr = &osc_env_info(env)->oti_attr;
2400
2401                 cl_object_attr_lock(obj);
2402                 rc = cl_object_attr_get(env, obj, attr);
2403                 cl_object_attr_unlock(obj);
2404
2405                 qid[USRQUOTA] = attr->cat_uid;
2406                 qid[GRPQUOTA] = attr->cat_gid;
2407                 qid[PRJQUOTA] = attr->cat_projid;
2408                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2409                         rc = -EDQUOT;
2410                 if (rc)
2411                         RETURN(rc);
2412         }
2413
2414         oap->oap_cmd = cmd;
2415         oap->oap_page_off = ops->ops_from;
2416         oap->oap_count = ops->ops_to - ops->ops_from;
2417         /* No need to hold a lock here,
2418          * since this page is not in any list yet. */
2419         oap->oap_async_flags = 0;
2420         oap->oap_brw_flags = brw_flags;
2421
2422         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2423                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2424
2425         index = osc_index(oap2osc(oap));
2426
2427         /* Add this page into extent by the following steps:
2428          * 1. if there exists an active extent for this IO, mostly this page
2429          *    can be added to the active extent and sometimes we need to
2430          *    expand extent to accomodate this page;
2431          * 2. otherwise, a new extent will be allocated. */
2432
2433         ext = oio->oi_active;
2434         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2435                 /* one chunk plus extent overhead must be enough to write this
2436                  * page */
2437                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2438                 if (ext->oe_end >= index)
2439                         grants = 0;
2440
2441                 /* it doesn't need any grant to dirty this page */
2442                 spin_lock(&cli->cl_loi_list_lock);
2443                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2444                 if (rc == 0) { /* try failed */
2445                         grants = 0;
2446                         need_release = 1;
2447                 } else if (ext->oe_end < index) {
2448                         tmp = grants;
2449                         /* try to expand this extent */
2450                         rc = osc_extent_expand(ext, index, &tmp);
2451                         if (rc < 0) {
2452                                 need_release = 1;
2453                                 /* don't free reserved grant */
2454                         } else {
2455                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2456                                                 "expanded for %lu.\n", index);
2457                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2458                                 grants = 0;
2459                         }
2460                 }
2461                 spin_unlock(&cli->cl_loi_list_lock);
2462                 rc = 0;
2463         } else if (ext != NULL) {
2464                 /* index is located outside of active extent */
2465                 need_release = 1;
2466         }
2467         if (need_release) {
2468                 osc_extent_release(env, ext);
2469                 oio->oi_active = NULL;
2470                 ext = NULL;
2471         }
2472
2473         if (ext == NULL) {
2474                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2475
2476                 /* try to find new extent to cover this page */
2477                 LASSERT(oio->oi_active == NULL);
2478                 /* we may have allocated grant for this page if we failed
2479                  * to expand the previous active extent. */
2480                 LASSERT(ergo(grants > 0, grants >= tmp));
2481
2482                 rc = 0;
2483                 if (grants == 0) {
2484                         /* We haven't allocated grant for this page, and we
2485                          * must not hold a page lock while we do enter_cache,
2486                          * so we must mark dirty & unlock any pages in the
2487                          * write commit pagevec. */
2488                         if (pagevec_count(pvec)) {
2489                                 cb(env, io, pvec);
2490                                 pagevec_reinit(pvec);
2491                         }
2492                         rc = osc_enter_cache(env, cli, oap, tmp);
2493                         if (rc == 0)
2494                                 grants = tmp;
2495                 }
2496
2497                 tmp = grants;
2498                 if (rc == 0) {
2499                         ext = osc_extent_find(env, osc, index, &tmp);
2500                         if (IS_ERR(ext)) {
2501                                 LASSERT(tmp == grants);
2502                                 osc_exit_cache(cli, oap);
2503                                 rc = PTR_ERR(ext);
2504                                 ext = NULL;
2505                         } else {
2506                                 oio->oi_active = ext;
2507                         }
2508                 }
2509                 if (grants > 0)
2510                         osc_unreserve_grant(cli, grants, tmp);
2511         }
2512
2513         LASSERT(ergo(rc == 0, ext != NULL));
2514         if (ext != NULL) {
2515                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2516                          ext, "index = %lu.\n", index);
2517                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2518
2519                 osc_object_lock(osc);
2520                 if (ext->oe_nr_pages == 0)
2521                         ext->oe_srvlock = ops->ops_srvlock;
2522                 else
2523                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2524                 ++ext->oe_nr_pages;
2525                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2526                 osc_object_unlock(osc);
2527
2528                 if (!ext->oe_layout_version)
2529                         ext->oe_layout_version = io->ci_layout_version;
2530         }
2531
2532         RETURN(rc);
2533 }
2534
2535 int osc_teardown_async_page(const struct lu_env *env,
2536                             struct osc_object *obj, struct osc_page *ops)
2537 {
2538         struct osc_async_page *oap = &ops->ops_oap;
2539         int rc = 0;
2540         ENTRY;
2541
2542         LASSERT(oap->oap_magic == OAP_MAGIC);
2543
2544         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2545                oap, ops, osc_index(oap2osc(oap)));
2546
2547         if (!list_empty(&oap->oap_rpc_item)) {
2548                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2549                 rc = -EBUSY;
2550         } else if (!list_empty(&oap->oap_pending_item)) {
2551                 struct osc_extent *ext = NULL;
2552
2553                 osc_object_lock(obj);
2554                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2555                 osc_object_unlock(obj);
2556                 /* only truncated pages are allowed to be taken out.
2557                  * See osc_extent_truncate() and osc_cache_truncate_start()
2558                  * for details. */
2559                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2560                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2561                                         osc_index(oap2osc(oap)));
2562                         rc = -EBUSY;
2563                 }
2564                 if (ext != NULL)
2565                         osc_extent_put(env, ext);
2566         }
2567         RETURN(rc);
2568 }
2569
2570 /**
2571  * This is called when a page is picked up by kernel to write out.
2572  *
2573  * We should find out the corresponding extent and add the whole extent
2574  * into urgent list. The extent may be being truncated or used, handle it
2575  * carefully.
2576  */
2577 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2578                          struct osc_page *ops)
2579 {
2580         struct osc_extent *ext   = NULL;
2581         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2582         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2583         pgoff_t            index = osc_index(ops);
2584         struct osc_async_page *oap = &ops->ops_oap;
2585         bool unplug = false;
2586         int rc = 0;
2587         ENTRY;
2588
2589         osc_object_lock(obj);
2590         ext = osc_extent_lookup(obj, index);
2591         if (ext == NULL) {
2592                 osc_extent_tree_dump(D_ERROR, obj);
2593                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2594         }
2595
2596         switch (ext->oe_state) {
2597         case OES_RPC:
2598         case OES_LOCK_DONE:
2599                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2600                 LASSERT(0);
2601                 break;
2602         case OES_LOCKING:
2603                 /* If we know this extent is being written out, we should abort
2604                  * so that the writer can make this page ready. Otherwise, there
2605                  * exists a deadlock problem because other process can wait for
2606                  * page writeback bit holding page lock; and meanwhile in
2607                  * vvp_page_make_ready(), we need to grab page lock before
2608                  * really sending the RPC. */
2609         case OES_TRUNC:
2610                 /* race with truncate, page will be redirtied */
2611         case OES_ACTIVE:
2612                 /* The extent is active so we need to abort and let the caller
2613                  * re-dirty the page. If we continued on here, and we were the
2614                  * one making the extent active, we could deadlock waiting for
2615                  * the page writeback to clear but it won't because the extent
2616                  * is active and won't be written out. */
2617                 GOTO(out, rc = -EAGAIN);
2618         default:
2619                 break;
2620         }
2621
2622         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2623         if (rc)
2624                 GOTO(out, rc);
2625
2626         spin_lock(&oap->oap_lock);
2627         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2628         spin_unlock(&oap->oap_lock);
2629
2630         if (memory_pressure_get())
2631                 ext->oe_memalloc = 1;
2632
2633         ext->oe_urgent = 1;
2634         if (ext->oe_state == OES_CACHE) {
2635                 OSC_EXTENT_DUMP(D_CACHE, ext,
2636                                 "flush page %p make it urgent.\n", oap);
2637                 if (list_empty(&ext->oe_link))
2638                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2639                 unplug = true;
2640         }
2641         rc = 0;
2642         EXIT;
2643
2644 out:
2645         osc_object_unlock(obj);
2646         osc_extent_put(env, ext);
2647         if (unplug)
2648                 osc_io_unplug_async(env, osc_cli(obj), obj);
2649         return rc;
2650 }
2651
2652 /**
2653  * this is called when a sync waiter receives an interruption.  Its job is to
2654  * get the caller woken as soon as possible.  If its page hasn't been put in an
2655  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2656  * desiring interruption which will forcefully complete the rpc once the rpc
2657  * has timed out.
2658  */
2659 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2660 {
2661         struct osc_async_page *oap = &ops->ops_oap;
2662         struct osc_object     *obj = oap->oap_obj;
2663         struct client_obd     *cli = osc_cli(obj);
2664         struct osc_extent     *ext;
2665         struct osc_extent     *found = NULL;
2666         struct list_head            *plist;
2667         pgoff_t index = osc_index(ops);
2668         int     rc = -EBUSY;
2669         int     cmd;
2670         ENTRY;
2671
2672         LASSERT(!oap->oap_interrupted);
2673         oap->oap_interrupted = 1;
2674
2675         /* Find out the caching extent */
2676         osc_object_lock(obj);
2677         if (oap->oap_cmd & OBD_BRW_WRITE) {
2678                 plist = &obj->oo_urgent_exts;
2679                 cmd   = OBD_BRW_WRITE;
2680         } else {
2681                 plist = &obj->oo_reading_exts;
2682                 cmd   = OBD_BRW_READ;
2683         }
2684         list_for_each_entry(ext, plist, oe_link) {
2685                 if (ext->oe_start <= index && ext->oe_end >= index) {
2686                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2687                         /* For OES_LOCK_DONE state extent, it has already held
2688                          * a refcount for RPC. */
2689                         found = osc_extent_get(ext);
2690                         break;
2691                 }
2692         }
2693         if (found != NULL) {
2694                 list_del_init(&found->oe_link);
2695                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2696                 osc_object_unlock(obj);
2697
2698                 osc_extent_finish(env, found, 0, -EINTR);
2699                 osc_extent_put(env, found);
2700                 rc = 0;
2701         } else {
2702                 osc_object_unlock(obj);
2703                 /* ok, it's been put in an rpc. only one oap gets a request
2704                  * reference */
2705                 if (oap->oap_request != NULL) {
2706                         ptlrpc_mark_interrupted(oap->oap_request);
2707                         ptlrpcd_wake(oap->oap_request);
2708                         ptlrpc_req_finished(oap->oap_request);
2709                         oap->oap_request = NULL;
2710                 }
2711         }
2712
2713         osc_list_maint(cli, obj);
2714         RETURN(rc);
2715 }
2716
2717 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2718                          struct osc_object *obj, struct list_head *list,
2719                          int brw_flags)
2720 {
2721         struct client_obd     *cli = osc_cli(obj);
2722         struct osc_extent     *ext;
2723         struct osc_async_page *oap;
2724         int     page_count = 0;
2725         int     mppr       = cli->cl_max_pages_per_rpc;
2726         bool    can_merge   = true;
2727         pgoff_t start      = CL_PAGE_EOF;
2728         pgoff_t end        = 0;
2729         ENTRY;
2730
2731         list_for_each_entry(oap, list, oap_pending_item) {
2732                 struct osc_page *opg = oap2osc_page(oap);
2733                 pgoff_t index = osc_index(opg);
2734
2735                 if (index > end)
2736                         end = index;
2737                 if (index < start)
2738                         start = index;
2739                 ++page_count;
2740                 mppr <<= (page_count > mppr);
2741
2742                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2743                         can_merge = false;
2744         }
2745
2746         ext = osc_extent_alloc(obj);
2747         if (ext == NULL) {
2748                 struct osc_async_page *tmp;
2749
2750                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2751                         list_del_init(&oap->oap_pending_item);
2752                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2753                 }
2754                 RETURN(-ENOMEM);
2755         }
2756
2757         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2758         ext->oe_sync = 1;
2759         ext->oe_no_merge = !can_merge;
2760         ext->oe_urgent = 1;
2761         ext->oe_start = start;
2762         ext->oe_end = ext->oe_max_end = end;
2763         ext->oe_obj = obj;
2764         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2765         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2766         ext->oe_nr_pages = page_count;
2767         ext->oe_mppr = mppr;
2768         list_splice_init(list, &ext->oe_pages);
2769         ext->oe_layout_version = io->ci_layout_version;
2770
2771         osc_object_lock(obj);
2772         /* Reuse the initial refcount for RPC, don't drop it */
2773         osc_extent_state_set(ext, OES_LOCK_DONE);
2774         if (!ext->oe_rw) { /* write */
2775                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2776                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2777         } else {
2778                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2779                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2780         }
2781         osc_object_unlock(obj);
2782
2783         osc_io_unplug_async(env, cli, obj);
2784         RETURN(0);
2785 }
2786
2787 /**
2788  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2789  */
2790 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2791                              __u64 size, struct osc_extent **extp)
2792 {
2793         struct client_obd *cli = osc_cli(obj);
2794         struct osc_extent *ext;
2795         struct osc_extent *waiting = NULL;
2796         pgoff_t index;
2797         LIST_HEAD(list);
2798         int result = 0;
2799         bool partial;
2800         ENTRY;
2801
2802         /* pages with index greater or equal to index will be truncated. */
2803         index = cl_index(osc2cl(obj), size);
2804         partial = size > cl_offset(osc2cl(obj), index);
2805
2806 again:
2807         osc_object_lock(obj);
2808         ext = osc_extent_search(obj, index);
2809         if (ext == NULL)
2810                 ext = first_extent(obj);
2811         else if (ext->oe_end < index)
2812                 ext = next_extent(ext);
2813         while (ext != NULL) {
2814                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2815
2816                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2817                         /* if ext is in urgent state, it means there must exist
2818                          * a page already having been flushed by write_page().
2819                          * We have to wait for this extent because we can't
2820                          * truncate that page. */
2821                         OSC_EXTENT_DUMP(D_CACHE, ext,
2822                                         "waiting for busy extent\n");
2823                         waiting = osc_extent_get(ext);
2824                         break;
2825                 }
2826
2827                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2828
2829                 osc_extent_get(ext);
2830                 if (ext->oe_state == OES_ACTIVE) {
2831                         /* though we grab inode mutex for write path, but we
2832                          * release it before releasing extent(in osc_io_end()),
2833                          * so there is a race window that an extent is still
2834                          * in OES_ACTIVE when truncate starts. */
2835                         LASSERT(!ext->oe_trunc_pending);
2836                         ext->oe_trunc_pending = 1;
2837                 } else {
2838                         EASSERT(ext->oe_state == OES_CACHE, ext);
2839                         osc_extent_state_set(ext, OES_TRUNC);
2840                         osc_update_pending(obj, OBD_BRW_WRITE,
2841                                            -ext->oe_nr_pages);
2842                 }
2843                 /* This extent could be on the full extents list, that's OK */
2844                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2845                 if (!list_empty(&ext->oe_link))
2846                         list_move_tail(&ext->oe_link, &list);
2847                 else
2848                         list_add_tail(&ext->oe_link, &list);
2849
2850                 ext = next_extent(ext);
2851         }
2852         osc_object_unlock(obj);
2853
2854         osc_list_maint(cli, obj);
2855
2856         while (!list_empty(&list)) {
2857                 int rc;
2858
2859                 ext = list_entry(list.next, struct osc_extent, oe_link);
2860                 list_del_init(&ext->oe_link);
2861
2862                 /* extent may be in OES_ACTIVE state because inode mutex
2863                  * is released before osc_io_end() in file write case */
2864                 if (ext->oe_state != OES_TRUNC)
2865                         osc_extent_wait(env, ext, OES_TRUNC);
2866
2867                 rc = osc_extent_truncate(ext, index, partial);
2868                 if (rc < 0) {
2869                         if (result == 0)
2870                                 result = rc;
2871
2872                         OSC_EXTENT_DUMP(D_ERROR, ext,
2873                                         "truncate error %d\n", rc);
2874                 } else if (ext->oe_nr_pages == 0) {
2875                         osc_extent_remove(ext);
2876                 } else {
2877                         /* this must be an overlapped extent which means only
2878                          * part of pages in this extent have been truncated.
2879                          */
2880                         EASSERTF(ext->oe_start <= index, ext,
2881                                  "trunc index = %lu/%d.\n", index, partial);
2882                         /* fix index to skip this partially truncated extent */
2883                         index = ext->oe_end + 1;
2884                         partial = false;
2885
2886                         /* we need to hold this extent in OES_TRUNC state so
2887                          * that no writeback will happen. This is to avoid
2888                          * BUG 17397.
2889                          * Only partial truncate can reach here, if @size is
2890                          * not zero, the caller should provide a valid @extp. */
2891                         LASSERT(*extp == NULL);
2892                         *extp = osc_extent_get(ext);
2893                         OSC_EXTENT_DUMP(D_CACHE, ext,
2894                                         "trunc at %llu\n", size);
2895                 }
2896                 osc_extent_put(env, ext);
2897         }
2898         if (waiting != NULL) {
2899                 int rc;
2900
2901                 /* ignore the result of osc_extent_wait the write initiator
2902                  * should take care of it. */
2903                 rc = osc_extent_wait(env, waiting, OES_INV);
2904                 if (rc < 0)
2905                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2906
2907                 osc_extent_put(env, waiting);
2908                 waiting = NULL;
2909                 goto again;
2910         }
2911         RETURN(result);
2912 }
2913 EXPORT_SYMBOL(osc_cache_truncate_start);
2914
2915 /**
2916  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2917  */
2918 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2919 {
2920         if (ext != NULL) {
2921                 struct osc_object *obj = ext->oe_obj;
2922                 bool unplug = false;
2923
2924                 EASSERT(ext->oe_nr_pages > 0, ext);
2925                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2926                 EASSERT(!ext->oe_urgent, ext);
2927
2928                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2929                 osc_object_lock(obj);
2930                 osc_extent_state_set(ext, OES_CACHE);
2931                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2932                         ext->oe_urgent = 1;
2933                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2934                         unplug = true;
2935                 }
2936                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2937                 osc_object_unlock(obj);
2938                 osc_extent_put(env, ext);
2939
2940                 if (unplug)
2941                         osc_io_unplug_async(env, osc_cli(obj), obj);
2942         }
2943 }
2944
2945 /**
2946  * Wait for extents in a specific range to be written out.
2947  * The caller must have called osc_cache_writeback_range() to issue IO
2948  * otherwise it will take a long time for this function to finish.
2949  *
2950  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2951  * nobody else can dirty this range of file while we're waiting for
2952  * extents to be written.
2953  */
2954 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2955                          pgoff_t start, pgoff_t end)
2956 {
2957         struct osc_extent *ext;
2958         pgoff_t index = start;
2959         int     result = 0;
2960         ENTRY;
2961
2962 again:
2963         osc_object_lock(obj);
2964         ext = osc_extent_search(obj, index);
2965         if (ext == NULL)
2966                 ext = first_extent(obj);
2967         else if (ext->oe_end < index)
2968                 ext = next_extent(ext);
2969         while (ext != NULL) {
2970                 int rc;
2971
2972                 if (ext->oe_start > end)
2973                         break;
2974
2975                 if (!ext->oe_fsync_wait) {
2976                         ext = next_extent(ext);
2977                         continue;
2978                 }
2979
2980                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2981                              ext->oe_hp || ext->oe_urgent), ext);
2982                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2983                              !ext->oe_hp && ext->oe_urgent), ext);
2984
2985                 index = ext->oe_end + 1;
2986                 osc_extent_get(ext);
2987                 osc_object_unlock(obj);
2988
2989                 rc = osc_extent_wait(env, ext, OES_INV);
2990                 if (result == 0)
2991                         result = rc;
2992                 osc_extent_put(env, ext);
2993                 goto again;
2994         }
2995         osc_object_unlock(obj);
2996
2997         OSC_IO_DEBUG(obj, "sync file range.\n");
2998         RETURN(result);
2999 }
3000 EXPORT_SYMBOL(osc_cache_wait_range);
3001
3002 /**
3003  * Called to write out a range of osc object.
3004  *
3005  * @hp     : should be set this is caused by lock cancel;
3006  * @discard: is set if dirty pages should be dropped - file will be deleted or
3007  *         truncated, this implies there is no partially discarding extents.
3008  *
3009  * Return how many pages will be issued, or error code if error occurred.
3010  */
3011 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3012                               pgoff_t start, pgoff_t end, int hp, int discard)
3013 {
3014         struct osc_extent *ext;
3015         LIST_HEAD(discard_list);
3016         bool unplug = false;
3017         int result = 0;
3018         ENTRY;
3019
3020         osc_object_lock(obj);
3021         ext = osc_extent_search(obj, start);
3022         if (ext == NULL)
3023                 ext = first_extent(obj);
3024         else if (ext->oe_end < start)
3025                 ext = next_extent(ext);
3026         while (ext != NULL) {
3027                 if (ext->oe_start > end)
3028                         break;
3029
3030                 ext->oe_fsync_wait = 1;
3031                 switch (ext->oe_state) {
3032                 case OES_CACHE:
3033                         result += ext->oe_nr_pages;
3034                         if (!discard) {
3035                                 struct list_head *list = NULL;
3036                                 if (hp) {
3037                                         EASSERT(!ext->oe_hp, ext);
3038                                         ext->oe_hp = 1;
3039                                         list = &obj->oo_hp_exts;
3040                                 } else if (!ext->oe_urgent) {
3041                                         ext->oe_urgent = 1;
3042                                         list = &obj->oo_urgent_exts;
3043                                 }
3044                                 if (list != NULL)
3045                                         list_move_tail(&ext->oe_link, list);
3046                                 unplug = true;
3047                         } else {
3048                                 struct client_obd *cli = osc_cli(obj);
3049                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
3050                                 pgoff_t align_by = (1 << pcc_bits);
3051                                 pgoff_t a_start = round_down(start, align_by);
3052                                 pgoff_t a_end = round_up(end, align_by);
3053
3054                                 /* overflow case */
3055                                 if (end && !a_end)
3056                                         a_end = CL_PAGE_EOF;
3057                                 /* the only discarder is lock cancelling, so
3058                                  * [start, end], aligned by chunk size, must
3059                                  * contain this extent */
3060                                 LASSERTF(ext->oe_start >= a_start &&
3061                                          ext->oe_end <= a_end,
3062                                          "ext [%lu, %lu] reg [%lu, %lu] "
3063                                          "orig [%lu %lu] align %lu bits "
3064                                          "%d\n", ext->oe_start, ext->oe_end,
3065                                          a_start, a_end, start, end,
3066                                          align_by, pcc_bits);
3067                                 osc_extent_state_set(ext, OES_LOCKING);
3068                                 ext->oe_owner = current;
3069                                 list_move_tail(&ext->oe_link,
3070                                                    &discard_list);
3071                                 osc_update_pending(obj, OBD_BRW_WRITE,
3072                                                    -ext->oe_nr_pages);
3073                         }
3074                         break;
3075                 case OES_ACTIVE:
3076                         /* It's pretty bad to wait for ACTIVE extents, because
3077                          * we don't know how long we will wait for it to be
3078                          * flushed since it may be blocked at awaiting more
3079                          * grants. We do this for the correctness of fsync. */
3080                         LASSERT(hp == 0 && discard == 0);
3081                         ext->oe_urgent = 1;
3082                         break;
3083                 case OES_TRUNC:
3084                         /* this extent is being truncated, can't do anything
3085                          * for it now. it will be set to urgent after truncate
3086                          * is finished in osc_cache_truncate_end(). */
3087                 default:
3088                         break;
3089                 }
3090                 ext = next_extent(ext);
3091         }
3092         osc_object_unlock(obj);
3093
3094         LASSERT(ergo(!discard, list_empty(&discard_list)));
3095         if (!list_empty(&discard_list)) {
3096                 struct osc_extent *tmp;
3097                 int rc;
3098
3099                 osc_list_maint(osc_cli(obj), obj);
3100                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3101                         list_del_init(&ext->oe_link);
3102                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3103
3104                         /* Discard caching pages. We don't actually write this
3105                          * extent out but we complete it as if we did. */
3106                         rc = osc_extent_make_ready(env, ext);
3107                         if (unlikely(rc < 0)) {
3108                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3109                                                 "make_ready returned %d\n", rc);
3110                                 if (result >= 0)
3111                                         result = rc;
3112                         }
3113
3114                         /* finish the extent as if the pages were sent */
3115                         osc_extent_finish(env, ext, 0, 0);
3116                 }
3117         }
3118
3119         if (unplug)
3120                 osc_io_unplug(env, osc_cli(obj), obj);
3121
3122         if (hp || discard) {
3123                 int rc;
3124                 rc = osc_cache_wait_range(env, obj, start, end);
3125                 if (result >= 0 && rc < 0)
3126                         result = rc;
3127         }
3128
3129         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3130         RETURN(result);
3131 }
3132 EXPORT_SYMBOL(osc_cache_writeback_range);
3133
3134 /**
3135  * Returns a list of pages by a given [start, end] of \a obj.
3136  *
3137  * \param resched If not NULL, then we give up before hogging CPU for too
3138  * long and set *resched = 1, in that case caller should implement a retry
3139  * logic.
3140  *
3141  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3142  * crucial in the face of [offset, EOF] locks.
3143  *
3144  * Return at least one page in @queue unless there is no covered page.
3145  */
3146 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3147                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3148                         osc_page_gang_cbt cb, void *cbdata)
3149 {
3150         struct osc_page *ops;
3151         struct pagevec  *pagevec;
3152         void            **pvec;
3153         pgoff_t         idx;
3154         unsigned int    nr;
3155         unsigned int    i;
3156         unsigned int    j;
3157         int             res = CLP_GANG_OKAY;
3158         bool            tree_lock = true;
3159         ENTRY;
3160
3161         idx = start;
3162         pvec = osc_env_info(env)->oti_pvec;
3163         pagevec = &osc_env_info(env)->oti_pagevec;
3164         ll_pagevec_init(pagevec, 0);
3165         spin_lock(&osc->oo_tree_lock);
3166         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3167                                             idx, OTI_PVEC_SIZE)) > 0) {
3168