Whamcloud - gitweb
LU-5577 obd: change brw_page->count to unsigned
[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, 2013, Intel Corporation.
31  *
32  */
33 /*
34  * This file is part of Lustre, http://www.lustre.org/
35  * Lustre is a trademark of Sun Microsystems, Inc.
36  *
37  * osc cache management.
38  *
39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_OSC
43
44 #include "osc_cl_internal.h"
45 #include "osc_internal.h"
46
47 static int extent_debug; /* set it to be true for more debug */
48
49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
51                            int state);
52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
53                               struct osc_async_page *oap, int sent, int rc);
54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
55                           int cmd);
56 static int osc_refresh_count(const struct lu_env *env,
57                              struct osc_async_page *oap, int cmd);
58 static int osc_io_unplug_async(const struct lu_env *env,
59                                struct client_obd *cli, struct osc_object *osc);
60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
61                            unsigned int lost_grant);
62
63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
64                                   const char *func, int line);
65 #define osc_extent_tree_dump(lvl, obj) \
66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
67
68 /** \addtogroup osc
69  *  @{
70  */
71
72 /* ------------------ osc extent ------------------ */
73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
74 {
75         char *buf = flags;
76         *buf++ = ext->oe_rw ? 'r' : 'w';
77         if (ext->oe_intree)
78                 *buf++ = 'i';
79         if (ext->oe_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         int 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                                           int *grants)
602
603 {
604         struct client_obd *cli = osc_cli(obj);
605         struct osc_lock   *olck;
606         struct cl_lock_descr *descr;
607         struct osc_extent *cur;
608         struct osc_extent *ext;
609         struct osc_extent *conflict = NULL;
610         struct osc_extent *found = NULL;
611         pgoff_t    chunk;
612         pgoff_t    max_end;
613         int        max_pages; /* max_pages_per_rpc */
614         int        chunksize;
615         int        ppc_bits; /* pages per chunk bits */
616         int        chunk_mask;
617         int        rc;
618         ENTRY;
619
620         cur = osc_extent_alloc(obj);
621         if (cur == NULL)
622                 RETURN(ERR_PTR(-ENOMEM));
623
624         olck = osc_env_io(env)->oi_write_osclock;
625         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
626         LASSERT(olck->ols_state == OLS_GRANTED);
627
628         descr = &olck->ols_cl.cls_lock->cll_descr;
629         LASSERT(descr->cld_mode >= CLM_WRITE);
630
631         LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
632         ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
633         chunk_mask = ~((1 << ppc_bits) - 1);
634         chunksize  = 1 << cli->cl_chunkbits;
635         chunk      = index >> ppc_bits;
636
637         /* align end to rpc edge, rpc size may not be a power 2 integer. */
638         max_pages = cli->cl_max_pages_per_rpc;
639         LASSERT((max_pages & ~chunk_mask) == 0);
640         max_end = index - (index % max_pages) + max_pages - 1;
641         max_end = min_t(pgoff_t, max_end, descr->cld_end);
642
643         /* initialize new extent by parameters so far */
644         cur->oe_max_end = max_end;
645         cur->oe_start   = index & chunk_mask;
646         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
647         if (cur->oe_start < descr->cld_start)
648                 cur->oe_start = descr->cld_start;
649         if (cur->oe_end > max_end)
650                 cur->oe_end = max_end;
651         cur->oe_grants  = 0;
652         cur->oe_mppr    = max_pages;
653         if (olck->ols_dlmlock != NULL) {
654                 LASSERT(olck->ols_hold);
655                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
656                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
657         }
658
659         /* grants has been allocated by caller */
660         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
661                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
662         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
663                  EXTPARA(cur));
664
665 restart:
666         osc_object_lock(obj);
667         ext = osc_extent_search(obj, cur->oe_start);
668         if (ext == NULL)
669                 ext = first_extent(obj);
670         while (ext != NULL) {
671                 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
672                 loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
673
674                 LASSERT(sanity_check_nolock(ext) == 0);
675                 if (chunk > ext_chk_end + 1)
676                         break;
677
678                 /* if covering by different locks, no chance to match */
679                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
680                         EASSERTF(!overlapped(ext, cur), ext,
681                                  EXTSTR"\n", EXTPARA(cur));
682
683                         ext = next_extent(ext);
684                         continue;
685                 }
686
687                 /* discontiguous chunks? */
688                 if (chunk + 1 < ext_chk_start) {
689                         ext = next_extent(ext);
690                         continue;
691                 }
692
693                 /* ok, from now on, ext and cur have these attrs:
694                  * 1. covered by the same lock
695                  * 2. contiguous at chunk level or overlapping. */
696
697                 if (overlapped(ext, cur)) {
698                         /* cur is the minimum unit, so overlapping means
699                          * full contain. */
700                         EASSERTF((ext->oe_start <= cur->oe_start &&
701                                   ext->oe_end >= cur->oe_end),
702                                  ext, EXTSTR"\n", EXTPARA(cur));
703
704                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
705                                 /* for simplicity, we wait for this extent to
706                                  * finish before going forward. */
707                                 conflict = osc_extent_get(ext);
708                                 break;
709                         }
710
711                         found = osc_extent_hold(ext);
712                         break;
713                 }
714
715                 /* non-overlapped extent */
716                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
717                         /* we can't do anything for a non OES_CACHE extent, or
718                          * if there is someone waiting for this extent to be
719                          * flushed, try next one. */
720                         ext = next_extent(ext);
721                         continue;
722                 }
723
724                 /* check if they belong to the same rpc slot before trying to
725                  * merge. the extents are not overlapped and contiguous at
726                  * chunk level to get here. */
727                 if (ext->oe_max_end != max_end) {
728                         /* if they don't belong to the same RPC slot or
729                          * max_pages_per_rpc has ever changed, do not merge. */
730                         ext = next_extent(ext);
731                         continue;
732                 }
733
734                 /* it's required that an extent must be contiguous at chunk
735                  * level so that we know the whole extent is covered by grant
736                  * (the pages in the extent are NOT required to be contiguous).
737                  * Otherwise, it will be too much difficult to know which
738                  * chunks have grants allocated. */
739
740                 /* try to do front merge - extend ext's start */
741                 if (chunk + 1 == ext_chk_start) {
742                         /* ext must be chunk size aligned */
743                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
744
745                         /* pull ext's start back to cover cur */
746                         ext->oe_start   = cur->oe_start;
747                         ext->oe_grants += chunksize;
748                         *grants -= chunksize;
749
750                         found = osc_extent_hold(ext);
751                 } else if (chunk == ext_chk_end + 1) {
752                         /* rear merge */
753                         ext->oe_end     = cur->oe_end;
754                         ext->oe_grants += chunksize;
755                         *grants -= chunksize;
756
757                         /* try to merge with the next one because we just fill
758                          * in a gap */
759                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
760                                 /* we can save extent tax from next extent */
761                                 *grants += cli->cl_extent_tax;
762
763                         found = osc_extent_hold(ext);
764                 }
765                 if (found != NULL)
766                         break;
767
768                 ext = next_extent(ext);
769         }
770
771         osc_extent_tree_dump(D_CACHE, obj);
772         if (found != NULL) {
773                 LASSERT(conflict == NULL);
774                 if (!IS_ERR(found)) {
775                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
776                         OSC_EXTENT_DUMP(D_CACHE, found,
777                                         "found caching ext for %lu.\n", index);
778                 }
779         } else if (conflict == NULL) {
780                 /* create a new extent */
781                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
782                 cur->oe_grants = chunksize + cli->cl_extent_tax;
783                 *grants -= cur->oe_grants;
784                 LASSERT(*grants >= 0);
785
786                 cur->oe_state = OES_CACHE;
787                 found = osc_extent_hold(cur);
788                 osc_extent_insert(obj, cur);
789                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
790                                 index, descr->cld_end);
791         }
792         osc_object_unlock(obj);
793
794         if (conflict != NULL) {
795                 LASSERT(found == NULL);
796
797                 /* waiting for IO to finish. Please notice that it's impossible
798                  * to be an OES_TRUNC extent. */
799                 rc = osc_extent_wait(env, conflict, OES_INV);
800                 osc_extent_put(env, conflict);
801                 conflict = NULL;
802                 if (rc < 0)
803                         GOTO(out, found = ERR_PTR(rc));
804
805                 goto restart;
806         }
807         EXIT;
808
809 out:
810         osc_extent_put(env, cur);
811         LASSERT(*grants >= 0);
812         return found;
813 }
814
815 /**
816  * Called when IO is finished to an extent.
817  */
818 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
819                       int sent, int rc)
820 {
821         struct client_obd *cli = osc_cli(ext->oe_obj);
822         struct osc_async_page *oap;
823         struct osc_async_page *tmp;
824         int nr_pages = ext->oe_nr_pages;
825         int lost_grant = 0;
826         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
827         __u64 last_off = 0;
828         int last_count = -1;
829         ENTRY;
830
831         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
832
833         ext->oe_rc = rc ?: ext->oe_nr_pages;
834         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
835
836         osc_lru_add_batch(cli, &ext->oe_pages);
837         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
838                                      oap_pending_item) {
839                 list_del_init(&oap->oap_rpc_item);
840                 list_del_init(&oap->oap_pending_item);
841                 if (last_off <= oap->oap_obj_off) {
842                         last_off = oap->oap_obj_off;
843                         last_count = oap->oap_count;
844                 }
845
846                 --ext->oe_nr_pages;
847                 osc_ap_completion(env, cli, oap, sent, rc);
848         }
849         EASSERT(ext->oe_nr_pages == 0, ext);
850
851         if (!sent) {
852                 lost_grant = ext->oe_grants;
853         } else if (blocksize < PAGE_CACHE_SIZE &&
854                    last_count != PAGE_CACHE_SIZE) {
855                 /* For short writes we shouldn't count parts of pages that
856                  * span a whole chunk on the OST side, or our accounting goes
857                  * wrong.  Should match the code in filter_grant_check. */
858                 int offset = last_off & ~CFS_PAGE_MASK;
859                 int count = last_count + (offset & (blocksize - 1));
860                 int end = (offset + last_count) & (blocksize - 1);
861                 if (end)
862                         count += blocksize - end;
863
864                 lost_grant = PAGE_CACHE_SIZE - count;
865         }
866         if (ext->oe_grants > 0)
867                 osc_free_grant(cli, nr_pages, lost_grant);
868
869         osc_extent_remove(ext);
870         /* put the refcount for RPC */
871         osc_extent_put(env, ext);
872         RETURN(0);
873 }
874
875 static int extent_wait_cb(struct osc_extent *ext, int state)
876 {
877         int ret;
878
879         osc_object_lock(ext->oe_obj);
880         ret = ext->oe_state == state;
881         osc_object_unlock(ext->oe_obj);
882
883         return ret;
884 }
885
886 /**
887  * Wait for the extent's state to become @state.
888  */
889 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
890                            int state)
891 {
892         struct osc_object *obj = ext->oe_obj;
893         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
894                                                   LWI_ON_SIGNAL_NOOP, NULL);
895         int rc = 0;
896         ENTRY;
897
898         osc_object_lock(obj);
899         LASSERT(sanity_check_nolock(ext) == 0);
900         /* `Kick' this extent only if the caller is waiting for it to be
901          * written out. */
902         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
903                 if (ext->oe_state == OES_ACTIVE) {
904                         ext->oe_urgent = 1;
905                 } else if (ext->oe_state == OES_CACHE) {
906                         ext->oe_urgent = 1;
907                         osc_extent_hold(ext);
908                         rc = 1;
909                 }
910         }
911         osc_object_unlock(obj);
912         if (rc == 1)
913                 osc_extent_release(env, ext);
914
915         /* wait for the extent until its state becomes @state */
916         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
917         if (rc == -ETIMEDOUT) {
918                 OSC_EXTENT_DUMP(D_ERROR, ext,
919                         "%s: wait ext to %d timedout, recovery in progress?\n",
920                         osc_export(obj)->exp_obd->obd_name, state);
921
922                 lwi = LWI_INTR(NULL, NULL);
923                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
924                                   &lwi);
925         }
926         if (rc == 0 && ext->oe_rc < 0)
927                 rc = ext->oe_rc;
928         RETURN(rc);
929 }
930
931 /**
932  * Discard pages with index greater than @size. If @ext is overlapped with
933  * @size, then partial truncate happens.
934  */
935 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
936                                 bool partial)
937 {
938         struct cl_env_nest     nest;
939         struct lu_env         *env;
940         struct cl_io          *io;
941         struct osc_object     *obj = ext->oe_obj;
942         struct client_obd     *cli = osc_cli(obj);
943         struct osc_async_page *oap;
944         struct osc_async_page *tmp;
945         int                    pages_in_chunk = 0;
946         int                    ppc_bits    = cli->cl_chunkbits -
947                                              PAGE_CACHE_SHIFT;
948         __u64                  trunc_chunk = trunc_index >> ppc_bits;
949         int                    grants   = 0;
950         int                    nr_pages = 0;
951         int                    rc       = 0;
952         ENTRY;
953
954         LASSERT(sanity_check(ext) == 0);
955         LASSERT(ext->oe_state == OES_TRUNC);
956         LASSERT(!ext->oe_urgent);
957
958         /* Request new lu_env.
959          * We can't use that env from osc_cache_truncate_start() because
960          * it's from lov_io_sub and not fully initialized. */
961         env = cl_env_nested_get(&nest);
962         io  = &osc_env_info(env)->oti_io;
963         io->ci_obj = cl_object_top(osc2cl(obj));
964         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
965         if (rc < 0)
966                 GOTO(out, rc);
967
968         /* discard all pages with index greater then trunc_index */
969         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
970                                      oap_pending_item) {
971                 pgoff_t index = osc_index(oap2osc(oap));
972                 struct cl_page  *page = oap2cl_page(oap);
973
974                 LASSERT(list_empty(&oap->oap_rpc_item));
975
976                 /* only discard the pages with their index greater than
977                  * trunc_index, and ... */
978                 if (index < trunc_index ||
979                     (index == trunc_index && partial)) {
980                         /* accounting how many pages remaining in the chunk
981                          * so that we can calculate grants correctly. */
982                         if (index >> ppc_bits == trunc_chunk)
983                                 ++pages_in_chunk;
984                         continue;
985                 }
986
987                 list_del_init(&oap->oap_pending_item);
988
989                 cl_page_get(page);
990                 lu_ref_add(&page->cp_reference, "truncate", current);
991
992                 if (cl_page_own(env, io, page) == 0) {
993                         cl_page_discard(env, io, page);
994                         cl_page_disown(env, io, page);
995                 } else {
996                         LASSERT(page->cp_state == CPS_FREEING);
997                         LASSERT(0);
998                 }
999
1000                 lu_ref_del(&page->cp_reference, "truncate", current);
1001                 cl_page_put(env, page);
1002
1003                 --ext->oe_nr_pages;
1004                 ++nr_pages;
1005         }
1006         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1007                       ext->oe_nr_pages == 0),
1008                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1009
1010         osc_object_lock(obj);
1011         if (ext->oe_nr_pages == 0) {
1012                 LASSERT(pages_in_chunk == 0);
1013                 grants = ext->oe_grants;
1014                 ext->oe_grants = 0;
1015         } else { /* calculate how many grants we can free */
1016                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1017                 pgoff_t last_index;
1018
1019
1020                 /* if there is no pages in this chunk, we can also free grants
1021                  * for the last chunk */
1022                 if (pages_in_chunk == 0) {
1023                         /* if this is the 1st chunk and no pages in this chunk,
1024                          * ext->oe_nr_pages must be zero, so we should be in
1025                          * the other if-clause. */
1026                         LASSERT(trunc_chunk > 0);
1027                         --trunc_chunk;
1028                         ++chunks;
1029                 }
1030
1031                 /* this is what we can free from this extent */
1032                 grants          = chunks << cli->cl_chunkbits;
1033                 ext->oe_grants -= grants;
1034                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1035                 ext->oe_end     = min(last_index, ext->oe_max_end);
1036                 LASSERT(ext->oe_end >= ext->oe_start);
1037                 LASSERT(ext->oe_grants > 0);
1038         }
1039         osc_object_unlock(obj);
1040
1041         if (grants > 0 || nr_pages > 0)
1042                 osc_free_grant(cli, nr_pages, grants);
1043
1044 out:
1045         cl_io_fini(env, io);
1046         cl_env_nested_put(&nest, env);
1047         RETURN(rc);
1048 }
1049
1050 /**
1051  * This function is used to make the extent prepared for transfer.
1052  * A race with flusing page - ll_writepage() has to be handled cautiously.
1053  */
1054 static int osc_extent_make_ready(const struct lu_env *env,
1055                                  struct osc_extent *ext)
1056 {
1057         struct osc_async_page *oap;
1058         struct osc_async_page *last = NULL;
1059         struct osc_object *obj = ext->oe_obj;
1060         unsigned int page_count = 0;
1061         int rc;
1062         ENTRY;
1063
1064         /* we're going to grab page lock, so object lock must not be taken. */
1065         LASSERT(sanity_check(ext) == 0);
1066         /* in locking state, any process should not touch this extent. */
1067         EASSERT(ext->oe_state == OES_LOCKING, ext);
1068         EASSERT(ext->oe_owner != NULL, ext);
1069
1070         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1071
1072         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1073                 ++page_count;
1074                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1075                         last = oap;
1076
1077                 /* checking ASYNC_READY is race safe */
1078                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1079                         continue;
1080
1081                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1082                 switch (rc) {
1083                 case 0:
1084                         spin_lock(&oap->oap_lock);
1085                         oap->oap_async_flags |= ASYNC_READY;
1086                         spin_unlock(&oap->oap_lock);
1087                         break;
1088                 case -EALREADY:
1089                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1090                         break;
1091                 default:
1092                         LASSERTF(0, "unknown return code: %d\n", rc);
1093                 }
1094         }
1095
1096         LASSERT(page_count == ext->oe_nr_pages);
1097         LASSERT(last != NULL);
1098         /* the last page is the only one we need to refresh its count by
1099          * the size of file. */
1100         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1101                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1102                 LASSERT(last_oap_count > 0);
1103                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_CACHE_SIZE);
1104                 last->oap_count = last_oap_count;
1105                 spin_lock(&last->oap_lock);
1106                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1107                 spin_unlock(&last->oap_lock);
1108         }
1109
1110         /* for the rest of pages, we don't need to call osf_refresh_count()
1111          * because it's known they are not the last page */
1112         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1113                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1114                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1115                         spin_lock(&oap->oap_lock);
1116                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1117                         spin_unlock(&oap->oap_lock);
1118                 }
1119         }
1120
1121         osc_object_lock(obj);
1122         osc_extent_state_set(ext, OES_RPC);
1123         osc_object_unlock(obj);
1124         /* get a refcount for RPC. */
1125         osc_extent_get(ext);
1126
1127         RETURN(0);
1128 }
1129
1130 /**
1131  * Quick and simple version of osc_extent_find(). This function is frequently
1132  * called to expand the extent for the same IO. To expand the extent, the
1133  * page index must be in the same or next chunk of ext->oe_end.
1134  */
1135 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, 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         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                 int 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         int 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 = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1539                                                   LWI_ON_SIGNAL_NOOP, NULL);
1540         int rc = -EDQUOT;
1541         ENTRY;
1542
1543         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1544
1545         spin_lock(&cli->cl_loi_list_lock);
1546
1547         /* force the caller to try sync io.  this can jump the list
1548          * of queued writes and create a discontiguous rpc stream */
1549         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1550             cli->cl_dirty_max_pages == 0 ||
1551             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1552                 GOTO(out, rc = -EDQUOT);
1553
1554         /* Hopefully normal case - cache space and write credits available */
1555         if (osc_enter_cache_try(cli, oap, bytes, 0))
1556                 GOTO(out, rc = 0);
1557
1558         /* We can get here for two reasons: too many dirty pages in cache, or
1559          * run out of grants. In both cases we should write dirty pages out.
1560          * Adding a cache waiter will trigger urgent write-out no matter what
1561          * RPC size will be.
1562          * The exiting condition is no avail grants and no dirty pages caching,
1563          * that really means there is no space on the OST. */
1564         init_waitqueue_head(&ocw.ocw_waitq);
1565         ocw.ocw_oap   = oap;
1566         ocw.ocw_grant = bytes;
1567         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1568                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1569                 ocw.ocw_rc = 0;
1570                 spin_unlock(&cli->cl_loi_list_lock);
1571
1572                 osc_io_unplug_async(env, cli, NULL);
1573
1574                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1575                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1576
1577                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1578
1579                 spin_lock(&cli->cl_loi_list_lock);
1580
1581                 /* l_wait_event is interrupted by signal, or timed out */
1582                 if (rc < 0) {
1583                         switch (rc) {
1584                         case -ETIMEDOUT:
1585                                 OSC_DUMP_GRANT(D_ERROR, cli,
1586                                                 "try to reserve %d.\n", bytes);
1587                                 osc_extent_tree_dump(D_ERROR, osc);
1588                                 rc = -EDQUOT;
1589                                 break;
1590                         case -EINTR:
1591                                 /* Ensures restartability - LU-3581 */
1592                                 rc = -ERESTARTSYS;
1593                                 break;
1594                         default:
1595                                 CDEBUG(D_CACHE, "%s: event for cache space @"
1596                                        " %p never arrived due to %d\n",
1597                                        cli->cl_import->imp_obd->obd_name,
1598                                        &ocw, rc);
1599                                 break;
1600                         }
1601                         list_del_init(&ocw.ocw_entry);
1602                         GOTO(out, rc);
1603                 }
1604
1605                 LASSERT(list_empty(&ocw.ocw_entry));
1606                 rc = ocw.ocw_rc;
1607
1608                 if (rc != -EDQUOT)
1609                         GOTO(out, rc);
1610                 if (osc_enter_cache_try(cli, oap, bytes, 0))
1611                         GOTO(out, rc = 0);
1612         }
1613         EXIT;
1614 out:
1615         spin_unlock(&cli->cl_loi_list_lock);
1616         OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1617         RETURN(rc);
1618 }
1619
1620 /* caller must hold loi_list_lock */
1621 void osc_wake_cache_waiters(struct client_obd *cli)
1622 {
1623         struct list_head *l, *tmp;
1624         struct osc_cache_waiter *ocw;
1625
1626         ENTRY;
1627         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1628                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1629                 list_del_init(&ocw->ocw_entry);
1630
1631                 ocw->ocw_rc = -EDQUOT;
1632                 /* we can't dirty more */
1633                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1634                     (1 + atomic_long_read(&obd_dirty_pages) >
1635                      obd_max_dirty_pages)) {
1636                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1637                                "osc max %ld, sys max %ld\n",
1638                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1639                                obd_max_dirty_pages);
1640                         goto wakeup;
1641                 }
1642
1643                 ocw->ocw_rc = 0;
1644                 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1645                         ocw->ocw_rc = -EDQUOT;
1646
1647 wakeup:
1648                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1649                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1650
1651                 wake_up(&ocw->ocw_waitq);
1652         }
1653
1654         EXIT;
1655 }
1656
1657 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1658 {
1659         int hprpc = !!list_empty(&osc->oo_hp_exts);
1660         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1661 }
1662
1663 /* This maintains the lists of pending pages to read/write for a given object
1664  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1665  * to quickly find objects that are ready to send an RPC. */
1666 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1667                          int cmd)
1668 {
1669         int invalid_import = 0;
1670         ENTRY;
1671
1672         /* if we have an invalid import we want to drain the queued pages
1673          * by forcing them through rpcs that immediately fail and complete
1674          * the pages.  recovery relies on this to empty the queued pages
1675          * before canceling the locks and evicting down the llite pages */
1676         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1677                 invalid_import = 1;
1678
1679         if (cmd & OBD_BRW_WRITE) {
1680                 if (atomic_read(&osc->oo_nr_writes) == 0)
1681                         RETURN(0);
1682                 if (invalid_import) {
1683                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1684                         RETURN(1);
1685                 }
1686                 if (!list_empty(&osc->oo_hp_exts)) {
1687                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1688                         RETURN(1);
1689                 }
1690                 if (!list_empty(&osc->oo_urgent_exts)) {
1691                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1692                         RETURN(1);
1693                 }
1694                 /* trigger a write rpc stream as long as there are dirtiers
1695                  * waiting for space.  as they're waiting, they're not going to
1696                  * create more pages to coalesce with what's waiting.. */
1697                 if (!list_empty(&cli->cl_cache_waiters)) {
1698                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1699                         RETURN(1);
1700                 }
1701                 if (atomic_read(&osc->oo_nr_writes) >=
1702                     cli->cl_max_pages_per_rpc)
1703                         RETURN(1);
1704         } else {
1705                 if (atomic_read(&osc->oo_nr_reads) == 0)
1706                         RETURN(0);
1707                 if (invalid_import) {
1708                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1709                         RETURN(1);
1710                 }
1711                 /* all read are urgent. */
1712                 if (!list_empty(&osc->oo_reading_exts))
1713                         RETURN(1);
1714         }
1715
1716         RETURN(0);
1717 }
1718
1719 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1720 {
1721         struct client_obd *cli = osc_cli(obj);
1722         if (cmd & OBD_BRW_WRITE) {
1723                 atomic_add(delta, &obj->oo_nr_writes);
1724                 atomic_add(delta, &cli->cl_pending_w_pages);
1725                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1726         } else {
1727                 atomic_add(delta, &obj->oo_nr_reads);
1728                 atomic_add(delta, &cli->cl_pending_r_pages);
1729                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1730         }
1731         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1732 }
1733
1734 static int osc_makes_hprpc(struct osc_object *obj)
1735 {
1736         return !list_empty(&obj->oo_hp_exts);
1737 }
1738
1739 static void on_list(struct list_head *item, struct list_head *list,
1740                     int should_be_on)
1741 {
1742         if (list_empty(item) && should_be_on)
1743                 list_add_tail(item, list);
1744         else if (!list_empty(item) && !should_be_on)
1745                 list_del_init(item);
1746 }
1747
1748 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1749  * can find pages to build into rpcs quickly */
1750 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1751 {
1752         if (osc_makes_hprpc(osc)) {
1753                 /* HP rpc */
1754                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1755                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1756         } else {
1757                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1758                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1759                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1760                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1761         }
1762
1763         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1764                 atomic_read(&osc->oo_nr_writes) > 0);
1765
1766         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1767                 atomic_read(&osc->oo_nr_reads) > 0);
1768
1769         return osc_is_ready(osc);
1770 }
1771
1772 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1773 {
1774         int is_ready;
1775
1776         spin_lock(&cli->cl_loi_list_lock);
1777         is_ready = __osc_list_maint(cli, osc);
1778         spin_unlock(&cli->cl_loi_list_lock);
1779
1780         return is_ready;
1781 }
1782
1783 /* this is trying to propogate async writeback errors back up to the
1784  * application.  As an async write fails we record the error code for later if
1785  * the app does an fsync.  As long as errors persist we force future rpcs to be
1786  * sync so that the app can get a sync error and break the cycle of queueing
1787  * pages for which writeback will fail. */
1788 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1789                            int rc)
1790 {
1791         if (rc) {
1792                 if (!ar->ar_rc)
1793                         ar->ar_rc = rc;
1794
1795                 ar->ar_force_sync = 1;
1796                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1797                 return;
1798
1799         }
1800
1801         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1802                 ar->ar_force_sync = 0;
1803 }
1804
1805 /* this must be called holding the loi list lock to give coverage to exit_cache,
1806  * async_flag maintenance, and oap_request */
1807 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1808                               struct osc_async_page *oap, int sent, int rc)
1809 {
1810         struct osc_object *osc = oap->oap_obj;
1811         struct lov_oinfo  *loi = osc->oo_oinfo;
1812         __u64 xid = 0;
1813
1814         ENTRY;
1815         if (oap->oap_request != NULL) {
1816                 xid = ptlrpc_req_xid(oap->oap_request);
1817                 ptlrpc_req_finished(oap->oap_request);
1818                 oap->oap_request = NULL;
1819         }
1820
1821         /* As the transfer for this page is being done, clear the flags */
1822         spin_lock(&oap->oap_lock);
1823         oap->oap_async_flags = 0;
1824         spin_unlock(&oap->oap_lock);
1825         oap->oap_interrupted = 0;
1826
1827         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1828                 spin_lock(&cli->cl_loi_list_lock);
1829                 osc_process_ar(&cli->cl_ar, xid, rc);
1830                 osc_process_ar(&loi->loi_ar, xid, rc);
1831                 spin_unlock(&cli->cl_loi_list_lock);
1832         }
1833
1834         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1835         if (rc)
1836                 CERROR("completion on oap %p obj %p returns %d.\n",
1837                        oap, osc, rc);
1838
1839         EXIT;
1840 }
1841
1842 /**
1843  * Try to add extent to one RPC. We need to think about the following things:
1844  * - # of pages must not be over max_pages_per_rpc
1845  * - extent must be compatible with previous ones
1846  */
1847 static int try_to_add_extent_for_io(struct client_obd *cli,
1848                                     struct osc_extent *ext,
1849                                     struct list_head *rpclist,
1850                                     int *pc, unsigned int *max_pages)
1851 {
1852         struct osc_extent *tmp;
1853         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1854                                                       struct osc_async_page,
1855                                                       oap_pending_item);
1856         ENTRY;
1857
1858         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1859                 ext);
1860
1861         *max_pages = max(ext->oe_mppr, *max_pages);
1862         if (*pc + ext->oe_nr_pages > *max_pages)
1863                 RETURN(0);
1864
1865         list_for_each_entry(tmp, rpclist, oe_link) {
1866                 struct osc_async_page *oap2;
1867                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1868                                         oap_pending_item);
1869                 EASSERT(tmp->oe_owner == current, tmp);
1870 #if 0
1871                 if (overlapped(tmp, ext)) {
1872                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1873                         EASSERT(0, ext);
1874                 }
1875 #endif
1876                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1877                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1878                                         " for a same RPC\n");
1879                         RETURN(0);
1880                 }
1881
1882                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1883                     !tmp->oe_grants != !ext->oe_grants)
1884                         RETURN(0);
1885
1886                 /* remove break for strict check */
1887                 break;
1888         }
1889
1890         *pc += ext->oe_nr_pages;
1891         list_move_tail(&ext->oe_link, rpclist);
1892         ext->oe_owner = current;
1893         RETURN(1);
1894 }
1895
1896 /**
1897  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1898  * get_write_extent() takes all appropriate extents in atomic.
1899  *
1900  * The following policy is used to collect extents for IO:
1901  * 1. Add as many HP extents as possible;
1902  * 2. Add the first urgent extent in urgent extent list and take it out of
1903  *    urgent list;
1904  * 3. Add subsequent extents of this urgent extent;
1905  * 4. If urgent list is not empty, goto 2;
1906  * 5. Traverse the extent tree from the 1st extent;
1907  * 6. Above steps exit if there is no space in this RPC.
1908  */
1909 static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1910 {
1911         struct client_obd *cli = osc_cli(obj);
1912         struct osc_extent *ext;
1913         int page_count = 0;
1914         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1915
1916         LASSERT(osc_object_is_locked(obj));
1917         while (!list_empty(&obj->oo_hp_exts)) {
1918                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1919                                  oe_link);
1920                 LASSERT(ext->oe_state == OES_CACHE);
1921                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1922                                               &max_pages))
1923                         return page_count;
1924                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1925         }
1926         if (page_count == max_pages)
1927                 return page_count;
1928
1929         while (!list_empty(&obj->oo_urgent_exts)) {
1930                 ext = list_entry(obj->oo_urgent_exts.next,
1931                                  struct osc_extent, oe_link);
1932                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1933                                               &max_pages))
1934                         return page_count;
1935
1936                 if (!ext->oe_intree)
1937                         continue;
1938
1939                 while ((ext = next_extent(ext)) != NULL) {
1940                         if ((ext->oe_state != OES_CACHE) ||
1941                             (!list_empty(&ext->oe_link) &&
1942                              ext->oe_owner != NULL))
1943                                 continue;
1944
1945                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1946                                                       &page_count, &max_pages))
1947                                 return page_count;
1948                 }
1949         }
1950         if (page_count == max_pages)
1951                 return page_count;
1952
1953         ext = first_extent(obj);
1954         while (ext != NULL) {
1955                 if ((ext->oe_state != OES_CACHE) ||
1956                     /* this extent may be already in current rpclist */
1957                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1958                         ext = next_extent(ext);
1959                         continue;
1960                 }
1961
1962                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1963                                               &max_pages))
1964                         return page_count;
1965
1966                 ext = next_extent(ext);
1967         }
1968         return page_count;
1969 }
1970
1971 static int
1972 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1973                    struct osc_object *osc, pdl_policy_t pol)
1974 __must_hold(osc)
1975 {
1976         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
1977         struct osc_extent *ext;
1978         struct osc_extent *tmp;
1979         struct osc_extent *first = NULL;
1980         obd_count page_count = 0;
1981         int srvlock = 0;
1982         int rc = 0;
1983         ENTRY;
1984
1985         LASSERT(osc_object_is_locked(osc));
1986
1987         page_count = get_write_extents(osc, &rpclist);
1988         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1989
1990         if (list_empty(&rpclist))
1991                 RETURN(0);
1992
1993         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1994
1995         list_for_each_entry(ext, &rpclist, oe_link) {
1996                 LASSERT(ext->oe_state == OES_CACHE ||
1997                         ext->oe_state == OES_LOCK_DONE);
1998                 if (ext->oe_state == OES_CACHE)
1999                         osc_extent_state_set(ext, OES_LOCKING);
2000                 else
2001                         osc_extent_state_set(ext, OES_RPC);
2002         }
2003
2004         /* we're going to grab page lock, so release object lock because
2005          * lock order is page lock -> object lock. */
2006         osc_object_unlock(osc);
2007
2008         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2009                 if (ext->oe_state == OES_LOCKING) {
2010                         rc = osc_extent_make_ready(env, ext);
2011                         if (unlikely(rc < 0)) {
2012                                 list_del_init(&ext->oe_link);
2013                                 osc_extent_finish(env, ext, 0, rc);
2014                                 continue;
2015                         }
2016                 }
2017                 if (first == NULL) {
2018                         first = ext;
2019                         srvlock = ext->oe_srvlock;
2020                 } else {
2021                         LASSERT(srvlock == ext->oe_srvlock);
2022                 }
2023         }
2024
2025         if (!list_empty(&rpclist)) {
2026                 LASSERT(page_count > 0);
2027                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2028                 LASSERT(list_empty(&rpclist));
2029         }
2030
2031         osc_object_lock(osc);
2032         RETURN(rc);
2033 }
2034
2035 /**
2036  * prepare pages for ASYNC io and put pages in send queue.
2037  *
2038  * \param cmd OBD_BRW_* macroses
2039  * \param lop pending pages
2040  *
2041  * \return zero if no page added to send queue.
2042  * \return 1 if pages successfully added to send queue.
2043  * \return negative on errors.
2044  */
2045 static int
2046 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2047                   struct osc_object *osc, pdl_policy_t pol)
2048 __must_hold(osc)
2049 {
2050         struct osc_extent *ext;
2051         struct osc_extent *next;
2052         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2053         int page_count = 0;
2054         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2055         int rc = 0;
2056         ENTRY;
2057
2058         LASSERT(osc_object_is_locked(osc));
2059         list_for_each_entry_safe(ext, next,
2060                                      &osc->oo_reading_exts, oe_link) {
2061                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2062                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2063                                               &max_pages))
2064                         break;
2065                 osc_extent_state_set(ext, OES_RPC);
2066                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2067         }
2068         LASSERT(page_count <= max_pages);
2069
2070         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2071
2072         if (!list_empty(&rpclist)) {
2073                 osc_object_unlock(osc);
2074
2075                 LASSERT(page_count > 0);
2076                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2077                 LASSERT(list_empty(&rpclist));
2078
2079                 osc_object_lock(osc);
2080         }
2081         RETURN(rc);
2082 }
2083
2084 #define list_to_obj(list, item) ({                                            \
2085         struct list_head *__tmp = (list)->next;                               \
2086         list_del_init(__tmp);                                         \
2087         list_entry(__tmp, struct osc_object, oo_##item);                      \
2088 })
2089
2090 /* This is called by osc_check_rpcs() to find which objects have pages that
2091  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2092 static struct osc_object *osc_next_obj(struct client_obd *cli)
2093 {
2094         ENTRY;
2095
2096         /* First return objects that have blocked locks so that they
2097          * will be flushed quickly and other clients can get the lock,
2098          * then objects which have pages ready to be stuffed into RPCs */
2099         if (!list_empty(&cli->cl_loi_hp_ready_list))
2100                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2101         if (!list_empty(&cli->cl_loi_ready_list))
2102                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2103
2104         /* then if we have cache waiters, return all objects with queued
2105          * writes.  This is especially important when many small files
2106          * have filled up the cache and not been fired into rpcs because
2107          * they don't pass the nr_pending/object threshhold */
2108         if (!list_empty(&cli->cl_cache_waiters) &&
2109             !list_empty(&cli->cl_loi_write_list))
2110                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2111
2112         /* then return all queued objects when we have an invalid import
2113          * so that they get flushed */
2114         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2115                 if (!list_empty(&cli->cl_loi_write_list))
2116                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2117                                            write_item));
2118                 if (!list_empty(&cli->cl_loi_read_list))
2119                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2120                                            read_item));
2121         }
2122         RETURN(NULL);
2123 }
2124
2125 /* called with the loi list lock held */
2126 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2127                            pdl_policy_t pol)
2128 __must_hold(&cli->cl_loi_list_lock)
2129 {
2130         struct osc_object *osc;
2131         int rc = 0;
2132         ENTRY;
2133
2134         while ((osc = osc_next_obj(cli)) != NULL) {
2135                 struct cl_object *obj = osc2cl(osc);
2136                 struct lu_ref_link link;
2137
2138                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2139
2140                 if (osc_max_rpc_in_flight(cli, osc)) {
2141                         __osc_list_maint(cli, osc);
2142                         break;
2143                 }
2144
2145                 cl_object_get(obj);
2146                 spin_unlock(&cli->cl_loi_list_lock);
2147                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2148
2149                 /* attempt some read/write balancing by alternating between
2150                  * reads and writes in an object.  The makes_rpc checks here
2151                  * would be redundant if we were getting read/write work items
2152                  * instead of objects.  we don't want send_oap_rpc to drain a
2153                  * partial read pending queue when we're given this object to
2154                  * do io on writes while there are cache waiters */
2155                 osc_object_lock(osc);
2156                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2157                         rc = osc_send_write_rpc(env, cli, osc, pol);
2158                         if (rc < 0) {
2159                                 CERROR("Write request failed with %d\n", rc);
2160
2161                                 /* osc_send_write_rpc failed, mostly because of
2162                                  * memory pressure.
2163                                  *
2164                                  * It can't break here, because if:
2165                                  *  - a page was submitted by osc_io_submit, so
2166                                  *    page locked;
2167                                  *  - no request in flight
2168                                  *  - no subsequent request
2169                                  * The system will be in live-lock state,
2170                                  * because there is no chance to call
2171                                  * osc_io_unplug() and osc_check_rpcs() any
2172                                  * more. pdflush can't help in this case,
2173                                  * because it might be blocked at grabbing
2174                                  * the page lock as we mentioned.
2175                                  *
2176                                  * Anyway, continue to drain pages. */
2177                                 /* break; */
2178                         }
2179                 }
2180                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2181                         rc = osc_send_read_rpc(env, cli, osc, pol);
2182                         if (rc < 0)
2183                                 CERROR("Read request failed with %d\n", rc);
2184                 }
2185                 osc_object_unlock(osc);
2186
2187                 osc_list_maint(cli, osc);
2188                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2189                 cl_object_put(env, obj);
2190
2191                 spin_lock(&cli->cl_loi_list_lock);
2192         }
2193 }
2194
2195 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2196                           struct osc_object *osc, pdl_policy_t pol, int async)
2197 {
2198         int rc = 0;
2199
2200         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2201                 return 0;
2202
2203         if (!async) {
2204                 /* disable osc_lru_shrink() temporarily to avoid
2205                  * potential stack overrun problem. LU-2859 */
2206                 atomic_inc(&cli->cl_lru_shrinkers);
2207                 spin_lock(&cli->cl_loi_list_lock);
2208                 osc_check_rpcs(env, cli, pol);
2209                 spin_unlock(&cli->cl_loi_list_lock);
2210                 atomic_dec(&cli->cl_lru_shrinkers);
2211         } else {
2212                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2213                 LASSERT(cli->cl_writeback_work != NULL);
2214                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2215         }
2216         return rc;
2217 }
2218
2219 static int osc_io_unplug_async(const struct lu_env *env,
2220                                 struct client_obd *cli, struct osc_object *osc)
2221 {
2222         /* XXX: policy is no use actually. */
2223         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2224 }
2225
2226 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2227                    struct osc_object *osc, pdl_policy_t pol)
2228 {
2229         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2230 }
2231
2232 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2233                         struct page *page, loff_t offset)
2234 {
2235         struct obd_export     *exp = osc_export(osc);
2236         struct osc_async_page *oap = &ops->ops_oap;
2237         ENTRY;
2238
2239         if (!page)
2240                 return cfs_size_round(sizeof(*oap));
2241
2242         oap->oap_magic = OAP_MAGIC;
2243         oap->oap_cli = &exp->exp_obd->u.cli;
2244         oap->oap_obj = osc;
2245
2246         oap->oap_page = page;
2247         oap->oap_obj_off = offset;
2248         LASSERT(!(offset & ~CFS_PAGE_MASK));
2249
2250         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2251                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2252
2253         INIT_LIST_HEAD(&oap->oap_pending_item);
2254         INIT_LIST_HEAD(&oap->oap_rpc_item);
2255
2256         spin_lock_init(&oap->oap_lock);
2257         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2258                oap, page, oap->oap_obj_off);
2259         RETURN(0);
2260 }
2261
2262 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2263                        struct osc_page *ops)
2264 {
2265         struct osc_io *oio = osc_env_io(env);
2266         struct osc_extent     *ext = NULL;
2267         struct osc_async_page *oap = &ops->ops_oap;
2268         struct client_obd     *cli = oap->oap_cli;
2269         struct osc_object     *osc = oap->oap_obj;
2270         pgoff_t index;
2271         int    grants = 0;
2272         int    brw_flags = OBD_BRW_ASYNC;
2273         int    cmd = OBD_BRW_WRITE;
2274         int    need_release = 0;
2275         int    rc = 0;
2276         ENTRY;
2277
2278         if (oap->oap_magic != OAP_MAGIC)
2279                 RETURN(-EINVAL);
2280
2281         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2282                 RETURN(-EIO);
2283
2284         if (!list_empty(&oap->oap_pending_item) ||
2285             !list_empty(&oap->oap_rpc_item))
2286                 RETURN(-EBUSY);
2287
2288         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2289         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2290         if (!client_is_remote(osc_export(osc)) &&
2291             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2292                 brw_flags |= OBD_BRW_NOQUOTA;
2293                 cmd |= OBD_BRW_NOQUOTA;
2294         }
2295
2296         /* check if the file's owner/group is over quota */
2297         if (!(cmd & OBD_BRW_NOQUOTA)) {
2298                 struct cl_object *obj;
2299                 struct cl_attr   *attr;
2300                 unsigned int qid[MAXQUOTAS];
2301
2302                 obj = cl_object_top(&osc->oo_cl);
2303                 attr = &osc_env_info(env)->oti_attr;
2304
2305                 cl_object_attr_lock(obj);
2306                 rc = cl_object_attr_get(env, obj, attr);
2307                 cl_object_attr_unlock(obj);
2308
2309                 qid[USRQUOTA] = attr->cat_uid;
2310                 qid[GRPQUOTA] = attr->cat_gid;
2311                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2312                         rc = -EDQUOT;
2313                 if (rc)
2314                         RETURN(rc);
2315         }
2316
2317         oap->oap_cmd = cmd;
2318         oap->oap_page_off = ops->ops_from;
2319         oap->oap_count = ops->ops_to - ops->ops_from;
2320         /* No need to hold a lock here,
2321          * since this page is not in any list yet. */
2322         oap->oap_async_flags = 0;
2323         oap->oap_brw_flags = brw_flags;
2324
2325         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2326                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2327
2328         index = osc_index(oap2osc(oap));
2329
2330         /* Add this page into extent by the following steps:
2331          * 1. if there exists an active extent for this IO, mostly this page
2332          *    can be added to the active extent and sometimes we need to
2333          *    expand extent to accomodate this page;
2334          * 2. otherwise, a new extent will be allocated. */
2335
2336         ext = oio->oi_active;
2337         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2338                 /* one chunk plus extent overhead must be enough to write this
2339                  * page */
2340                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2341                 if (ext->oe_end >= index)
2342                         grants = 0;
2343
2344                 /* it doesn't need any grant to dirty this page */
2345                 spin_lock(&cli->cl_loi_list_lock);
2346                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2347                 spin_unlock(&cli->cl_loi_list_lock);
2348                 if (rc == 0) { /* try failed */
2349                         grants = 0;
2350                         need_release = 1;
2351                 } else if (ext->oe_end < index) {
2352                         int tmp = grants;
2353                         /* try to expand this extent */
2354                         rc = osc_extent_expand(ext, index, &tmp);
2355                         if (rc < 0) {
2356                                 need_release = 1;
2357                                 /* don't free reserved grant */
2358                         } else {
2359                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2360                                                 "expanded for %lu.\n", index);
2361                                 osc_unreserve_grant(cli, grants, tmp);
2362                                 grants = 0;
2363                         }
2364                 }
2365                 rc = 0;
2366         } else if (ext != NULL) {
2367                 /* index is located outside of active extent */
2368                 need_release = 1;
2369         }
2370         if (need_release) {
2371                 osc_extent_release(env, ext);
2372                 oio->oi_active = NULL;
2373                 ext = NULL;
2374         }
2375
2376         if (ext == NULL) {
2377                 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2378
2379                 /* try to find new extent to cover this page */
2380                 LASSERT(oio->oi_active == NULL);
2381                 /* we may have allocated grant for this page if we failed
2382                  * to expand the previous active extent. */
2383                 LASSERT(ergo(grants > 0, grants >= tmp));
2384
2385                 rc = 0;
2386                 if (grants == 0) {
2387                         /* we haven't allocated grant for this page. */
2388                         rc = osc_enter_cache(env, cli, oap, tmp);
2389                         if (rc == 0)
2390                                 grants = tmp;
2391                 }
2392
2393                 tmp = grants;
2394                 if (rc == 0) {
2395                         ext = osc_extent_find(env, osc, index, &tmp);
2396                         if (IS_ERR(ext)) {
2397                                 LASSERT(tmp == grants);
2398                                 osc_exit_cache(cli, oap);
2399                                 rc = PTR_ERR(ext);
2400                                 ext = NULL;
2401                         } else {
2402                                 oio->oi_active = ext;
2403                         }
2404                 }
2405                 if (grants > 0)
2406                         osc_unreserve_grant(cli, grants, tmp);
2407         }
2408
2409         LASSERT(ergo(rc == 0, ext != NULL));
2410         if (ext != NULL) {
2411                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2412                          ext, "index = %lu.\n", index);
2413                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2414
2415                 osc_object_lock(osc);
2416                 if (ext->oe_nr_pages == 0)
2417                         ext->oe_srvlock = ops->ops_srvlock;
2418                 else
2419                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2420                 ++ext->oe_nr_pages;
2421                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2422                 osc_object_unlock(osc);
2423         }
2424         RETURN(rc);
2425 }
2426
2427 int osc_teardown_async_page(const struct lu_env *env,
2428                             struct osc_object *obj, struct osc_page *ops)
2429 {
2430         struct osc_async_page *oap = &ops->ops_oap;
2431         struct osc_extent     *ext = NULL;
2432         int rc = 0;
2433         ENTRY;
2434
2435         LASSERT(oap->oap_magic == OAP_MAGIC);
2436
2437         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2438                oap, ops, osc_index(oap2osc(oap)));
2439
2440         osc_object_lock(obj);
2441         if (!list_empty(&oap->oap_rpc_item)) {
2442                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2443                 rc = -EBUSY;
2444         } else if (!list_empty(&oap->oap_pending_item)) {
2445                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2446                 /* only truncated pages are allowed to be taken out.
2447                  * See osc_extent_truncate() and osc_cache_truncate_start()
2448                  * for details. */
2449                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2450                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2451                                         osc_index(oap2osc(oap)));
2452                         rc = -EBUSY;
2453                 }
2454         }
2455         osc_object_unlock(obj);
2456         if (ext != NULL)
2457                 osc_extent_put(env, ext);
2458         RETURN(rc);
2459 }
2460
2461 /**
2462  * This is called when a page is picked up by kernel to write out.
2463  *
2464  * We should find out the corresponding extent and add the whole extent
2465  * into urgent list. The extent may be being truncated or used, handle it
2466  * carefully.
2467  */
2468 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2469                          struct osc_page *ops)
2470 {
2471         struct osc_extent *ext   = NULL;
2472         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2473         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2474         pgoff_t            index = osc_index(ops);
2475         struct osc_async_page *oap = &ops->ops_oap;
2476         bool unplug = false;
2477         int rc = 0;
2478         ENTRY;
2479
2480         osc_object_lock(obj);
2481         ext = osc_extent_lookup(obj, index);
2482         if (ext == NULL) {
2483                 osc_extent_tree_dump(D_ERROR, obj);
2484                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2485         }
2486
2487         switch (ext->oe_state) {
2488         case OES_RPC:
2489         case OES_LOCK_DONE:
2490                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2491                 LASSERT(0);
2492                 break;
2493         case OES_LOCKING:
2494                 /* If we know this extent is being written out, we should abort
2495                  * so that the writer can make this page ready. Otherwise, there
2496                  * exists a deadlock problem because other process can wait for
2497                  * page writeback bit holding page lock; and meanwhile in
2498                  * vvp_page_make_ready(), we need to grab page lock before
2499                  * really sending the RPC. */
2500         case OES_TRUNC:
2501                 /* race with truncate, page will be redirtied */
2502         case OES_ACTIVE:
2503                 /* The extent is active so we need to abort and let the caller
2504                  * re-dirty the page. If we continued on here, and we were the
2505                  * one making the extent active, we could deadlock waiting for
2506                  * the page writeback to clear but it won't because the extent
2507                  * is active and won't be written out. */
2508                 GOTO(out, rc = -EAGAIN);
2509         default:
2510                 break;
2511         }
2512
2513         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2514         if (rc)
2515                 GOTO(out, rc);
2516
2517         spin_lock(&oap->oap_lock);
2518         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2519         spin_unlock(&oap->oap_lock);
2520
2521         if (memory_pressure_get())
2522                 ext->oe_memalloc = 1;
2523
2524         ext->oe_urgent = 1;
2525         if (ext->oe_state == OES_CACHE) {
2526                 OSC_EXTENT_DUMP(D_CACHE, ext,
2527                                 "flush page %p make it urgent.\n", oap);
2528                 if (list_empty(&ext->oe_link))
2529                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2530                 unplug = true;
2531         }
2532         rc = 0;
2533         EXIT;
2534
2535 out:
2536         osc_object_unlock(obj);
2537         osc_extent_put(env, ext);
2538         if (unplug)
2539                 osc_io_unplug_async(env, osc_cli(obj), obj);
2540         return rc;
2541 }
2542
2543 /**
2544  * this is called when a sync waiter receives an interruption.  Its job is to
2545  * get the caller woken as soon as possible.  If its page hasn't been put in an
2546  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2547  * desiring interruption which will forcefully complete the rpc once the rpc
2548  * has timed out.
2549  */
2550 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2551 {
2552         struct osc_async_page *oap = &ops->ops_oap;
2553         struct osc_object     *obj = oap->oap_obj;
2554         struct client_obd     *cli = osc_cli(obj);
2555         struct osc_extent     *ext;
2556         struct osc_extent     *found = NULL;
2557         struct list_head            *plist;
2558         pgoff_t index = osc_index(ops);
2559         int     rc = -EBUSY;
2560         int     cmd;
2561         ENTRY;
2562
2563         LASSERT(!oap->oap_interrupted);
2564         oap->oap_interrupted = 1;
2565
2566         /* Find out the caching extent */
2567         osc_object_lock(obj);
2568         if (oap->oap_cmd & OBD_BRW_WRITE) {
2569                 plist = &obj->oo_urgent_exts;
2570                 cmd   = OBD_BRW_WRITE;
2571         } else {
2572                 plist = &obj->oo_reading_exts;
2573                 cmd   = OBD_BRW_READ;
2574         }
2575         list_for_each_entry(ext, plist, oe_link) {
2576                 if (ext->oe_start <= index && ext->oe_end >= index) {
2577                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2578                         /* For OES_LOCK_DONE state extent, it has already held
2579                          * a refcount for RPC. */
2580                         found = osc_extent_get(ext);
2581                         break;
2582                 }
2583         }
2584         if (found != NULL) {
2585                 list_del_init(&found->oe_link);
2586                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2587                 osc_object_unlock(obj);
2588
2589                 osc_extent_finish(env, found, 0, -EINTR);
2590                 osc_extent_put(env, found);
2591                 rc = 0;
2592         } else {
2593                 osc_object_unlock(obj);
2594                 /* ok, it's been put in an rpc. only one oap gets a request
2595                  * reference */
2596                 if (oap->oap_request != NULL) {
2597                         ptlrpc_mark_interrupted(oap->oap_request);
2598                         ptlrpcd_wake(oap->oap_request);
2599                         ptlrpc_req_finished(oap->oap_request);
2600                         oap->oap_request = NULL;
2601                 }
2602         }
2603
2604         osc_list_maint(cli, obj);
2605         RETURN(rc);
2606 }
2607
2608 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2609                          struct list_head *list, int cmd, int brw_flags)
2610 {
2611         struct client_obd     *cli = osc_cli(obj);
2612         struct osc_extent     *ext;
2613         struct osc_async_page *oap;
2614         int     page_count = 0;
2615         int     mppr       = cli->cl_max_pages_per_rpc;
2616         pgoff_t start      = CL_PAGE_EOF;
2617         pgoff_t end        = 0;
2618         ENTRY;
2619
2620         list_for_each_entry(oap, list, oap_pending_item) {
2621                 pgoff_t index = osc_index(oap2osc(oap));
2622                 if (index > end)
2623                         end = index;
2624                 if (index < start)
2625                         start = index;
2626                 ++page_count;
2627                 mppr <<= (page_count > mppr);
2628         }
2629
2630         ext = osc_extent_alloc(obj);
2631         if (ext == NULL) {
2632                 list_for_each_entry(oap, list, oap_pending_item) {
2633                         list_del_init(&oap->oap_pending_item);
2634                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2635                 }
2636                 RETURN(-ENOMEM);
2637         }
2638
2639         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2640         ext->oe_sync = 1;
2641         ext->oe_urgent = 1;
2642         ext->oe_start = start;
2643         ext->oe_end = ext->oe_max_end = end;
2644         ext->oe_obj = obj;
2645         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2646         ext->oe_nr_pages = page_count;
2647         ext->oe_mppr = mppr;
2648         list_splice_init(list, &ext->oe_pages);
2649
2650         osc_object_lock(obj);
2651         /* Reuse the initial refcount for RPC, don't drop it */
2652         osc_extent_state_set(ext, OES_LOCK_DONE);
2653         if (cmd & OBD_BRW_WRITE) {
2654                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2655                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2656         } else {
2657                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2658                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2659         }
2660         osc_object_unlock(obj);
2661
2662         osc_io_unplug_async(env, cli, obj);
2663         RETURN(0);
2664 }
2665
2666 /**
2667  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2668  */
2669 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2670                              struct osc_object *obj, __u64 size)
2671 {
2672         struct client_obd *cli = osc_cli(obj);
2673         struct osc_extent *ext;
2674         struct osc_extent *waiting = NULL;
2675         pgoff_t index;
2676         struct list_head list = LIST_HEAD_INIT(list);
2677         int result = 0;
2678         bool partial;
2679         ENTRY;
2680
2681         /* pages with index greater or equal to index will be truncated. */
2682         index = cl_index(osc2cl(obj), size);
2683         partial = size > cl_offset(osc2cl(obj), index);
2684
2685 again:
2686         osc_object_lock(obj);
2687         ext = osc_extent_search(obj, index);
2688         if (ext == NULL)
2689                 ext = first_extent(obj);
2690         else if (ext->oe_end < index)
2691                 ext = next_extent(ext);
2692         while (ext != NULL) {
2693                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2694
2695                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2696                         /* if ext is in urgent state, it means there must exist
2697                          * a page already having been flushed by write_page().
2698                          * We have to wait for this extent because we can't
2699                          * truncate that page. */
2700                         LASSERT(!ext->oe_hp);
2701                         OSC_EXTENT_DUMP(D_CACHE, ext,
2702                                         "waiting for busy extent\n");
2703                         waiting = osc_extent_get(ext);
2704                         break;
2705                 }
2706
2707                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2708
2709                 osc_extent_get(ext);
2710                 if (ext->oe_state == OES_ACTIVE) {
2711                         /* though we grab inode mutex for write path, but we
2712                          * release it before releasing extent(in osc_io_end()),
2713                          * so there is a race window that an extent is still
2714                          * in OES_ACTIVE when truncate starts. */
2715                         LASSERT(!ext->oe_trunc_pending);
2716                         ext->oe_trunc_pending = 1;
2717                 } else {
2718                         EASSERT(ext->oe_state == OES_CACHE, ext);
2719                         osc_extent_state_set(ext, OES_TRUNC);
2720                         osc_update_pending(obj, OBD_BRW_WRITE,
2721                                            -ext->oe_nr_pages);
2722                 }
2723                 EASSERT(list_empty(&ext->oe_link), ext);
2724                 list_add_tail(&ext->oe_link, &list);
2725
2726                 ext = next_extent(ext);
2727         }
2728         osc_object_unlock(obj);
2729
2730         osc_list_maint(cli, obj);
2731
2732         while (!list_empty(&list)) {
2733                 int rc;
2734
2735                 ext = list_entry(list.next, struct osc_extent, oe_link);
2736                 list_del_init(&ext->oe_link);
2737
2738                 /* extent may be in OES_ACTIVE state because inode mutex
2739                  * is released before osc_io_end() in file write case */
2740                 if (ext->oe_state != OES_TRUNC)
2741                         osc_extent_wait(env, ext, OES_TRUNC);
2742
2743                 rc = osc_extent_truncate(ext, index, partial);
2744                 if (rc < 0) {
2745                         if (result == 0)
2746                                 result = rc;
2747
2748                         OSC_EXTENT_DUMP(D_ERROR, ext,
2749                                         "truncate error %d\n", rc);
2750                 } else if (ext->oe_nr_pages == 0) {
2751                         osc_extent_remove(ext);
2752                 } else {
2753                         /* this must be an overlapped extent which means only
2754                          * part of pages in this extent have been truncated.
2755                          */
2756                         EASSERTF(ext->oe_start <= index, ext,
2757                                  "trunc index = %lu/%d.\n", index, partial);
2758                         /* fix index to skip this partially truncated extent */
2759                         index = ext->oe_end + 1;
2760                         partial = false;
2761
2762                         /* we need to hold this extent in OES_TRUNC state so
2763                          * that no writeback will happen. This is to avoid
2764                          * BUG 17397. */
2765                         LASSERT(oio->oi_trunc == NULL);
2766                         oio->oi_trunc = osc_extent_get(ext);
2767                         OSC_EXTENT_DUMP(D_CACHE, ext,
2768                                         "trunc at "LPU64"\n", size);
2769                 }
2770                 osc_extent_put(env, ext);
2771         }
2772         if (waiting != NULL) {
2773                 int rc;
2774
2775                 /* ignore the result of osc_extent_wait the write initiator
2776                  * should take care of it. */
2777                 rc = osc_extent_wait(env, waiting, OES_INV);
2778                 if (rc < 0)
2779                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2780
2781                 osc_extent_put(env, waiting);
2782                 waiting = NULL;
2783                 goto again;
2784         }
2785         RETURN(result);
2786 }
2787
2788 /**
2789  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2790  */
2791 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2792                             struct osc_object *obj)
2793 {
2794         struct osc_extent *ext = oio->oi_trunc;
2795
2796         oio->oi_trunc = NULL;
2797         if (ext != NULL) {
2798                 bool unplug = false;
2799
2800                 EASSERT(ext->oe_nr_pages > 0, ext);
2801                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2802                 EASSERT(!ext->oe_urgent, ext);
2803
2804                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2805                 osc_object_lock(obj);
2806                 osc_extent_state_set(ext, OES_CACHE);
2807                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2808                         ext->oe_urgent = 1;
2809                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2810                         unplug = true;
2811                 }
2812                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2813                 osc_object_unlock(obj);
2814                 osc_extent_put(env, ext);
2815
2816                 if (unplug)
2817                         osc_io_unplug_async(env, osc_cli(obj), obj);
2818         }
2819 }
2820
2821 /**
2822  * Wait for extents in a specific range to be written out.
2823  * The caller must have called osc_cache_writeback_range() to issue IO
2824  * otherwise it will take a long time for this function to finish.
2825  *
2826  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2827  * nobody else can dirty this range of file while we're waiting for
2828  * extents to be written.
2829  */
2830 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2831                          pgoff_t start, pgoff_t end)
2832 {
2833         struct osc_extent *ext;
2834         pgoff_t index = start;
2835         int     result = 0;
2836         ENTRY;
2837
2838 again:
2839         osc_object_lock(obj);
2840         ext = osc_extent_search(obj, index);
2841         if (ext == NULL)
2842                 ext = first_extent(obj);
2843         else if (ext->oe_end < index)
2844                 ext = next_extent(ext);
2845         while (ext != NULL) {
2846                 int rc;
2847
2848                 if (ext->oe_start > end)
2849                         break;
2850
2851                 if (!ext->oe_fsync_wait) {
2852                         ext = next_extent(ext);
2853                         continue;
2854                 }
2855
2856                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2857                              ext->oe_hp || ext->oe_urgent), ext);
2858                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2859                              !ext->oe_hp && ext->oe_urgent), ext);
2860
2861                 index = ext->oe_end + 1;
2862                 osc_extent_get(ext);
2863                 osc_object_unlock(obj);
2864
2865                 rc = osc_extent_wait(env, ext, OES_INV);
2866                 if (result == 0)
2867                         result = rc;
2868                 osc_extent_put(env, ext);
2869                 goto again;
2870         }
2871         osc_object_unlock(obj);
2872
2873         OSC_IO_DEBUG(obj, "sync file range.\n");
2874         RETURN(result);
2875 }
2876
2877 /**
2878  * Called to write out a range of osc object.
2879  *
2880  * @hp     : should be set this is caused by lock cancel;
2881  * @discard: is set if dirty pages should be dropped - file will be deleted or
2882  *         truncated, this implies there is no partially discarding extents.
2883  *
2884  * Return how many pages will be issued, or error code if error occurred.
2885  */
2886 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2887                               pgoff_t start, pgoff_t end, int hp, int discard)
2888 {
2889         struct osc_extent *ext;
2890         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2891         bool unplug = false;
2892         int result = 0;
2893         ENTRY;
2894
2895         osc_object_lock(obj);
2896         ext = osc_extent_search(obj, start);
2897         if (ext == NULL)
2898                 ext = first_extent(obj);
2899         else if (ext->oe_end < start)
2900                 ext = next_extent(ext);
2901         while (ext != NULL) {
2902                 if (ext->oe_start > end)
2903                         break;
2904
2905                 ext->oe_fsync_wait = 1;
2906                 switch (ext->oe_state) {
2907                 case OES_CACHE:
2908                         result += ext->oe_nr_pages;
2909                         if (!discard) {
2910                                 struct list_head *list = NULL;
2911                                 if (hp) {
2912                                         EASSERT(!ext->oe_hp, ext);
2913                                         ext->oe_hp = 1;
2914                                         list = &obj->oo_hp_exts;
2915                                 } else if (!ext->oe_urgent) {
2916                                         ext->oe_urgent = 1;
2917                                         list = &obj->oo_urgent_exts;
2918                                 }
2919                                 if (list != NULL)
2920                                         list_move_tail(&ext->oe_link, list);
2921                                 unplug = true;
2922                         } else {
2923                                 /* the only discarder is lock cancelling, so
2924                                  * [start, end] must contain this extent */
2925                                 EASSERT(ext->oe_start >= start &&
2926                                         ext->oe_max_end <= end, ext);
2927                                 osc_extent_state_set(ext, OES_LOCKING);
2928                                 ext->oe_owner = current;
2929                                 list_move_tail(&ext->oe_link,
2930                                                    &discard_list);
2931                                 osc_update_pending(obj, OBD_BRW_WRITE,
2932                                                    -ext->oe_nr_pages);
2933                         }
2934                         break;
2935                 case OES_ACTIVE:
2936                         /* It's pretty bad to wait for ACTIVE extents, because
2937                          * we don't know how long we will wait for it to be
2938                          * flushed since it may be blocked at awaiting more
2939                          * grants. We do this for the correctness of fsync. */
2940                         LASSERT(hp == 0 && discard == 0);
2941                         ext->oe_urgent = 1;
2942                         break;
2943                 case OES_TRUNC:
2944                         /* this extent is being truncated, can't do anything
2945                          * for it now. it will be set to urgent after truncate
2946                          * is finished in osc_cache_truncate_end(). */
2947                 default:
2948                         break;
2949                 }
2950                 ext = next_extent(ext);
2951         }
2952         osc_object_unlock(obj);
2953
2954         LASSERT(ergo(!discard, list_empty(&discard_list)));
2955         if (!list_empty(&discard_list)) {
2956                 struct osc_extent *tmp;
2957                 int rc;
2958
2959                 osc_list_maint(osc_cli(obj), obj);
2960                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2961                         list_del_init(&ext->oe_link);
2962                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2963
2964                         /* Discard caching pages. We don't actually write this
2965                          * extent out but we complete it as if we did. */
2966                         rc = osc_extent_make_ready(env, ext);
2967                         if (unlikely(rc < 0)) {
2968                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2969                                                 "make_ready returned %d\n", rc);
2970                                 if (result >= 0)
2971                                         result = rc;
2972                         }
2973
2974                         /* finish the extent as if the pages were sent */
2975                         osc_extent_finish(env, ext, 0, 0);
2976                 }
2977         }
2978
2979         if (unplug)
2980                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2981
2982         if (hp || discard) {
2983                 int rc;
2984                 rc = osc_cache_wait_range(env, obj, start, end);
2985                 if (result >= 0 && rc < 0)
2986                         result = rc;
2987         }
2988
2989         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2990         RETURN(result);
2991 }
2992
2993 /**
2994  * Returns a list of pages by a given [start, end] of \a obj.
2995  *
2996  * \param resched If not NULL, then we give up before hogging CPU for too
2997  * long and set *resched = 1, in that case caller should implement a retry
2998  * logic.
2999  *
3000  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3001  * crucial in the face of [offset, EOF] locks.
3002  *
3003  * Return at least one page in @queue unless there is no covered page.
3004  */
3005 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3006                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3007                         osc_page_gang_cbt cb, void *cbdata)
3008 {
3009         struct osc_page *ops;
3010         void            **pvec;
3011         pgoff_t         idx;
3012         unsigned int    nr;
3013         unsigned int    i;
3014         unsigned int    j;
3015         int             res = CLP_GANG_OKAY;
3016         bool            tree_lock = true;
3017         ENTRY;
3018
3019         idx = start;
3020         pvec = osc_env_info(env)->oti_pvec;
3021         spin_lock(&osc->oo_tree_lock);
3022         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3023                                             idx, OTI_PVEC_SIZE)) > 0) {
3024                 struct cl_page *page;
3025                 bool end_of_region = false;
3026
3027                 for (i = 0, j = 0; i < nr; ++i) {
3028                         ops = pvec[i];
3029                         pvec[i] = NULL;
3030
3031                         idx = osc_index(ops);
3032                         if (idx > end) {
3033                                 end_of_region = true;
3034                                 break;
3035                         }
3036
3037                         page = ops->ops_cl.cpl_page;
3038                         LASSERT(page->cp_type == CPT_CACHEABLE);
3039                         if (page->cp_state == CPS_FREEING)
3040                                 continue;
3041
3042                         cl_page_get(page);
3043                         lu_ref_add_atomic(&page->cp_reference,
3044                                           "gang_lookup", current);
3045                         pvec[j++] = ops;
3046                 }
3047                 ++idx;
3048
3049                 /*
3050                  * Here a delicate locking dance is performed. Current thread
3051                  * holds a reference to a page, but has to own it before it
3052                  * can be placed into queue. Owning implies waiting, so
3053                  * radix-tree lock is to be released. After a wait one has to
3054                  * check that pages weren't truncated (cl_page_own() returns
3055                  * error in the latter case).
3056                  */
3057                 spin_unlock(&osc->oo_tree_lock);
3058                 tree_lock = false;
3059
3060                 for (i = 0; i < j; ++i) {
3061                         ops = pvec[i];
3062                         if (res == CLP_GANG_OKAY)
3063                                 res = (*cb)(env, io, ops, cbdata);
3064
3065                         page = ops->ops_cl.cpl_page;
3066                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3067                         cl_page_put(env, page);
3068                 }
3069                 if (nr < OTI_PVEC_SIZE || end_of_region)
3070                         break;
3071
3072                 if (res == CLP_GANG_OKAY && need_resched())
3073                         res = CLP_GANG_RESCHED;
3074                 if (res != CLP_GANG_OKAY)
3075                         break;
3076
3077                 spin_lock(&osc->oo_tree_lock);
3078                 tree_lock = true;
3079         }
3080         if (tree_lock)
3081                 spin_unlock(&osc->oo_tree_lock);
3082         RETURN(res);
3083 }
3084
3085 /**
3086  * Check if page @page is covered by an extra lock or discard it.
3087  */
3088 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3089                                 struct osc_page *ops, void *cbdata)
3090 {
3091         struct osc_thread_info *info = osc_env_info(env);
3092         struct osc_object *osc = cbdata;
3093         pgoff_t index;
3094
3095         index = osc_index(ops);
3096         if (index >= info->oti_fn_index) {
3097                 struct ldlm_lock *tmp;
3098                 struct cl_page *page = ops->ops_cl.cpl_page;
3099
3100                 /* refresh non-overlapped index */
3101                 tmp = osc_dlmlock_at_pgoff(env, osc, index, 0, 0);
3102                 if (tmp != NULL) {
3103                         __u64 end = tmp->l_policy_data.l_extent.end;
3104                         /* Cache the first-non-overlapped index so as to skip
3105                          * all pages within [index, oti_fn_index). This is safe
3106                          * because if tmp lock is canceled, it will discard
3107                          * these pages. */
3108                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3109                         if (end == OBD_OBJECT_EOF)
3110                                 info->oti_fn_index = CL_PAGE_EOF;
3111                         LDLM_LOCK_PUT(tmp);
3112                 } else if (cl_page_own(env, io, page) == 0) {
3113                         /* discard the page */
3114                         cl_page_discard(env, io, page);
3115                         cl_page_disown(env, io, page);
3116                 } else {
3117                         LASSERT(page->cp_state == CPS_FREEING);
3118                 }
3119         }
3120
3121         info->oti_next_index = index + 1;
3122         return CLP_GANG_OKAY;
3123 }
3124
3125 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3126                       struct osc_page *ops, void *cbdata)
3127 {
3128         struct osc_thread_info *info = osc_env_info(env);
3129         struct cl_page *page = ops->ops_cl.cpl_page;
3130
3131         /* page is top page. */
3132         info->oti_next_index = osc_index(ops) + 1;
3133         if (cl_page_own(env, io, page) == 0) {
3134                 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3135                               !PageDirty(cl_page_vmpage(page))));
3136
3137                 /* discard the page */
3138                 cl_page_discard(env, io, page);
3139                 cl_page_disown(env, io, page);
3140         } else {
3141                 LASSERT(page->cp_state == CPS_FREEING);
3142         }
3143
3144         return CLP_GANG_OKAY;
3145 }
3146
3147 /**
3148  * Discard pages protected by the given lock. This function traverses radix
3149  * tree to find all covering pages and discard them. If a page is being covered
3150  * by other locks, it should remain in cache.
3151  *
3152  * If error happens on any step, the process continues anyway (the reasoning
3153  * behind this being that lock cancellation cannot be delayed indefinitely).
3154  */
3155 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3156                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3157 {
3158         struct osc_thread_info *info = osc_env_info(env);
3159         struct cl_io *io = &info->oti_io;
3160         osc_page_gang_cbt cb;
3161         int res;
3162         int result;
3163
3164         ENTRY;
3165
3166         io->ci_obj = cl_object_top(osc2cl(osc));
3167         io->ci_ignore_layout = 1;
3168         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3169         if (result != 0)
3170                 GOTO(out, result);
3171
3172         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3173         info->oti_fn_index = info->oti_next_index = start;
3174         do {
3175                 res = osc_page_gang_lookup(env, io, osc,
3176