Whamcloud - gitweb
LU-2722 clio: directIO thread races with completion thread
[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.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2012, Intel Corporation.
31  *
32  */
33 /*
34  * This file is part of Lustre, http://www.lustre.org/
35  * Lustre is a trademark of Sun Microsystems, Inc.
36  *
37  * osc cache management.
38  *
39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_OSC
43
44 #include "osc_cl_internal.h"
45 #include "osc_internal.h"
46
47 static int extent_debug; /* set it to be true for more debug */
48
49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
51                            int state);
52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
53                               struct osc_async_page *oap, int sent, int rc);
54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
55                           int cmd);
56 static int osc_refresh_count(const struct lu_env *env,
57                              struct osc_async_page *oap, int cmd);
58 static int osc_io_unplug_async(const struct lu_env *env,
59                                struct client_obd *cli, struct osc_object *osc);
60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
61                            unsigned int lost_grant);
62
63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
64                                   const char *func, int line);
65 #define osc_extent_tree_dump(lvl, obj) \
66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
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_srvlock)
80                 *buf++ = 's';
81         if (ext->oe_hp)
82                 *buf++ = 'h';
83         if (ext->oe_urgent)
84                 *buf++ = 'u';
85         if (ext->oe_memalloc)
86                 *buf++ = 'm';
87         if (ext->oe_trunc_pending)
88                 *buf++ = 't';
89         if (ext->oe_fsync_wait)
90                 *buf++ = 'Y';
91         *buf = 0;
92         return flags;
93 }
94
95 static inline char list_empty_marker(cfs_list_t *list)
96 {
97         return cfs_list_empty(list) ? '-' : '+';
98 }
99
100 #define EXTSTR       "[%lu -> %lu/%lu]"
101 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
102
103 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
104         struct osc_extent *__ext = (extent);                                  \
105         const char *__str[] = OES_STRINGS;                                    \
106         char __buf[16];                                                       \
107                                                                               \
108         CDEBUG(lvl,                                                           \
109                 "extent %p@{" EXTSTR ", "                                     \
110                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
111                 /* ----- extent part 0 ----- */                               \
112                 __ext, EXTPARA(__ext),                                        \
113                 /* ----- part 1 ----- */                                      \
114                 cfs_atomic_read(&__ext->oe_refc),                             \
115                 cfs_atomic_read(&__ext->oe_users),                            \
116                 list_empty_marker(&__ext->oe_link),                           \
117                 __str[__ext->oe_state], ext_flags(__ext, __buf),              \
118                 __ext->oe_obj,                                                \
119                 /* ----- part 2 ----- */                                      \
120                 __ext->oe_grants, __ext->oe_nr_pages,                         \
121                 list_empty_marker(&__ext->oe_pages),                          \
122                 cfs_waitq_active(&__ext->oe_waitq) ? '+' : '-',               \
123                 __ext->oe_osclock, __ext->oe_mppr, __ext->oe_owner,           \
124                 /* ----- part 4 ----- */                                      \
125                 ## __VA_ARGS__);                                              \
126 } while (0)
127
128 #undef EASSERTF
129 #define EASSERTF(expr, ext, fmt, args...) do {                          \
130         if (!(expr)) {                                                  \
131                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);            \
132                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);            \
133                 LASSERT(expr);                                          \
134         }                                                                    \
135 } while (0)
136
137 #undef EASSERT
138 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
139
140 static inline struct osc_extent *rb_extent(struct rb_node *n)
141 {
142         if (n == NULL)
143                 return NULL;
144
145         return container_of(n, struct osc_extent, oe_node);
146 }
147
148 static inline struct osc_extent *next_extent(struct osc_extent *ext)
149 {
150         if (ext == NULL)
151                 return NULL;
152
153         LASSERT(ext->oe_intree);
154         return rb_extent(rb_next(&ext->oe_node));
155 }
156
157 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
158 {
159         if (ext == NULL)
160                 return NULL;
161
162         LASSERT(ext->oe_intree);
163         return rb_extent(rb_prev(&ext->oe_node));
164 }
165
166 static inline struct osc_extent *first_extent(struct osc_object *obj)
167 {
168         return rb_extent(rb_first(&obj->oo_root));
169 }
170
171 /* object must be locked by caller. */
172 static int osc_extent_sanity_check0(struct osc_extent *ext,
173                                     const char *func, const int line)
174 {
175         struct osc_object *obj = ext->oe_obj;
176         struct osc_async_page *oap;
177         int page_count;
178         int rc = 0;
179
180         if (!osc_object_is_locked(obj))
181                 GOTO(out, rc = 9);
182
183         if (ext->oe_state >= OES_STATE_MAX)
184                 GOTO(out, rc = 10);
185
186         if (cfs_atomic_read(&ext->oe_refc) <= 0)
187                 GOTO(out, rc = 20);
188
189         if (cfs_atomic_read(&ext->oe_refc) < cfs_atomic_read(&ext->oe_users))
190                 GOTO(out, rc = 30);
191
192         switch (ext->oe_state) {
193         case OES_INV:
194                 if (ext->oe_nr_pages > 0 || !cfs_list_empty(&ext->oe_pages))
195                         GOTO(out, rc = 35);
196                 GOTO(out, rc = 0);
197                 break;
198         case OES_ACTIVE:
199                 if (cfs_atomic_read(&ext->oe_users) == 0)
200                         GOTO(out, rc = 40);
201                 if (ext->oe_hp)
202                         GOTO(out, rc = 50);
203                 if (ext->oe_fsync_wait && !ext->oe_urgent)
204                         GOTO(out, rc = 55);
205                 break;
206         case OES_CACHE:
207                 if (ext->oe_grants == 0)
208                         GOTO(out, rc = 60);
209                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
210                         GOTO(out, rc = 65);
211         default:
212                 if (cfs_atomic_read(&ext->oe_users) > 0)
213                         GOTO(out, rc = 70);
214         }
215
216         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
217                 GOTO(out, rc = 80);
218
219         if (ext->oe_osclock == NULL && ext->oe_grants > 0)
220                 GOTO(out, rc = 90);
221
222         if (ext->oe_osclock) {
223                 struct cl_lock_descr *descr;
224                 descr = &ext->oe_osclock->cll_descr;
225                 if (!(descr->cld_start <= ext->oe_start &&
226                       descr->cld_end >= ext->oe_max_end))
227                         GOTO(out, rc = 100);
228         }
229
230         if (ext->oe_nr_pages > ext->oe_mppr)
231                 GOTO(out, rc = 105);
232
233         /* Do not verify page list if extent is in RPC. This is because an
234          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
235         if (ext->oe_state > OES_CACHE)
236                 GOTO(out, rc = 0);
237
238         if (!extent_debug)
239                 GOTO(out, rc = 0);
240
241         page_count = 0;
242         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
243                 pgoff_t index = oap2cl_page(oap)->cp_index;
244                 ++page_count;
245                 if (index > ext->oe_end || index < ext->oe_start)
246                         GOTO(out, rc = 110);
247         }
248         if (page_count != ext->oe_nr_pages)
249                 GOTO(out, rc = 120);
250
251 out:
252         if (rc != 0)
253                 OSC_EXTENT_DUMP(D_ERROR, ext,
254                                 "%s:%d sanity check %p failed with rc = %d\n",
255                                 func, line, ext, rc);
256         return rc;
257 }
258
259 #define sanity_check_nolock(ext) \
260         osc_extent_sanity_check0(ext, __func__, __LINE__)
261
262 #define sanity_check(ext) ({                                                   \
263         int __res;                                                             \
264         osc_object_lock((ext)->oe_obj);                                        \
265         __res = sanity_check_nolock(ext);                                      \
266         osc_object_unlock((ext)->oe_obj);                                      \
267         __res;                                                                 \
268 })
269
270
271 /**
272  * sanity check - to make sure there is no overlapped extent in the tree.
273  */
274 static int osc_extent_is_overlapped(struct osc_object *obj,
275                                     struct osc_extent *ext)
276 {
277         struct osc_extent *tmp;
278
279         LASSERT(osc_object_is_locked(obj));
280
281         if (!extent_debug)
282                 return 0;
283
284         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
285                 if (tmp == ext)
286                         continue;
287                 if (tmp->oe_end >= ext->oe_start &&
288                     tmp->oe_start <= ext->oe_end)
289                         return 1;
290         }
291         return 0;
292 }
293
294 static void osc_extent_state_set(struct osc_extent *ext, int state)
295 {
296         LASSERT(osc_object_is_locked(ext->oe_obj));
297         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
298
299         /* Never try to sanity check a state changing extent :-) */
300         /* LASSERT(sanity_check_nolock(ext) == 0); */
301
302         /* TODO: validate the state machine */
303         ext->oe_state = state;
304         cfs_waitq_broadcast(&ext->oe_waitq);
305 }
306
307 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
308 {
309         struct osc_extent *ext;
310
311         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, CFS_ALLOC_STD);
312         if (ext == NULL)
313                 return NULL;
314
315         RB_CLEAR_NODE(&ext->oe_node);
316         ext->oe_obj = obj;
317         cfs_atomic_set(&ext->oe_refc, 1);
318         cfs_atomic_set(&ext->oe_users, 0);
319         CFS_INIT_LIST_HEAD(&ext->oe_link);
320         ext->oe_state = OES_INV;
321         CFS_INIT_LIST_HEAD(&ext->oe_pages);
322         cfs_waitq_init(&ext->oe_waitq);
323         ext->oe_osclock = NULL;
324
325         return ext;
326 }
327
328 static void osc_extent_free(struct osc_extent *ext)
329 {
330         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
331 }
332
333 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
334 {
335         LASSERT(cfs_atomic_read(&ext->oe_refc) >= 0);
336         cfs_atomic_inc(&ext->oe_refc);
337         return ext;
338 }
339
340 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
341 {
342         LASSERT(cfs_atomic_read(&ext->oe_refc) > 0);
343         if (cfs_atomic_dec_and_test(&ext->oe_refc)) {
344                 LASSERT(cfs_list_empty(&ext->oe_link));
345                 LASSERT(cfs_atomic_read(&ext->oe_users) == 0);
346                 LASSERT(ext->oe_state == OES_INV);
347                 LASSERT(!ext->oe_intree);
348
349                 if (ext->oe_osclock) {
350                         cl_lock_put(env, ext->oe_osclock);
351                         ext->oe_osclock = NULL;
352                 }
353                 osc_extent_free(ext);
354         }
355 }
356
357 /**
358  * osc_extent_put_trust() is a special version of osc_extent_put() when
359  * it's known that the caller is not the last user. This is to address the
360  * problem of lacking of lu_env ;-).
361  */
362 static void osc_extent_put_trust(struct osc_extent *ext)
363 {
364         LASSERT(cfs_atomic_read(&ext->oe_refc) > 1);
365         LASSERT(osc_object_is_locked(ext->oe_obj));
366         cfs_atomic_dec(&ext->oe_refc);
367 }
368
369 /**
370  * Return the extent which includes pgoff @index, or return the greatest
371  * previous extent in the tree.
372  */
373 static struct osc_extent *osc_extent_search(struct osc_object *obj,
374                                             pgoff_t index)
375 {
376         struct rb_node    *n = obj->oo_root.rb_node;
377         struct osc_extent *tmp, *p = NULL;
378
379         LASSERT(osc_object_is_locked(obj));
380         while (n != NULL) {
381                 tmp = rb_extent(n);
382                 if (index < tmp->oe_start) {
383                         n = n->rb_left;
384                 } else if (index > tmp->oe_end) {
385                         p = rb_extent(n);
386                         n = n->rb_right;
387                 } else {
388                         return tmp;
389                 }
390         }
391         return p;
392 }
393
394 /*
395  * Return the extent covering @index, otherwise return NULL.
396  * caller must have held object lock.
397  */
398 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
399                                             pgoff_t index)
400 {
401         struct osc_extent *ext;
402
403         ext = osc_extent_search(obj, index);
404         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
405                 return osc_extent_get(ext);
406         return NULL;
407 }
408
409 /* caller must have held object lock. */
410 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
411 {
412         struct rb_node   **n      = &obj->oo_root.rb_node;
413         struct rb_node    *parent = NULL;
414         struct osc_extent *tmp;
415
416         LASSERT(ext->oe_intree == 0);
417         LASSERT(ext->oe_obj == obj);
418         LASSERT(osc_object_is_locked(obj));
419         while (*n != NULL) {
420                 tmp = rb_extent(*n);
421                 parent = *n;
422
423                 if (ext->oe_end < tmp->oe_start)
424                         n = &(*n)->rb_left;
425                 else if (ext->oe_start > tmp->oe_end)
426                         n = &(*n)->rb_right;
427                 else
428                         EASSERTF(0, tmp, EXTSTR, EXTPARA(ext));
429         }
430         rb_link_node(&ext->oe_node, parent, n);
431         rb_insert_color(&ext->oe_node, &obj->oo_root);
432         osc_extent_get(ext);
433         ext->oe_intree = 1;
434 }
435
436 /* caller must have held object lock. */
437 static void osc_extent_erase(struct osc_extent *ext)
438 {
439         struct osc_object *obj = ext->oe_obj;
440         LASSERT(osc_object_is_locked(obj));
441         if (ext->oe_intree) {
442                 rb_erase(&ext->oe_node, &obj->oo_root);
443                 ext->oe_intree = 0;
444                 /* rbtree held a refcount */
445                 osc_extent_put_trust(ext);
446         }
447 }
448
449 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
450 {
451         struct osc_object *obj = ext->oe_obj;
452
453         LASSERT(osc_object_is_locked(obj));
454         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
455         if (ext->oe_state == OES_CACHE) {
456                 osc_extent_state_set(ext, OES_ACTIVE);
457                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
458         }
459         cfs_atomic_inc(&ext->oe_users);
460         cfs_list_del_init(&ext->oe_link);
461         return osc_extent_get(ext);
462 }
463
464 static void __osc_extent_remove(struct osc_extent *ext)
465 {
466         LASSERT(osc_object_is_locked(ext->oe_obj));
467         LASSERT(cfs_list_empty(&ext->oe_pages));
468         osc_extent_erase(ext);
469         cfs_list_del_init(&ext->oe_link);
470         osc_extent_state_set(ext, OES_INV);
471         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
472 }
473
474 static void osc_extent_remove(struct osc_extent *ext)
475 {
476         struct osc_object *obj = ext->oe_obj;
477
478         osc_object_lock(obj);
479         __osc_extent_remove(ext);
480         osc_object_unlock(obj);
481 }
482
483 /**
484  * This function is used to merge extents to get better performance. It checks
485  * if @cur and @victim are contiguous at chunk level.
486  */
487 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
488                             struct osc_extent *victim)
489 {
490         struct osc_object *obj = cur->oe_obj;
491         pgoff_t chunk_start;
492         pgoff_t chunk_end;
493         int ppc_bits;
494
495         LASSERT(cur->oe_state == OES_CACHE);
496         LASSERT(osc_object_is_locked(obj));
497         if (victim == NULL)
498                 return -EINVAL;
499
500         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
501                 return -EBUSY;
502
503         if (cur->oe_max_end != victim->oe_max_end)
504                 return -ERANGE;
505
506         LASSERT(cur->oe_osclock == victim->oe_osclock);
507         ppc_bits = osc_cli(obj)->cl_chunkbits - CFS_PAGE_SHIFT;
508         chunk_start = cur->oe_start >> ppc_bits;
509         chunk_end   = cur->oe_end   >> ppc_bits;
510         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
511             chunk_end + 1 != victim->oe_start >> ppc_bits)
512                 return -ERANGE;
513
514         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
515
516         cur->oe_start     = min(cur->oe_start, victim->oe_start);
517         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
518         cur->oe_grants   += victim->oe_grants;
519         cur->oe_nr_pages += victim->oe_nr_pages;
520         /* only the following bits are needed to merge */
521         cur->oe_urgent   |= victim->oe_urgent;
522         cur->oe_memalloc |= victim->oe_memalloc;
523         cfs_list_splice_init(&victim->oe_pages, &cur->oe_pages);
524         cfs_list_del_init(&victim->oe_link);
525         victim->oe_nr_pages = 0;
526
527         osc_extent_get(victim);
528         __osc_extent_remove(victim);
529         osc_extent_put(env, victim);
530
531         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
532         return 0;
533 }
534
535 /**
536  * Drop user count of osc_extent, and unplug IO asynchronously.
537  */
538 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
539 {
540         struct osc_object *obj = ext->oe_obj;
541         int rc = 0;
542         ENTRY;
543
544         LASSERT(cfs_atomic_read(&ext->oe_users) > 0);
545         LASSERT(sanity_check(ext) == 0);
546         LASSERT(ext->oe_grants > 0);
547
548         if (cfs_atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
549                 LASSERT(ext->oe_state == OES_ACTIVE);
550                 if (ext->oe_trunc_pending) {
551                         /* a truncate process is waiting for this extent.
552                          * This may happen due to a race, check
553                          * osc_cache_truncate_start(). */
554                         osc_extent_state_set(ext, OES_TRUNC);
555                         ext->oe_trunc_pending = 0;
556                 } else {
557                         osc_extent_state_set(ext, OES_CACHE);
558                         osc_update_pending(obj, OBD_BRW_WRITE,
559                                            ext->oe_nr_pages);
560
561                         /* try to merge the previous and next extent. */
562                         osc_extent_merge(env, ext, prev_extent(ext));
563                         osc_extent_merge(env, ext, next_extent(ext));
564
565                         if (ext->oe_urgent)
566                                 cfs_list_move_tail(&ext->oe_link,
567                                                    &obj->oo_urgent_exts);
568                 }
569                 osc_object_unlock(obj);
570
571                 osc_io_unplug_async(env, osc_cli(obj), obj);
572         }
573         osc_extent_put(env, ext);
574         RETURN(rc);
575 }
576
577 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
578 {
579         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
580 }
581
582 /**
583  * Find or create an extent which includes @index, core function to manage
584  * extent tree.
585  */
586 struct osc_extent *osc_extent_find(const struct lu_env *env,
587                                    struct osc_object *obj, pgoff_t index,
588                                    int *grants)
589
590 {
591         struct client_obd *cli = osc_cli(obj);
592         struct cl_lock    *lock;
593         struct osc_extent *cur;
594         struct osc_extent *ext;
595         struct osc_extent *conflict = NULL;
596         struct osc_extent *found = NULL;
597         pgoff_t    chunk;
598         pgoff_t    max_end;
599         int        max_pages; /* max_pages_per_rpc */
600         int        chunksize;
601         int        ppc_bits; /* pages per chunk bits */
602         int        chunk_mask;
603         int        rc;
604         ENTRY;
605
606         cur = osc_extent_alloc(obj);
607         if (cur == NULL)
608                 RETURN(ERR_PTR(-ENOMEM));
609
610         lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0);
611         LASSERT(lock != NULL);
612         LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
613
614         LASSERT(cli->cl_chunkbits >= CFS_PAGE_SHIFT);
615         ppc_bits   = cli->cl_chunkbits - CFS_PAGE_SHIFT;
616         chunk_mask = ~((1 << ppc_bits) - 1);
617         chunksize  = 1 << cli->cl_chunkbits;
618         chunk      = index >> ppc_bits;
619
620         /* align end to rpc edge, rpc size may not be a power 2 integer. */
621         max_pages = cli->cl_max_pages_per_rpc;
622         LASSERT((max_pages & ~chunk_mask) == 0);
623         max_end = index - (index % max_pages) + max_pages - 1;
624         max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end);
625
626         /* initialize new extent by parameters so far */
627         cur->oe_max_end = max_end;
628         cur->oe_start   = index & chunk_mask;
629         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
630         if (cur->oe_start < lock->cll_descr.cld_start)
631                 cur->oe_start = lock->cll_descr.cld_start;
632         if (cur->oe_end > max_end)
633                 cur->oe_end = max_end;
634         cur->oe_osclock = lock;
635         cur->oe_grants  = 0;
636         cur->oe_mppr    = max_pages;
637
638         /* grants has been allocated by caller */
639         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
640                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
641         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur));
642
643 restart:
644         osc_object_lock(obj);
645         ext = osc_extent_search(obj, cur->oe_start);
646         if (ext == NULL)
647                 ext = first_extent(obj);
648         while (ext != NULL) {
649                 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
650                 loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
651
652                 LASSERT(sanity_check_nolock(ext) == 0);
653                 if (chunk > ext_chk_end + 1)
654                         break;
655
656                 /* if covering by different locks, no chance to match */
657                 if (lock != ext->oe_osclock) {
658                         EASSERTF(!overlapped(ext, cur), ext,
659                                  EXTSTR, EXTPARA(cur));
660
661                         ext = next_extent(ext);
662                         continue;
663                 }
664
665                 /* discontiguous chunks? */
666                 if (chunk + 1 < ext_chk_start) {
667                         ext = next_extent(ext);
668                         continue;
669                 }
670
671                 /* ok, from now on, ext and cur have these attrs:
672                  * 1. covered by the same lock
673                  * 2. contiguous at chunk level or overlapping. */
674
675                 if (overlapped(ext, cur)) {
676                         /* cur is the minimum unit, so overlapping means
677                          * full contain. */
678                         EASSERTF((ext->oe_start <= cur->oe_start &&
679                                   ext->oe_end >= cur->oe_end),
680                                  ext, EXTSTR, EXTPARA(cur));
681
682                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
683                                 /* for simplicity, we wait for this extent to
684                                  * finish before going forward. */
685                                 conflict = osc_extent_get(ext);
686                                 break;
687                         }
688
689                         found = osc_extent_hold(ext);
690                         break;
691                 }
692
693                 /* non-overlapped extent */
694                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
695                         /* we can't do anything for a non OES_CACHE extent, or
696                          * if there is someone waiting for this extent to be
697                          * flushed, try next one. */
698                         ext = next_extent(ext);
699                         continue;
700                 }
701
702                 /* check if they belong to the same rpc slot before trying to
703                  * merge. the extents are not overlapped and contiguous at
704                  * chunk level to get here. */
705                 if (ext->oe_max_end != max_end) {
706                         /* if they don't belong to the same RPC slot or
707                          * max_pages_per_rpc has ever changed, do not merge. */
708                         ext = next_extent(ext);
709                         continue;
710                 }
711
712                 /* it's required that an extent must be contiguous at chunk
713                  * level so that we know the whole extent is covered by grant
714                  * (the pages in the extent are NOT required to be contiguous).
715                  * Otherwise, it will be too much difficult to know which
716                  * chunks have grants allocated. */
717
718                 /* try to do front merge - extend ext's start */
719                 if (chunk + 1 == ext_chk_start) {
720                         /* ext must be chunk size aligned */
721                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
722
723                         /* pull ext's start back to cover cur */
724                         ext->oe_start   = cur->oe_start;
725                         ext->oe_grants += chunksize;
726                         *grants -= chunksize;
727
728                         found = osc_extent_hold(ext);
729                 } else if (chunk == ext_chk_end + 1) {
730                         /* rear merge */
731                         ext->oe_end     = cur->oe_end;
732                         ext->oe_grants += chunksize;
733                         *grants -= chunksize;
734
735                         /* try to merge with the next one because we just fill
736                          * in a gap */
737                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
738                                 /* we can save extent tax from next extent */
739                                 *grants += cli->cl_extent_tax;
740
741                         found = osc_extent_hold(ext);
742                 }
743                 if (found != NULL)
744                         break;
745
746                 ext = next_extent(ext);
747         }
748
749         osc_extent_tree_dump(D_CACHE, obj);
750         if (found != NULL) {
751                 LASSERT(conflict == NULL);
752                 if (!IS_ERR(found)) {
753                         LASSERT(found->oe_osclock == cur->oe_osclock);
754                         OSC_EXTENT_DUMP(D_CACHE, found,
755                                         "found caching ext for %lu.\n", index);
756                 }
757         } else if (conflict == NULL) {
758                 /* create a new extent */
759                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
760                 cur->oe_grants = chunksize + cli->cl_extent_tax;
761                 *grants -= cur->oe_grants;
762                 LASSERT(*grants >= 0);
763
764                 cur->oe_state = OES_CACHE;
765                 found = osc_extent_hold(cur);
766                 osc_extent_insert(obj, cur);
767                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
768                                 index, lock->cll_descr.cld_end);
769         }
770         osc_object_unlock(obj);
771
772         if (conflict != NULL) {
773                 LASSERT(found == NULL);
774
775                 /* waiting for IO to finish. Please notice that it's impossible
776                  * to be an OES_TRUNC extent. */
777                 rc = osc_extent_wait(env, conflict, OES_INV);
778                 osc_extent_put(env, conflict);
779                 conflict = NULL;
780                 if (rc < 0)
781                         GOTO(out, found = ERR_PTR(rc));
782
783                 goto restart;
784         }
785         EXIT;
786
787 out:
788         osc_extent_put(env, cur);
789         LASSERT(*grants >= 0);
790         return found;
791 }
792
793 /**
794  * Called when IO is finished to an extent.
795  */
796 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
797                       int sent, int rc)
798 {
799         struct client_obd *cli = osc_cli(ext->oe_obj);
800         struct osc_async_page *oap;
801         struct osc_async_page *tmp;
802         int nr_pages = ext->oe_nr_pages;
803         int lost_grant = 0;
804         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
805         __u64 last_off = 0;
806         int last_count = -1;
807         ENTRY;
808
809         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
810
811         ext->oe_rc = rc ?: ext->oe_nr_pages;
812         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
813         cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
814                                      oap_pending_item) {
815                 cfs_list_del_init(&oap->oap_rpc_item);
816                 cfs_list_del_init(&oap->oap_pending_item);
817                 if (last_off <= oap->oap_obj_off) {
818                         last_off = oap->oap_obj_off;
819                         last_count = oap->oap_count;
820                 }
821
822                 --ext->oe_nr_pages;
823                 osc_ap_completion(env, cli, oap, sent, rc);
824         }
825         EASSERT(ext->oe_nr_pages == 0, ext);
826
827         if (!sent) {
828                 lost_grant = ext->oe_grants;
829         } else if (blocksize < CFS_PAGE_SIZE &&
830                    last_count != CFS_PAGE_SIZE) {
831                 /* For short writes we shouldn't count parts of pages that
832                  * span a whole chunk on the OST side, or our accounting goes
833                  * wrong.  Should match the code in filter_grant_check. */
834                 int offset = oap->oap_page_off & ~CFS_PAGE_MASK;
835                 int count = oap->oap_count + (offset & (blocksize - 1));
836                 int end = (offset + oap->oap_count) & (blocksize - 1);
837                 if (end)
838                         count += blocksize - end;
839
840                 lost_grant = CFS_PAGE_SIZE - count;
841         }
842         if (ext->oe_grants > 0)
843                 osc_free_grant(cli, nr_pages, lost_grant);
844
845         osc_extent_remove(ext);
846         /* put the refcount for RPC */
847         osc_extent_put(env, ext);
848         RETURN(0);
849 }
850
851 static int extent_wait_cb(struct osc_extent *ext, int state)
852 {
853         int ret;
854
855         osc_object_lock(ext->oe_obj);
856         ret = ext->oe_state == state;
857         osc_object_unlock(ext->oe_obj);
858
859         return ret;
860 }
861
862 /**
863  * Wait for the extent's state to become @state.
864  */
865 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
866                            int state)
867 {
868         struct osc_object *obj = ext->oe_obj;
869         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
870                                                   LWI_ON_SIGNAL_NOOP, NULL);
871         int rc = 0;
872         ENTRY;
873
874         osc_object_lock(obj);
875         LASSERT(sanity_check_nolock(ext) == 0);
876         /* `Kick' this extent only if the caller is waiting for it to be
877          * written out. */
878         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
879                 if (ext->oe_state == OES_ACTIVE) {
880                         ext->oe_urgent = 1;
881                 } else if (ext->oe_state == OES_CACHE) {
882                         ext->oe_urgent = 1;
883                         osc_extent_hold(ext);
884                         rc = 1;
885                 }
886         }
887         osc_object_unlock(obj);
888         if (rc == 1)
889                 osc_extent_release(env, ext);
890
891         /* wait for the extent until its state becomes @state */
892         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
893         if (rc == -ETIMEDOUT) {
894                 OSC_EXTENT_DUMP(D_ERROR, ext,
895                         "%s: wait ext to %d timedout, recovery in progress?\n",
896                         osc_export(obj)->exp_obd->obd_name, state);
897
898                 lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
899                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
900                                   &lwi);
901         }
902         if (rc == 0 && ext->oe_rc < 0)
903                 rc = ext->oe_rc;
904         RETURN(rc);
905 }
906
907 /**
908  * Discard pages with index greater than @size. If @ext is overlapped with
909  * @size, then partial truncate happens.
910  */
911 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
912                                 bool partial)
913 {
914         struct cl_env_nest     nest;
915         struct lu_env         *env;
916         struct cl_io          *io;
917         struct osc_object     *obj = ext->oe_obj;
918         struct client_obd     *cli = osc_cli(obj);
919         struct osc_async_page *oap;
920         struct osc_async_page *tmp;
921         int                    pages_in_chunk = 0;
922         int                    ppc_bits    = cli->cl_chunkbits - CFS_PAGE_SHIFT;
923         __u64                  trunc_chunk = trunc_index >> ppc_bits;
924         int                    grants   = 0;
925         int                    nr_pages = 0;
926         int                    rc       = 0;
927         ENTRY;
928
929         LASSERT(sanity_check(ext) == 0);
930         LASSERT(ext->oe_state == OES_TRUNC);
931         LASSERT(!ext->oe_urgent);
932
933         /* Request new lu_env.
934          * We can't use that env from osc_cache_truncate_start() because
935          * it's from lov_io_sub and not fully initialized. */
936         env = cl_env_nested_get(&nest);
937         io  = &osc_env_info(env)->oti_io;
938         io->ci_obj = cl_object_top(osc2cl(obj));
939         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
940         if (rc < 0)
941                 GOTO(out, rc);
942
943         /* discard all pages with index greater then trunc_index */
944         cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
945                                      oap_pending_item) {
946                 struct cl_page  *sub  = oap2cl_page(oap);
947                 struct cl_page  *page = cl_page_top(sub);
948
949                 LASSERT(cfs_list_empty(&oap->oap_rpc_item));
950
951                 /* only discard the pages with their index greater than
952                  * trunc_index, and ... */
953                 if (sub->cp_index < trunc_index ||
954                     (sub->cp_index == trunc_index && partial)) {
955                         /* accounting how many pages remaining in the chunk
956                          * so that we can calculate grants correctly. */
957                         if (sub->cp_index >> ppc_bits == trunc_chunk)
958                                 ++pages_in_chunk;
959                         continue;
960                 }
961
962                 cfs_list_del_init(&oap->oap_pending_item);
963
964                 cl_page_get(page);
965                 lu_ref_add(&page->cp_reference, "truncate", cfs_current());
966
967                 if (cl_page_own(env, io, page) == 0) {
968                         cl_page_unmap(env, io, page);
969                         cl_page_discard(env, io, page);
970                         cl_page_disown(env, io, page);
971                 } else {
972                         LASSERT(page->cp_state == CPS_FREEING);
973                         LASSERT(0);
974                 }
975
976                 lu_ref_del(&page->cp_reference, "truncate", cfs_current());
977                 cl_page_put(env, page);
978
979                 --ext->oe_nr_pages;
980                 ++nr_pages;
981         }
982         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
983                       ext->oe_nr_pages == 0),
984                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
985
986         osc_object_lock(obj);
987         if (ext->oe_nr_pages == 0) {
988                 LASSERT(pages_in_chunk == 0);
989                 grants = ext->oe_grants;
990                 ext->oe_grants = 0;
991         } else { /* calculate how many grants we can free */
992                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
993                 pgoff_t last_index;
994
995
996                 /* if there is no pages in this chunk, we can also free grants
997                  * for the last chunk */
998                 if (pages_in_chunk == 0) {
999                         /* if this is the 1st chunk and no pages in this chunk,
1000                          * ext->oe_nr_pages must be zero, so we should be in
1001                          * the other if-clause. */
1002                         LASSERT(trunc_chunk > 0);
1003                         --trunc_chunk;
1004                         ++chunks;
1005                 }
1006
1007                 /* this is what we can free from this extent */
1008                 grants          = chunks << cli->cl_chunkbits;
1009                 ext->oe_grants -= grants;
1010                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1011                 ext->oe_end     = min(last_index, ext->oe_max_end);
1012                 LASSERT(ext->oe_end >= ext->oe_start);
1013                 LASSERT(ext->oe_grants > 0);
1014         }
1015         osc_object_unlock(obj);
1016
1017         if (grants > 0 || nr_pages > 0)
1018                 osc_free_grant(cli, nr_pages, grants);
1019
1020 out:
1021         cl_io_fini(env, io);
1022         cl_env_nested_put(&nest, env);
1023         RETURN(rc);
1024 }
1025
1026 /**
1027  * This function is used to make the extent prepared for transfer.
1028  * A race with flusing page - ll_writepage() has to be handled cautiously.
1029  */
1030 static int osc_extent_make_ready(const struct lu_env *env,
1031                                  struct osc_extent *ext)
1032 {
1033         struct osc_async_page *oap;
1034         struct osc_async_page *last = NULL;
1035         struct osc_object *obj = ext->oe_obj;
1036         int page_count = 0;
1037         int rc;
1038         ENTRY;
1039
1040         /* we're going to grab page lock, so object lock must not be taken. */
1041         LASSERT(sanity_check(ext) == 0);
1042         /* in locking state, any process should not touch this extent. */
1043         EASSERT(ext->oe_state == OES_LOCKING, ext);
1044         EASSERT(ext->oe_owner != NULL, ext);
1045
1046         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1047
1048         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1049                 ++page_count;
1050                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1051                         last = oap;
1052
1053                 /* checking ASYNC_READY is race safe */
1054                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1055                         continue;
1056
1057                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1058                 switch (rc) {
1059                 case 0:
1060                         spin_lock(&oap->oap_lock);
1061                         oap->oap_async_flags |= ASYNC_READY;
1062                         spin_unlock(&oap->oap_lock);
1063                         break;
1064                 case -EALREADY:
1065                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1066                         break;
1067                 default:
1068                         LASSERTF(0, "unknown return code: %d\n", rc);
1069                 }
1070         }
1071
1072         LASSERT(page_count == ext->oe_nr_pages);
1073         LASSERT(last != NULL);
1074         /* the last page is the only one we need to refresh its count by
1075          * the size of file. */
1076         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1077                 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1078                 LASSERT(last->oap_count > 0);
1079                 LASSERT(last->oap_page_off + last->oap_count <= CFS_PAGE_SIZE);
1080                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1081         }
1082
1083         /* for the rest of pages, we don't need to call osf_refresh_count()
1084          * because it's known they are not the last page */
1085         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1086                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1087                         oap->oap_count = CFS_PAGE_SIZE - oap->oap_page_off;
1088                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1089                 }
1090         }
1091
1092         osc_object_lock(obj);
1093         osc_extent_state_set(ext, OES_RPC);
1094         osc_object_unlock(obj);
1095         /* get a refcount for RPC. */
1096         osc_extent_get(ext);
1097
1098         RETURN(0);
1099 }
1100
1101 /**
1102  * Quick and simple version of osc_extent_find(). This function is frequently
1103  * called to expand the extent for the same IO. To expand the extent, the
1104  * page index must be in the same or next chunk of ext->oe_end.
1105  */
1106 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1107 {
1108         struct osc_object *obj = ext->oe_obj;
1109         struct client_obd *cli = osc_cli(obj);
1110         struct osc_extent *next;
1111         int ppc_bits = cli->cl_chunkbits - CFS_PAGE_SHIFT;
1112         pgoff_t chunk = index >> ppc_bits;
1113         pgoff_t end_chunk;
1114         pgoff_t end_index;
1115         int chunksize = 1 << cli->cl_chunkbits;
1116         int rc = 0;
1117         ENTRY;
1118
1119         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1120         osc_object_lock(obj);
1121         LASSERT(sanity_check_nolock(ext) == 0);
1122         end_chunk = ext->oe_end >> ppc_bits;
1123         if (chunk > end_chunk + 1)
1124                 GOTO(out, rc = -ERANGE);
1125
1126         if (end_chunk >= chunk)
1127                 GOTO(out, rc = 0);
1128
1129         LASSERT(end_chunk + 1 == chunk);
1130         /* try to expand this extent to cover @index */
1131         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1132
1133         next = next_extent(ext);
1134         if (next != NULL && next->oe_start <= end_index)
1135                 /* complex mode - overlapped with the next extent,
1136                  * this case will be handled by osc_extent_find() */
1137                 GOTO(out, rc = -EAGAIN);
1138
1139         ext->oe_end = end_index;
1140         ext->oe_grants += chunksize;
1141         *grants -= chunksize;
1142         LASSERT(*grants >= 0);
1143         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1144                  "overlapped after expanding for %lu.\n", index);
1145         EXIT;
1146
1147 out:
1148         osc_object_unlock(obj);
1149         RETURN(rc);
1150 }
1151
1152 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1153                                   const char *func, int line)
1154 {
1155         struct osc_extent *ext;
1156         int cnt;
1157
1158         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1159                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1160
1161         /* osc_object_lock(obj); */
1162         cnt = 1;
1163         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1164                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1165
1166         cnt = 1;
1167         cfs_list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1168                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1169
1170         cnt = 1;
1171         cfs_list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1172                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1173
1174         cnt = 1;
1175         cfs_list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1176                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1177         /* osc_object_unlock(obj); */
1178 }
1179
1180 /* ------------------ osc extent end ------------------ */
1181
1182 static inline int osc_is_ready(struct osc_object *osc)
1183 {
1184         return !cfs_list_empty(&osc->oo_ready_item) ||
1185                !cfs_list_empty(&osc->oo_hp_ready_item);
1186 }
1187
1188 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1189         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1190                (OSC), osc_is_ready(OSC),                                       \
1191                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1192                list_empty_marker(&(OSC)->oo_ready_item),                       \
1193                cfs_atomic_read(&(OSC)->oo_nr_writes),                          \
1194                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1195                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1196                cfs_atomic_read(&(OSC)->oo_nr_reads),                           \
1197                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1198                ##args)
1199
1200 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1201                           int cmd)
1202 {
1203         struct osc_page *opg  = oap2osc_page(oap);
1204         struct cl_page  *page = cl_page_top(oap2cl_page(oap));
1205         int result;
1206
1207         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1208
1209         ENTRY;
1210         result = cl_page_make_ready(env, page, CRT_WRITE);
1211         if (result == 0)
1212                 opg->ops_submit_time = cfs_time_current();
1213         RETURN(result);
1214 }
1215
1216 static int osc_refresh_count(const struct lu_env *env,
1217                              struct osc_async_page *oap, int cmd)
1218 {
1219         struct osc_page  *opg = oap2osc_page(oap);
1220         struct cl_page   *page = oap2cl_page(oap);
1221         struct cl_object *obj;
1222         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1223
1224         int result;
1225         loff_t kms;
1226
1227         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1228         LASSERT(!(cmd & OBD_BRW_READ));
1229         LASSERT(opg != NULL);
1230         obj = opg->ops_cl.cpl_obj;
1231
1232         cl_object_attr_lock(obj);
1233         result = cl_object_attr_get(env, obj, attr);
1234         cl_object_attr_unlock(obj);
1235         if (result < 0)
1236                 return result;
1237         kms = attr->cat_kms;
1238         if (cl_offset(obj, page->cp_index) >= kms)
1239                 /* catch race with truncate */
1240                 return 0;
1241         else if (cl_offset(obj, page->cp_index + 1) > kms)
1242                 /* catch sub-page write at end of file */
1243                 return kms % CFS_PAGE_SIZE;
1244         else
1245                 return CFS_PAGE_SIZE;
1246 }
1247
1248 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1249                           int cmd, int rc)
1250 {
1251         struct osc_page   *opg  = oap2osc_page(oap);
1252         struct cl_page    *page = cl_page_top(oap2cl_page(oap));
1253         struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
1254         enum cl_req_type   crt;
1255         int srvlock;
1256
1257         ENTRY;
1258
1259         cmd &= ~OBD_BRW_NOQUOTA;
1260         LASSERT(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ));
1261         LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
1262         LASSERT(opg->ops_transfer_pinned);
1263
1264         /*
1265          * page->cp_req can be NULL if io submission failed before
1266          * cl_req was allocated.
1267          */
1268         if (page->cp_req != NULL)
1269                 cl_req_page_done(env, page);
1270         LASSERT(page->cp_req == NULL);
1271
1272         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1273         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1274         opg->ops_transfer_pinned = 0;
1275
1276         spin_lock(&obj->oo_seatbelt);
1277         LASSERT(opg->ops_submitter != NULL);
1278         LASSERT(!cfs_list_empty(&opg->ops_inflight));
1279         cfs_list_del_init(&opg->ops_inflight);
1280         opg->ops_submitter = NULL;
1281         spin_unlock(&obj->oo_seatbelt);
1282
1283         opg->ops_submit_time = 0;
1284         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1285
1286         /* statistic */
1287         if (rc == 0 && srvlock) {
1288                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1289                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1290                 int bytes = oap->oap_count;
1291
1292                 if (crt == CRT_READ)
1293                         stats->os_lockless_reads += bytes;
1294                 else
1295                         stats->os_lockless_writes += bytes;
1296         }
1297
1298         /*
1299          * This has to be the last operation with the page, as locks are
1300          * released in cl_page_completion() and nothing except for the
1301          * reference counter protects page from concurrent reclaim.
1302          */
1303         lu_ref_del(&page->cp_reference, "transfer", page);
1304
1305         cl_page_completion(env, page, crt, rc);
1306
1307         RETURN(0);
1308 }
1309
1310 #define OSC_DUMP_GRANT(cli, fmt, args...) do {                                \
1311         struct client_obd *__tmp = (cli);                                     \
1312         CDEBUG(D_CACHE, "%s: { dirty: %ld/%ld dirty_pages: %d/%d "            \
1313                "dropped: %ld avail: %ld, reserved: %ld, flight: %d } " fmt,   \
1314                __tmp->cl_import->imp_obd->obd_name,                           \
1315                __tmp->cl_dirty, __tmp->cl_dirty_max,                          \
1316                cfs_atomic_read(&obd_dirty_pages), obd_max_dirty_pages,        \
1317                __tmp->cl_lost_grant, __tmp->cl_avail_grant,                   \
1318                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, ##args);      \
1319 } while (0)
1320
1321 /* caller must hold loi_list_lock */
1322 static void osc_consume_write_grant(struct client_obd *cli,
1323                                     struct brw_page *pga)
1324 {
1325         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
1326         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1327         cfs_atomic_inc(&obd_dirty_pages);
1328         cli->cl_dirty += CFS_PAGE_SIZE;
1329         pga->flag |= OBD_BRW_FROM_GRANT;
1330         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1331                CFS_PAGE_SIZE, pga, pga->pg);
1332         osc_update_next_shrink(cli);
1333 }
1334
1335 /* the companion to osc_consume_write_grant, called when a brw has completed.
1336  * must be called with the loi lock held. */
1337 static void osc_release_write_grant(struct client_obd *cli,
1338                                     struct brw_page *pga)
1339 {
1340         ENTRY;
1341
1342         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
1343         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1344                 EXIT;
1345                 return;
1346         }
1347
1348         pga->flag &= ~OBD_BRW_FROM_GRANT;
1349         cfs_atomic_dec(&obd_dirty_pages);
1350         cli->cl_dirty -= CFS_PAGE_SIZE;
1351         if (pga->flag & OBD_BRW_NOCACHE) {
1352                 pga->flag &= ~OBD_BRW_NOCACHE;
1353                 cfs_atomic_dec(&obd_dirty_transit_pages);
1354                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
1355         }
1356         EXIT;
1357 }
1358
1359 /**
1360  * To avoid sleeping with object lock held, it's good for us allocate enough
1361  * grants before entering into critical section.
1362  *
1363  * client_obd_list_lock held by caller
1364  */
1365 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1366 {
1367         int rc = -EDQUOT;
1368
1369         if (cli->cl_avail_grant >= bytes) {
1370                 cli->cl_avail_grant    -= bytes;
1371                 cli->cl_reserved_grant += bytes;
1372                 rc = 0;
1373         }
1374         return rc;
1375 }
1376
1377 static void __osc_unreserve_grant(struct client_obd *cli,
1378                                   unsigned int reserved, unsigned int unused)
1379 {
1380         /* it's quite normal for us to get more grant than reserved.
1381          * Thinking about a case that two extents merged by adding a new
1382          * chunk, we can save one extent tax. If extent tax is greater than
1383          * one chunk, we can save more grant by adding a new chunk */
1384         cli->cl_reserved_grant -= reserved;
1385         if (unused > reserved) {
1386                 cli->cl_avail_grant += reserved;
1387                 cli->cl_lost_grant  += unused - reserved;
1388         } else {
1389                 cli->cl_avail_grant += unused;
1390         }
1391 }
1392
1393 void osc_unreserve_grant(struct client_obd *cli,
1394                          unsigned int reserved, unsigned int unused)
1395 {
1396         client_obd_list_lock(&cli->cl_loi_list_lock);
1397         __osc_unreserve_grant(cli, reserved, unused);
1398         if (unused > 0)
1399                 osc_wake_cache_waiters(cli);
1400         client_obd_list_unlock(&cli->cl_loi_list_lock);
1401 }
1402
1403 /**
1404  * Free grant after IO is finished or canceled.
1405  *
1406  * @lost_grant is used to remember how many grants we have allocated but not
1407  * used, we should return these grants to OST. There're two cases where grants
1408  * can be lost:
1409  * 1. truncate;
1410  * 2. blocksize at OST is less than CFS_PAGE_SIZE and a partial page was
1411  *    written. In this case OST may use less chunks to serve this partial
1412  *    write. OSTs don't actually know the page size on the client side. so
1413  *    clients have to calculate lost grant by the blocksize on the OST.
1414  *    See filter_grant_check() for details.
1415  */
1416 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1417                            unsigned int lost_grant)
1418 {
1419         int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1420
1421         client_obd_list_lock(&cli->cl_loi_list_lock);
1422         cfs_atomic_sub(nr_pages, &obd_dirty_pages);
1423         cli->cl_dirty -= nr_pages << CFS_PAGE_SHIFT;
1424         cli->cl_lost_grant += lost_grant;
1425         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1426                 /* borrow some grant from truncate to avoid the case that
1427                  * truncate uses up all avail grant */
1428                 cli->cl_lost_grant -= grant;
1429                 cli->cl_avail_grant += grant;
1430         }
1431         osc_wake_cache_waiters(cli);
1432         client_obd_list_unlock(&cli->cl_loi_list_lock);
1433         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1434                lost_grant, cli->cl_lost_grant,
1435                cli->cl_avail_grant, cli->cl_dirty);
1436 }
1437
1438 /**
1439  * The companion to osc_enter_cache(), called when @oap is no longer part of
1440  * the dirty accounting due to error.
1441  */
1442 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1443 {
1444         client_obd_list_lock(&cli->cl_loi_list_lock);
1445         osc_release_write_grant(cli, &oap->oap_brw_page);
1446         client_obd_list_unlock(&cli->cl_loi_list_lock);
1447 }
1448
1449 /**
1450  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1451  * is available.
1452  */
1453 static int osc_enter_cache_try(struct client_obd *cli,
1454                                struct osc_async_page *oap,
1455                                int bytes, int transient)
1456 {
1457         int rc;
1458
1459         OSC_DUMP_GRANT(cli, "need:%d.\n", bytes);
1460
1461         rc = osc_reserve_grant(cli, bytes);
1462         if (rc < 0)
1463                 return 0;
1464
1465         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
1466             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) {
1467                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1468                 if (transient) {
1469                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
1470                         cfs_atomic_inc(&obd_dirty_transit_pages);
1471                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1472                 }
1473                 rc = 1;
1474         } else {
1475                 __osc_unreserve_grant(cli, bytes, bytes);
1476                 rc = 0;
1477         }
1478         return rc;
1479 }
1480
1481 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1482 {
1483         int rc;
1484         client_obd_list_lock(&cli->cl_loi_list_lock);
1485         rc = cfs_list_empty(&ocw->ocw_entry) || cli->cl_w_in_flight == 0;
1486         client_obd_list_unlock(&cli->cl_loi_list_lock);
1487         return rc;
1488 }
1489
1490 /**
1491  * The main entry to reserve dirty page accounting. Usually the grant reserved
1492  * in this function will be freed in bulk in osc_free_grant() unless it fails
1493  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1494  *
1495  * The process will be put into sleep if it's already run out of grant.
1496  */
1497 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1498                            struct osc_async_page *oap, int bytes)
1499 {
1500         struct osc_object *osc = oap->oap_obj;
1501         struct lov_oinfo  *loi = osc->oo_oinfo;
1502         struct osc_cache_waiter ocw;
1503         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1504         int rc = -EDQUOT;
1505         ENTRY;
1506
1507         OSC_DUMP_GRANT(cli, "need:%d.\n", bytes);
1508
1509         client_obd_list_lock(&cli->cl_loi_list_lock);
1510
1511         /* force the caller to try sync io.  this can jump the list
1512          * of queued writes and create a discontiguous rpc stream */
1513         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1514             cli->cl_dirty_max < CFS_PAGE_SIZE     ||
1515             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1516                 GOTO(out, rc = -EDQUOT);
1517
1518         /* Hopefully normal case - cache space and write credits available */
1519         if (osc_enter_cache_try(cli, oap, bytes, 0))
1520                 GOTO(out, rc = 0);
1521
1522         /* We can get here for two reasons: too many dirty pages in cache, or
1523          * run out of grants. In both cases we should write dirty pages out.
1524          * Adding a cache waiter will trigger urgent write-out no matter what
1525          * RPC size will be.
1526          * The exiting condition is no avail grants and no dirty pages caching,
1527          * that really means there is no space on the OST. */
1528         cfs_waitq_init(&ocw.ocw_waitq);
1529         ocw.ocw_oap   = oap;
1530         ocw.ocw_grant = bytes;
1531         if (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
1532                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1533                 ocw.ocw_rc = 0;
1534                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1535
1536                 /* First osc_io_unplug() tries to put current object
1537                  * on ready list, second osc_io_unplug() makes sure that
1538                  * dirty flush can still be triggered even if current
1539                  * object hasn't any dirty pages */
1540                 osc_io_unplug(env, cli, osc, PDL_POLICY_ROUND);
1541                 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
1542
1543                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1544                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1545
1546                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1547
1548                 client_obd_list_lock(&cli->cl_loi_list_lock);
1549
1550                 /* l_wait_event is interrupted by signal */
1551                 if (rc < 0) {
1552                         cfs_list_del_init(&ocw.ocw_entry);
1553                         GOTO(out, rc);
1554                 }
1555
1556                 /* If ocw_entry isn't empty, which means it's not waked up
1557                  * by osc_wake_cache_waiters(), then the page must not be
1558                  * granted yet. */
1559                 if (!cfs_list_empty(&ocw.ocw_entry)) {
1560                         rc = -EDQUOT;
1561                         cfs_list_del_init(&ocw.ocw_entry);
1562                 } else {
1563                         rc = ocw.ocw_rc;
1564                 }
1565
1566                 if (rc != -EDQUOT)
1567                         GOTO(out, rc);
1568                 if (osc_enter_cache_try(cli, oap, bytes, 0))
1569                         rc = 0;
1570         }
1571         EXIT;
1572 out:
1573         client_obd_list_unlock(&cli->cl_loi_list_lock);
1574         OSC_DUMP_GRANT(cli, "returned %d.\n", rc);
1575         RETURN(rc);
1576 }
1577
1578 /* caller must hold loi_list_lock */
1579 void osc_wake_cache_waiters(struct client_obd *cli)
1580 {
1581         cfs_list_t *l, *tmp;
1582         struct osc_cache_waiter *ocw;
1583
1584         ENTRY;
1585         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1586                 /* if we can't dirty more, we must wait until some is written */
1587                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
1588                     (cfs_atomic_read(&obd_dirty_pages) + 1 >
1589                      obd_max_dirty_pages)) {
1590                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1591                                "osc max %ld, sys max %d\n", cli->cl_dirty,
1592                                cli->cl_dirty_max, obd_max_dirty_pages);
1593                         return;
1594                 }
1595
1596                 /* if still dirty cache but no grant wait for pending RPCs that
1597                  * may yet return us some grant before doing sync writes */
1598                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
1599                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
1600                                cli->cl_w_in_flight);
1601                         return;
1602                 }
1603
1604                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
1605                 cfs_list_del_init(&ocw->ocw_entry);
1606
1607                 ocw->ocw_rc = 0;
1608                 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1609                         ocw->ocw_rc = -EDQUOT;
1610
1611                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1612                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1613
1614                 cfs_waitq_signal(&ocw->ocw_waitq);
1615         }
1616
1617         EXIT;
1618 }
1619
1620 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1621 {
1622         int hprpc = !!cfs_list_empty(&osc->oo_hp_exts);
1623         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1624 }
1625
1626 /* This maintains the lists of pending pages to read/write for a given object
1627  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1628  * to quickly find objects that are ready to send an RPC. */
1629 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1630                          int cmd)
1631 {
1632         int invalid_import = 0;
1633         ENTRY;
1634
1635         /* if we have an invalid import we want to drain the queued pages
1636          * by forcing them through rpcs that immediately fail and complete
1637          * the pages.  recovery relies on this to empty the queued pages
1638          * before canceling the locks and evicting down the llite pages */
1639         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1640                 invalid_import = 1;
1641
1642         if (cmd & OBD_BRW_WRITE) {
1643                 if (cfs_atomic_read(&osc->oo_nr_writes) == 0)
1644                         RETURN(0);
1645                 if (invalid_import) {
1646                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1647                         RETURN(1);
1648                 }
1649                 if (!cfs_list_empty(&osc->oo_hp_exts)) {
1650                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1651                         RETURN(1);
1652                 }
1653                 if (!cfs_list_empty(&osc->oo_urgent_exts)) {
1654                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1655                         RETURN(1);
1656                 }
1657                 /* trigger a write rpc stream as long as there are dirtiers
1658                  * waiting for space.  as they're waiting, they're not going to
1659                  * create more pages to coalesce with what's waiting.. */
1660                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1661                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1662                         RETURN(1);
1663                 }
1664                 if (cfs_atomic_read(&osc->oo_nr_writes) >=
1665                     cli->cl_max_pages_per_rpc)
1666                         RETURN(1);
1667         } else {
1668                 if (cfs_atomic_read(&osc->oo_nr_reads) == 0)
1669                         RETURN(0);
1670                 if (invalid_import) {
1671                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1672                         RETURN(1);
1673                 }
1674                 /* all read are urgent. */
1675                 if (!cfs_list_empty(&osc->oo_reading_exts))
1676                         RETURN(1);
1677         }
1678
1679         RETURN(0);
1680 }
1681
1682 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1683 {
1684         struct client_obd *cli = osc_cli(obj);
1685         if (cmd & OBD_BRW_WRITE) {
1686                 cfs_atomic_add(delta, &obj->oo_nr_writes);
1687                 cfs_atomic_add(delta, &cli->cl_pending_w_pages);
1688                 LASSERT(cfs_atomic_read(&obj->oo_nr_writes) >= 0);
1689         } else {
1690                 cfs_atomic_add(delta, &obj->oo_nr_reads);
1691                 cfs_atomic_add(delta, &cli->cl_pending_r_pages);
1692                 LASSERT(cfs_atomic_read(&obj->oo_nr_reads) >= 0);
1693         }
1694         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1695 }
1696
1697 static int osc_makes_hprpc(struct osc_object *obj)
1698 {
1699         return !cfs_list_empty(&obj->oo_hp_exts);
1700 }
1701
1702 static void on_list(cfs_list_t *item, cfs_list_t *list, int should_be_on)
1703 {
1704         if (cfs_list_empty(item) && should_be_on)
1705                 cfs_list_add_tail(item, list);
1706         else if (!cfs_list_empty(item) && !should_be_on)
1707                 cfs_list_del_init(item);
1708 }
1709
1710 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1711  * can find pages to build into rpcs quickly */
1712 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1713 {
1714         if (osc_makes_hprpc(osc)) {
1715                 /* HP rpc */
1716                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1717                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1718         } else {
1719                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1720                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1721                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1722                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1723         }
1724
1725         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1726                 cfs_atomic_read(&osc->oo_nr_writes) > 0);
1727
1728         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1729                 cfs_atomic_read(&osc->oo_nr_reads) > 0);
1730
1731         return osc_is_ready(osc);
1732 }
1733
1734 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1735 {
1736         int is_ready;
1737
1738         client_obd_list_lock(&cli->cl_loi_list_lock);
1739         is_ready = __osc_list_maint(cli, osc);
1740         client_obd_list_unlock(&cli->cl_loi_list_lock);
1741
1742         return is_ready;
1743 }
1744
1745 /* this is trying to propogate async writeback errors back up to the
1746  * application.  As an async write fails we record the error code for later if
1747  * the app does an fsync.  As long as errors persist we force future rpcs to be
1748  * sync so that the app can get a sync error and break the cycle of queueing
1749  * pages for which writeback will fail. */
1750 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1751                            int rc)
1752 {
1753         if (rc) {
1754                 if (!ar->ar_rc)
1755                         ar->ar_rc = rc;
1756
1757                 ar->ar_force_sync = 1;
1758                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1759                 return;
1760
1761         }
1762
1763         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1764                 ar->ar_force_sync = 0;
1765 }
1766
1767 /* this must be called holding the loi list lock to give coverage to exit_cache,
1768  * async_flag maintenance, and oap_request */
1769 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1770                               struct osc_async_page *oap, int sent, int rc)
1771 {
1772         struct osc_object *osc = oap->oap_obj;
1773         struct lov_oinfo  *loi = osc->oo_oinfo;
1774         __u64 xid = 0;
1775
1776         ENTRY;
1777         if (oap->oap_request != NULL) {
1778                 xid = ptlrpc_req_xid(oap->oap_request);
1779                 ptlrpc_req_finished(oap->oap_request);
1780                 oap->oap_request = NULL;
1781         }
1782
1783         /* As the transfer for this page is being done, clear the flags */
1784         spin_lock(&oap->oap_lock);
1785         oap->oap_async_flags = 0;
1786         spin_unlock(&oap->oap_lock);
1787         oap->oap_interrupted = 0;
1788
1789         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1790                 client_obd_list_lock(&cli->cl_loi_list_lock);
1791                 osc_process_ar(&cli->cl_ar, xid, rc);
1792                 osc_process_ar(&loi->loi_ar, xid, rc);
1793                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1794         }
1795
1796         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1797         if (rc)
1798                 CERROR("completion on oap %p obj %p returns %d.\n",
1799                        oap, osc, rc);
1800
1801         EXIT;
1802 }
1803
1804 /**
1805  * Try to add extent to one RPC. We need to think about the following things:
1806  * - # of pages must not be over max_pages_per_rpc
1807  * - extent must be compatible with previous ones
1808  */
1809 static int try_to_add_extent_for_io(struct client_obd *cli,
1810                                     struct osc_extent *ext, cfs_list_t *rpclist,
1811                                     int *pc, unsigned int *max_pages)
1812 {
1813         struct osc_extent *tmp;
1814         ENTRY;
1815
1816         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1817                 ext);
1818
1819         *max_pages = max(ext->oe_mppr, *max_pages);
1820         if (*pc + ext->oe_nr_pages > *max_pages)
1821                 RETURN(0);
1822
1823         cfs_list_for_each_entry(tmp, rpclist, oe_link) {
1824                 EASSERT(tmp->oe_owner == cfs_current(), tmp);
1825 #if 0
1826                 if (overlapped(tmp, ext)) {
1827                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1828                         EASSERT(0, ext);
1829                 }
1830 #endif
1831
1832                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1833                     !tmp->oe_grants != !ext->oe_grants)
1834                         RETURN(0);
1835
1836                 /* remove break for strict check */
1837                 break;
1838         }
1839
1840         *pc += ext->oe_nr_pages;
1841         cfs_list_move_tail(&ext->oe_link, rpclist);
1842         ext->oe_owner = cfs_current();
1843         RETURN(1);
1844 }
1845
1846 /**
1847  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1848  * get_write_extent() takes all appropriate extents in atomic.
1849  *
1850  * The following policy is used to collect extents for IO:
1851  * 1. Add as many HP extents as possible;
1852  * 2. Add the first urgent extent in urgent extent list and take it out of
1853  *    urgent list;
1854  * 3. Add subsequent extents of this urgent extent;
1855  * 4. If urgent list is not empty, goto 2;
1856  * 5. Traverse the extent tree from the 1st extent;
1857  * 6. Above steps exit if there is no space in this RPC.
1858  */
1859 static int get_write_extents(struct osc_object *obj, cfs_list_t *rpclist)
1860 {
1861         struct client_obd *cli = osc_cli(obj);
1862         struct osc_extent *ext;
1863         int page_count = 0;
1864         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1865
1866         LASSERT(osc_object_is_locked(obj));
1867         while (!cfs_list_empty(&obj->oo_hp_exts)) {
1868                 ext = cfs_list_entry(obj->oo_hp_exts.next, struct osc_extent,
1869                                      oe_link);
1870                 LASSERT(ext->oe_state == OES_CACHE);
1871                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1872                                               &max_pages))
1873                         return page_count;
1874                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1875         }
1876         if (page_count == max_pages)
1877                 return page_count;
1878
1879         while (!cfs_list_empty(&obj->oo_urgent_exts)) {
1880                 ext = cfs_list_entry(obj->oo_urgent_exts.next,
1881                                      struct osc_extent, oe_link);
1882                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1883                                               &max_pages))
1884                         return page_count;
1885
1886                 if (!ext->oe_intree)
1887                         continue;
1888
1889                 while ((ext = next_extent(ext)) != NULL) {
1890                         if ((ext->oe_state != OES_CACHE) ||
1891                             (!cfs_list_empty(&ext->oe_link) &&
1892                              ext->oe_owner != NULL))
1893                                 continue;
1894
1895                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1896                                                       &page_count, &max_pages))
1897                                 return page_count;
1898                 }
1899         }
1900         if (page_count == max_pages)
1901                 return page_count;
1902
1903         ext = first_extent(obj);
1904         while (ext != NULL) {
1905                 if ((ext->oe_state != OES_CACHE) ||
1906                     /* this extent may be already in current rpclist */
1907                     (!cfs_list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1908                         ext = next_extent(ext);
1909                         continue;
1910                 }
1911
1912                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1913                                               &max_pages))
1914                         return page_count;
1915
1916                 ext = next_extent(ext);
1917         }
1918         return page_count;
1919 }
1920
1921 static int
1922 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1923                    struct osc_object *osc, pdl_policy_t pol)
1924 {
1925         CFS_LIST_HEAD(rpclist);
1926         struct osc_extent *ext;
1927         struct osc_extent *tmp;
1928         struct osc_extent *first = NULL;
1929         obd_count page_count = 0;
1930         int srvlock = 0;
1931         int rc = 0;
1932         ENTRY;
1933
1934         LASSERT(osc_object_is_locked(osc));
1935
1936         page_count = get_write_extents(osc, &rpclist);
1937         LASSERT(equi(page_count == 0, cfs_list_empty(&rpclist)));
1938
1939         if (cfs_list_empty(&rpclist))
1940                 RETURN(0);
1941
1942         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1943
1944         cfs_list_for_each_entry(ext, &rpclist, oe_link) {
1945                 LASSERT(ext->oe_state == OES_CACHE ||
1946                         ext->oe_state == OES_LOCK_DONE);
1947                 if (ext->oe_state == OES_CACHE)
1948                         osc_extent_state_set(ext, OES_LOCKING);
1949                 else
1950                         osc_extent_state_set(ext, OES_RPC);
1951         }
1952
1953         /* we're going to grab page lock, so release object lock because
1954          * lock order is page lock -> object lock. */
1955         osc_object_unlock(osc);
1956
1957         cfs_list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
1958                 if (ext->oe_state == OES_LOCKING) {
1959                         rc = osc_extent_make_ready(env, ext);
1960                         if (unlikely(rc < 0)) {
1961                                 cfs_list_del_init(&ext->oe_link);
1962                                 osc_extent_finish(env, ext, 0, rc);
1963                                 continue;
1964                         }
1965                 }
1966                 if (first == NULL) {
1967                         first = ext;
1968                         srvlock = ext->oe_srvlock;
1969                 } else {
1970                         LASSERT(srvlock == ext->oe_srvlock);
1971                 }
1972         }
1973
1974         if (!cfs_list_empty(&rpclist)) {
1975                 LASSERT(page_count > 0);
1976                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
1977                 LASSERT(cfs_list_empty(&rpclist));
1978         }
1979
1980         osc_object_lock(osc);
1981         RETURN(rc);
1982 }
1983
1984 /**
1985  * prepare pages for ASYNC io and put pages in send queue.
1986  *
1987  * \param cmd OBD_BRW_* macroses
1988  * \param lop pending pages
1989  *
1990  * \return zero if no page added to send queue.
1991  * \return 1 if pages successfully added to send queue.
1992  * \return negative on errors.
1993  */
1994 static int
1995 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
1996                   struct osc_object *osc, pdl_policy_t pol)
1997 {
1998         struct osc_extent *ext;
1999         struct osc_extent *next;
2000         CFS_LIST_HEAD(rpclist);
2001         int page_count = 0;
2002         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2003         int rc = 0;
2004         ENTRY;
2005
2006         LASSERT(osc_object_is_locked(osc));
2007         cfs_list_for_each_entry_safe(ext, next,
2008                                      &osc->oo_reading_exts, oe_link) {
2009                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2010                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2011                                               &max_pages))
2012                         break;
2013                 osc_extent_state_set(ext, OES_RPC);
2014                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2015         }
2016         LASSERT(page_count <= max_pages);
2017
2018         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2019
2020         if (!cfs_list_empty(&rpclist)) {
2021                 osc_object_unlock(osc);
2022
2023                 LASSERT(page_count > 0);
2024                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2025                 LASSERT(cfs_list_empty(&rpclist));
2026
2027                 osc_object_lock(osc);
2028         }
2029         RETURN(rc);
2030 }
2031
2032 #define list_to_obj(list, item) ({                                            \
2033         cfs_list_t *__tmp = (list)->next;                                     \
2034         cfs_list_del_init(__tmp);                                             \
2035         cfs_list_entry(__tmp, struct osc_object, oo_##item);                  \
2036 })
2037
2038 /* This is called by osc_check_rpcs() to find which objects have pages that
2039  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2040 static struct osc_object *osc_next_obj(struct client_obd *cli)
2041 {
2042         ENTRY;
2043
2044         /* First return objects that have blocked locks so that they
2045          * will be flushed quickly and other clients can get the lock,
2046          * then objects which have pages ready to be stuffed into RPCs */
2047         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2048                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2049         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2050                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2051
2052         /* then if we have cache waiters, return all objects with queued
2053          * writes.  This is especially important when many small files
2054          * have filled up the cache and not been fired into rpcs because
2055          * they don't pass the nr_pending/object threshhold */
2056         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2057             !cfs_list_empty(&cli->cl_loi_write_list))
2058                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2059
2060         /* then return all queued objects when we have an invalid import
2061          * so that they get flushed */
2062         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2063                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2064                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2065                                            write_item));
2066                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2067                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2068                                            read_item));
2069         }
2070         RETURN(NULL);
2071 }
2072
2073 /* called with the loi list lock held */
2074 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2075                            pdl_policy_t pol)
2076 {
2077         struct osc_object *osc;
2078         int rc = 0;
2079         ENTRY;
2080
2081         while ((osc = osc_next_obj(cli)) != NULL) {
2082                 struct cl_object *obj = osc2cl(osc);
2083                 struct lu_ref_link *link;
2084
2085                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2086
2087                 if (osc_max_rpc_in_flight(cli, osc)) {
2088                         __osc_list_maint(cli, osc);
2089                         break;
2090                 }
2091
2092                 cl_object_get(obj);
2093                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2094                 link = lu_object_ref_add(&obj->co_lu, "check", cfs_current());
2095
2096                 /* attempt some read/write balancing by alternating between
2097                  * reads and writes in an object.  The makes_rpc checks here
2098                  * would be redundant if we were getting read/write work items
2099                  * instead of objects.  we don't want send_oap_rpc to drain a
2100                  * partial read pending queue when we're given this object to
2101                  * do io on writes while there are cache waiters */
2102                 osc_object_lock(osc);
2103                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2104                         rc = osc_send_write_rpc(env, cli, osc, pol);
2105                         if (rc < 0) {
2106                                 CERROR("Write request failed with %d\n", rc);
2107
2108                                 /* osc_send_write_rpc failed, mostly because of
2109                                  * memory pressure.
2110                                  *
2111                                  * It can't break here, because if:
2112                                  *  - a page was submitted by osc_io_submit, so
2113                                  *    page locked;
2114                                  *  - no request in flight
2115                                  *  - no subsequent request
2116                                  * The system will be in live-lock state,
2117                                  * because there is no chance to call
2118                                  * osc_io_unplug() and osc_check_rpcs() any
2119                                  * more. pdflush can't help in this case,
2120                                  * because it might be blocked at grabbing
2121                                  * the page lock as we mentioned.
2122                                  *
2123                                  * Anyway, continue to drain pages. */
2124                                 /* break; */
2125                         }
2126                 }
2127                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2128                         rc = osc_send_read_rpc(env, cli, osc, pol);
2129                         if (rc < 0)
2130                                 CERROR("Read request failed with %d\n", rc);
2131                 }
2132                 osc_object_unlock(osc);
2133
2134                 osc_list_maint(cli, osc);
2135                 lu_object_ref_del_at(&obj->co_lu, link, "check", cfs_current());
2136                 cl_object_put(env, obj);
2137
2138                 client_obd_list_lock(&cli->cl_loi_list_lock);
2139         }
2140 }
2141
2142 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2143                           struct osc_object *osc, pdl_policy_t pol, int async)
2144 {
2145         int has_rpcs = 1;
2146         int rc = 0;
2147
2148         client_obd_list_lock(&cli->cl_loi_list_lock);
2149         if (osc != NULL)
2150                 has_rpcs = __osc_list_maint(cli, osc);
2151         if (has_rpcs) {
2152                 if (!async) {
2153                         osc_check_rpcs(env, cli, pol);
2154                 } else {
2155                         CDEBUG(D_CACHE, "Queue writeback work for client %p.\n",
2156                                cli);
2157                         LASSERT(cli->cl_writeback_work != NULL);
2158                         rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2159                 }
2160         }
2161         client_obd_list_unlock(&cli->cl_loi_list_lock);
2162         return rc;
2163 }
2164
2165 static int osc_io_unplug_async(const struct lu_env *env,
2166                                 struct client_obd *cli, struct osc_object *osc)
2167 {
2168         /* XXX: policy is no use actually. */
2169         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2170 }
2171
2172 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2173                    struct osc_object *osc, pdl_policy_t pol)
2174 {
2175         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2176 }
2177
2178 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2179                         cfs_page_t *page, loff_t offset)
2180 {
2181         struct obd_export     *exp = osc_export(osc);
2182         struct osc_async_page *oap = &ops->ops_oap;
2183         ENTRY;
2184
2185         if (!page)
2186                 return cfs_size_round(sizeof(*oap));
2187
2188         oap->oap_magic = OAP_MAGIC;
2189         oap->oap_cli = &exp->exp_obd->u.cli;
2190         oap->oap_obj = osc;
2191
2192         oap->oap_page = page;
2193         oap->oap_obj_off = offset;
2194         LASSERT(!(offset & ~CFS_PAGE_MASK));
2195
2196         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2197                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2198
2199         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2200         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2201
2202         spin_lock_init(&oap->oap_lock);
2203         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2204                oap, page, oap->oap_obj_off);
2205         RETURN(0);
2206 }
2207
2208 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2209                        struct osc_page *ops)
2210 {
2211         struct osc_io *oio = osc_env_io(env);
2212         struct osc_extent     *ext = NULL;
2213         struct osc_async_page *oap = &ops->ops_oap;
2214         struct client_obd     *cli = oap->oap_cli;
2215         struct osc_object     *osc = oap->oap_obj;
2216         pgoff_t index;
2217         int    grants = 0;
2218         int    brw_flags = OBD_BRW_ASYNC;
2219         int    cmd = OBD_BRW_WRITE;
2220         int    need_release = 0;
2221         int    rc = 0;
2222         ENTRY;
2223
2224         if (oap->oap_magic != OAP_MAGIC)
2225                 RETURN(-EINVAL);
2226
2227         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2228                 RETURN(-EIO);
2229
2230         if (!cfs_list_empty(&oap->oap_pending_item) ||
2231             !cfs_list_empty(&oap->oap_rpc_item))
2232                 RETURN(-EBUSY);
2233
2234         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2235         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2236         if (!client_is_remote(osc_export(osc)) &&
2237             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2238                 brw_flags |= OBD_BRW_NOQUOTA;
2239                 cmd |= OBD_BRW_NOQUOTA;
2240         }
2241
2242         /* check if the file's owner/group is over quota */
2243         if (!(cmd & OBD_BRW_NOQUOTA)) {
2244                 struct cl_object *obj;
2245                 struct cl_attr   *attr;
2246                 unsigned int qid[MAXQUOTAS];
2247
2248                 obj = cl_object_top(&osc->oo_cl);
2249                 attr = &osc_env_info(env)->oti_attr;
2250
2251                 cl_object_attr_lock(obj);
2252                 rc = cl_object_attr_get(env, obj, attr);
2253                 cl_object_attr_unlock(obj);
2254
2255                 qid[USRQUOTA] = attr->cat_uid;
2256                 qid[GRPQUOTA] = attr->cat_gid;
2257                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2258                         rc = -EDQUOT;
2259                 if (rc)
2260                         RETURN(rc);
2261         }
2262
2263         oap->oap_cmd = cmd;
2264         oap->oap_page_off = ops->ops_from;
2265         oap->oap_count = ops->ops_to - ops->ops_from;
2266         oap->oap_async_flags = 0;
2267         oap->oap_brw_flags = brw_flags;
2268
2269         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2270                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2271
2272         index = oap2cl_page(oap)->cp_index;
2273
2274         /* Add this page into extent by the following steps:
2275          * 1. if there exists an active extent for this IO, mostly this page
2276          *    can be added to the active extent and sometimes we need to
2277          *    expand extent to accomodate this page;
2278          * 2. otherwise, a new extent will be allocated. */
2279
2280         ext = oio->oi_active;
2281         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2282                 /* one chunk plus extent overhead must be enough to write this
2283                  * page */
2284                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2285                 if (ext->oe_end >= index)
2286                         grants = 0;
2287
2288                 /* it doesn't need any grant to dirty this page */
2289                 client_obd_list_lock(&cli->cl_loi_list_lock);
2290                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2291                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2292                 if (rc == 0) { /* try failed */
2293                         grants = 0;
2294                         need_release = 1;
2295                 } else if (ext->oe_end < index) {
2296                         int tmp = grants;
2297                         /* try to expand this extent */
2298                         rc = osc_extent_expand(ext, index, &tmp);
2299                         if (rc < 0) {
2300                                 need_release = 1;
2301                                 /* don't free reserved grant */
2302                         } else {
2303                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2304                                                 "expanded for %lu.\n", index);
2305                                 osc_unreserve_grant(cli, grants, tmp);
2306                                 grants = 0;
2307                         }
2308                 }
2309                 rc = 0;
2310         } else if (ext != NULL) {
2311                 /* index is located outside of active extent */
2312                 need_release = 1;
2313         }
2314         if (need_release) {
2315                 osc_extent_release(env, ext);
2316                 oio->oi_active = NULL;
2317                 ext = NULL;
2318         }
2319
2320         if (ext == NULL) {
2321                 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2322
2323                 /* try to find new extent to cover this page */
2324                 LASSERT(oio->oi_active == NULL);
2325                 /* we may have allocated grant for this page if we failed
2326                  * to expand the previous active extent. */
2327                 LASSERT(ergo(grants > 0, grants >= tmp));
2328
2329                 rc = 0;
2330                 if (grants == 0) {
2331                         /* we haven't allocated grant for this page. */
2332                         rc = osc_enter_cache(env, cli, oap, tmp);
2333                         if (rc == 0)
2334                                 grants = tmp;
2335                 }
2336
2337                 tmp = grants;
2338                 if (rc == 0) {
2339                         ext = osc_extent_find(env, osc, index, &tmp);
2340                         if (IS_ERR(ext)) {
2341                                 LASSERT(tmp == grants);
2342                                 osc_exit_cache(cli, oap);
2343                                 rc = PTR_ERR(ext);
2344                                 ext = NULL;
2345                         } else {
2346                                 oio->oi_active = ext;
2347                         }
2348                 }
2349                 if (grants > 0)
2350                         osc_unreserve_grant(cli, grants, tmp);
2351         }
2352
2353         LASSERT(ergo(rc == 0, ext != NULL));
2354         if (ext != NULL) {
2355                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2356                          ext, "index = %lu.\n", index);
2357                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2358
2359                 osc_object_lock(osc);
2360                 if (ext->oe_nr_pages == 0)
2361                         ext->oe_srvlock = ops->ops_srvlock;
2362                 else
2363                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2364                 ++ext->oe_nr_pages;
2365                 cfs_list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2366                 osc_object_unlock(osc);
2367         }
2368         RETURN(rc);
2369 }
2370
2371 int osc_teardown_async_page(const struct lu_env *env,
2372                             struct osc_object *obj, struct osc_page *ops)
2373 {
2374         struct osc_async_page *oap = &ops->ops_oap;
2375         struct osc_extent     *ext = NULL;
2376         int rc = 0;
2377         ENTRY;
2378
2379         LASSERT(oap->oap_magic == OAP_MAGIC);
2380
2381         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2382                oap, ops, oap2cl_page(oap)->cp_index);
2383
2384         osc_object_lock(obj);
2385         if (!cfs_list_empty(&oap->oap_rpc_item)) {
2386                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2387                 rc = -EBUSY;
2388         } else if (!cfs_list_empty(&oap->oap_pending_item)) {
2389                 ext = osc_extent_lookup(obj, oap2cl_page(oap)->cp_index);
2390                 /* only truncated pages are allowed to be taken out.
2391                  * See osc_extent_truncate() and osc_cache_truncate_start()
2392                  * for details. */
2393                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2394                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2395                                         oap2cl_page(oap)->cp_index);
2396                         rc = -EBUSY;
2397                 }
2398         }
2399         osc_object_unlock(obj);
2400         if (ext != NULL)
2401                 osc_extent_put(env, ext);
2402         RETURN(rc);
2403 }
2404
2405 /**
2406  * This is called when a page is picked up by kernel to write out.
2407  *
2408  * We should find out the corresponding extent and add the whole extent
2409  * into urgent list. The extent may be being truncated or used, handle it
2410  * carefully.
2411  */
2412 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2413                          struct osc_page *ops)
2414 {
2415         struct osc_extent *ext   = NULL;
2416         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2417         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2418         pgoff_t            index = cp->cp_index;
2419         struct osc_async_page *oap = &ops->ops_oap;
2420         bool unplug = false;
2421         int rc = 0;
2422         ENTRY;
2423
2424         osc_object_lock(obj);
2425         ext = osc_extent_lookup(obj, index);
2426         if (ext == NULL) {
2427                 osc_extent_tree_dump(D_ERROR, obj);
2428                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2429         }
2430
2431         switch (ext->oe_state) {
2432         case OES_RPC:
2433         case OES_LOCK_DONE:
2434                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(cp),
2435                               "flush an in-rpc page?\n");
2436                 LASSERT(0);
2437                 break;
2438         case OES_LOCKING:
2439                 /* If we know this extent is being written out, we should abort
2440                  * so that the writer can make this page ready. Otherwise, there
2441                  * exists a deadlock problem because other process can wait for
2442                  * page writeback bit holding page lock; and meanwhile in
2443                  * vvp_page_make_ready(), we need to grab page lock before
2444                  * really sending the RPC. */
2445         case OES_TRUNC:
2446                 /* race with truncate, page will be redirtied */
2447                 GOTO(out, rc = -EAGAIN);
2448         default:
2449                 break;
2450         }
2451
2452         rc = cl_page_prep(env, io, cl_page_top(cp), CRT_WRITE);
2453         if (rc)
2454                 GOTO(out, rc);
2455
2456         spin_lock(&oap->oap_lock);
2457         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2458         spin_unlock(&oap->oap_lock);
2459
2460         if (cfs_memory_pressure_get())
2461                 ext->oe_memalloc = 1;
2462
2463         ext->oe_urgent = 1;
2464         if (ext->oe_state == OES_CACHE) {
2465                 OSC_EXTENT_DUMP(D_CACHE, ext,
2466                                 "flush page %p make it urgent.\n", oap);
2467                 if (cfs_list_empty(&ext->oe_link))
2468                         cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2469                 unplug = true;
2470         }
2471         rc = 0;
2472         EXIT;
2473
2474 out:
2475         osc_object_unlock(obj);
2476         osc_extent_put(env, ext);
2477         if (unplug)
2478                 osc_io_unplug_async(env, osc_cli(obj), obj);
2479         return rc;
2480 }
2481
2482 /**
2483  * this is called when a sync waiter receives an interruption.  Its job is to
2484  * get the caller woken as soon as possible.  If its page hasn't been put in an
2485  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2486  * desiring interruption which will forcefully complete the rpc once the rpc
2487  * has timed out.
2488  */
2489 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2490 {
2491         struct osc_async_page *oap = &ops->ops_oap;
2492         struct osc_object     *obj = oap->oap_obj;
2493         struct client_obd     *cli = osc_cli(obj);
2494         struct osc_extent     *ext;
2495         struct osc_extent     *found = NULL;
2496         cfs_list_t            *plist;
2497         pgoff_t index = oap2cl_page(oap)->cp_index;
2498         int     rc = -EBUSY;
2499         int     cmd;
2500         ENTRY;
2501
2502         LASSERT(!oap->oap_interrupted);
2503         oap->oap_interrupted = 1;
2504
2505         /* Find out the caching extent */
2506         osc_object_lock(obj);
2507         if (oap->oap_cmd & OBD_BRW_WRITE) {
2508                 plist = &obj->oo_urgent_exts;
2509                 cmd   = OBD_BRW_WRITE;
2510         } else {
2511                 plist = &obj->oo_reading_exts;
2512                 cmd   = OBD_BRW_READ;
2513         }
2514         cfs_list_for_each_entry(ext, plist, oe_link) {
2515                 if (ext->oe_start <= index && ext->oe_end >= index) {
2516                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2517                         /* For OES_LOCK_DONE state extent, it has already held
2518                          * a refcount for RPC. */
2519                         found = osc_extent_get(ext);
2520                         break;
2521                 }
2522         }
2523         if (found != NULL) {
2524                 cfs_list_del_init(&found->oe_link);
2525                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2526                 osc_object_unlock(obj);
2527
2528                 osc_extent_finish(env, found, 0, -EINTR);
2529                 osc_extent_put(env, found);
2530                 rc = 0;
2531         } else {
2532                 osc_object_unlock(obj);
2533                 /* ok, it's been put in an rpc. only one oap gets a request
2534                  * reference */
2535                 if (oap->oap_request != NULL) {
2536                         ptlrpc_mark_interrupted(oap->oap_request);
2537                         ptlrpcd_wake(oap->oap_request);
2538                         ptlrpc_req_finished(oap->oap_request);
2539                         oap->oap_request = NULL;
2540                 }
2541         }
2542
2543         osc_list_maint(cli, obj);
2544         RETURN(rc);
2545 }
2546
2547 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2548                          cfs_list_t *list, int cmd, int brw_flags)
2549 {
2550         struct client_obd     *cli = osc_cli(obj);
2551         struct osc_extent     *ext;
2552         struct osc_async_page *oap;
2553         int     page_count = 0;
2554         int     mppr       = cli->cl_max_pages_per_rpc;
2555         pgoff_t start      = CL_PAGE_EOF;
2556         pgoff_t end        = 0;
2557         ENTRY;
2558
2559         cfs_list_for_each_entry(oap, list, oap_pending_item) {
2560                 struct cl_page *cp = oap2cl_page(oap);
2561                 if (cp->cp_index > end)
2562                         end = cp->cp_index;
2563                 if (cp->cp_index < start)
2564                         start = cp->cp_index;
2565                 ++page_count;
2566                 mppr <<= (page_count > mppr);
2567         }
2568
2569         ext = osc_extent_alloc(obj);
2570         if (ext == NULL) {
2571                 cfs_list_for_each_entry(oap, list, oap_pending_item) {
2572                         cfs_list_del_init(&oap->oap_pending_item);
2573                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2574                 }
2575                 RETURN(-ENOMEM);
2576         }
2577
2578         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2579         ext->oe_urgent = 1;
2580         ext->oe_start = start;
2581         ext->oe_end = ext->oe_max_end = end;
2582         ext->oe_obj = obj;
2583         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2584         ext->oe_nr_pages = page_count;
2585         ext->oe_mppr = mppr;
2586         cfs_list_splice_init(list, &ext->oe_pages);
2587
2588         osc_object_lock(obj);
2589         /* Reuse the initial refcount for RPC, don't drop it */
2590         osc_extent_state_set(ext, OES_LOCK_DONE);
2591         if (cmd & OBD_BRW_WRITE) {
2592                 cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2593                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2594         } else {
2595                 cfs_list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2596                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2597         }
2598         osc_object_unlock(obj);
2599
2600         osc_io_unplug(env, cli, obj, PDL_POLICY_ROUND);
2601         RETURN(0);
2602 }
2603
2604 /**
2605  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2606  */
2607 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2608                              struct osc_object *obj, __u64 size)
2609 {
2610         struct client_obd *cli = osc_cli(obj);
2611         struct osc_extent *ext;
2612         struct osc_extent *waiting = NULL;
2613         pgoff_t index;
2614         CFS_LIST_HEAD(list);
2615         int result = 0;
2616         bool partial;
2617         ENTRY;
2618
2619         /* pages with index greater or equal to index will be truncated. */
2620         index = cl_index(osc2cl(obj), size);
2621         partial = size > cl_offset(osc2cl(obj), index);
2622
2623 again:
2624         osc_object_lock(obj);
2625         ext = osc_extent_search(obj, index);
2626         if (ext == NULL)
2627                 ext = first_extent(obj);
2628         else if (ext->oe_end < index)
2629                 ext = next_extent(ext);
2630         while (ext != NULL) {
2631                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2632
2633                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2634                         /* if ext is in urgent state, it means there must exist
2635                          * a page already having been flushed by write_page().
2636                          * We have to wait for this extent because we can't
2637                          * truncate that page. */
2638                         LASSERT(!ext->oe_hp);
2639                         OSC_EXTENT_DUMP(D_CACHE, ext,
2640                                         "waiting for busy extent\n");
2641                         waiting = osc_extent_get(ext);
2642                         break;
2643                 }
2644
2645                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2646
2647                 osc_extent_get(ext);
2648                 if (ext->oe_state == OES_ACTIVE) {
2649                         /* though we grab inode mutex for write path, but we
2650                          * release it before releasing extent(in osc_io_end()),
2651                          * so there is a race window that an extent is still
2652                          * in OES_ACTIVE when truncate starts. */
2653                         LASSERT(!ext->oe_trunc_pending);
2654                         ext->oe_trunc_pending = 1;
2655                 } else {
2656                         EASSERT(ext->oe_state == OES_CACHE, ext);
2657                         osc_extent_state_set(ext, OES_TRUNC);
2658                         osc_update_pending(obj, OBD_BRW_WRITE,
2659                                            -ext->oe_nr_pages);
2660                 }
2661                 EASSERT(cfs_list_empty(&ext->oe_link), ext);
2662                 cfs_list_add_tail(&ext->oe_link, &list);
2663
2664                 ext = next_extent(ext);
2665         }
2666         osc_object_unlock(obj);
2667
2668         osc_list_maint(cli, obj);
2669
2670         while (!cfs_list_empty(&list)) {
2671                 int rc;
2672
2673                 ext = cfs_list_entry(list.next, struct osc_extent, oe_link);
2674                 cfs_list_del_init(&ext->oe_link);
2675
2676                 /* extent may be in OES_ACTIVE state because inode mutex
2677                  * is released before osc_io_end() in file write case */
2678                 if (ext->oe_state != OES_TRUNC)
2679                         osc_extent_wait(env, ext, OES_TRUNC);
2680
2681                 rc = osc_extent_truncate(ext, index, partial);
2682                 if (rc < 0) {
2683                         if (result == 0)
2684                                 result = rc;
2685
2686                         OSC_EXTENT_DUMP(D_ERROR, ext,
2687                                         "truncate error %d\n", rc);
2688                 } else if (ext->oe_nr_pages == 0) {
2689                         osc_extent_remove(ext);
2690                 } else {
2691                         /* this must be an overlapped extent which means only
2692                          * part of pages in this extent have been truncated.
2693                          */
2694                         EASSERTF(ext->oe_start <= index, ext,
2695                                  "trunc index = %lu/%d.\n", index, partial);
2696                         /* fix index to skip this partially truncated extent */
2697                         index = ext->oe_end + 1;
2698                         partial = false;
2699
2700                         /* we need to hold this extent in OES_TRUNC state so
2701                          * that no writeback will happen. This is to avoid
2702                          * BUG 17397. */
2703                         LASSERT(oio->oi_trunc == NULL);
2704                         oio->oi_trunc = osc_extent_get(ext);
2705                         OSC_EXTENT_DUMP(D_CACHE, ext,
2706                                         "trunc at "LPU64"\n", size);
2707                 }
2708                 osc_extent_put(env, ext);
2709         }
2710         if (waiting != NULL) {
2711                 int rc;
2712
2713                 /* ignore the result of osc_extent_wait the write initiator
2714                  * should take care of it. */
2715                 rc = osc_extent_wait(env, waiting, OES_INV);
2716                 if (rc < 0)
2717                         OSC_EXTENT_DUMP(D_CACHE, ext, "wait error: %d.\n", rc);
2718
2719                 osc_extent_put(env, waiting);
2720                 waiting = NULL;
2721                 goto again;
2722         }
2723         RETURN(result);
2724 }
2725
2726 /**
2727  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2728  */
2729 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2730                             struct osc_object *obj)
2731 {
2732         struct osc_extent *ext = oio->oi_trunc;
2733
2734         oio->oi_trunc = NULL;
2735         if (ext != NULL) {
2736                 bool unplug = false;
2737
2738                 EASSERT(ext->oe_nr_pages > 0, ext);
2739                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2740                 EASSERT(!ext->oe_urgent, ext);
2741
2742                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2743                 osc_object_lock(obj);
2744                 osc_extent_state_set(ext, OES_CACHE);
2745                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2746                         ext->oe_urgent = 1;
2747                         cfs_list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2748                         unplug = true;
2749                 }
2750                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2751                 osc_object_unlock(obj);
2752                 osc_extent_put(env, ext);
2753
2754                 if (unplug)
2755                         osc_io_unplug_async(env, osc_cli(obj), obj);
2756         }
2757 }
2758
2759 /**
2760  * Wait for extents in a specific range to be written out.
2761  * The caller must have called osc_cache_writeback_range() to issue IO
2762  * otherwise it will take a long time for this function to finish.
2763  *
2764  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2765  * nobody else can dirty this range of file while we're waiting for
2766  * extents to be written.
2767  */
2768 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2769                          pgoff_t start, pgoff_t end)
2770 {
2771         struct osc_extent *ext;
2772         pgoff_t index = start;
2773         int     result = 0;
2774         ENTRY;
2775
2776 again:
2777         osc_object_lock(obj);
2778         ext = osc_extent_search(obj, index);
2779         if (ext == NULL)
2780                 ext = first_extent(obj);
2781         else if (ext->oe_end < index)
2782                 ext = next_extent(ext);
2783         while (ext != NULL) {
2784                 int rc;
2785
2786                 if (ext->oe_start > end)
2787                         break;
2788
2789                 if (!ext->oe_fsync_wait) {
2790                         ext = next_extent(ext);
2791                         continue;
2792                 }
2793
2794                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2795                              ext->oe_hp || ext->oe_urgent), ext);
2796                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2797                              !ext->oe_hp && ext->oe_urgent), ext);
2798
2799                 index = ext->oe_end + 1;
2800                 osc_extent_get(ext);
2801                 osc_object_unlock(obj);
2802
2803                 rc = osc_extent_wait(env, ext, OES_INV);
2804                 if (result == 0)
2805                         result = rc;
2806                 osc_extent_put(env, ext);
2807                 goto again;
2808         }
2809         osc_object_unlock(obj);
2810
2811         OSC_IO_DEBUG(obj, "sync file range.\n");
2812         RETURN(result);
2813 }
2814
2815 /**
2816  * Called to write out a range of osc object.
2817  *
2818  * @hp     : should be set this is caused by lock cancel;
2819  * @discard: is set if dirty pages should be dropped - file will be deleted or
2820  *         truncated, this implies there is no partially discarding extents.
2821  *
2822  * Return how many pages will be issued, or error code if error occurred.
2823  */
2824 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2825                               pgoff_t start, pgoff_t end, int hp, int discard)
2826 {
2827         struct osc_extent *ext;
2828         CFS_LIST_HEAD(discard_list);
2829         bool unplug = false;
2830         int result = 0;
2831         ENTRY;
2832
2833         osc_object_lock(obj);
2834         ext = osc_extent_search(obj, start);
2835         if (ext == NULL)
2836                 ext = first_extent(obj);
2837         else if (ext->oe_end < start)
2838                 ext = next_extent(ext);
2839         while (ext != NULL) {
2840                 if (ext->oe_start > end)
2841                         break;
2842
2843                 ext->oe_fsync_wait = 1;
2844                 switch (ext->oe_state) {
2845                 case OES_CACHE:
2846                         result += ext->oe_nr_pages;
2847                         if (!discard) {
2848                                 cfs_list_t *list = NULL;
2849                                 if (hp) {
2850                                         EASSERT(!ext->oe_hp, ext);
2851                                         ext->oe_hp = 1;
2852                                         list = &obj->oo_hp_exts;
2853                                 } else if (!ext->oe_urgent) {
2854                                         ext->oe_urgent = 1;
2855                                         list = &obj->oo_urgent_exts;
2856                                 }
2857                                 if (list != NULL)
2858                                         cfs_list_move_tail(&ext->oe_link, list);
2859                                 unplug = true;
2860                         } else {
2861                                 /* the only discarder is lock cancelling, so
2862                                  * [start, end] must contain this extent */
2863                                 EASSERT(ext->oe_start >= start &&
2864                                         ext->oe_max_end <= end, ext);
2865                                 osc_extent_state_set(ext, OES_LOCKING);
2866                                 ext->oe_owner = cfs_current();
2867                                 cfs_list_move_tail(&ext->oe_link,
2868                                                    &discard_list);
2869                                 osc_update_pending(obj, OBD_BRW_WRITE,
2870                                                    -ext->oe_nr_pages);
2871                         }
2872                         break;
2873                 case OES_ACTIVE:
2874                         /* It's pretty bad to wait for ACTIVE extents, because
2875                          * we don't know how long we will wait for it to be
2876                          * flushed since it may be blocked at awaiting more
2877                          * grants. We do this for the correctness of fsync. */
2878                         LASSERT(hp == 0 && discard == 0);
2879                         ext->oe_urgent = 1;
2880                         break;
2881                 case OES_TRUNC:
2882                         /* this extent is being truncated, can't do anything
2883                          * for it now. it will be set to urgent after truncate
2884                          * is finished in osc_cache_truncate_end(). */
2885                 default:
2886                         break;
2887                 }
2888                 ext = next_extent(ext);
2889         }
2890         osc_object_unlock(obj);
2891
2892         LASSERT(ergo(!discard, cfs_list_empty(&discard_list)));
2893         if (!cfs_list_empty(&discard_list)) {
2894                 struct osc_extent *tmp;
2895                 int rc;
2896
2897                 osc_list_maint(osc_cli(obj), obj);
2898                 cfs_list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2899                         cfs_list_del_init(&ext->oe_link);
2900                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2901
2902                         /* Discard caching pages. We don't actually write this
2903                          * extent out but we complete it as if we did. */
2904                         rc = osc_extent_make_ready(env, ext);
2905                         if (unlikely(rc < 0)) {
2906                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2907                                                 "make_ready returned %d\n", rc);
2908                                 if (result >= 0)
2909                                         result = rc;
2910                         }
2911
2912                         /* finish the extent as if the pages were sent */
2913                         osc_extent_finish(env, ext, 0, 0);
2914                 }
2915         }
2916
2917         if (unplug)
2918                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2919
2920         if (hp || discard) {
2921                 int rc;
2922                 rc = osc_cache_wait_range(env, obj, start, end);
2923                 if (result >= 0 && rc < 0)
2924                         result = rc;
2925         }
2926
2927         OSC_IO_DEBUG(obj, "cache page out.\n");
2928         RETURN(result);
2929 }
2930
2931 /** @} osc */