Whamcloud - gitweb
LU-5521 grant: quiet message on grant waiting timeout
[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, 2014, 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                            enum osc_extent_state 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_sync)
80                 *buf++ = 'S';
81         if (ext->oe_srvlock)
82                 *buf++ = 's';
83         if (ext->oe_hp)
84                 *buf++ = 'h';
85         if (ext->oe_urgent)
86                 *buf++ = 'u';
87         if (ext->oe_memalloc)
88                 *buf++ = 'm';
89         if (ext->oe_trunc_pending)
90                 *buf++ = 't';
91         if (ext->oe_fsync_wait)
92                 *buf++ = 'Y';
93         *buf = 0;
94         return flags;
95 }
96
97 static inline char list_empty_marker(struct list_head *list)
98 {
99         return list_empty(list) ? '-' : '+';
100 }
101
102 #define EXTSTR       "[%lu -> %lu/%lu]"
103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
104 static const char *oes_strings[] = {
105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
106
107 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
108         struct osc_extent *__ext = (extent);                                  \
109         char __buf[16];                                                       \
110                                                                               \
111         CDEBUG(lvl,                                                           \
112                 "extent %p@{" EXTSTR ", "                                     \
113                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
114                 /* ----- extent part 0 ----- */                               \
115                 __ext, EXTPARA(__ext),                                        \
116                 /* ----- part 1 ----- */                                      \
117                 atomic_read(&__ext->oe_refc),                         \
118                 atomic_read(&__ext->oe_users),                        \
119                 list_empty_marker(&__ext->oe_link),                           \
120                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
121                 __ext->oe_obj,                                                \
122                 /* ----- part 2 ----- */                                      \
123                 __ext->oe_grants, __ext->oe_nr_pages,                         \
124                 list_empty_marker(&__ext->oe_pages),                          \
125                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
126                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
127                 /* ----- part 4 ----- */                                      \
128                 ## __VA_ARGS__);                                              \
129         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
130                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
131         else                                                                  \
132                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
133 } while (0)
134
135 #undef EASSERTF
136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
137         if (!(expr)) {                                                  \
138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
140                 LASSERT(expr);                                          \
141         }                                                               \
142 } while (0)
143
144 #undef EASSERT
145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
146
147 static inline struct osc_extent *rb_extent(struct rb_node *n)
148 {
149         if (n == NULL)
150                 return NULL;
151
152         return container_of(n, struct osc_extent, oe_node);
153 }
154
155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
156 {
157         if (ext == NULL)
158                 return NULL;
159
160         LASSERT(ext->oe_intree);
161         return rb_extent(rb_next(&ext->oe_node));
162 }
163
164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
165 {
166         if (ext == NULL)
167                 return NULL;
168
169         LASSERT(ext->oe_intree);
170         return rb_extent(rb_prev(&ext->oe_node));
171 }
172
173 static inline struct osc_extent *first_extent(struct osc_object *obj)
174 {
175         return rb_extent(rb_first(&obj->oo_root));
176 }
177
178 /* object must be locked by caller. */
179 static int osc_extent_sanity_check0(struct osc_extent *ext,
180                                     const char *func, const int line)
181 {
182         struct osc_object *obj = ext->oe_obj;
183         struct osc_async_page *oap;
184         size_t page_count;
185         int rc = 0;
186
187         if (!osc_object_is_locked(obj))
188                 GOTO(out, rc = 9);
189
190         if (ext->oe_state >= OES_STATE_MAX)
191                 GOTO(out, rc = 10);
192
193         if (atomic_read(&ext->oe_refc) <= 0)
194                 GOTO(out, rc = 20);
195
196         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
197                 GOTO(out, rc = 30);
198
199         switch (ext->oe_state) {
200         case OES_INV:
201                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
202                         GOTO(out, rc = 35);
203                 GOTO(out, rc = 0);
204                 break;
205         case OES_ACTIVE:
206                 if (atomic_read(&ext->oe_users) == 0)
207                         GOTO(out, rc = 40);
208                 if (ext->oe_hp)
209                         GOTO(out, rc = 50);
210                 if (ext->oe_fsync_wait && !ext->oe_urgent)
211                         GOTO(out, rc = 55);
212                 break;
213         case OES_CACHE:
214                 if (ext->oe_grants == 0)
215                         GOTO(out, rc = 60);
216                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
217                         GOTO(out, rc = 65);
218         default:
219                 if (atomic_read(&ext->oe_users) > 0)
220                         GOTO(out, rc = 70);
221         }
222
223         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
224                 GOTO(out, rc = 80);
225
226         if (ext->oe_sync && ext->oe_grants > 0)
227                 GOTO(out, rc = 90);
228
229         if (ext->oe_dlmlock != NULL) {
230                 struct ldlm_extent *extent;
231
232                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
233                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
234                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
235                         GOTO(out, rc = 100);
236
237                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
238                         GOTO(out, rc = 102);
239         }
240
241         if (ext->oe_nr_pages > ext->oe_mppr)
242                 GOTO(out, rc = 105);
243
244         /* Do not verify page list if extent is in RPC. This is because an
245          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
246         if (ext->oe_state > OES_CACHE)
247                 GOTO(out, rc = 0);
248
249         if (!extent_debug)
250                 GOTO(out, rc = 0);
251
252         page_count = 0;
253         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
254                 pgoff_t index = osc_index(oap2osc(oap));
255                 ++page_count;
256                 if (index > ext->oe_end || index < ext->oe_start)
257                         GOTO(out, rc = 110);
258         }
259         if (page_count != ext->oe_nr_pages)
260                 GOTO(out, rc = 120);
261
262 out:
263         if (rc != 0)
264                 OSC_EXTENT_DUMP(D_ERROR, ext,
265                                 "%s:%d sanity check %p failed with rc = %d\n",
266                                 func, line, ext, rc);
267         return rc;
268 }
269
270 #define sanity_check_nolock(ext) \
271         osc_extent_sanity_check0(ext, __func__, __LINE__)
272
273 #define sanity_check(ext) ({                                                   \
274         int __res;                                                             \
275         osc_object_lock((ext)->oe_obj);                                        \
276         __res = sanity_check_nolock(ext);                                      \
277         osc_object_unlock((ext)->oe_obj);                                      \
278         __res;                                                                 \
279 })
280
281
282 /**
283  * sanity check - to make sure there is no overlapped extent in the tree.
284  */
285 static int osc_extent_is_overlapped(struct osc_object *obj,
286                                     struct osc_extent *ext)
287 {
288         struct osc_extent *tmp;
289
290         LASSERT(osc_object_is_locked(obj));
291
292         if (!extent_debug)
293                 return 0;
294
295         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
296                 if (tmp == ext)
297                         continue;
298                 if (tmp->oe_end >= ext->oe_start &&
299                     tmp->oe_start <= ext->oe_end)
300                         return 1;
301         }
302         return 0;
303 }
304
305 static void osc_extent_state_set(struct osc_extent *ext, int state)
306 {
307         LASSERT(osc_object_is_locked(ext->oe_obj));
308         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
309
310         /* Never try to sanity check a state changing extent :-) */
311         /* LASSERT(sanity_check_nolock(ext) == 0); */
312
313         /* TODO: validate the state machine */
314         ext->oe_state = state;
315         wake_up_all(&ext->oe_waitq);
316 }
317
318 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
319 {
320         struct osc_extent *ext;
321
322         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
323         if (ext == NULL)
324                 return NULL;
325
326         RB_CLEAR_NODE(&ext->oe_node);
327         ext->oe_obj = obj;
328         atomic_set(&ext->oe_refc, 1);
329         atomic_set(&ext->oe_users, 0);
330         INIT_LIST_HEAD(&ext->oe_link);
331         ext->oe_state = OES_INV;
332         INIT_LIST_HEAD(&ext->oe_pages);
333         init_waitqueue_head(&ext->oe_waitq);
334         ext->oe_dlmlock = NULL;
335
336         return ext;
337 }
338
339 static void osc_extent_free(struct osc_extent *ext)
340 {
341         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
342 }
343
344 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
345 {
346         LASSERT(atomic_read(&ext->oe_refc) >= 0);
347         atomic_inc(&ext->oe_refc);
348         return ext;
349 }
350
351 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
352 {
353         LASSERT(atomic_read(&ext->oe_refc) > 0);
354         if (atomic_dec_and_test(&ext->oe_refc)) {
355                 LASSERT(list_empty(&ext->oe_link));
356                 LASSERT(atomic_read(&ext->oe_users) == 0);
357                 LASSERT(ext->oe_state == OES_INV);
358                 LASSERT(!ext->oe_intree);
359
360                 if (ext->oe_dlmlock != NULL) {
361                         lu_ref_add(&ext->oe_dlmlock->l_reference,
362                                    "osc_extent", ext);
363                         LDLM_LOCK_PUT(ext->oe_dlmlock);
364                         ext->oe_dlmlock = NULL;
365                 }
366                 osc_extent_free(ext);
367         }
368 }
369
370 /**
371  * osc_extent_put_trust() is a special version of osc_extent_put() when
372  * it's known that the caller is not the last user. This is to address the
373  * problem of lacking of lu_env ;-).
374  */
375 static void osc_extent_put_trust(struct osc_extent *ext)
376 {
377         LASSERT(atomic_read(&ext->oe_refc) > 1);
378         LASSERT(osc_object_is_locked(ext->oe_obj));
379         atomic_dec(&ext->oe_refc);
380 }
381
382 /**
383  * Return the extent which includes pgoff @index, or return the greatest
384  * previous extent in the tree.
385  */
386 static struct osc_extent *osc_extent_search(struct osc_object *obj,
387                                             pgoff_t index)
388 {
389         struct rb_node    *n = obj->oo_root.rb_node;
390         struct osc_extent *tmp, *p = NULL;
391
392         LASSERT(osc_object_is_locked(obj));
393         while (n != NULL) {
394                 tmp = rb_extent(n);
395                 if (index < tmp->oe_start) {
396                         n = n->rb_left;
397                 } else if (index > tmp->oe_end) {
398                         p = rb_extent(n);
399                         n = n->rb_right;
400                 } else {
401                         return tmp;
402                 }
403         }
404         return p;
405 }
406
407 /*
408  * Return the extent covering @index, otherwise return NULL.
409  * caller must have held object lock.
410  */
411 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
412                                             pgoff_t index)
413 {
414         struct osc_extent *ext;
415
416         ext = osc_extent_search(obj, index);
417         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
418                 return osc_extent_get(ext);
419         return NULL;
420 }
421
422 /* caller must have held object lock. */
423 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
424 {
425         struct rb_node   **n      = &obj->oo_root.rb_node;
426         struct rb_node    *parent = NULL;
427         struct osc_extent *tmp;
428
429         LASSERT(ext->oe_intree == 0);
430         LASSERT(ext->oe_obj == obj);
431         LASSERT(osc_object_is_locked(obj));
432         while (*n != NULL) {
433                 tmp = rb_extent(*n);
434                 parent = *n;
435
436                 if (ext->oe_end < tmp->oe_start)
437                         n = &(*n)->rb_left;
438                 else if (ext->oe_start > tmp->oe_end)
439                         n = &(*n)->rb_right;
440                 else
441                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
442         }
443         rb_link_node(&ext->oe_node, parent, n);
444         rb_insert_color(&ext->oe_node, &obj->oo_root);
445         osc_extent_get(ext);
446         ext->oe_intree = 1;
447 }
448
449 /* caller must have held object lock. */
450 static void osc_extent_erase(struct osc_extent *ext)
451 {
452         struct osc_object *obj = ext->oe_obj;
453         LASSERT(osc_object_is_locked(obj));
454         if (ext->oe_intree) {
455                 rb_erase(&ext->oe_node, &obj->oo_root);
456                 ext->oe_intree = 0;
457                 /* rbtree held a refcount */
458                 osc_extent_put_trust(ext);
459         }
460 }
461
462 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
463 {
464         struct osc_object *obj = ext->oe_obj;
465
466         LASSERT(osc_object_is_locked(obj));
467         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
468         if (ext->oe_state == OES_CACHE) {
469                 osc_extent_state_set(ext, OES_ACTIVE);
470                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
471         }
472         atomic_inc(&ext->oe_users);
473         list_del_init(&ext->oe_link);
474         return osc_extent_get(ext);
475 }
476
477 static void __osc_extent_remove(struct osc_extent *ext)
478 {
479         LASSERT(osc_object_is_locked(ext->oe_obj));
480         LASSERT(list_empty(&ext->oe_pages));
481         osc_extent_erase(ext);
482         list_del_init(&ext->oe_link);
483         osc_extent_state_set(ext, OES_INV);
484         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
485 }
486
487 static void osc_extent_remove(struct osc_extent *ext)
488 {
489         struct osc_object *obj = ext->oe_obj;
490
491         osc_object_lock(obj);
492         __osc_extent_remove(ext);
493         osc_object_unlock(obj);
494 }
495
496 /**
497  * This function is used to merge extents to get better performance. It checks
498  * if @cur and @victim are contiguous at chunk level.
499  */
500 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
501                             struct osc_extent *victim)
502 {
503         struct osc_object *obj = cur->oe_obj;
504         pgoff_t chunk_start;
505         pgoff_t chunk_end;
506         int ppc_bits;
507
508         LASSERT(cur->oe_state == OES_CACHE);
509         LASSERT(osc_object_is_locked(obj));
510         if (victim == NULL)
511                 return -EINVAL;
512
513         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
514                 return -EBUSY;
515
516         if (cur->oe_max_end != victim->oe_max_end)
517                 return -ERANGE;
518
519         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
520         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
521         chunk_start = cur->oe_start >> ppc_bits;
522         chunk_end   = cur->oe_end   >> ppc_bits;
523         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
524             chunk_end + 1 != victim->oe_start >> ppc_bits)
525                 return -ERANGE;
526
527         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
528
529         cur->oe_start     = min(cur->oe_start, victim->oe_start);
530         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
531         cur->oe_grants   += victim->oe_grants;
532         cur->oe_nr_pages += victim->oe_nr_pages;
533         /* only the following bits are needed to merge */
534         cur->oe_urgent   |= victim->oe_urgent;
535         cur->oe_memalloc |= victim->oe_memalloc;
536         list_splice_init(&victim->oe_pages, &cur->oe_pages);
537         list_del_init(&victim->oe_link);
538         victim->oe_nr_pages = 0;
539
540         osc_extent_get(victim);
541         __osc_extent_remove(victim);
542         osc_extent_put(env, victim);
543
544         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
545         return 0;
546 }
547
548 /**
549  * Drop user count of osc_extent, and unplug IO asynchronously.
550  */
551 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
552 {
553         struct osc_object *obj = ext->oe_obj;
554         int rc = 0;
555         ENTRY;
556
557         LASSERT(atomic_read(&ext->oe_users) > 0);
558         LASSERT(sanity_check(ext) == 0);
559         LASSERT(ext->oe_grants > 0);
560
561         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
562                 LASSERT(ext->oe_state == OES_ACTIVE);
563                 if (ext->oe_trunc_pending) {
564                         /* a truncate process is waiting for this extent.
565                          * This may happen due to a race, check
566                          * osc_cache_truncate_start(). */
567                         osc_extent_state_set(ext, OES_TRUNC);
568                         ext->oe_trunc_pending = 0;
569                 } else {
570                         osc_extent_state_set(ext, OES_CACHE);
571                         osc_update_pending(obj, OBD_BRW_WRITE,
572                                            ext->oe_nr_pages);
573
574                         /* try to merge the previous and next extent. */
575                         osc_extent_merge(env, ext, prev_extent(ext));
576                         osc_extent_merge(env, ext, next_extent(ext));
577
578                         if (ext->oe_urgent)
579                                 list_move_tail(&ext->oe_link,
580                                                &obj->oo_urgent_exts);
581                 }
582                 osc_object_unlock(obj);
583
584                 osc_io_unplug_async(env, osc_cli(obj), obj);
585         }
586         osc_extent_put(env, ext);
587         RETURN(rc);
588 }
589
590 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
591 {
592         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
593 }
594
595 /**
596  * Find or create an extent which includes @index, core function to manage
597  * extent tree.
598  */
599 static struct osc_extent *osc_extent_find(const struct lu_env *env,
600                                           struct osc_object *obj, pgoff_t index,
601                                           unsigned int *grants)
602 {
603         struct client_obd *cli = osc_cli(obj);
604         struct osc_lock   *olck;
605         struct cl_lock_descr *descr;
606         struct osc_extent *cur;
607         struct osc_extent *ext;
608         struct osc_extent *conflict = NULL;
609         struct osc_extent *found = NULL;
610         pgoff_t    chunk;
611         pgoff_t    max_end;
612         unsigned int max_pages; /* max_pages_per_rpc */
613         unsigned int chunksize;
614         int        ppc_bits; /* pages per chunk bits */
615         pgoff_t    chunk_mask;
616         int        rc;
617         ENTRY;
618
619         cur = osc_extent_alloc(obj);
620         if (cur == NULL)
621                 RETURN(ERR_PTR(-ENOMEM));
622
623         olck = osc_env_io(env)->oi_write_osclock;
624         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
625         LASSERT(olck->ols_state == OLS_GRANTED);
626
627         descr = &olck->ols_cl.cls_lock->cll_descr;
628         LASSERT(descr->cld_mode >= CLM_WRITE);
629
630         LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
631         ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
632         chunk_mask = ~((1 << ppc_bits) - 1);
633         chunksize  = 1 << cli->cl_chunkbits;
634         chunk      = index >> ppc_bits;
635
636         /* align end to rpc edge, rpc size may not be a power 2 integer. */
637         max_pages = cli->cl_max_pages_per_rpc;
638         LASSERT((max_pages & ~chunk_mask) == 0);
639         max_end = index - (index % max_pages) + max_pages - 1;
640         max_end = min_t(pgoff_t, max_end, descr->cld_end);
641
642         /* initialize new extent by parameters so far */
643         cur->oe_max_end = max_end;
644         cur->oe_start   = index & chunk_mask;
645         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
646         if (cur->oe_start < descr->cld_start)
647                 cur->oe_start = descr->cld_start;
648         if (cur->oe_end > max_end)
649                 cur->oe_end = max_end;
650         cur->oe_grants  = 0;
651         cur->oe_mppr    = max_pages;
652         if (olck->ols_dlmlock != NULL) {
653                 LASSERT(olck->ols_hold);
654                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
655                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
656         }
657
658         /* grants has been allocated by caller */
659         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
660                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
661         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
662                  EXTPARA(cur));
663
664 restart:
665         osc_object_lock(obj);
666         ext = osc_extent_search(obj, cur->oe_start);
667         if (ext == NULL)
668                 ext = first_extent(obj);
669         while (ext != NULL) {
670                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
671                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
672
673                 LASSERT(sanity_check_nolock(ext) == 0);
674                 if (chunk > ext_chk_end + 1)
675                         break;
676
677                 /* if covering by different locks, no chance to match */
678                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
679                         EASSERTF(!overlapped(ext, cur), ext,
680                                  EXTSTR"\n", EXTPARA(cur));
681
682                         ext = next_extent(ext);
683                         continue;
684                 }
685
686                 /* discontiguous chunks? */
687                 if (chunk + 1 < ext_chk_start) {
688                         ext = next_extent(ext);
689                         continue;
690                 }
691
692                 /* ok, from now on, ext and cur have these attrs:
693                  * 1. covered by the same lock
694                  * 2. contiguous at chunk level or overlapping. */
695
696                 if (overlapped(ext, cur)) {
697                         /* cur is the minimum unit, so overlapping means
698                          * full contain. */
699                         EASSERTF((ext->oe_start <= cur->oe_start &&
700                                   ext->oe_end >= cur->oe_end),
701                                  ext, EXTSTR"\n", EXTPARA(cur));
702
703                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
704                                 /* for simplicity, we wait for this extent to
705                                  * finish before going forward. */
706                                 conflict = osc_extent_get(ext);
707                                 break;
708                         }
709
710                         found = osc_extent_hold(ext);
711                         break;
712                 }
713
714                 /* non-overlapped extent */
715                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
716                         /* we can't do anything for a non OES_CACHE extent, or
717                          * if there is someone waiting for this extent to be
718                          * flushed, try next one. */
719                         ext = next_extent(ext);
720                         continue;
721                 }
722
723                 /* check if they belong to the same rpc slot before trying to
724                  * merge. the extents are not overlapped and contiguous at
725                  * chunk level to get here. */
726                 if (ext->oe_max_end != max_end) {
727                         /* if they don't belong to the same RPC slot or
728                          * max_pages_per_rpc has ever changed, do not merge. */
729                         ext = next_extent(ext);
730                         continue;
731                 }
732
733                 /* it's required that an extent must be contiguous at chunk
734                  * level so that we know the whole extent is covered by grant
735                  * (the pages in the extent are NOT required to be contiguous).
736                  * Otherwise, it will be too much difficult to know which
737                  * chunks have grants allocated. */
738
739                 /* try to do front merge - extend ext's start */
740                 if (chunk + 1 == ext_chk_start) {
741                         /* ext must be chunk size aligned */
742                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
743
744                         /* pull ext's start back to cover cur */
745                         ext->oe_start   = cur->oe_start;
746                         ext->oe_grants += chunksize;
747                         *grants -= chunksize;
748
749                         found = osc_extent_hold(ext);
750                 } else if (chunk == ext_chk_end + 1) {
751                         /* rear merge */
752                         ext->oe_end     = cur->oe_end;
753                         ext->oe_grants += chunksize;
754                         *grants -= chunksize;
755
756                         /* try to merge with the next one because we just fill
757                          * in a gap */
758                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
759                                 /* we can save extent tax from next extent */
760                                 *grants += cli->cl_extent_tax;
761
762                         found = osc_extent_hold(ext);
763                 }
764                 if (found != NULL)
765                         break;
766
767                 ext = next_extent(ext);
768         }
769
770         osc_extent_tree_dump(D_CACHE, obj);
771         if (found != NULL) {
772                 LASSERT(conflict == NULL);
773                 if (!IS_ERR(found)) {
774                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
775                         OSC_EXTENT_DUMP(D_CACHE, found,
776                                         "found caching ext for %lu.\n", index);
777                 }
778         } else if (conflict == NULL) {
779                 /* create a new extent */
780                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
781                 cur->oe_grants = chunksize + cli->cl_extent_tax;
782                 *grants -= cur->oe_grants;
783                 LASSERT(*grants >= 0);
784
785                 cur->oe_state = OES_CACHE;
786                 found = osc_extent_hold(cur);
787                 osc_extent_insert(obj, cur);
788                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
789                                 index, descr->cld_end);
790         }
791         osc_object_unlock(obj);
792
793         if (conflict != NULL) {
794                 LASSERT(found == NULL);
795
796                 /* waiting for IO to finish. Please notice that it's impossible
797                  * to be an OES_TRUNC extent. */
798                 rc = osc_extent_wait(env, conflict, OES_INV);
799                 osc_extent_put(env, conflict);
800                 conflict = NULL;
801                 if (rc < 0)
802                         GOTO(out, found = ERR_PTR(rc));
803
804                 goto restart;
805         }
806         EXIT;
807
808 out:
809         osc_extent_put(env, cur);
810         LASSERT(*grants >= 0);
811         return found;
812 }
813
814 /**
815  * Called when IO is finished to an extent.
816  */
817 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
818                       int sent, int rc)
819 {
820         struct client_obd *cli = osc_cli(ext->oe_obj);
821         struct osc_async_page *oap;
822         struct osc_async_page *tmp;
823         int nr_pages = ext->oe_nr_pages;
824         int lost_grant = 0;
825         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
826         __u64 last_off = 0;
827         int last_count = -1;
828         ENTRY;
829
830         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
831
832         ext->oe_rc = rc ?: ext->oe_nr_pages;
833         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
834
835         osc_lru_add_batch(cli, &ext->oe_pages);
836         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
837                                      oap_pending_item) {
838                 list_del_init(&oap->oap_rpc_item);
839                 list_del_init(&oap->oap_pending_item);
840                 if (last_off <= oap->oap_obj_off) {
841                         last_off = oap->oap_obj_off;
842                         last_count = oap->oap_count;
843                 }
844
845                 --ext->oe_nr_pages;
846                 osc_ap_completion(env, cli, oap, sent, rc);
847         }
848         EASSERT(ext->oe_nr_pages == 0, ext);
849
850         if (!sent) {
851                 lost_grant = ext->oe_grants;
852         } else if (blocksize < PAGE_CACHE_SIZE &&
853                    last_count != PAGE_CACHE_SIZE) {
854                 /* For short writes we shouldn't count parts of pages that
855                  * span a whole chunk on the OST side, or our accounting goes
856                  * wrong.  Should match the code in filter_grant_check. */
857                 int offset = last_off & ~CFS_PAGE_MASK;
858                 int count = last_count + (offset & (blocksize - 1));
859                 int end = (offset + last_count) & (blocksize - 1);
860                 if (end)
861                         count += blocksize - end;
862
863                 lost_grant = PAGE_CACHE_SIZE - count;
864         }
865         if (ext->oe_grants > 0)
866                 osc_free_grant(cli, nr_pages, lost_grant);
867
868         osc_extent_remove(ext);
869         /* put the refcount for RPC */
870         osc_extent_put(env, ext);
871         RETURN(0);
872 }
873
874 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
875 {
876         int ret;
877
878         osc_object_lock(ext->oe_obj);
879         ret = ext->oe_state == state;
880         osc_object_unlock(ext->oe_obj);
881
882         return ret;
883 }
884
885 /**
886  * Wait for the extent's state to become @state.
887  */
888 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
889                            enum osc_extent_state state)
890 {
891         struct osc_object *obj = ext->oe_obj;
892         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
893                                                   LWI_ON_SIGNAL_NOOP, NULL);
894         int rc = 0;
895         ENTRY;
896
897         osc_object_lock(obj);
898         LASSERT(sanity_check_nolock(ext) == 0);
899         /* `Kick' this extent only if the caller is waiting for it to be
900          * written out. */
901         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
902                 if (ext->oe_state == OES_ACTIVE) {
903                         ext->oe_urgent = 1;
904                 } else if (ext->oe_state == OES_CACHE) {
905                         ext->oe_urgent = 1;
906                         osc_extent_hold(ext);
907                         rc = 1;
908                 }
909         }
910         osc_object_unlock(obj);
911         if (rc == 1)
912                 osc_extent_release(env, ext);
913
914         /* wait for the extent until its state becomes @state */
915         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
916         if (rc == -ETIMEDOUT) {
917                 OSC_EXTENT_DUMP(D_ERROR, ext,
918                         "%s: wait ext to %u timedout, recovery in progress?\n",
919                         osc_export(obj)->exp_obd->obd_name, state);
920
921                 lwi = LWI_INTR(NULL, NULL);
922                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
923                                   &lwi);
924         }
925         if (rc == 0 && ext->oe_rc < 0)
926                 rc = ext->oe_rc;
927         RETURN(rc);
928 }
929
930 /**
931  * Discard pages with index greater than @size. If @ext is overlapped with
932  * @size, then partial truncate happens.
933  */
934 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
935                                 bool partial)
936 {
937         struct cl_env_nest     nest;
938         struct lu_env         *env;
939         struct cl_io          *io;
940         struct osc_object     *obj = ext->oe_obj;
941         struct client_obd     *cli = osc_cli(obj);
942         struct osc_async_page *oap;
943         struct osc_async_page *tmp;
944         int                    pages_in_chunk = 0;
945         int                    ppc_bits    = cli->cl_chunkbits -
946                                              PAGE_CACHE_SHIFT;
947         __u64                  trunc_chunk = trunc_index >> ppc_bits;
948         int                    grants   = 0;
949         int                    nr_pages = 0;
950         int                    rc       = 0;
951         ENTRY;
952
953         LASSERT(sanity_check(ext) == 0);
954         LASSERT(ext->oe_state == OES_TRUNC);
955         LASSERT(!ext->oe_urgent);
956
957         /* Request new lu_env.
958          * We can't use that env from osc_cache_truncate_start() because
959          * it's from lov_io_sub and not fully initialized. */
960         env = cl_env_nested_get(&nest);
961         io  = &osc_env_info(env)->oti_io;
962         io->ci_obj = cl_object_top(osc2cl(obj));
963         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
964         if (rc < 0)
965                 GOTO(out, rc);
966
967         /* discard all pages with index greater then trunc_index */
968         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
969                                      oap_pending_item) {
970                 pgoff_t index = osc_index(oap2osc(oap));
971                 struct cl_page  *page = oap2cl_page(oap);
972
973                 LASSERT(list_empty(&oap->oap_rpc_item));
974
975                 /* only discard the pages with their index greater than
976                  * trunc_index, and ... */
977                 if (index < trunc_index ||
978                     (index == trunc_index && partial)) {
979                         /* accounting how many pages remaining in the chunk
980                          * so that we can calculate grants correctly. */
981                         if (index >> ppc_bits == trunc_chunk)
982                                 ++pages_in_chunk;
983                         continue;
984                 }
985
986                 list_del_init(&oap->oap_pending_item);
987
988                 cl_page_get(page);
989                 lu_ref_add(&page->cp_reference, "truncate", current);
990
991                 if (cl_page_own(env, io, page) == 0) {
992                         cl_page_discard(env, io, page);
993                         cl_page_disown(env, io, page);
994                 } else {
995                         LASSERT(page->cp_state == CPS_FREEING);
996                         LASSERT(0);
997                 }
998
999                 lu_ref_del(&page->cp_reference, "truncate", current);
1000                 cl_page_put(env, page);
1001
1002                 --ext->oe_nr_pages;
1003                 ++nr_pages;
1004         }
1005         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1006                       ext->oe_nr_pages == 0),
1007                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1008
1009         osc_object_lock(obj);
1010         if (ext->oe_nr_pages == 0) {
1011                 LASSERT(pages_in_chunk == 0);
1012                 grants = ext->oe_grants;
1013                 ext->oe_grants = 0;
1014         } else { /* calculate how many grants we can free */
1015                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1016                 pgoff_t last_index;
1017
1018
1019                 /* if there is no pages in this chunk, we can also free grants
1020                  * for the last chunk */
1021                 if (pages_in_chunk == 0) {
1022                         /* if this is the 1st chunk and no pages in this chunk,
1023                          * ext->oe_nr_pages must be zero, so we should be in
1024                          * the other if-clause. */
1025                         LASSERT(trunc_chunk > 0);
1026                         --trunc_chunk;
1027                         ++chunks;
1028                 }
1029
1030                 /* this is what we can free from this extent */
1031                 grants          = chunks << cli->cl_chunkbits;
1032                 ext->oe_grants -= grants;
1033                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1034                 ext->oe_end     = min(last_index, ext->oe_max_end);
1035                 LASSERT(ext->oe_end >= ext->oe_start);
1036                 LASSERT(ext->oe_grants > 0);
1037         }
1038         osc_object_unlock(obj);
1039
1040         if (grants > 0 || nr_pages > 0)
1041                 osc_free_grant(cli, nr_pages, grants);
1042
1043 out:
1044         cl_io_fini(env, io);
1045         cl_env_nested_put(&nest, env);
1046         RETURN(rc);
1047 }
1048
1049 /**
1050  * This function is used to make the extent prepared for transfer.
1051  * A race with flusing page - ll_writepage() has to be handled cautiously.
1052  */
1053 static int osc_extent_make_ready(const struct lu_env *env,
1054                                  struct osc_extent *ext)
1055 {
1056         struct osc_async_page *oap;
1057         struct osc_async_page *last = NULL;
1058         struct osc_object *obj = ext->oe_obj;
1059         unsigned int page_count = 0;
1060         int rc;
1061         ENTRY;
1062
1063         /* we're going to grab page lock, so object lock must not be taken. */
1064         LASSERT(sanity_check(ext) == 0);
1065         /* in locking state, any process should not touch this extent. */
1066         EASSERT(ext->oe_state == OES_LOCKING, ext);
1067         EASSERT(ext->oe_owner != NULL, ext);
1068
1069         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1070
1071         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1072                 ++page_count;
1073                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1074                         last = oap;
1075
1076                 /* checking ASYNC_READY is race safe */
1077                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1078                         continue;
1079
1080                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1081                 switch (rc) {
1082                 case 0:
1083                         spin_lock(&oap->oap_lock);
1084                         oap->oap_async_flags |= ASYNC_READY;
1085                         spin_unlock(&oap->oap_lock);
1086                         break;
1087                 case -EALREADY:
1088                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1089                         break;
1090                 default:
1091                         LASSERTF(0, "unknown return code: %d\n", rc);
1092                 }
1093         }
1094
1095         LASSERT(page_count == ext->oe_nr_pages);
1096         LASSERT(last != NULL);
1097         /* the last page is the only one we need to refresh its count by
1098          * the size of file. */
1099         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1100                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1101                 LASSERT(last_oap_count > 0);
1102                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_CACHE_SIZE);
1103                 last->oap_count = last_oap_count;
1104                 spin_lock(&last->oap_lock);
1105                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1106                 spin_unlock(&last->oap_lock);
1107         }
1108
1109         /* for the rest of pages, we don't need to call osf_refresh_count()
1110          * because it's known they are not the last page */
1111         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1112                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1113                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1114                         spin_lock(&oap->oap_lock);
1115                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1116                         spin_unlock(&oap->oap_lock);
1117                 }
1118         }
1119
1120         osc_object_lock(obj);
1121         osc_extent_state_set(ext, OES_RPC);
1122         osc_object_unlock(obj);
1123         /* get a refcount for RPC. */
1124         osc_extent_get(ext);
1125
1126         RETURN(0);
1127 }
1128
1129 /**
1130  * Quick and simple version of osc_extent_find(). This function is frequently
1131  * called to expand the extent for the same IO. To expand the extent, the
1132  * page index must be in the same or next chunk of ext->oe_end.
1133  */
1134 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1135                              unsigned int *grants)
1136 {
1137         struct osc_object *obj = ext->oe_obj;
1138         struct client_obd *cli = osc_cli(obj);
1139         struct osc_extent *next;
1140         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1141         pgoff_t chunk = index >> ppc_bits;
1142         pgoff_t end_chunk;
1143         pgoff_t end_index;
1144         unsigned int chunksize = 1 << cli->cl_chunkbits;
1145         int rc = 0;
1146         ENTRY;
1147
1148         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1149         osc_object_lock(obj);
1150         LASSERT(sanity_check_nolock(ext) == 0);
1151         end_chunk = ext->oe_end >> ppc_bits;
1152         if (chunk > end_chunk + 1)
1153                 GOTO(out, rc = -ERANGE);
1154
1155         if (end_chunk >= chunk)
1156                 GOTO(out, rc = 0);
1157
1158         LASSERT(end_chunk + 1 == chunk);
1159         /* try to expand this extent to cover @index */
1160         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1161
1162         next = next_extent(ext);
1163         if (next != NULL && next->oe_start <= end_index)
1164                 /* complex mode - overlapped with the next extent,
1165                  * this case will be handled by osc_extent_find() */
1166                 GOTO(out, rc = -EAGAIN);
1167
1168         ext->oe_end = end_index;
1169         ext->oe_grants += chunksize;
1170         *grants -= chunksize;
1171         LASSERT(*grants >= 0);
1172         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1173                  "overlapped after expanding for %lu.\n", index);
1174         EXIT;
1175
1176 out:
1177         osc_object_unlock(obj);
1178         RETURN(rc);
1179 }
1180
1181 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1182                                   const char *func, int line)
1183 {
1184         struct osc_extent *ext;
1185         int cnt;
1186
1187         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1188                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1189
1190         /* osc_object_lock(obj); */
1191         cnt = 1;
1192         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1193                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1194
1195         cnt = 1;
1196         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1197                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1198
1199         cnt = 1;
1200         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1201                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1202
1203         cnt = 1;
1204         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1205                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1206         /* osc_object_unlock(obj); */
1207 }
1208
1209 /* ------------------ osc extent end ------------------ */
1210
1211 static inline int osc_is_ready(struct osc_object *osc)
1212 {
1213         return !list_empty(&osc->oo_ready_item) ||
1214                !list_empty(&osc->oo_hp_ready_item);
1215 }
1216
1217 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1218         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1219                (OSC), osc_is_ready(OSC),                                       \
1220                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1221                list_empty_marker(&(OSC)->oo_ready_item),                       \
1222                atomic_read(&(OSC)->oo_nr_writes),                              \
1223                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1224                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1225                atomic_read(&(OSC)->oo_nr_reads),                               \
1226                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1227                ##args)
1228
1229 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1230                           int cmd)
1231 {
1232         struct osc_page *opg  = oap2osc_page(oap);
1233         struct cl_page  *page = oap2cl_page(oap);
1234         int result;
1235
1236         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1237
1238         ENTRY;
1239         result = cl_page_make_ready(env, page, CRT_WRITE);
1240         if (result == 0)
1241                 opg->ops_submit_time = cfs_time_current();
1242         RETURN(result);
1243 }
1244
1245 static int osc_refresh_count(const struct lu_env *env,
1246                              struct osc_async_page *oap, int cmd)
1247 {
1248         struct osc_page  *opg = oap2osc_page(oap);
1249         pgoff_t index = osc_index(oap2osc(oap));
1250         struct cl_object *obj;
1251         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1252
1253         int result;
1254         loff_t kms;
1255
1256         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1257         LASSERT(!(cmd & OBD_BRW_READ));
1258         LASSERT(opg != NULL);
1259         obj = opg->ops_cl.cpl_obj;
1260
1261         cl_object_attr_lock(obj);
1262         result = cl_object_attr_get(env, obj, attr);
1263         cl_object_attr_unlock(obj);
1264         if (result < 0)
1265                 return result;
1266         kms = attr->cat_kms;
1267         if (cl_offset(obj, index) >= kms)
1268                 /* catch race with truncate */
1269                 return 0;
1270         else if (cl_offset(obj, index + 1) > kms)
1271                 /* catch sub-page write at end of file */
1272                 return kms % PAGE_CACHE_SIZE;
1273         else
1274                 return PAGE_CACHE_SIZE;
1275 }
1276
1277 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1278                           int cmd, int rc)
1279 {
1280         struct osc_page   *opg  = oap2osc_page(oap);
1281         struct cl_page    *page = oap2cl_page(oap);
1282         struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
1283         enum cl_req_type   crt;
1284         int srvlock;
1285
1286         ENTRY;
1287
1288         cmd &= ~OBD_BRW_NOQUOTA;
1289         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1290                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1291         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1292                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1293         LASSERT(opg->ops_transfer_pinned);
1294
1295         /*
1296          * page->cp_req can be NULL if io submission failed before
1297          * cl_req was allocated.
1298          */
1299         if (page->cp_req != NULL)
1300                 cl_req_page_done(env, page);
1301         LASSERT(page->cp_req == NULL);
1302
1303         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1304         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1305         opg->ops_transfer_pinned = 0;
1306
1307         spin_lock(&obj->oo_seatbelt);
1308         LASSERT(opg->ops_submitter != NULL);
1309         LASSERT(!list_empty(&opg->ops_inflight));
1310         list_del_init(&opg->ops_inflight);
1311         opg->ops_submitter = NULL;
1312         spin_unlock(&obj->oo_seatbelt);
1313
1314         opg->ops_submit_time = 0;
1315         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1316
1317         /* statistic */
1318         if (rc == 0 && srvlock) {
1319                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1320                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1321                 size_t bytes = oap->oap_count;
1322
1323                 if (crt == CRT_READ)
1324                         stats->os_lockless_reads += bytes;
1325                 else
1326                         stats->os_lockless_writes += bytes;
1327         }
1328
1329         /*
1330          * This has to be the last operation with the page, as locks are
1331          * released in cl_page_completion() and nothing except for the
1332          * reference counter protects page from concurrent reclaim.
1333          */
1334         lu_ref_del(&page->cp_reference, "transfer", page);
1335
1336         cl_page_completion(env, page, crt, rc);
1337
1338         RETURN(0);
1339 }
1340
1341 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1342         struct client_obd *__tmp = (cli);                               \
1343         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1344                "dropped: %ld avail: %ld, reserved: %ld, flight: %d }"   \
1345                "lru {in list: %ld, left: %ld, waiters: %d }"fmt"\n",    \
1346                __tmp->cl_import->imp_obd->obd_name,                     \
1347                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1348                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1349                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1350                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1351                atomic_long_read(&__tmp->cl_lru_in_list),                \
1352                atomic_long_read(&__tmp->cl_lru_busy),                   \
1353                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1354 } while (0)
1355
1356 /* caller must hold loi_list_lock */
1357 static void osc_consume_write_grant(struct client_obd *cli,
1358                                     struct brw_page *pga)
1359 {
1360         assert_spin_locked(&cli->cl_loi_list_lock);
1361         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1362         atomic_long_inc(&obd_dirty_pages);
1363         cli->cl_dirty_pages++;
1364         pga->flag |= OBD_BRW_FROM_GRANT;
1365         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1366                PAGE_CACHE_SIZE, pga, pga->pg);
1367         osc_update_next_shrink(cli);
1368 }
1369
1370 /* the companion to osc_consume_write_grant, called when a brw has completed.
1371  * must be called with the loi lock held. */
1372 static void osc_release_write_grant(struct client_obd *cli,
1373                                     struct brw_page *pga)
1374 {
1375         ENTRY;
1376
1377         assert_spin_locked(&cli->cl_loi_list_lock);
1378         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1379                 EXIT;
1380                 return;
1381         }
1382
1383         pga->flag &= ~OBD_BRW_FROM_GRANT;
1384         atomic_long_dec(&obd_dirty_pages);
1385         cli->cl_dirty_pages--;
1386         if (pga->flag & OBD_BRW_NOCACHE) {
1387                 pga->flag &= ~OBD_BRW_NOCACHE;
1388                 atomic_long_dec(&obd_dirty_transit_pages);
1389                 cli->cl_dirty_transit--;
1390         }
1391         EXIT;
1392 }
1393
1394 /**
1395  * To avoid sleeping with object lock held, it's good for us allocate enough
1396  * grants before entering into critical section.
1397  *
1398  * client_obd_list_lock held by caller
1399  */
1400 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1401 {
1402         int rc = -EDQUOT;
1403
1404         if (cli->cl_avail_grant >= bytes) {
1405                 cli->cl_avail_grant    -= bytes;
1406                 cli->cl_reserved_grant += bytes;
1407                 rc = 0;
1408         }
1409         return rc;
1410 }
1411
1412 static void __osc_unreserve_grant(struct client_obd *cli,
1413                                   unsigned int reserved, unsigned int unused)
1414 {
1415         /* it's quite normal for us to get more grant than reserved.
1416          * Thinking about a case that two extents merged by adding a new
1417          * chunk, we can save one extent tax. If extent tax is greater than
1418          * one chunk, we can save more grant by adding a new chunk */
1419         cli->cl_reserved_grant -= reserved;
1420         if (unused > reserved) {
1421                 cli->cl_avail_grant += reserved;
1422                 cli->cl_lost_grant  += unused - reserved;
1423         } else {
1424                 cli->cl_avail_grant += unused;
1425         }
1426 }
1427
1428 static void osc_unreserve_grant(struct client_obd *cli,
1429                                 unsigned int reserved, unsigned int unused)
1430 {
1431         spin_lock(&cli->cl_loi_list_lock);
1432         __osc_unreserve_grant(cli, reserved, unused);
1433         if (unused > 0)
1434                 osc_wake_cache_waiters(cli);
1435         spin_unlock(&cli->cl_loi_list_lock);
1436 }
1437
1438 /**
1439  * Free grant after IO is finished or canceled.
1440  *
1441  * @lost_grant is used to remember how many grants we have allocated but not
1442  * used, we should return these grants to OST. There're two cases where grants
1443  * can be lost:
1444  * 1. truncate;
1445  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1446  *    written. In this case OST may use less chunks to serve this partial
1447  *    write. OSTs don't actually know the page size on the client side. so
1448  *    clients have to calculate lost grant by the blocksize on the OST.
1449  *    See filter_grant_check() for details.
1450  */
1451 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1452                            unsigned int lost_grant)
1453 {
1454         unsigned long grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1455
1456         spin_lock(&cli->cl_loi_list_lock);
1457         atomic_long_sub(nr_pages, &obd_dirty_pages);
1458         cli->cl_dirty_pages -= nr_pages;
1459         cli->cl_lost_grant += lost_grant;
1460         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1461                 /* borrow some grant from truncate to avoid the case that
1462                  * truncate uses up all avail grant */
1463                 cli->cl_lost_grant -= grant;
1464                 cli->cl_avail_grant += grant;
1465         }
1466         osc_wake_cache_waiters(cli);
1467         spin_unlock(&cli->cl_loi_list_lock);
1468         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1469                lost_grant, cli->cl_lost_grant,
1470                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
1471 }
1472
1473 /**
1474  * The companion to osc_enter_cache(), called when @oap is no longer part of
1475  * the dirty accounting due to error.
1476  */
1477 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1478 {
1479         spin_lock(&cli->cl_loi_list_lock);
1480         osc_release_write_grant(cli, &oap->oap_brw_page);
1481         spin_unlock(&cli->cl_loi_list_lock);
1482 }
1483
1484 /**
1485  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1486  * is available.
1487  */
1488 static int osc_enter_cache_try(struct client_obd *cli,
1489                                struct osc_async_page *oap,
1490                                int bytes, int transient)
1491 {
1492         int rc;
1493
1494         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1495
1496         rc = osc_reserve_grant(cli, bytes);
1497         if (rc < 0)
1498                 return 0;
1499
1500         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1501             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1502                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1503                 if (transient) {
1504                         cli->cl_dirty_transit++;
1505                         atomic_long_inc(&obd_dirty_transit_pages);
1506                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1507                 }
1508                 rc = 1;
1509         } else {
1510                 __osc_unreserve_grant(cli, bytes, bytes);
1511                 rc = 0;
1512         }
1513         return rc;
1514 }
1515
1516 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1517 {
1518         int rc;
1519         spin_lock(&cli->cl_loi_list_lock);
1520         rc = list_empty(&ocw->ocw_entry);
1521         spin_unlock(&cli->cl_loi_list_lock);
1522         return rc;
1523 }
1524
1525 /**
1526  * The main entry to reserve dirty page accounting. Usually the grant reserved
1527  * in this function will be freed in bulk in osc_free_grant() unless it fails
1528  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1529  *
1530  * The process will be put into sleep if it's already run out of grant.
1531  */
1532 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1533                            struct osc_async_page *oap, int bytes)
1534 {
1535         struct osc_object       *osc = oap->oap_obj;
1536         struct lov_oinfo        *loi = osc->oo_oinfo;
1537         struct osc_cache_waiter  ocw;
1538         struct l_wait_info       lwi;
1539         int                      rc = -EDQUOT;
1540         ENTRY;
1541
1542         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1543                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1544
1545         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1546
1547         spin_lock(&cli->cl_loi_list_lock);
1548
1549         /* force the caller to try sync io.  this can jump the list
1550          * of queued writes and create a discontiguous rpc stream */
1551         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1552             cli->cl_dirty_max_pages == 0 ||
1553             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1554                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1555                 GOTO(out, rc = -EDQUOT);
1556         }
1557
1558         /* Hopefully normal case - cache space and write credits available */
1559         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1560                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1561                 GOTO(out, rc = 0);
1562         }
1563
1564         /* We can get here for two reasons: too many dirty pages in cache, or
1565          * run out of grants. In both cases we should write dirty pages out.
1566          * Adding a cache waiter will trigger urgent write-out no matter what
1567          * RPC size will be.
1568          * The exiting condition is no avail grants and no dirty pages caching,
1569          * that really means there is no space on the OST. */
1570         init_waitqueue_head(&ocw.ocw_waitq);
1571         ocw.ocw_oap   = oap;
1572         ocw.ocw_grant = bytes;
1573         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1574                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1575                 ocw.ocw_rc = 0;
1576                 spin_unlock(&cli->cl_loi_list_lock);
1577
1578                 osc_io_unplug_async(env, cli, NULL);
1579
1580                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1581                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1582
1583                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1584
1585                 spin_lock(&cli->cl_loi_list_lock);
1586
1587                 if (rc < 0) {
1588                         /* l_wait_event is interrupted by signal or timed out */
1589                         list_del_init(&ocw.ocw_entry);
1590                         break;
1591                 }
1592                 LASSERT(list_empty(&ocw.ocw_entry));
1593                 rc = ocw.ocw_rc;
1594
1595                 if (rc != -EDQUOT)
1596                         break;
1597                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1598                         rc = 0;
1599                         break;
1600                 }
1601         }
1602
1603         switch (rc) {
1604         case 0:
1605                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1606                 break;
1607         case -ETIMEDOUT:
1608                 OSC_DUMP_GRANT(D_CACHE, cli,
1609                                "timeout, fall back to sync i/o\n");
1610                 osc_extent_tree_dump(D_CACHE, osc);
1611                 /* fall back to synchronous I/O */
1612                 rc = -EDQUOT;
1613                 break;
1614         case -EINTR:
1615                 /* Ensures restartability - LU-3581 */
1616                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1617                 rc = -ERESTARTSYS;
1618                 break;
1619         case -EDQUOT:
1620                 OSC_DUMP_GRANT(D_CACHE, cli,
1621                                "no grant space, fall back to sync i/o\n");
1622                 break;
1623         default:
1624                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1625                        "due to %d, fall back to sync i/o\n",
1626                        cli->cl_import->imp_obd->obd_name, &ocw, rc);
1627                 break;
1628         }
1629         EXIT;
1630 out:
1631         spin_unlock(&cli->cl_loi_list_lock);
1632         RETURN(rc);
1633 }
1634
1635 /* caller must hold loi_list_lock */
1636 void osc_wake_cache_waiters(struct client_obd *cli)
1637 {
1638         struct list_head *l, *tmp;
1639         struct osc_cache_waiter *ocw;
1640
1641         ENTRY;
1642         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1643                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1644                 list_del_init(&ocw->ocw_entry);
1645
1646                 ocw->ocw_rc = -EDQUOT;
1647                 /* we can't dirty more */
1648                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1649                     (1 + atomic_long_read(&obd_dirty_pages) >
1650                      obd_max_dirty_pages)) {
1651                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1652                                "osc max %ld, sys max %ld\n",
1653                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1654                                obd_max_dirty_pages);
1655                         goto wakeup;
1656                 }
1657
1658                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1659                         ocw->ocw_rc = 0;
1660 wakeup:
1661                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1662                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1663
1664                 wake_up(&ocw->ocw_waitq);
1665         }
1666
1667         EXIT;
1668 }
1669
1670 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1671 {
1672         int hprpc = !!list_empty(&osc->oo_hp_exts);
1673         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1674 }
1675
1676 /* This maintains the lists of pending pages to read/write for a given object
1677  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1678  * to quickly find objects that are ready to send an RPC. */
1679 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1680                          int cmd)
1681 {
1682         int invalid_import = 0;
1683         ENTRY;
1684
1685         /* if we have an invalid import we want to drain the queued pages
1686          * by forcing them through rpcs that immediately fail and complete
1687          * the pages.  recovery relies on this to empty the queued pages
1688          * before canceling the locks and evicting down the llite pages */
1689         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1690                 invalid_import = 1;
1691
1692         if (cmd & OBD_BRW_WRITE) {
1693                 if (atomic_read(&osc->oo_nr_writes) == 0)
1694                         RETURN(0);
1695                 if (invalid_import) {
1696                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1697                         RETURN(1);
1698                 }
1699                 if (!list_empty(&osc->oo_hp_exts)) {
1700                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1701                         RETURN(1);
1702                 }
1703                 if (!list_empty(&osc->oo_urgent_exts)) {
1704                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1705                         RETURN(1);
1706                 }
1707                 /* trigger a write rpc stream as long as there are dirtiers
1708                  * waiting for space.  as they're waiting, they're not going to
1709                  * create more pages to coalesce with what's waiting.. */
1710                 if (!list_empty(&cli->cl_cache_waiters)) {
1711                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1712                         RETURN(1);
1713                 }
1714                 if (atomic_read(&osc->oo_nr_writes) >=
1715                     cli->cl_max_pages_per_rpc)
1716                         RETURN(1);
1717         } else {
1718                 if (atomic_read(&osc->oo_nr_reads) == 0)
1719                         RETURN(0);
1720                 if (invalid_import) {
1721                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1722                         RETURN(1);
1723                 }
1724                 /* all read are urgent. */
1725                 if (!list_empty(&osc->oo_reading_exts))
1726                         RETURN(1);
1727         }
1728
1729         RETURN(0);
1730 }
1731
1732 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1733 {
1734         struct client_obd *cli = osc_cli(obj);
1735         if (cmd & OBD_BRW_WRITE) {
1736                 atomic_add(delta, &obj->oo_nr_writes);
1737                 atomic_add(delta, &cli->cl_pending_w_pages);
1738                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1739         } else {
1740                 atomic_add(delta, &obj->oo_nr_reads);
1741                 atomic_add(delta, &cli->cl_pending_r_pages);
1742                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1743         }
1744         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1745 }
1746
1747 static int osc_makes_hprpc(struct osc_object *obj)
1748 {
1749         return !list_empty(&obj->oo_hp_exts);
1750 }
1751
1752 static void on_list(struct list_head *item, struct list_head *list,
1753                     int should_be_on)
1754 {
1755         if (list_empty(item) && should_be_on)
1756                 list_add_tail(item, list);
1757         else if (!list_empty(item) && !should_be_on)
1758                 list_del_init(item);
1759 }
1760
1761 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1762  * can find pages to build into rpcs quickly */
1763 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1764 {
1765         if (osc_makes_hprpc(osc)) {
1766                 /* HP rpc */
1767                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1768                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1769         } else {
1770                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1771                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1772                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1773                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1774         }
1775
1776         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1777                 atomic_read(&osc->oo_nr_writes) > 0);
1778
1779         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1780                 atomic_read(&osc->oo_nr_reads) > 0);
1781
1782         return osc_is_ready(osc);
1783 }
1784
1785 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1786 {
1787         int is_ready;
1788
1789         spin_lock(&cli->cl_loi_list_lock);
1790         is_ready = __osc_list_maint(cli, osc);
1791         spin_unlock(&cli->cl_loi_list_lock);
1792
1793         return is_ready;
1794 }
1795
1796 /* this is trying to propogate async writeback errors back up to the
1797  * application.  As an async write fails we record the error code for later if
1798  * the app does an fsync.  As long as errors persist we force future rpcs to be
1799  * sync so that the app can get a sync error and break the cycle of queueing
1800  * pages for which writeback will fail. */
1801 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1802                            int rc)
1803 {
1804         if (rc) {
1805                 if (!ar->ar_rc)
1806                         ar->ar_rc = rc;
1807
1808                 ar->ar_force_sync = 1;
1809                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1810                 return;
1811
1812         }
1813
1814         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1815                 ar->ar_force_sync = 0;
1816 }
1817
1818 /* this must be called holding the loi list lock to give coverage to exit_cache,
1819  * async_flag maintenance, and oap_request */
1820 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1821                               struct osc_async_page *oap, int sent, int rc)
1822 {
1823         struct osc_object *osc = oap->oap_obj;
1824         struct lov_oinfo  *loi = osc->oo_oinfo;
1825         __u64 xid = 0;
1826
1827         ENTRY;
1828         if (oap->oap_request != NULL) {
1829                 xid = ptlrpc_req_xid(oap->oap_request);
1830                 ptlrpc_req_finished(oap->oap_request);
1831                 oap->oap_request = NULL;
1832         }
1833
1834         /* As the transfer for this page is being done, clear the flags */
1835         spin_lock(&oap->oap_lock);
1836         oap->oap_async_flags = 0;
1837         spin_unlock(&oap->oap_lock);
1838         oap->oap_interrupted = 0;
1839
1840         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1841                 spin_lock(&cli->cl_loi_list_lock);
1842                 osc_process_ar(&cli->cl_ar, xid, rc);
1843                 osc_process_ar(&loi->loi_ar, xid, rc);
1844                 spin_unlock(&cli->cl_loi_list_lock);
1845         }
1846
1847         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1848         if (rc)
1849                 CERROR("completion on oap %p obj %p returns %d.\n",
1850                        oap, osc, rc);
1851
1852         EXIT;
1853 }
1854
1855 /**
1856  * Try to add extent to one RPC. We need to think about the following things:
1857  * - # of pages must not be over max_pages_per_rpc
1858  * - extent must be compatible with previous ones
1859  */
1860 static int try_to_add_extent_for_io(struct client_obd *cli,
1861                                     struct osc_extent *ext,
1862                                     struct list_head *rpclist,
1863                                     unsigned int *pc, unsigned int *max_pages)
1864 {
1865         struct osc_extent *tmp;
1866         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1867                                                       struct osc_async_page,
1868                                                       oap_pending_item);
1869         ENTRY;
1870
1871         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1872                 ext);
1873
1874         *max_pages = max(ext->oe_mppr, *max_pages);
1875         if (*pc + ext->oe_nr_pages > *max_pages)
1876                 RETURN(0);
1877
1878         list_for_each_entry(tmp, rpclist, oe_link) {
1879                 struct osc_async_page *oap2;
1880                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1881                                         oap_pending_item);
1882                 EASSERT(tmp->oe_owner == current, tmp);
1883 #if 0
1884                 if (overlapped(tmp, ext)) {
1885                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1886                         EASSERT(0, ext);
1887                 }
1888 #endif
1889                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1890                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1891                                         " for a same RPC\n");
1892                         RETURN(0);
1893                 }
1894
1895                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1896                     !tmp->oe_grants != !ext->oe_grants)
1897                         RETURN(0);
1898
1899                 /* remove break for strict check */
1900                 break;
1901         }
1902
1903         *pc += ext->oe_nr_pages;
1904         list_move_tail(&ext->oe_link, rpclist);
1905         ext->oe_owner = current;
1906         RETURN(1);
1907 }
1908
1909 /**
1910  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1911  * get_write_extent() takes all appropriate extents in atomic.
1912  *
1913  * The following policy is used to collect extents for IO:
1914  * 1. Add as many HP extents as possible;
1915  * 2. Add the first urgent extent in urgent extent list and take it out of
1916  *    urgent list;
1917  * 3. Add subsequent extents of this urgent extent;
1918  * 4. If urgent list is not empty, goto 2;
1919  * 5. Traverse the extent tree from the 1st extent;
1920  * 6. Above steps exit if there is no space in this RPC.
1921  */
1922 static unsigned int get_write_extents(struct osc_object *obj,
1923                                       struct list_head *rpclist)
1924 {
1925         struct client_obd *cli = osc_cli(obj);
1926         struct osc_extent *ext;
1927         unsigned int page_count = 0;
1928         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1929
1930         LASSERT(osc_object_is_locked(obj));
1931         while (!list_empty(&obj->oo_hp_exts)) {
1932                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1933                                  oe_link);
1934                 LASSERT(ext->oe_state == OES_CACHE);
1935                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1936                                               &max_pages))
1937                         return page_count;
1938                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1939         }
1940         if (page_count == max_pages)
1941                 return page_count;
1942
1943         while (!list_empty(&obj->oo_urgent_exts)) {
1944                 ext = list_entry(obj->oo_urgent_exts.next,
1945                                  struct osc_extent, oe_link);
1946                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1947                                               &max_pages))
1948                         return page_count;
1949
1950                 if (!ext->oe_intree)
1951                         continue;
1952
1953                 while ((ext = next_extent(ext)) != NULL) {
1954                         if ((ext->oe_state != OES_CACHE) ||
1955                             (!list_empty(&ext->oe_link) &&
1956                              ext->oe_owner != NULL))
1957                                 continue;
1958
1959                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1960                                                       &page_count, &max_pages))
1961                                 return page_count;
1962                 }
1963         }
1964         if (page_count == max_pages)
1965                 return page_count;
1966
1967         ext = first_extent(obj);
1968         while (ext != NULL) {
1969                 if ((ext->oe_state != OES_CACHE) ||
1970                     /* this extent may be already in current rpclist */
1971                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1972                         ext = next_extent(ext);
1973                         continue;
1974                 }
1975
1976                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1977                                               &max_pages))
1978                         return page_count;
1979
1980                 ext = next_extent(ext);
1981         }
1982         return page_count;
1983 }
1984
1985 static int
1986 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1987                    struct osc_object *osc, pdl_policy_t pol)
1988 __must_hold(osc)
1989 {
1990         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
1991         struct osc_extent *ext;
1992         struct osc_extent *tmp;
1993         struct osc_extent *first = NULL;
1994         unsigned int page_count = 0;
1995         int srvlock = 0;
1996         int rc = 0;
1997         ENTRY;
1998
1999         LASSERT(osc_object_is_locked(osc));
2000
2001         page_count = get_write_extents(osc, &rpclist);
2002         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2003
2004         if (list_empty(&rpclist))
2005                 RETURN(0);
2006
2007         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2008
2009         list_for_each_entry(ext, &rpclist, oe_link) {
2010                 LASSERT(ext->oe_state == OES_CACHE ||
2011                         ext->oe_state == OES_LOCK_DONE);
2012                 if (ext->oe_state == OES_CACHE)
2013                         osc_extent_state_set(ext, OES_LOCKING);
2014                 else
2015                         osc_extent_state_set(ext, OES_RPC);
2016         }
2017
2018         /* we're going to grab page lock, so release object lock because
2019          * lock order is page lock -> object lock. */
2020         osc_object_unlock(osc);
2021
2022         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2023                 if (ext->oe_state == OES_LOCKING) {
2024                         rc = osc_extent_make_ready(env, ext);
2025                         if (unlikely(rc < 0)) {
2026                                 list_del_init(&ext->oe_link);
2027                                 osc_extent_finish(env, ext, 0, rc);
2028                                 continue;
2029                         }
2030                 }
2031                 if (first == NULL) {
2032                         first = ext;
2033                         srvlock = ext->oe_srvlock;
2034                 } else {
2035                         LASSERT(srvlock == ext->oe_srvlock);
2036                 }
2037         }
2038
2039         if (!list_empty(&rpclist)) {
2040                 LASSERT(page_count > 0);
2041                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2042                 LASSERT(list_empty(&rpclist));
2043         }
2044
2045         osc_object_lock(osc);
2046         RETURN(rc);
2047 }
2048
2049 /**
2050  * prepare pages for ASYNC io and put pages in send queue.
2051  *
2052  * \param cmd OBD_BRW_* macroses
2053  * \param lop pending pages
2054  *
2055  * \return zero if no page added to send queue.
2056  * \return 1 if pages successfully added to send queue.
2057  * \return negative on errors.
2058  */
2059 static int
2060 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2061                   struct osc_object *osc, pdl_policy_t pol)
2062 __must_hold(osc)
2063 {
2064         struct osc_extent *ext;
2065         struct osc_extent *next;
2066         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2067         unsigned int page_count = 0;
2068         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2069         int rc = 0;
2070         ENTRY;
2071
2072         LASSERT(osc_object_is_locked(osc));
2073         list_for_each_entry_safe(ext, next,
2074                                      &osc->oo_reading_exts, oe_link) {
2075                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2076                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2077                                               &max_pages))
2078                         break;
2079                 osc_extent_state_set(ext, OES_RPC);
2080                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2081         }
2082         LASSERT(page_count <= max_pages);
2083
2084         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2085
2086         if (!list_empty(&rpclist)) {
2087                 osc_object_unlock(osc);
2088
2089                 LASSERT(page_count > 0);
2090                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2091                 LASSERT(list_empty(&rpclist));
2092
2093                 osc_object_lock(osc);
2094         }
2095         RETURN(rc);
2096 }
2097
2098 #define list_to_obj(list, item) ({                                            \
2099         struct list_head *__tmp = (list)->next;                               \
2100         list_del_init(__tmp);                                         \
2101         list_entry(__tmp, struct osc_object, oo_##item);                      \
2102 })
2103
2104 /* This is called by osc_check_rpcs() to find which objects have pages that
2105  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2106 static struct osc_object *osc_next_obj(struct client_obd *cli)
2107 {
2108         ENTRY;
2109
2110         /* First return objects that have blocked locks so that they
2111          * will be flushed quickly and other clients can get the lock,
2112          * then objects which have pages ready to be stuffed into RPCs */
2113         if (!list_empty(&cli->cl_loi_hp_ready_list))
2114                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2115         if (!list_empty(&cli->cl_loi_ready_list))
2116                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2117
2118         /* then if we have cache waiters, return all objects with queued
2119          * writes.  This is especially important when many small files
2120          * have filled up the cache and not been fired into rpcs because
2121          * they don't pass the nr_pending/object threshhold */
2122         if (!list_empty(&cli->cl_cache_waiters) &&
2123             !list_empty(&cli->cl_loi_write_list))
2124                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2125
2126         /* then return all queued objects when we have an invalid import
2127          * so that they get flushed */
2128         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2129                 if (!list_empty(&cli->cl_loi_write_list))
2130                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2131                                            write_item));
2132                 if (!list_empty(&cli->cl_loi_read_list))
2133                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2134                                            read_item));
2135         }
2136         RETURN(NULL);
2137 }
2138
2139 /* called with the loi list lock held */
2140 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2141                            pdl_policy_t pol)
2142 __must_hold(&cli->cl_loi_list_lock)
2143 {
2144         struct osc_object *osc;
2145         int rc = 0;
2146         ENTRY;
2147
2148         while ((osc = osc_next_obj(cli)) != NULL) {
2149                 struct cl_object *obj = osc2cl(osc);
2150                 struct lu_ref_link link;
2151
2152                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2153
2154                 if (osc_max_rpc_in_flight(cli, osc)) {
2155                         __osc_list_maint(cli, osc);
2156                         break;
2157                 }
2158
2159                 cl_object_get(obj);
2160                 spin_unlock(&cli->cl_loi_list_lock);
2161                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2162
2163                 /* attempt some read/write balancing by alternating between
2164                  * reads and writes in an object.  The makes_rpc checks here
2165                  * would be redundant if we were getting read/write work items
2166                  * instead of objects.  we don't want send_oap_rpc to drain a
2167                  * partial read pending queue when we're given this object to
2168                  * do io on writes while there are cache waiters */
2169                 osc_object_lock(osc);
2170                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2171                         rc = osc_send_write_rpc(env, cli, osc, pol);
2172                         if (rc < 0) {
2173                                 CERROR("Write request failed with %d\n", rc);
2174
2175                                 /* osc_send_write_rpc failed, mostly because of
2176                                  * memory pressure.
2177                                  *
2178                                  * It can't break here, because if:
2179                                  *  - a page was submitted by osc_io_submit, so
2180                                  *    page locked;
2181                                  *  - no request in flight
2182                                  *  - no subsequent request
2183                                  * The system will be in live-lock state,
2184                                  * because there is no chance to call
2185                                  * osc_io_unplug() and osc_check_rpcs() any
2186                                  * more. pdflush can't help in this case,
2187                                  * because it might be blocked at grabbing
2188                                  * the page lock as we mentioned.
2189                                  *
2190                                  * Anyway, continue to drain pages. */
2191                                 /* break; */
2192                         }
2193                 }
2194                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2195                         rc = osc_send_read_rpc(env, cli, osc, pol);
2196                         if (rc < 0)
2197                                 CERROR("Read request failed with %d\n", rc);
2198                 }
2199                 osc_object_unlock(osc);
2200
2201                 osc_list_maint(cli, osc);
2202                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2203                 cl_object_put(env, obj);
2204
2205                 spin_lock(&cli->cl_loi_list_lock);
2206         }
2207 }
2208
2209 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2210                           struct osc_object *osc, pdl_policy_t pol, int async)
2211 {
2212         int rc = 0;
2213
2214         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2215                 return 0;
2216
2217         if (!async) {
2218                 /* disable osc_lru_shrink() temporarily to avoid
2219                  * potential stack overrun problem. LU-2859 */
2220                 atomic_inc(&cli->cl_lru_shrinkers);
2221                 spin_lock(&cli->cl_loi_list_lock);
2222                 osc_check_rpcs(env, cli, pol);
2223                 spin_unlock(&cli->cl_loi_list_lock);
2224                 atomic_dec(&cli->cl_lru_shrinkers);
2225         } else {
2226                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2227                 LASSERT(cli->cl_writeback_work != NULL);
2228                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2229         }
2230         return rc;
2231 }
2232
2233 static int osc_io_unplug_async(const struct lu_env *env,
2234                                 struct client_obd *cli, struct osc_object *osc)
2235 {
2236         /* XXX: policy is no use actually. */
2237         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2238 }
2239
2240 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2241                    struct osc_object *osc, pdl_policy_t pol)
2242 {
2243         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2244 }
2245
2246 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2247                         struct page *page, loff_t offset)
2248 {
2249         struct obd_export     *exp = osc_export(osc);
2250         struct osc_async_page *oap = &ops->ops_oap;
2251         ENTRY;
2252
2253         if (!page)
2254                 return cfs_size_round(sizeof(*oap));
2255
2256         oap->oap_magic = OAP_MAGIC;
2257         oap->oap_cli = &exp->exp_obd->u.cli;
2258         oap->oap_obj = osc;
2259
2260         oap->oap_page = page;
2261         oap->oap_obj_off = offset;
2262         LASSERT(!(offset & ~CFS_PAGE_MASK));
2263
2264         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2265                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2266
2267         INIT_LIST_HEAD(&oap->oap_pending_item);
2268         INIT_LIST_HEAD(&oap->oap_rpc_item);
2269
2270         spin_lock_init(&oap->oap_lock);
2271         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2272                oap, page, oap->oap_obj_off);
2273         RETURN(0);
2274 }
2275
2276 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2277                        struct osc_page *ops)
2278 {
2279         struct osc_io *oio = osc_env_io(env);
2280         struct osc_extent     *ext = NULL;
2281         struct osc_async_page *oap = &ops->ops_oap;
2282         struct client_obd     *cli = oap->oap_cli;
2283         struct osc_object     *osc = oap->oap_obj;
2284         pgoff_t index;
2285         unsigned int tmp;
2286         unsigned int grants = 0;
2287         int    brw_flags = OBD_BRW_ASYNC;
2288         int    cmd = OBD_BRW_WRITE;
2289         int    need_release = 0;
2290         int    rc = 0;
2291         ENTRY;
2292
2293         if (oap->oap_magic != OAP_MAGIC)
2294                 RETURN(-EINVAL);
2295
2296         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2297                 RETURN(-EIO);
2298
2299         if (!list_empty(&oap->oap_pending_item) ||
2300             !list_empty(&oap->oap_rpc_item))
2301                 RETURN(-EBUSY);
2302
2303         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2304         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2305         if (!client_is_remote(osc_export(osc)) &&
2306             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2307                 brw_flags |= OBD_BRW_NOQUOTA;
2308                 cmd |= OBD_BRW_NOQUOTA;
2309         }
2310
2311         /* check if the file's owner/group is over quota */
2312         if (!(cmd & OBD_BRW_NOQUOTA)) {
2313                 struct cl_object *obj;
2314                 struct cl_attr   *attr;
2315                 unsigned int qid[MAXQUOTAS];
2316
2317                 obj = cl_object_top(&osc->oo_cl);
2318                 attr = &osc_env_info(env)->oti_attr;
2319
2320                 cl_object_attr_lock(obj);
2321                 rc = cl_object_attr_get(env, obj, attr);
2322                 cl_object_attr_unlock(obj);
2323
2324                 qid[USRQUOTA] = attr->cat_uid;
2325                 qid[GRPQUOTA] = attr->cat_gid;
2326                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2327                         rc = -EDQUOT;
2328                 if (rc)
2329                         RETURN(rc);
2330         }
2331
2332         oap->oap_cmd = cmd;
2333         oap->oap_page_off = ops->ops_from;
2334         oap->oap_count = ops->ops_to - ops->ops_from;
2335         /* No need to hold a lock here,
2336          * since this page is not in any list yet. */
2337         oap->oap_async_flags = 0;
2338         oap->oap_brw_flags = brw_flags;
2339
2340         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2341                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2342
2343         index = osc_index(oap2osc(oap));
2344
2345         /* Add this page into extent by the following steps:
2346          * 1. if there exists an active extent for this IO, mostly this page
2347          *    can be added to the active extent and sometimes we need to
2348          *    expand extent to accomodate this page;
2349          * 2. otherwise, a new extent will be allocated. */
2350
2351         ext = oio->oi_active;
2352         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2353                 /* one chunk plus extent overhead must be enough to write this
2354                  * page */
2355                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2356                 if (ext->oe_end >= index)
2357                         grants = 0;
2358
2359                 /* it doesn't need any grant to dirty this page */
2360                 spin_lock(&cli->cl_loi_list_lock);
2361                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2362                 spin_unlock(&cli->cl_loi_list_lock);
2363                 if (rc == 0) { /* try failed */
2364                         grants = 0;
2365                         need_release = 1;
2366                 } else if (ext->oe_end < index) {
2367                         tmp = grants;
2368                         /* try to expand this extent */
2369                         rc = osc_extent_expand(ext, index, &tmp);
2370                         if (rc < 0) {
2371                                 need_release = 1;
2372                                 /* don't free reserved grant */
2373                         } else {
2374                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2375                                                 "expanded for %lu.\n", index);
2376                                 osc_unreserve_grant(cli, grants, tmp);
2377                                 grants = 0;
2378                         }
2379                 }
2380                 rc = 0;
2381         } else if (ext != NULL) {
2382                 /* index is located outside of active extent */
2383                 need_release = 1;
2384         }
2385         if (need_release) {
2386                 osc_extent_release(env, ext);
2387                 oio->oi_active = NULL;
2388                 ext = NULL;
2389         }
2390
2391         if (ext == NULL) {
2392                 tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2393
2394                 /* try to find new extent to cover this page */
2395                 LASSERT(oio->oi_active == NULL);
2396                 /* we may have allocated grant for this page if we failed
2397                  * to expand the previous active extent. */
2398                 LASSERT(ergo(grants > 0, grants >= tmp));
2399
2400                 rc = 0;
2401                 if (grants == 0) {
2402                         /* we haven't allocated grant for this page. */
2403                         rc = osc_enter_cache(env, cli, oap, tmp);
2404                         if (rc == 0)
2405                                 grants = tmp;
2406                 }
2407
2408                 tmp = grants;
2409                 if (rc == 0) {
2410                         ext = osc_extent_find(env, osc, index, &tmp);
2411                         if (IS_ERR(ext)) {
2412                                 LASSERT(tmp == grants);
2413                                 osc_exit_cache(cli, oap);
2414                                 rc = PTR_ERR(ext);
2415                                 ext = NULL;
2416                         } else {
2417                                 oio->oi_active = ext;
2418                         }
2419                 }
2420                 if (grants > 0)
2421                         osc_unreserve_grant(cli, grants, tmp);
2422         }
2423
2424         LASSERT(ergo(rc == 0, ext != NULL));
2425         if (ext != NULL) {
2426                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2427                          ext, "index = %lu.\n", index);
2428                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2429
2430                 osc_object_lock(osc);
2431                 if (ext->oe_nr_pages == 0)
2432                         ext->oe_srvlock = ops->ops_srvlock;
2433                 else
2434                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2435                 ++ext->oe_nr_pages;
2436                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2437                 osc_object_unlock(osc);
2438         }
2439         RETURN(rc);
2440 }
2441
2442 int osc_teardown_async_page(const struct lu_env *env,
2443                             struct osc_object *obj, struct osc_page *ops)
2444 {
2445         struct osc_async_page *oap = &ops->ops_oap;
2446         struct osc_extent     *ext = NULL;
2447         int rc = 0;
2448         ENTRY;
2449
2450         LASSERT(oap->oap_magic == OAP_MAGIC);
2451
2452         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2453                oap, ops, osc_index(oap2osc(oap)));
2454
2455         osc_object_lock(obj);
2456         if (!list_empty(&oap->oap_rpc_item)) {
2457                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2458                 rc = -EBUSY;
2459         } else if (!list_empty(&oap->oap_pending_item)) {
2460                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2461                 /* only truncated pages are allowed to be taken out.
2462                  * See osc_extent_truncate() and osc_cache_truncate_start()
2463                  * for details. */
2464                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2465                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2466                                         osc_index(oap2osc(oap)));
2467                         rc = -EBUSY;
2468                 }
2469         }
2470         osc_object_unlock(obj);
2471         if (ext != NULL)
2472                 osc_extent_put(env, ext);
2473         RETURN(rc);
2474 }
2475
2476 /**
2477  * This is called when a page is picked up by kernel to write out.
2478  *
2479  * We should find out the corresponding extent and add the whole extent
2480  * into urgent list. The extent may be being truncated or used, handle it
2481  * carefully.
2482  */
2483 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2484                          struct osc_page *ops)
2485 {
2486         struct osc_extent *ext   = NULL;
2487         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2488         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2489         pgoff_t            index = osc_index(ops);
2490         struct osc_async_page *oap = &ops->ops_oap;
2491         bool unplug = false;
2492         int rc = 0;
2493         ENTRY;
2494
2495         osc_object_lock(obj);
2496         ext = osc_extent_lookup(obj, index);
2497         if (ext == NULL) {
2498                 osc_extent_tree_dump(D_ERROR, obj);
2499                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2500         }
2501
2502         switch (ext->oe_state) {
2503         case OES_RPC:
2504         case OES_LOCK_DONE:
2505                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2506                 LASSERT(0);
2507                 break;
2508         case OES_LOCKING:
2509                 /* If we know this extent is being written out, we should abort
2510                  * so that the writer can make this page ready. Otherwise, there
2511                  * exists a deadlock problem because other process can wait for
2512                  * page writeback bit holding page lock; and meanwhile in
2513                  * vvp_page_make_ready(), we need to grab page lock before
2514                  * really sending the RPC. */
2515         case OES_TRUNC:
2516                 /* race with truncate, page will be redirtied */
2517         case OES_ACTIVE:
2518                 /* The extent is active so we need to abort and let the caller
2519                  * re-dirty the page. If we continued on here, and we were the
2520                  * one making the extent active, we could deadlock waiting for
2521                  * the page writeback to clear but it won't because the extent
2522                  * is active and won't be written out. */
2523                 GOTO(out, rc = -EAGAIN);
2524         default:
2525                 break;
2526         }
2527
2528         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2529         if (rc)
2530                 GOTO(out, rc);
2531
2532         spin_lock(&oap->oap_lock);
2533         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2534         spin_unlock(&oap->oap_lock);
2535
2536         if (memory_pressure_get())
2537                 ext->oe_memalloc = 1;
2538
2539         ext->oe_urgent = 1;
2540         if (ext->oe_state == OES_CACHE) {
2541                 OSC_EXTENT_DUMP(D_CACHE, ext,
2542                                 "flush page %p make it urgent.\n", oap);
2543                 if (list_empty(&ext->oe_link))
2544                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2545                 unplug = true;
2546         }
2547         rc = 0;
2548         EXIT;
2549
2550 out:
2551         osc_object_unlock(obj);
2552         osc_extent_put(env, ext);
2553         if (unplug)
2554                 osc_io_unplug_async(env, osc_cli(obj), obj);
2555         return rc;
2556 }
2557
2558 /**
2559  * this is called when a sync waiter receives an interruption.  Its job is to
2560  * get the caller woken as soon as possible.  If its page hasn't been put in an
2561  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2562  * desiring interruption which will forcefully complete the rpc once the rpc
2563  * has timed out.
2564  */
2565 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2566 {
2567         struct osc_async_page *oap = &ops->ops_oap;
2568         struct osc_object     *obj = oap->oap_obj;
2569         struct client_obd     *cli = osc_cli(obj);
2570         struct osc_extent     *ext;
2571         struct osc_extent     *found = NULL;
2572         struct list_head            *plist;
2573         pgoff_t index = osc_index(ops);
2574         int     rc = -EBUSY;
2575         int     cmd;
2576         ENTRY;
2577
2578         LASSERT(!oap->oap_interrupted);
2579         oap->oap_interrupted = 1;
2580
2581         /* Find out the caching extent */
2582         osc_object_lock(obj);
2583         if (oap->oap_cmd & OBD_BRW_WRITE) {
2584                 plist = &obj->oo_urgent_exts;
2585                 cmd   = OBD_BRW_WRITE;
2586         } else {
2587                 plist = &obj->oo_reading_exts;
2588                 cmd   = OBD_BRW_READ;
2589         }
2590         list_for_each_entry(ext, plist, oe_link) {
2591                 if (ext->oe_start <= index && ext->oe_end >= index) {
2592                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2593                         /* For OES_LOCK_DONE state extent, it has already held
2594                          * a refcount for RPC. */
2595                         found = osc_extent_get(ext);
2596                         break;
2597                 }
2598         }
2599         if (found != NULL) {
2600                 list_del_init(&found->oe_link);
2601                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2602                 osc_object_unlock(obj);
2603
2604                 osc_extent_finish(env, found, 0, -EINTR);
2605                 osc_extent_put(env, found);
2606                 rc = 0;
2607         } else {
2608                 osc_object_unlock(obj);
2609                 /* ok, it's been put in an rpc. only one oap gets a request
2610                  * reference */
2611                 if (oap->oap_request != NULL) {
2612                         ptlrpc_mark_interrupted(oap->oap_request);
2613                         ptlrpcd_wake(oap->oap_request);
2614                         ptlrpc_req_finished(oap->oap_request);
2615                         oap->oap_request = NULL;
2616                 }
2617         }
2618
2619         osc_list_maint(cli, obj);
2620         RETURN(rc);
2621 }
2622
2623 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2624                          struct list_head *list, int cmd, int brw_flags)
2625 {
2626         struct client_obd     *cli = osc_cli(obj);
2627         struct osc_extent     *ext;
2628         struct osc_async_page *oap;
2629         int     page_count = 0;
2630         int     mppr       = cli->cl_max_pages_per_rpc;
2631         pgoff_t start      = CL_PAGE_EOF;
2632         pgoff_t end        = 0;
2633         ENTRY;
2634
2635         list_for_each_entry(oap, list, oap_pending_item) {
2636                 pgoff_t index = osc_index(oap2osc(oap));
2637                 if (index > end)
2638                         end = index;
2639                 if (index < start)
2640                         start = index;
2641                 ++page_count;
2642                 mppr <<= (page_count > mppr);
2643         }
2644
2645         ext = osc_extent_alloc(obj);
2646         if (ext == NULL) {
2647                 list_for_each_entry(oap, list, oap_pending_item) {
2648                         list_del_init(&oap->oap_pending_item);
2649                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2650                 }
2651                 RETURN(-ENOMEM);
2652         }
2653
2654         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2655         ext->oe_sync = 1;
2656         ext->oe_urgent = 1;
2657         ext->oe_start = start;
2658         ext->oe_end = ext->oe_max_end = end;
2659         ext->oe_obj = obj;
2660         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2661         ext->oe_nr_pages = page_count;
2662         ext->oe_mppr = mppr;
2663         list_splice_init(list, &ext->oe_pages);
2664
2665         osc_object_lock(obj);
2666         /* Reuse the initial refcount for RPC, don't drop it */
2667         osc_extent_state_set(ext, OES_LOCK_DONE);
2668         if (cmd & OBD_BRW_WRITE) {
2669                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2670                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2671         } else {
2672                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2673                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2674         }
2675         osc_object_unlock(obj);
2676
2677         osc_io_unplug_async(env, cli, obj);
2678         RETURN(0);
2679 }
2680
2681 /**
2682  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2683  */
2684 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2685                              struct osc_object *obj, __u64 size)
2686 {
2687         struct client_obd *cli = osc_cli(obj);
2688         struct osc_extent *ext;
2689         struct osc_extent *waiting = NULL;
2690         pgoff_t index;
2691         struct list_head list = LIST_HEAD_INIT(list);
2692         int result = 0;
2693         bool partial;
2694         ENTRY;
2695
2696         /* pages with index greater or equal to index will be truncated. */
2697         index = cl_index(osc2cl(obj), size);
2698         partial = size > cl_offset(osc2cl(obj), index);
2699
2700 again:
2701         osc_object_lock(obj);
2702         ext = osc_extent_search(obj, index);
2703         if (ext == NULL)
2704                 ext = first_extent(obj);
2705         else if (ext->oe_end < index)
2706                 ext = next_extent(ext);
2707         while (ext != NULL) {
2708                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2709
2710                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2711                         /* if ext is in urgent state, it means there must exist
2712                          * a page already having been flushed by write_page().
2713                          * We have to wait for this extent because we can't
2714                          * truncate that page. */
2715                         LASSERT(!ext->oe_hp);
2716                         OSC_EXTENT_DUMP(D_CACHE, ext,
2717                                         "waiting for busy extent\n");
2718                         waiting = osc_extent_get(ext);
2719                         break;
2720                 }
2721
2722                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2723
2724                 osc_extent_get(ext);
2725                 if (ext->oe_state == OES_ACTIVE) {
2726                         /* though we grab inode mutex for write path, but we
2727                          * release it before releasing extent(in osc_io_end()),
2728                          * so there is a race window that an extent is still
2729                          * in OES_ACTIVE when truncate starts. */
2730                         LASSERT(!ext->oe_trunc_pending);
2731                         ext->oe_trunc_pending = 1;
2732                 } else {
2733                         EASSERT(ext->oe_state == OES_CACHE, ext);
2734                         osc_extent_state_set(ext, OES_TRUNC);
2735                         osc_update_pending(obj, OBD_BRW_WRITE,
2736                                            -ext->oe_nr_pages);
2737                 }
2738                 EASSERT(list_empty(&ext->oe_link), ext);
2739                 list_add_tail(&ext->oe_link, &list);
2740
2741                 ext = next_extent(ext);
2742         }
2743         osc_object_unlock(obj);
2744
2745         osc_list_maint(cli, obj);
2746
2747         while (!list_empty(&list)) {
2748                 int rc;
2749
2750                 ext = list_entry(list.next, struct osc_extent, oe_link);
2751                 list_del_init(&ext->oe_link);
2752
2753                 /* extent may be in OES_ACTIVE state because inode mutex
2754                  * is released before osc_io_end() in file write case */
2755                 if (ext->oe_state != OES_TRUNC)
2756                         osc_extent_wait(env, ext, OES_TRUNC);
2757
2758                 rc = osc_extent_truncate(ext, index, partial);
2759                 if (rc < 0) {
2760                         if (result == 0)
2761                                 result = rc;
2762
2763                         OSC_EXTENT_DUMP(D_ERROR, ext,
2764                                         "truncate error %d\n", rc);
2765                 } else if (ext->oe_nr_pages == 0) {
2766                         osc_extent_remove(ext);
2767                 } else {
2768                         /* this must be an overlapped extent which means only
2769                          * part of pages in this extent have been truncated.
2770                          */
2771                         EASSERTF(ext->oe_start <= index, ext,
2772                                  "trunc index = %lu/%d.\n", index, partial);
2773                         /* fix index to skip this partially truncated extent */
2774                         index = ext->oe_end + 1;
2775                         partial = false;
2776
2777                         /* we need to hold this extent in OES_TRUNC state so
2778                          * that no writeback will happen. This is to avoid
2779                          * BUG 17397. */
2780                         LASSERT(oio->oi_trunc == NULL);
2781                         oio->oi_trunc = osc_extent_get(ext);
2782                         OSC_EXTENT_DUMP(D_CACHE, ext,
2783                                         "trunc at "LPU64"\n", size);
2784                 }
2785                 osc_extent_put(env, ext);
2786         }
2787         if (waiting != NULL) {
2788                 int rc;
2789
2790                 /* ignore the result of osc_extent_wait the write initiator
2791                  * should take care of it. */
2792                 rc = osc_extent_wait(env, waiting, OES_INV);
2793                 if (rc < 0)
2794                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2795
2796                 osc_extent_put(env, waiting);
2797                 waiting = NULL;
2798                 goto again;
2799         }
2800         RETURN(result);
2801 }
2802
2803 /**
2804  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2805  */
2806 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2807                             struct osc_object *obj)
2808 {
2809         struct osc_extent *ext = oio->oi_trunc;
2810
2811         oio->oi_trunc = NULL;
2812         if (ext != NULL) {
2813                 bool unplug = false;
2814
2815                 EASSERT(ext->oe_nr_pages > 0, ext);
2816                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2817                 EASSERT(!ext->oe_urgent, ext);
2818
2819                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2820                 osc_object_lock(obj);
2821                 osc_extent_state_set(ext, OES_CACHE);
2822                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2823                         ext->oe_urgent = 1;
2824                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2825                         unplug = true;
2826                 }
2827                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2828                 osc_object_unlock(obj);
2829                 osc_extent_put(env, ext);
2830
2831                 if (unplug)
2832                         osc_io_unplug_async(env, osc_cli(obj), obj);
2833         }
2834 }
2835
2836 /**
2837  * Wait for extents in a specific range to be written out.
2838  * The caller must have called osc_cache_writeback_range() to issue IO
2839  * otherwise it will take a long time for this function to finish.
2840  *
2841  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2842  * nobody else can dirty this range of file while we're waiting for
2843  * extents to be written.
2844  */
2845 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2846                          pgoff_t start, pgoff_t end)
2847 {
2848         struct osc_extent *ext;
2849         pgoff_t index = start;
2850         int     result = 0;
2851         ENTRY;
2852
2853 again:
2854         osc_object_lock(obj);
2855         ext = osc_extent_search(obj, index);
2856         if (ext == NULL)
2857                 ext = first_extent(obj);
2858         else if (ext->oe_end < index)
2859                 ext = next_extent(ext);
2860         while (ext != NULL) {
2861                 int rc;
2862
2863                 if (ext->oe_start > end)
2864                         break;
2865
2866                 if (!ext->oe_fsync_wait) {
2867                         ext = next_extent(ext);
2868                         continue;
2869                 }
2870
2871                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2872                              ext->oe_hp || ext->oe_urgent), ext);
2873                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2874                              !ext->oe_hp && ext->oe_urgent), ext);
2875
2876                 index = ext->oe_end + 1;
2877                 osc_extent_get(ext);
2878                 osc_object_unlock(obj);
2879
2880                 rc = osc_extent_wait(env, ext, OES_INV);
2881                 if (result == 0)
2882                         result = rc;
2883                 osc_extent_put(env, ext);
2884                 goto again;
2885         }
2886         osc_object_unlock(obj);
2887
2888         OSC_IO_DEBUG(obj, "sync file range.\n");
2889         RETURN(result);
2890 }
2891
2892 /**
2893  * Called to write out a range of osc object.
2894  *
2895  * @hp     : should be set this is caused by lock cancel;
2896  * @discard: is set if dirty pages should be dropped - file will be deleted or
2897  *         truncated, this implies there is no partially discarding extents.
2898  *
2899  * Return how many pages will be issued, or error code if error occurred.
2900  */
2901 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2902                               pgoff_t start, pgoff_t end, int hp, int discard)
2903 {
2904         struct osc_extent *ext;
2905         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2906         bool unplug = false;
2907         int result = 0;
2908         ENTRY;
2909
2910         osc_object_lock(obj);
2911         ext = osc_extent_search(obj, start);
2912         if (ext == NULL)
2913                 ext = first_extent(obj);
2914         else if (ext->oe_end < start)
2915                 ext = next_extent(ext);
2916         while (ext != NULL) {
2917                 if (ext->oe_start > end)
2918                         break;
2919
2920                 ext->oe_fsync_wait = 1;
2921                 switch (ext->oe_state) {
2922                 case OES_CACHE:
2923                         result += ext->oe_nr_pages;
2924                         if (!discard) {
2925                                 struct list_head *list = NULL;
2926                                 if (hp) {
2927                                         EASSERT(!ext->oe_hp, ext);
2928                                         ext->oe_hp = 1;
2929                                         list = &obj->oo_hp_exts;
2930                                 } else if (!ext->oe_urgent) {
2931                                         ext->oe_urgent = 1;
2932                                         list = &obj->oo_urgent_exts;
2933                                 }
2934                                 if (list != NULL)
2935                                         list_move_tail(&ext->oe_link, list);
2936                                 unplug = true;
2937                         } else {
2938                                 /* the only discarder is lock cancelling, so
2939                                  * [start, end] must contain this extent */
2940                                 EASSERT(ext->oe_start >= start &&
2941                                         ext->oe_max_end <= end, ext);
2942                                 osc_extent_state_set(ext, OES_LOCKING);
2943                                 ext->oe_owner = current;
2944                                 list_move_tail(&ext->oe_link,
2945                                                    &discard_list);
2946                                 osc_update_pending(obj, OBD_BRW_WRITE,
2947                                                    -ext->oe_nr_pages);
2948                         }
2949                         break;
2950                 case OES_ACTIVE:
2951                         /* It's pretty bad to wait for ACTIVE extents, because
2952                          * we don't know how long we will wait for it to be
2953                          * flushed since it may be blocked at awaiting more
2954                          * grants. We do this for the correctness of fsync. */
2955                         LASSERT(hp == 0 && discard == 0);
2956                         ext->oe_urgent = 1;
2957                         break;
2958                 case OES_TRUNC:
2959                         /* this extent is being truncated, can't do anything
2960                          * for it now. it will be set to urgent after truncate
2961                          * is finished in osc_cache_truncate_end(). */
2962                 default:
2963                         break;
2964                 }
2965                 ext = next_extent(ext);
2966         }
2967         osc_object_unlock(obj);
2968
2969         LASSERT(ergo(!discard, list_empty(&discard_list)));
2970         if (!list_empty(&discard_list)) {
2971                 struct osc_extent *tmp;
2972                 int rc;
2973
2974                 osc_list_maint(osc_cli(obj), obj);
2975                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2976                         list_del_init(&ext->oe_link);
2977                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2978
2979                         /* Discard caching pages. We don't actually write this
2980                          * extent out but we complete it as if we did. */
2981                         rc = osc_extent_make_ready(env, ext);
2982                         if (unlikely(rc < 0)) {
2983                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2984                                                 "make_ready returned %d\n", rc);
2985                                 if (result >= 0)
2986                                         result = rc;
2987                         }
2988
2989                         /* finish the extent as if the pages were sent */
2990                         osc_extent_finish(env, ext, 0, 0);
2991                 }
2992         }
2993
2994         if (unplug)
2995                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2996
2997         if (hp || discard) {
2998                 int rc;
2999                 rc = osc_cache_wait_range(env, obj, start, end);
3000                 if (result >= 0 && rc < 0)
3001                         result = rc;
3002         }
3003
3004         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3005         RETURN(result);
3006 }
3007
3008 /**
3009  * Returns a list of pages by a given [start, end] of \a obj.
3010  *
3011  * \param resched If not NULL, then we give up before hogging CPU for too
3012  * long and set *resched = 1, in that case caller should implement a retry
3013  * logic.
3014  *
3015  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3016  * crucial in the face of [offset, EOF] locks.
3017  *
3018  * Return at least one page in @queue unless there is no covered page.
3019  */
3020 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3021                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3022                         osc_page_gang_cbt cb, void *cbdata)
3023 {
3024         struct osc_page *ops;
3025         void            **pvec;
3026         pgoff_t         idx;
3027         unsigned int    nr;
3028         unsigned int    i;
3029         unsigned int    j;
3030         int             res = CLP_GANG_OKAY;
3031         bool            tree_lock = true;
3032         ENTRY;
3033
3034         idx = start;
3035         pvec = osc_env_info(env)->oti_pvec;
3036         spin_lock(&osc->oo_tree_lock);
3037         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3038                                             idx, OTI_PVEC_SIZE)) > 0) {
3039                 struct cl_page *page;
3040                 bool end_of_region = false;
3041
3042                 for (i = 0, j = 0; i < nr; ++i) {
3043                         ops = pvec[i];
3044                         pvec[i] = NULL;
3045
3046                         idx = osc_index(ops);
3047                         if (idx > end) {
3048                                 end_of_region = true;
3049                                 break;
3050                         }
3051
3052                         page = ops->ops_cl.cpl_page;
3053                         LASSERT(page->cp_type == CPT_CACHEABLE);
3054                         if (page->cp_state == CPS_FREEING)
3055                                 continue;
3056
3057                         cl_page_get(page);
3058                         lu_ref_add_atomic(&page->cp_reference,
3059                                           "gang_lookup", current);
3060                         pvec[j++] = ops;
3061                 }
3062                 ++idx;
3063
3064                 /*
3065                  * Here a delicate locking dance is performed. Current thread
3066                  * holds a reference to a page, but has to own it before it
3067                  * can be placed into queue. Owning implies waiting, so
3068                  * radix-tree lock is to be released. After a wait one has to
3069                  * check that pages weren't truncated (cl_page_own() returns
3070                  * error in the latter case).
3071                  */
3072                 spin_unlock(&osc->oo_tree_lock);
3073                 tree_lock = false;
3074
3075                 for (i = 0; i < j; ++i) {
3076                         ops = pvec[i];
3077                         if (res == CLP_GANG_OKAY)
3078                                 res = (*cb)(env, io, ops, cbdata);
3079
3080                         page = ops->ops_cl.cpl_page;
3081                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3082                         cl_page_put(env, page);
3083                 }
3084                 if (nr < OTI_PVEC_SIZE || end_of_region)
3085                         break;
3086
3087                 if (res == CLP_GANG_OKAY && need_resched())
3088                         res = CLP_GANG_RESCHED;
3089                 if (res != CLP_GANG_OKAY)
3090                         break;
3091
3092                 spin_lock(&osc->oo_tree_lock);
3093                 tree_lock = true;
3094         }
3095         if (tree_lock)
3096                 spin_unlock(&osc->oo_tree_lock);
3097         RETURN(res);
3098 }
3099
3100 /**
3101  * Check if page @page is covered by an extra lock or discard it.
3102  */
3103 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3104                                 struct osc_page *ops, void *cbdata)
3105 {
3106         struct osc_thread_info *info = osc_env_info(env);
3107         struct osc_object *osc = cbdata;
3108         pgoff_t index;
3109
3110         index = osc_index(ops);
3111         if (index >= info->oti_fn_index) {
3112                 struct ldlm_lock *tmp;
3113                 struct cl_page *page = ops->ops_cl.cpl_page;
3114
3115                 /* refresh non-overlapped index */
3116                 tmp = osc_dlmlock_at_pgoff(env, osc, index, 0, 0);
3117                 if (tmp != NULL) {
3118                         __u64 end = tmp->l_policy_data.l_extent.end;
3119                         /* Cache the first-non-overlapped index so as to skip
3120                          * all pages within [index, oti_fn_index). This is safe
3121                          * because if tmp lock is canceled, it will discard
3122                          * these pages. */
3123                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3124                         if (end == OBD_OBJECT_EOF)
3125                                 info->oti_fn_index = CL_PAGE_EOF;
3126                         LDLM_LOCK_PUT(tmp);
3127                 } else if (cl_page_own(env, io, page) == 0) {
3128                         /* discard the page */
3129                         cl_page_discard(env, io, page);
3130                         cl_page_disown(env, io, page);
3131                 } else {
3132                         LASSERT(page->cp_state == CPS_FREEING);
3133                 }
3134         }
3135
3136         info->oti_next_index = index + 1;
3137         return CLP_GANG_OKAY;
3138 }
3139
3140 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3141                       struct osc_page *ops, void *cbdata)
3142 {
3143         struct osc_thread_info *info = osc_env_info(env);
3144         struct cl_page *page = ops->ops_cl.cpl_page;
3145
3146         /* page is top page. */
3147         info->oti_next_index = osc_index(ops) + 1;
3148         if (cl_page_own(env, io, page) == 0) {
3149                 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3150                               !PageDirty(cl_page_vmpage(page))));
3151
3152                 /* discard the page */
3153                 cl_page_discard(env, io, page);
3154                 cl_page_disown(env, io, page);
3155         } else {
3156                 LASSERT(page->cp_state == CPS_FREEING);
3157         }
3158
3159         return CLP_GANG_OKAY;
3160 }
3161
3162 /**
3163  * Discard pages protected by the given lock. This function traverses radix
3164  * tree to find all covering pages and discard them. If a page is being covered
3165  * by other locks, it should remain in cache.
3166  *
3167  * If error happens on any step, the process continues anyway (the reasoning
3168  * behind this being that lock cancellation cannot be delayed indefinitely).
3169  */
3170 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3171                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3172 {
3173         struct osc_thread_info *info = osc_env_info(env);
3174         struct cl_io *io = &info->oti_io;
3175         osc_page_gang_cbt cb;
3176         int res;
3177         int result;