Whamcloud - gitweb
LU-6271 osc: handle osc eviction correctly
[fs/lustre-release.git] / lustre / osc / osc_cache.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2012, 2014, Intel Corporation.
31  *
32  */
33 /*
34  * This file is part of Lustre, http://www.lustre.org/
35  * Lustre is a trademark of Sun Microsystems, Inc.
36  *
37  * osc cache management.
38  *
39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_OSC
43
44 #include "osc_cl_internal.h"
45 #include "osc_internal.h"
46
47 static int extent_debug; /* set it to be true for more debug */
48
49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
51                            enum osc_extent_state state);
52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
53                               struct osc_async_page *oap, int sent, int rc);
54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
55                           int cmd);
56 static int osc_refresh_count(const struct lu_env *env,
57                              struct osc_async_page *oap, int cmd);
58 static int osc_io_unplug_async(const struct lu_env *env,
59                                struct client_obd *cli, struct osc_object *osc);
60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
61                            unsigned int lost_grant);
62
63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
64                                   const char *func, int line);
65 #define osc_extent_tree_dump(lvl, obj) \
66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
67
68 /** \addtogroup osc
69  *  @{
70  */
71
72 /* ------------------ osc extent ------------------ */
73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
74 {
75         char *buf = flags;
76         *buf++ = ext->oe_rw ? 'r' : 'w';
77         if (ext->oe_intree)
78                 *buf++ = 'i';
79         if (ext->oe_sync)
80                 *buf++ = 'S';
81         if (ext->oe_srvlock)
82                 *buf++ = 's';
83         if (ext->oe_hp)
84                 *buf++ = 'h';
85         if (ext->oe_urgent)
86                 *buf++ = 'u';
87         if (ext->oe_memalloc)
88                 *buf++ = 'm';
89         if (ext->oe_trunc_pending)
90                 *buf++ = 't';
91         if (ext->oe_fsync_wait)
92                 *buf++ = 'Y';
93         *buf = 0;
94         return flags;
95 }
96
97 static inline char list_empty_marker(struct list_head *list)
98 {
99         return list_empty(list) ? '-' : '+';
100 }
101
102 #define EXTSTR       "[%lu -> %lu/%lu]"
103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
104 static const char *oes_strings[] = {
105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
106
107 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
108         struct osc_extent *__ext = (extent);                                  \
109         char __buf[16];                                                       \
110                                                                               \
111         CDEBUG(lvl,                                                           \
112                 "extent %p@{" EXTSTR ", "                                     \
113                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
114                 /* ----- extent part 0 ----- */                               \
115                 __ext, EXTPARA(__ext),                                        \
116                 /* ----- part 1 ----- */                                      \
117                 atomic_read(&__ext->oe_refc),                         \
118                 atomic_read(&__ext->oe_users),                        \
119                 list_empty_marker(&__ext->oe_link),                           \
120                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
121                 __ext->oe_obj,                                                \
122                 /* ----- part 2 ----- */                                      \
123                 __ext->oe_grants, __ext->oe_nr_pages,                         \
124                 list_empty_marker(&__ext->oe_pages),                          \
125                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
126                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
127                 /* ----- part 4 ----- */                                      \
128                 ## __VA_ARGS__);                                              \
129         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
130                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
131         else                                                                  \
132                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
133 } while (0)
134
135 #undef EASSERTF
136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
137         if (!(expr)) {                                                  \
138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
140                 LASSERT(expr);                                          \
141         }                                                               \
142 } while (0)
143
144 #undef EASSERT
145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
146
147 static inline struct osc_extent *rb_extent(struct rb_node *n)
148 {
149         if (n == NULL)
150                 return NULL;
151
152         return container_of(n, struct osc_extent, oe_node);
153 }
154
155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
156 {
157         if (ext == NULL)
158                 return NULL;
159
160         LASSERT(ext->oe_intree);
161         return rb_extent(rb_next(&ext->oe_node));
162 }
163
164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
165 {
166         if (ext == NULL)
167                 return NULL;
168
169         LASSERT(ext->oe_intree);
170         return rb_extent(rb_prev(&ext->oe_node));
171 }
172
173 static inline struct osc_extent *first_extent(struct osc_object *obj)
174 {
175         return rb_extent(rb_first(&obj->oo_root));
176 }
177
178 /* object must be locked by caller. */
179 static int osc_extent_sanity_check0(struct osc_extent *ext,
180                                     const char *func, const int line)
181 {
182         struct osc_object *obj = ext->oe_obj;
183         struct osc_async_page *oap;
184         size_t page_count;
185         int rc = 0;
186
187         if (!osc_object_is_locked(obj))
188                 GOTO(out, rc = 9);
189
190         if (ext->oe_state >= OES_STATE_MAX)
191                 GOTO(out, rc = 10);
192
193         if (atomic_read(&ext->oe_refc) <= 0)
194                 GOTO(out, rc = 20);
195
196         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
197                 GOTO(out, rc = 30);
198
199         switch (ext->oe_state) {
200         case OES_INV:
201                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
202                         GOTO(out, rc = 35);
203                 GOTO(out, rc = 0);
204                 break;
205         case OES_ACTIVE:
206                 if (atomic_read(&ext->oe_users) == 0)
207                         GOTO(out, rc = 40);
208                 if (ext->oe_hp)
209                         GOTO(out, rc = 50);
210                 if (ext->oe_fsync_wait && !ext->oe_urgent)
211                         GOTO(out, rc = 55);
212                 break;
213         case OES_CACHE:
214                 if (ext->oe_grants == 0)
215                         GOTO(out, rc = 60);
216                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
217                         GOTO(out, rc = 65);
218         default:
219                 if (atomic_read(&ext->oe_users) > 0)
220                         GOTO(out, rc = 70);
221         }
222
223         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
224                 GOTO(out, rc = 80);
225
226         if (ext->oe_sync && ext->oe_grants > 0)
227                 GOTO(out, rc = 90);
228
229         if (ext->oe_dlmlock != NULL && !ldlm_is_failed(ext->oe_dlmlock)) {
230                 struct ldlm_extent *extent;
231
232                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
233                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
234                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
235                         GOTO(out, rc = 100);
236
237                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
238                         GOTO(out, rc = 102);
239         }
240
241         if (ext->oe_nr_pages > ext->oe_mppr)
242                 GOTO(out, rc = 105);
243
244         /* Do not verify page list if extent is in RPC. This is because an
245          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
246         if (ext->oe_state > OES_CACHE)
247                 GOTO(out, rc = 0);
248
249         if (!extent_debug)
250                 GOTO(out, rc = 0);
251
252         page_count = 0;
253         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
254                 pgoff_t index = osc_index(oap2osc(oap));
255                 ++page_count;
256                 if (index > ext->oe_end || index < ext->oe_start)
257                         GOTO(out, rc = 110);
258         }
259         if (page_count != ext->oe_nr_pages)
260                 GOTO(out, rc = 120);
261
262 out:
263         if (rc != 0)
264                 OSC_EXTENT_DUMP(D_ERROR, ext,
265                                 "%s:%d sanity check %p failed with rc = %d\n",
266                                 func, line, ext, rc);
267         return rc;
268 }
269
270 #define sanity_check_nolock(ext) \
271         osc_extent_sanity_check0(ext, __func__, __LINE__)
272
273 #define sanity_check(ext) ({                                                   \
274         int __res;                                                             \
275         osc_object_lock((ext)->oe_obj);                                        \
276         __res = sanity_check_nolock(ext);                                      \
277         osc_object_unlock((ext)->oe_obj);                                      \
278         __res;                                                                 \
279 })
280
281
282 /**
283  * sanity check - to make sure there is no overlapped extent in the tree.
284  */
285 static int osc_extent_is_overlapped(struct osc_object *obj,
286                                     struct osc_extent *ext)
287 {
288         struct osc_extent *tmp;
289
290         LASSERT(osc_object_is_locked(obj));
291
292         if (!extent_debug)
293                 return 0;
294
295         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
296                 if (tmp == ext)
297                         continue;
298                 if (tmp->oe_end >= ext->oe_start &&
299                     tmp->oe_start <= ext->oe_end)
300                         return 1;
301         }
302         return 0;
303 }
304
305 static void osc_extent_state_set(struct osc_extent *ext, int state)
306 {
307         LASSERT(osc_object_is_locked(ext->oe_obj));
308         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
309
310         /* Never try to sanity check a state changing extent :-) */
311         /* LASSERT(sanity_check_nolock(ext) == 0); */
312
313         /* TODO: validate the state machine */
314         ext->oe_state = state;
315         wake_up_all(&ext->oe_waitq);
316 }
317
318 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
319 {
320         struct osc_extent *ext;
321
322         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
323         if (ext == NULL)
324                 return NULL;
325
326         RB_CLEAR_NODE(&ext->oe_node);
327         ext->oe_obj = obj;
328         atomic_set(&ext->oe_refc, 1);
329         atomic_set(&ext->oe_users, 0);
330         INIT_LIST_HEAD(&ext->oe_link);
331         ext->oe_state = OES_INV;
332         INIT_LIST_HEAD(&ext->oe_pages);
333         init_waitqueue_head(&ext->oe_waitq);
334         ext->oe_dlmlock = NULL;
335
336         return ext;
337 }
338
339 static void osc_extent_free(struct osc_extent *ext)
340 {
341         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
342 }
343
344 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
345 {
346         LASSERT(atomic_read(&ext->oe_refc) >= 0);
347         atomic_inc(&ext->oe_refc);
348         return ext;
349 }
350
351 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
352 {
353         LASSERT(atomic_read(&ext->oe_refc) > 0);
354         if (atomic_dec_and_test(&ext->oe_refc)) {
355                 LASSERT(list_empty(&ext->oe_link));
356                 LASSERT(atomic_read(&ext->oe_users) == 0);
357                 LASSERT(ext->oe_state == OES_INV);
358                 LASSERT(!ext->oe_intree);
359
360                 if (ext->oe_dlmlock != NULL) {
361                         lu_ref_add(&ext->oe_dlmlock->l_reference,
362                                    "osc_extent", ext);
363                         LDLM_LOCK_PUT(ext->oe_dlmlock);
364                         ext->oe_dlmlock = NULL;
365                 }
366                 osc_extent_free(ext);
367         }
368 }
369
370 /**
371  * osc_extent_put_trust() is a special version of osc_extent_put() when
372  * it's known that the caller is not the last user. This is to address the
373  * problem of lacking of lu_env ;-).
374  */
375 static void osc_extent_put_trust(struct osc_extent *ext)
376 {
377         LASSERT(atomic_read(&ext->oe_refc) > 1);
378         LASSERT(osc_object_is_locked(ext->oe_obj));
379         atomic_dec(&ext->oe_refc);
380 }
381
382 /**
383  * Return the extent which includes pgoff @index, or return the greatest
384  * previous extent in the tree.
385  */
386 static struct osc_extent *osc_extent_search(struct osc_object *obj,
387                                             pgoff_t index)
388 {
389         struct rb_node    *n = obj->oo_root.rb_node;
390         struct osc_extent *tmp, *p = NULL;
391
392         LASSERT(osc_object_is_locked(obj));
393         while (n != NULL) {
394                 tmp = rb_extent(n);
395                 if (index < tmp->oe_start) {
396                         n = n->rb_left;
397                 } else if (index > tmp->oe_end) {
398                         p = rb_extent(n);
399                         n = n->rb_right;
400                 } else {
401                         return tmp;
402                 }
403         }
404         return p;
405 }
406
407 /*
408  * Return the extent covering @index, otherwise return NULL.
409  * caller must have held object lock.
410  */
411 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
412                                             pgoff_t index)
413 {
414         struct osc_extent *ext;
415
416         ext = osc_extent_search(obj, index);
417         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
418                 return osc_extent_get(ext);
419         return NULL;
420 }
421
422 /* caller must have held object lock. */
423 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
424 {
425         struct rb_node   **n      = &obj->oo_root.rb_node;
426         struct rb_node    *parent = NULL;
427         struct osc_extent *tmp;
428
429         LASSERT(ext->oe_intree == 0);
430         LASSERT(ext->oe_obj == obj);
431         LASSERT(osc_object_is_locked(obj));
432         while (*n != NULL) {
433                 tmp = rb_extent(*n);
434                 parent = *n;
435
436                 if (ext->oe_end < tmp->oe_start)
437                         n = &(*n)->rb_left;
438                 else if (ext->oe_start > tmp->oe_end)
439                         n = &(*n)->rb_right;
440                 else
441                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
442         }
443         rb_link_node(&ext->oe_node, parent, n);
444         rb_insert_color(&ext->oe_node, &obj->oo_root);
445         osc_extent_get(ext);
446         ext->oe_intree = 1;
447 }
448
449 /* caller must have held object lock. */
450 static void osc_extent_erase(struct osc_extent *ext)
451 {
452         struct osc_object *obj = ext->oe_obj;
453         LASSERT(osc_object_is_locked(obj));
454         if (ext->oe_intree) {
455                 rb_erase(&ext->oe_node, &obj->oo_root);
456                 ext->oe_intree = 0;
457                 /* rbtree held a refcount */
458                 osc_extent_put_trust(ext);
459         }
460 }
461
462 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
463 {
464         struct osc_object *obj = ext->oe_obj;
465
466         LASSERT(osc_object_is_locked(obj));
467         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
468         if (ext->oe_state == OES_CACHE) {
469                 osc_extent_state_set(ext, OES_ACTIVE);
470                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
471         }
472         atomic_inc(&ext->oe_users);
473         list_del_init(&ext->oe_link);
474         return osc_extent_get(ext);
475 }
476
477 static void __osc_extent_remove(struct osc_extent *ext)
478 {
479         LASSERT(osc_object_is_locked(ext->oe_obj));
480         LASSERT(list_empty(&ext->oe_pages));
481         osc_extent_erase(ext);
482         list_del_init(&ext->oe_link);
483         osc_extent_state_set(ext, OES_INV);
484         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
485 }
486
487 static void osc_extent_remove(struct osc_extent *ext)
488 {
489         struct osc_object *obj = ext->oe_obj;
490
491         osc_object_lock(obj);
492         __osc_extent_remove(ext);
493         osc_object_unlock(obj);
494 }
495
496 /**
497  * This function is used to merge extents to get better performance. It checks
498  * if @cur and @victim are contiguous at chunk level.
499  */
500 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
501                             struct osc_extent *victim)
502 {
503         struct osc_object *obj = cur->oe_obj;
504         pgoff_t chunk_start;
505         pgoff_t chunk_end;
506         int ppc_bits;
507
508         LASSERT(cur->oe_state == OES_CACHE);
509         LASSERT(osc_object_is_locked(obj));
510         if (victim == NULL)
511                 return -EINVAL;
512
513         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
514                 return -EBUSY;
515
516         if (cur->oe_max_end != victim->oe_max_end)
517                 return -ERANGE;
518
519         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
520         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
521         chunk_start = cur->oe_start >> ppc_bits;
522         chunk_end   = cur->oe_end   >> ppc_bits;
523         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
524             chunk_end + 1 != victim->oe_start >> ppc_bits)
525                 return -ERANGE;
526
527         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
528
529         cur->oe_start     = min(cur->oe_start, victim->oe_start);
530         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
531         cur->oe_grants   += victim->oe_grants;
532         cur->oe_nr_pages += victim->oe_nr_pages;
533         /* only the following bits are needed to merge */
534         cur->oe_urgent   |= victim->oe_urgent;
535         cur->oe_memalloc |= victim->oe_memalloc;
536         list_splice_init(&victim->oe_pages, &cur->oe_pages);
537         list_del_init(&victim->oe_link);
538         victim->oe_nr_pages = 0;
539
540         osc_extent_get(victim);
541         __osc_extent_remove(victim);
542         osc_extent_put(env, victim);
543
544         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
545         return 0;
546 }
547
548 /**
549  * Drop user count of osc_extent, and unplug IO asynchronously.
550  */
551 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
552 {
553         struct osc_object *obj = ext->oe_obj;
554         int rc = 0;
555         ENTRY;
556
557         LASSERT(atomic_read(&ext->oe_users) > 0);
558         LASSERT(sanity_check(ext) == 0);
559         LASSERT(ext->oe_grants > 0);
560
561         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
562                 LASSERT(ext->oe_state == OES_ACTIVE);
563                 if (ext->oe_trunc_pending) {
564                         /* a truncate process is waiting for this extent.
565                          * This may happen due to a race, check
566                          * osc_cache_truncate_start(). */
567                         osc_extent_state_set(ext, OES_TRUNC);
568                         ext->oe_trunc_pending = 0;
569                 } else {
570                         osc_extent_state_set(ext, OES_CACHE);
571                         osc_update_pending(obj, OBD_BRW_WRITE,
572                                            ext->oe_nr_pages);
573
574                         /* try to merge the previous and next extent. */
575                         osc_extent_merge(env, ext, prev_extent(ext));
576                         osc_extent_merge(env, ext, next_extent(ext));
577
578                         if (ext->oe_urgent)
579                                 list_move_tail(&ext->oe_link,
580                                                &obj->oo_urgent_exts);
581                 }
582                 osc_object_unlock(obj);
583
584                 osc_io_unplug_async(env, osc_cli(obj), obj);
585         }
586         osc_extent_put(env, ext);
587         RETURN(rc);
588 }
589
590 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
591 {
592         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
593 }
594
595 /**
596  * Find or create an extent which includes @index, core function to manage
597  * extent tree.
598  */
599 static struct osc_extent *osc_extent_find(const struct lu_env *env,
600                                           struct osc_object *obj, pgoff_t index,
601                                           unsigned int *grants)
602 {
603         struct client_obd *cli = osc_cli(obj);
604         struct osc_lock   *olck;
605         struct cl_lock_descr *descr;
606         struct osc_extent *cur;
607         struct osc_extent *ext;
608         struct osc_extent *conflict = NULL;
609         struct osc_extent *found = NULL;
610         pgoff_t    chunk;
611         pgoff_t    max_end;
612         unsigned int max_pages; /* max_pages_per_rpc */
613         unsigned int chunksize;
614         int        ppc_bits; /* pages per chunk bits */
615         pgoff_t    chunk_mask;
616         int        rc;
617         ENTRY;
618
619         cur = osc_extent_alloc(obj);
620         if (cur == NULL)
621                 RETURN(ERR_PTR(-ENOMEM));
622
623         olck = osc_env_io(env)->oi_write_osclock;
624         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
625         LASSERT(olck->ols_state == OLS_GRANTED);
626
627         descr = &olck->ols_cl.cls_lock->cll_descr;
628         LASSERT(descr->cld_mode >= CLM_WRITE);
629
630         LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
631         ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
632         chunk_mask = ~((1 << ppc_bits) - 1);
633         chunksize  = 1 << cli->cl_chunkbits;
634         chunk      = index >> ppc_bits;
635
636         /* align end to rpc edge, rpc size may not be a power 2 integer. */
637         max_pages = cli->cl_max_pages_per_rpc;
638         LASSERT((max_pages & ~chunk_mask) == 0);
639         max_end = index - (index % max_pages) + max_pages - 1;
640         max_end = min_t(pgoff_t, max_end, descr->cld_end);
641
642         /* initialize new extent by parameters so far */
643         cur->oe_max_end = max_end;
644         cur->oe_start   = index & chunk_mask;
645         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
646         if (cur->oe_start < descr->cld_start)
647                 cur->oe_start = descr->cld_start;
648         if (cur->oe_end > max_end)
649                 cur->oe_end = max_end;
650         cur->oe_grants  = 0;
651         cur->oe_mppr    = max_pages;
652         if (olck->ols_dlmlock != NULL) {
653                 LASSERT(olck->ols_hold);
654                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
655                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
656         }
657
658         /* grants has been allocated by caller */
659         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
660                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
661         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
662                  EXTPARA(cur));
663
664 restart:
665         osc_object_lock(obj);
666         ext = osc_extent_search(obj, cur->oe_start);
667         if (ext == NULL)
668                 ext = first_extent(obj);
669         while (ext != NULL) {
670                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
671                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
672
673                 LASSERT(sanity_check_nolock(ext) == 0);
674                 if (chunk > ext_chk_end + 1)
675                         break;
676
677                 /* if covering by different locks, no chance to match */
678                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
679                         EASSERTF(!overlapped(ext, cur), ext,
680                                  EXTSTR"\n", EXTPARA(cur));
681
682                         ext = next_extent(ext);
683                         continue;
684                 }
685
686                 /* discontiguous chunks? */
687                 if (chunk + 1 < ext_chk_start) {
688                         ext = next_extent(ext);
689                         continue;
690                 }
691
692                 /* ok, from now on, ext and cur have these attrs:
693                  * 1. covered by the same lock
694                  * 2. contiguous at chunk level or overlapping. */
695
696                 if (overlapped(ext, cur)) {
697                         /* cur is the minimum unit, so overlapping means
698                          * full contain. */
699                         EASSERTF((ext->oe_start <= cur->oe_start &&
700                                   ext->oe_end >= cur->oe_end),
701                                  ext, EXTSTR"\n", EXTPARA(cur));
702
703                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
704                                 /* for simplicity, we wait for this extent to
705                                  * finish before going forward. */
706                                 conflict = osc_extent_get(ext);
707                                 break;
708                         }
709
710                         found = osc_extent_hold(ext);
711                         break;
712                 }
713
714                 /* non-overlapped extent */
715                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
716                         /* we can't do anything for a non OES_CACHE extent, or
717                          * if there is someone waiting for this extent to be
718                          * flushed, try next one. */
719                         ext = next_extent(ext);
720                         continue;
721                 }
722
723                 /* check if they belong to the same rpc slot before trying to
724                  * merge. the extents are not overlapped and contiguous at
725                  * chunk level to get here. */
726                 if (ext->oe_max_end != max_end) {
727                         /* if they don't belong to the same RPC slot or
728                          * max_pages_per_rpc has ever changed, do not merge. */
729                         ext = next_extent(ext);
730                         continue;
731                 }
732
733                 /* it's required that an extent must be contiguous at chunk
734                  * level so that we know the whole extent is covered by grant
735                  * (the pages in the extent are NOT required to be contiguous).
736                  * Otherwise, it will be too much difficult to know which
737                  * chunks have grants allocated. */
738
739                 /* try to do front merge - extend ext's start */
740                 if (chunk + 1 == ext_chk_start) {
741                         /* ext must be chunk size aligned */
742                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
743
744                         /* pull ext's start back to cover cur */
745                         ext->oe_start   = cur->oe_start;
746                         ext->oe_grants += chunksize;
747                         *grants -= chunksize;
748
749                         found = osc_extent_hold(ext);
750                 } else if (chunk == ext_chk_end + 1) {
751                         /* rear merge */
752                         ext->oe_end     = cur->oe_end;
753                         ext->oe_grants += chunksize;
754                         *grants -= chunksize;
755
756                         /* try to merge with the next one because we just fill
757                          * in a gap */
758                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
759                                 /* we can save extent tax from next extent */
760                                 *grants += cli->cl_extent_tax;
761
762                         found = osc_extent_hold(ext);
763                 }
764                 if (found != NULL)
765                         break;
766
767                 ext = next_extent(ext);
768         }
769
770         osc_extent_tree_dump(D_CACHE, obj);
771         if (found != NULL) {
772                 LASSERT(conflict == NULL);
773                 if (!IS_ERR(found)) {
774                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
775                         OSC_EXTENT_DUMP(D_CACHE, found,
776                                         "found caching ext for %lu.\n", index);
777                 }
778         } else if (conflict == NULL) {
779                 /* create a new extent */
780                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
781                 cur->oe_grants = chunksize + cli->cl_extent_tax;
782                 *grants -= cur->oe_grants;
783                 LASSERT(*grants >= 0);
784
785                 cur->oe_state = OES_CACHE;
786                 found = osc_extent_hold(cur);
787                 osc_extent_insert(obj, cur);
788                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
789                                 index, descr->cld_end);
790         }
791         osc_object_unlock(obj);
792
793         if (conflict != NULL) {
794                 LASSERT(found == NULL);
795
796                 /* waiting for IO to finish. Please notice that it's impossible
797                  * to be an OES_TRUNC extent. */
798                 rc = osc_extent_wait(env, conflict, OES_INV);
799                 osc_extent_put(env, conflict);
800                 conflict = NULL;
801                 if (rc < 0)
802                         GOTO(out, found = ERR_PTR(rc));
803
804                 goto restart;
805         }
806         EXIT;
807
808 out:
809         osc_extent_put(env, cur);
810         LASSERT(*grants >= 0);
811         return found;
812 }
813
814 /**
815  * Called when IO is finished to an extent.
816  */
817 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
818                       int sent, int rc)
819 {
820         struct client_obd *cli = osc_cli(ext->oe_obj);
821         struct osc_async_page *oap;
822         struct osc_async_page *tmp;
823         int nr_pages = ext->oe_nr_pages;
824         int lost_grant = 0;
825         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
826         loff_t last_off = 0;
827         int last_count = -1;
828         ENTRY;
829
830         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
831
832         ext->oe_rc = rc ?: ext->oe_nr_pages;
833         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
834
835         osc_lru_add_batch(cli, &ext->oe_pages);
836         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
837                                      oap_pending_item) {
838                 list_del_init(&oap->oap_rpc_item);
839                 list_del_init(&oap->oap_pending_item);
840                 if (last_off <= oap->oap_obj_off) {
841                         last_off = oap->oap_obj_off;
842                         last_count = oap->oap_count;
843                 }
844
845                 --ext->oe_nr_pages;
846                 osc_ap_completion(env, cli, oap, sent, rc);
847         }
848         EASSERT(ext->oe_nr_pages == 0, ext);
849
850         if (!sent) {
851                 lost_grant = ext->oe_grants;
852         } else if (blocksize < PAGE_CACHE_SIZE &&
853                    last_count != PAGE_CACHE_SIZE) {
854                 /* For short writes we shouldn't count parts of pages that
855                  * span a whole chunk on the OST side, or our accounting goes
856                  * wrong.  Should match the code in filter_grant_check. */
857                 int offset = last_off & ~PAGE_MASK;
858                 int count = last_count + (offset & (blocksize - 1));
859                 int end = (offset + last_count) & (blocksize - 1);
860                 if (end)
861                         count += blocksize - end;
862
863                 lost_grant = PAGE_CACHE_SIZE - count;
864         }
865         if (ext->oe_grants > 0)
866                 osc_free_grant(cli, nr_pages, lost_grant);
867
868         osc_extent_remove(ext);
869         /* put the refcount for RPC */
870         osc_extent_put(env, ext);
871         RETURN(0);
872 }
873
874 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
875 {
876         int ret;
877
878         osc_object_lock(ext->oe_obj);
879         ret = ext->oe_state == state;
880         osc_object_unlock(ext->oe_obj);
881
882         return ret;
883 }
884
885 /**
886  * Wait for the extent's state to become @state.
887  */
888 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
889                            enum osc_extent_state state)
890 {
891         struct osc_object *obj = ext->oe_obj;
892         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
893                                                   LWI_ON_SIGNAL_NOOP, NULL);
894         int rc = 0;
895         ENTRY;
896
897         osc_object_lock(obj);
898         LASSERT(sanity_check_nolock(ext) == 0);
899         /* `Kick' this extent only if the caller is waiting for it to be
900          * written out. */
901         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
902                 if (ext->oe_state == OES_ACTIVE) {
903                         ext->oe_urgent = 1;
904                 } else if (ext->oe_state == OES_CACHE) {
905                         ext->oe_urgent = 1;
906                         osc_extent_hold(ext);
907                         rc = 1;
908                 }
909         }
910         osc_object_unlock(obj);
911         if (rc == 1)
912                 osc_extent_release(env, ext);
913
914         /* wait for the extent until its state becomes @state */
915         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
916         if (rc == -ETIMEDOUT) {
917                 OSC_EXTENT_DUMP(D_ERROR, ext,
918                         "%s: wait ext to %u timedout, recovery in progress?\n",
919                         cli_name(osc_cli(obj)), state);
920
921                 lwi = LWI_INTR(NULL, NULL);
922                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
923                                   &lwi);
924         }
925         if (rc == 0 && ext->oe_rc < 0)
926                 rc = ext->oe_rc;
927         RETURN(rc);
928 }
929
930 /**
931  * Discard pages with index greater than @size. If @ext is overlapped with
932  * @size, then partial truncate happens.
933  */
934 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
935                                 bool partial)
936 {
937         struct cl_env_nest     nest;
938         struct lu_env         *env;
939         struct cl_io          *io;
940         struct osc_object     *obj = ext->oe_obj;
941         struct client_obd     *cli = osc_cli(obj);
942         struct osc_async_page *oap;
943         struct osc_async_page *tmp;
944         int                    pages_in_chunk = 0;
945         int                    ppc_bits    = cli->cl_chunkbits -
946                                              PAGE_CACHE_SHIFT;
947         __u64                  trunc_chunk = trunc_index >> ppc_bits;
948         int                    grants   = 0;
949         int                    nr_pages = 0;
950         int                    rc       = 0;
951         ENTRY;
952
953         LASSERT(sanity_check(ext) == 0);
954         LASSERT(ext->oe_state == OES_TRUNC);
955         LASSERT(!ext->oe_urgent);
956
957         /* Request new lu_env.
958          * We can't use that env from osc_cache_truncate_start() because
959          * it's from lov_io_sub and not fully initialized. */
960         env = cl_env_nested_get(&nest);
961         io  = &osc_env_info(env)->oti_io;
962         io->ci_obj = cl_object_top(osc2cl(obj));
963         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
964         if (rc < 0)
965                 GOTO(out, rc);
966
967         /* discard all pages with index greater than trunc_index */
968         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
969                                      oap_pending_item) {
970                 pgoff_t index = osc_index(oap2osc(oap));
971                 struct cl_page  *page = oap2cl_page(oap);
972
973                 LASSERT(list_empty(&oap->oap_rpc_item));
974
975                 /* only discard the pages with their index greater than
976                  * trunc_index, and ... */
977                 if (index < trunc_index ||
978                     (index == trunc_index && partial)) {
979                         /* accounting how many pages remaining in the chunk
980                          * so that we can calculate grants correctly. */
981                         if (index >> ppc_bits == trunc_chunk)
982                                 ++pages_in_chunk;
983                         continue;
984                 }
985
986                 list_del_init(&oap->oap_pending_item);
987
988                 cl_page_get(page);
989                 lu_ref_add(&page->cp_reference, "truncate", current);
990
991                 if (cl_page_own(env, io, page) == 0) {
992                         cl_page_discard(env, io, page);
993                         cl_page_disown(env, io, page);
994                 } else {
995                         LASSERT(page->cp_state == CPS_FREEING);
996                         LASSERT(0);
997                 }
998
999                 lu_ref_del(&page->cp_reference, "truncate", current);
1000                 cl_page_put(env, page);
1001
1002                 --ext->oe_nr_pages;
1003                 ++nr_pages;
1004         }
1005         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1006                       ext->oe_nr_pages == 0),
1007                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1008
1009         osc_object_lock(obj);
1010         if (ext->oe_nr_pages == 0) {
1011                 LASSERT(pages_in_chunk == 0);
1012                 grants = ext->oe_grants;
1013                 ext->oe_grants = 0;
1014         } else { /* calculate how many grants we can free */
1015                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1016                 pgoff_t last_index;
1017
1018
1019                 /* if there is no pages in this chunk, we can also free grants
1020                  * for the last chunk */
1021                 if (pages_in_chunk == 0) {
1022                         /* if this is the 1st chunk and no pages in this chunk,
1023                          * ext->oe_nr_pages must be zero, so we should be in
1024                          * the other if-clause. */
1025                         LASSERT(trunc_chunk > 0);
1026                         --trunc_chunk;
1027                         ++chunks;
1028                 }
1029
1030                 /* this is what we can free from this extent */
1031                 grants          = chunks << cli->cl_chunkbits;
1032                 ext->oe_grants -= grants;
1033                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1034                 ext->oe_end     = min(last_index, ext->oe_max_end);
1035                 LASSERT(ext->oe_end >= ext->oe_start);
1036                 LASSERT(ext->oe_grants > 0);
1037         }
1038         osc_object_unlock(obj);
1039
1040         if (grants > 0 || nr_pages > 0)
1041                 osc_free_grant(cli, nr_pages, grants);
1042
1043 out:
1044         cl_io_fini(env, io);
1045         cl_env_nested_put(&nest, env);
1046         RETURN(rc);
1047 }
1048
1049 /**
1050  * This function is used to make the extent prepared for transfer.
1051  * A race with flusing page - ll_writepage() has to be handled cautiously.
1052  */
1053 static int osc_extent_make_ready(const struct lu_env *env,
1054                                  struct osc_extent *ext)
1055 {
1056         struct osc_async_page *oap;
1057         struct osc_async_page *last = NULL;
1058         struct osc_object *obj = ext->oe_obj;
1059         unsigned int page_count = 0;
1060         int rc;
1061         ENTRY;
1062
1063         /* we're going to grab page lock, so object lock must not be taken. */
1064         LASSERT(sanity_check(ext) == 0);
1065         /* in locking state, any process should not touch this extent. */
1066         EASSERT(ext->oe_state == OES_LOCKING, ext);
1067         EASSERT(ext->oe_owner != NULL, ext);
1068
1069         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1070
1071         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1072                 ++page_count;
1073                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1074                         last = oap;
1075
1076                 /* checking ASYNC_READY is race safe */
1077                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1078                         continue;
1079
1080                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1081                 switch (rc) {
1082                 case 0:
1083                         spin_lock(&oap->oap_lock);
1084                         oap->oap_async_flags |= ASYNC_READY;
1085                         spin_unlock(&oap->oap_lock);
1086                         break;
1087                 case -EALREADY:
1088                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1089                         break;
1090                 default:
1091                         LASSERTF(0, "unknown return code: %d\n", rc);
1092                 }
1093         }
1094
1095         LASSERT(page_count == ext->oe_nr_pages);
1096         LASSERT(last != NULL);
1097         /* the last page is the only one we need to refresh its count by
1098          * the size of file. */
1099         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1100                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1101                 LASSERT(last_oap_count > 0);
1102                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_CACHE_SIZE);
1103                 last->oap_count = last_oap_count;
1104                 spin_lock(&last->oap_lock);
1105                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1106                 spin_unlock(&last->oap_lock);
1107         }
1108
1109         /* for the rest of pages, we don't need to call osf_refresh_count()
1110          * because it's known they are not the last page */
1111         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1112                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1113                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1114                         spin_lock(&oap->oap_lock);
1115                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1116                         spin_unlock(&oap->oap_lock);
1117                 }
1118         }
1119
1120         osc_object_lock(obj);
1121         osc_extent_state_set(ext, OES_RPC);
1122         osc_object_unlock(obj);
1123         /* get a refcount for RPC. */
1124         osc_extent_get(ext);
1125
1126         RETURN(0);
1127 }
1128
1129 /**
1130  * Quick and simple version of osc_extent_find(). This function is frequently
1131  * called to expand the extent for the same IO. To expand the extent, the
1132  * page index must be in the same or next chunk of ext->oe_end.
1133  */
1134 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1135                              unsigned int *grants)
1136 {
1137         struct osc_object *obj = ext->oe_obj;
1138         struct client_obd *cli = osc_cli(obj);
1139         struct osc_extent *next;
1140         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1141         pgoff_t chunk = index >> ppc_bits;
1142         pgoff_t end_chunk;
1143         pgoff_t end_index;
1144         unsigned int chunksize = 1 << cli->cl_chunkbits;
1145         int rc = 0;
1146         ENTRY;
1147
1148         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1149         osc_object_lock(obj);
1150         LASSERT(sanity_check_nolock(ext) == 0);
1151         end_chunk = ext->oe_end >> ppc_bits;
1152         if (chunk > end_chunk + 1)
1153                 GOTO(out, rc = -ERANGE);
1154
1155         if (end_chunk >= chunk)
1156                 GOTO(out, rc = 0);
1157
1158         LASSERT(end_chunk + 1 == chunk);
1159         /* try to expand this extent to cover @index */
1160         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1161
1162         next = next_extent(ext);
1163         if (next != NULL && next->oe_start <= end_index)
1164                 /* complex mode - overlapped with the next extent,
1165                  * this case will be handled by osc_extent_find() */
1166                 GOTO(out, rc = -EAGAIN);
1167
1168         ext->oe_end = end_index;
1169         ext->oe_grants += chunksize;
1170         *grants -= chunksize;
1171         LASSERT(*grants >= 0);
1172         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1173                  "overlapped after expanding for %lu.\n", index);
1174         EXIT;
1175
1176 out:
1177         osc_object_unlock(obj);
1178         RETURN(rc);
1179 }
1180
1181 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1182                                   const char *func, int line)
1183 {
1184         struct osc_extent *ext;
1185         int cnt;
1186
1187         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1188                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1189
1190         /* osc_object_lock(obj); */
1191         cnt = 1;
1192         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1193                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1194
1195         cnt = 1;
1196         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1197                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1198
1199         cnt = 1;
1200         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1201                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1202
1203         cnt = 1;
1204         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1205                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1206         /* osc_object_unlock(obj); */
1207 }
1208
1209 /* ------------------ osc extent end ------------------ */
1210
1211 static inline int osc_is_ready(struct osc_object *osc)
1212 {
1213         return !list_empty(&osc->oo_ready_item) ||
1214                !list_empty(&osc->oo_hp_ready_item);
1215 }
1216
1217 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1218         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1219                (OSC), osc_is_ready(OSC),                                       \
1220                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1221                list_empty_marker(&(OSC)->oo_ready_item),                       \
1222                atomic_read(&(OSC)->oo_nr_writes),                              \
1223                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1224                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1225                atomic_read(&(OSC)->oo_nr_reads),                               \
1226                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1227                ##args)
1228
1229 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1230                           int cmd)
1231 {
1232         struct osc_page *opg  = oap2osc_page(oap);
1233         struct cl_page  *page = oap2cl_page(oap);
1234         int result;
1235
1236         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1237
1238         ENTRY;
1239         result = cl_page_make_ready(env, page, CRT_WRITE);
1240         if (result == 0)
1241                 opg->ops_submit_time = cfs_time_current();
1242         RETURN(result);
1243 }
1244
1245 static int osc_refresh_count(const struct lu_env *env,
1246                              struct osc_async_page *oap, int cmd)
1247 {
1248         struct osc_page  *opg = oap2osc_page(oap);
1249         pgoff_t index = osc_index(oap2osc(oap));
1250         struct cl_object *obj;
1251         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1252
1253         int result;
1254         loff_t kms;
1255
1256         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1257         LASSERT(!(cmd & OBD_BRW_READ));
1258         LASSERT(opg != NULL);
1259         obj = opg->ops_cl.cpl_obj;
1260
1261         cl_object_attr_lock(obj);
1262         result = cl_object_attr_get(env, obj, attr);
1263         cl_object_attr_unlock(obj);
1264         if (result < 0)
1265                 return result;
1266         kms = attr->cat_kms;
1267         if (cl_offset(obj, index) >= kms)
1268                 /* catch race with truncate */
1269                 return 0;
1270         else if (cl_offset(obj, index + 1) > kms)
1271                 /* catch sub-page write at end of file */
1272                 return kms % PAGE_CACHE_SIZE;
1273         else
1274                 return PAGE_CACHE_SIZE;
1275 }
1276
1277 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1278                           int cmd, int rc)
1279 {
1280         struct osc_page   *opg  = oap2osc_page(oap);
1281         struct cl_page    *page = oap2cl_page(oap);
1282         enum cl_req_type   crt;
1283         int srvlock;
1284
1285         ENTRY;
1286
1287         cmd &= ~OBD_BRW_NOQUOTA;
1288         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1289                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1290         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1291                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1292         LASSERT(opg->ops_transfer_pinned);
1293
1294         /*
1295          * page->cp_req can be NULL if io submission failed before
1296          * cl_req was allocated.
1297          */
1298         if (page->cp_req != NULL)
1299                 cl_req_page_done(env, page);
1300         LASSERT(page->cp_req == NULL);
1301
1302         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1303         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1304         opg->ops_transfer_pinned = 0;
1305
1306         opg->ops_submit_time = 0;
1307         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1308
1309         /* statistic */
1310         if (rc == 0 && srvlock) {
1311                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1312                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1313                 size_t bytes = oap->oap_count;
1314
1315                 if (crt == CRT_READ)
1316                         stats->os_lockless_reads += bytes;
1317                 else
1318                         stats->os_lockless_writes += bytes;
1319         }
1320
1321         /*
1322          * This has to be the last operation with the page, as locks are
1323          * released in cl_page_completion() and nothing except for the
1324          * reference counter protects page from concurrent reclaim.
1325          */
1326         lu_ref_del(&page->cp_reference, "transfer", page);
1327
1328         cl_page_completion(env, page, crt, rc);
1329
1330         RETURN(0);
1331 }
1332
1333 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1334         struct client_obd *__tmp = (cli);                               \
1335         CDEBUG(lvl, "%s: grant { dirty: %lu/%lu dirty_pages: %ld/%lu "  \
1336                "dropped: %ld avail: %ld, reserved: %ld, flight: %d }"   \
1337                "lru {in list: %ld, left: %ld, waiters: %d }"fmt"\n",    \
1338                cli_name(__tmp),                                         \
1339                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1340                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1341                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1342                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1343                atomic_long_read(&__tmp->cl_lru_in_list),                \
1344                atomic_long_read(&__tmp->cl_lru_busy),                   \
1345                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1346 } while (0)
1347
1348 /* caller must hold loi_list_lock */
1349 static void osc_consume_write_grant(struct client_obd *cli,
1350                                     struct brw_page *pga)
1351 {
1352         assert_spin_locked(&cli->cl_loi_list_lock);
1353         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1354         atomic_long_inc(&obd_dirty_pages);
1355         cli->cl_dirty_pages++;
1356         pga->flag |= OBD_BRW_FROM_GRANT;
1357         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1358                PAGE_CACHE_SIZE, pga, pga->pg);
1359         osc_update_next_shrink(cli);
1360 }
1361
1362 /* the companion to osc_consume_write_grant, called when a brw has completed.
1363  * must be called with the loi lock held. */
1364 static void osc_release_write_grant(struct client_obd *cli,
1365                                     struct brw_page *pga)
1366 {
1367         ENTRY;
1368
1369         assert_spin_locked(&cli->cl_loi_list_lock);
1370         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1371                 EXIT;
1372                 return;
1373         }
1374
1375         pga->flag &= ~OBD_BRW_FROM_GRANT;
1376         atomic_long_dec(&obd_dirty_pages);
1377         cli->cl_dirty_pages--;
1378         if (pga->flag & OBD_BRW_NOCACHE) {
1379                 pga->flag &= ~OBD_BRW_NOCACHE;
1380                 atomic_long_dec(&obd_dirty_transit_pages);
1381                 cli->cl_dirty_transit--;
1382         }
1383         EXIT;
1384 }
1385
1386 /**
1387  * To avoid sleeping with object lock held, it's good for us allocate enough
1388  * grants before entering into critical section.
1389  *
1390  * client_obd_list_lock held by caller
1391  */
1392 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1393 {
1394         int rc = -EDQUOT;
1395
1396         if (cli->cl_avail_grant >= bytes) {
1397                 cli->cl_avail_grant    -= bytes;
1398                 cli->cl_reserved_grant += bytes;
1399                 rc = 0;
1400         }
1401         return rc;
1402 }
1403
1404 static void __osc_unreserve_grant(struct client_obd *cli,
1405                                   unsigned int reserved, unsigned int unused)
1406 {
1407         /* it's quite normal for us to get more grant than reserved.
1408          * Thinking about a case that two extents merged by adding a new
1409          * chunk, we can save one extent tax. If extent tax is greater than
1410          * one chunk, we can save more grant by adding a new chunk */
1411         cli->cl_reserved_grant -= reserved;
1412         if (unused > reserved) {
1413                 cli->cl_avail_grant += reserved;
1414                 cli->cl_lost_grant  += unused - reserved;
1415         } else {
1416                 cli->cl_avail_grant += unused;
1417         }
1418 }
1419
1420 static void osc_unreserve_grant(struct client_obd *cli,
1421                                 unsigned int reserved, unsigned int unused)
1422 {
1423         spin_lock(&cli->cl_loi_list_lock);
1424         __osc_unreserve_grant(cli, reserved, unused);
1425         if (unused > 0)
1426                 osc_wake_cache_waiters(cli);
1427         spin_unlock(&cli->cl_loi_list_lock);
1428 }
1429
1430 /**
1431  * Free grant after IO is finished or canceled.
1432  *
1433  * @lost_grant is used to remember how many grants we have allocated but not
1434  * used, we should return these grants to OST. There're two cases where grants
1435  * can be lost:
1436  * 1. truncate;
1437  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1438  *    written. In this case OST may use less chunks to serve this partial
1439  *    write. OSTs don't actually know the page size on the client side. so
1440  *    clients have to calculate lost grant by the blocksize on the OST.
1441  *    See filter_grant_check() for details.
1442  */
1443 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1444                            unsigned int lost_grant)
1445 {
1446         unsigned long grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1447
1448         spin_lock(&cli->cl_loi_list_lock);
1449         atomic_long_sub(nr_pages, &obd_dirty_pages);
1450         cli->cl_dirty_pages -= nr_pages;
1451         cli->cl_lost_grant += lost_grant;
1452         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1453                 /* borrow some grant from truncate to avoid the case that
1454                  * truncate uses up all avail grant */
1455                 cli->cl_lost_grant -= grant;
1456                 cli->cl_avail_grant += grant;
1457         }
1458         osc_wake_cache_waiters(cli);
1459         spin_unlock(&cli->cl_loi_list_lock);
1460         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1461                lost_grant, cli->cl_lost_grant,
1462                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
1463 }
1464
1465 /**
1466  * The companion to osc_enter_cache(), called when @oap is no longer part of
1467  * the dirty accounting due to error.
1468  */
1469 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1470 {
1471         spin_lock(&cli->cl_loi_list_lock);
1472         osc_release_write_grant(cli, &oap->oap_brw_page);
1473         spin_unlock(&cli->cl_loi_list_lock);
1474 }
1475
1476 /**
1477  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1478  * is available.
1479  */
1480 static int osc_enter_cache_try(struct client_obd *cli,
1481                                struct osc_async_page *oap,
1482                                int bytes, int transient)
1483 {
1484         int rc;
1485
1486         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1487
1488         rc = osc_reserve_grant(cli, bytes);
1489         if (rc < 0)
1490                 return 0;
1491
1492         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1493             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1494                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1495                 if (transient) {
1496                         cli->cl_dirty_transit++;
1497                         atomic_long_inc(&obd_dirty_transit_pages);
1498                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1499                 }
1500                 rc = 1;
1501         } else {
1502                 __osc_unreserve_grant(cli, bytes, bytes);
1503                 rc = 0;
1504         }
1505         return rc;
1506 }
1507
1508 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1509 {
1510         int rc;
1511         spin_lock(&cli->cl_loi_list_lock);
1512         rc = list_empty(&ocw->ocw_entry);
1513         spin_unlock(&cli->cl_loi_list_lock);
1514         return rc;
1515 }
1516
1517 /**
1518  * The main entry to reserve dirty page accounting. Usually the grant reserved
1519  * in this function will be freed in bulk in osc_free_grant() unless it fails
1520  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1521  *
1522  * The process will be put into sleep if it's already run out of grant.
1523  */
1524 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1525                            struct osc_async_page *oap, int bytes)
1526 {
1527         struct osc_object       *osc = oap->oap_obj;
1528         struct lov_oinfo        *loi = osc->oo_oinfo;
1529         struct osc_cache_waiter  ocw;
1530         struct l_wait_info       lwi;
1531         int                      rc = -EDQUOT;
1532         ENTRY;
1533
1534         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1535                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1536
1537         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1538
1539         spin_lock(&cli->cl_loi_list_lock);
1540
1541         /* force the caller to try sync io.  this can jump the list
1542          * of queued writes and create a discontiguous rpc stream */
1543         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1544             cli->cl_dirty_max_pages == 0 ||
1545             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1546                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1547                 GOTO(out, rc = -EDQUOT);
1548         }
1549
1550         /* Hopefully normal case - cache space and write credits available */
1551         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1552                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1553                 GOTO(out, rc = 0);
1554         }
1555
1556         /* We can get here for two reasons: too many dirty pages in cache, or
1557          * run out of grants. In both cases we should write dirty pages out.
1558          * Adding a cache waiter will trigger urgent write-out no matter what
1559          * RPC size will be.
1560          * The exiting condition is no avail grants and no dirty pages caching,
1561          * that really means there is no space on the OST. */
1562         init_waitqueue_head(&ocw.ocw_waitq);
1563         ocw.ocw_oap   = oap;
1564         ocw.ocw_grant = bytes;
1565         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1566                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1567                 ocw.ocw_rc = 0;
1568                 spin_unlock(&cli->cl_loi_list_lock);
1569
1570                 osc_io_unplug_async(env, cli, NULL);
1571
1572                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1573                        cli_name(cli), &ocw, oap);
1574
1575                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1576
1577                 spin_lock(&cli->cl_loi_list_lock);
1578
1579                 if (rc < 0) {
1580                         /* l_wait_event is interrupted by signal or timed out */
1581                         list_del_init(&ocw.ocw_entry);
1582                         break;
1583                 }
1584                 LASSERT(list_empty(&ocw.ocw_entry));
1585                 rc = ocw.ocw_rc;
1586
1587                 if (rc != -EDQUOT)
1588                         break;
1589                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1590                         rc = 0;
1591                         break;
1592                 }
1593         }
1594
1595         switch (rc) {
1596         case 0:
1597                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1598                 break;
1599         case -ETIMEDOUT:
1600                 OSC_DUMP_GRANT(D_CACHE, cli,
1601                                "timeout, fall back to sync i/o\n");
1602                 osc_extent_tree_dump(D_CACHE, osc);
1603                 /* fall back to synchronous I/O */
1604                 rc = -EDQUOT;
1605                 break;
1606         case -EINTR:
1607                 /* Ensures restartability - LU-3581 */
1608                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1609                 rc = -ERESTARTSYS;
1610                 break;
1611         case -EDQUOT:
1612                 OSC_DUMP_GRANT(D_CACHE, cli,
1613                                "no grant space, fall back to sync i/o\n");
1614                 break;
1615         default:
1616                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1617                        "due to %d, fall back to sync i/o\n",
1618                        cli_name(cli), &ocw, rc);
1619                 break;
1620         }
1621         EXIT;
1622 out:
1623         spin_unlock(&cli->cl_loi_list_lock);
1624         RETURN(rc);
1625 }
1626
1627 /* caller must hold loi_list_lock */
1628 void osc_wake_cache_waiters(struct client_obd *cli)
1629 {
1630         struct list_head *l, *tmp;
1631         struct osc_cache_waiter *ocw;
1632
1633         ENTRY;
1634         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1635                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1636                 list_del_init(&ocw->ocw_entry);
1637
1638                 ocw->ocw_rc = -EDQUOT;
1639                 /* we can't dirty more */
1640                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1641                     (1 + atomic_long_read(&obd_dirty_pages) >
1642                      obd_max_dirty_pages)) {
1643                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1644                                "osc max %ld, sys max %ld\n",
1645                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1646                                obd_max_dirty_pages);
1647                         goto wakeup;
1648                 }
1649
1650                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1651                         ocw->ocw_rc = 0;
1652 wakeup:
1653                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1654                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1655
1656                 wake_up(&ocw->ocw_waitq);
1657         }
1658
1659         EXIT;
1660 }
1661
1662 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1663 {
1664         int hprpc = !!list_empty(&osc->oo_hp_exts);
1665         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1666 }
1667
1668 /* This maintains the lists of pending pages to read/write for a given object
1669  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1670  * to quickly find objects that are ready to send an RPC. */
1671 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1672                          int cmd)
1673 {
1674         int invalid_import = 0;
1675         ENTRY;
1676
1677         /* if we have an invalid import we want to drain the queued pages
1678          * by forcing them through rpcs that immediately fail and complete
1679          * the pages.  recovery relies on this to empty the queued pages
1680          * before canceling the locks and evicting down the llite pages */
1681         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1682                 invalid_import = 1;
1683
1684         if (cmd & OBD_BRW_WRITE) {
1685                 if (atomic_read(&osc->oo_nr_writes) == 0)
1686                         RETURN(0);
1687                 if (invalid_import) {
1688                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1689                         RETURN(1);
1690                 }
1691                 if (!list_empty(&osc->oo_hp_exts)) {
1692                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1693                         RETURN(1);
1694                 }
1695                 if (!list_empty(&osc->oo_urgent_exts)) {
1696                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1697                         RETURN(1);
1698                 }
1699                 /* trigger a write rpc stream as long as there are dirtiers
1700                  * waiting for space.  as they're waiting, they're not going to
1701                  * create more pages to coalesce with what's waiting.. */
1702                 if (!list_empty(&cli->cl_cache_waiters)) {
1703                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1704                         RETURN(1);
1705                 }
1706                 if (atomic_read(&osc->oo_nr_writes) >=
1707                     cli->cl_max_pages_per_rpc)
1708                         RETURN(1);
1709         } else {
1710                 if (atomic_read(&osc->oo_nr_reads) == 0)
1711                         RETURN(0);
1712                 if (invalid_import) {
1713                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1714                         RETURN(1);
1715                 }
1716                 /* all read are urgent. */
1717                 if (!list_empty(&osc->oo_reading_exts))
1718                         RETURN(1);
1719         }
1720
1721         RETURN(0);
1722 }
1723
1724 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1725 {
1726         struct client_obd *cli = osc_cli(obj);
1727         if (cmd & OBD_BRW_WRITE) {
1728                 atomic_add(delta, &obj->oo_nr_writes);
1729                 atomic_add(delta, &cli->cl_pending_w_pages);
1730                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1731         } else {
1732                 atomic_add(delta, &obj->oo_nr_reads);
1733                 atomic_add(delta, &cli->cl_pending_r_pages);
1734                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1735         }
1736         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1737 }
1738
1739 static int osc_makes_hprpc(struct osc_object *obj)
1740 {
1741         return !list_empty(&obj->oo_hp_exts);
1742 }
1743
1744 static void on_list(struct list_head *item, struct list_head *list,
1745                     int should_be_on)
1746 {
1747         if (list_empty(item) && should_be_on)
1748                 list_add_tail(item, list);
1749         else if (!list_empty(item) && !should_be_on)
1750                 list_del_init(item);
1751 }
1752
1753 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1754  * can find pages to build into rpcs quickly */
1755 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1756 {
1757         if (osc_makes_hprpc(osc)) {
1758                 /* HP rpc */
1759                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1760                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1761         } else {
1762                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1763                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1764                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1765                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1766         }
1767
1768         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1769                 atomic_read(&osc->oo_nr_writes) > 0);
1770
1771         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1772                 atomic_read(&osc->oo_nr_reads) > 0);
1773
1774         return osc_is_ready(osc);
1775 }
1776
1777 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1778 {
1779         int is_ready;
1780
1781         spin_lock(&cli->cl_loi_list_lock);
1782         is_ready = __osc_list_maint(cli, osc);
1783         spin_unlock(&cli->cl_loi_list_lock);
1784
1785         return is_ready;
1786 }
1787
1788 /* this is trying to propogate async writeback errors back up to the
1789  * application.  As an async write fails we record the error code for later if
1790  * the app does an fsync.  As long as errors persist we force future rpcs to be
1791  * sync so that the app can get a sync error and break the cycle of queueing
1792  * pages for which writeback will fail. */
1793 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1794                            int rc)
1795 {
1796         if (rc) {
1797                 if (!ar->ar_rc)
1798                         ar->ar_rc = rc;
1799
1800                 ar->ar_force_sync = 1;
1801                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1802                 return;
1803
1804         }
1805
1806         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1807                 ar->ar_force_sync = 0;
1808 }
1809
1810 /* this must be called holding the loi list lock to give coverage to exit_cache,
1811  * async_flag maintenance, and oap_request */
1812 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1813                               struct osc_async_page *oap, int sent, int rc)
1814 {
1815         struct osc_object *osc = oap->oap_obj;
1816         struct lov_oinfo  *loi = osc->oo_oinfo;
1817         __u64 xid = 0;
1818
1819         ENTRY;
1820         if (oap->oap_request != NULL) {
1821                 xid = ptlrpc_req_xid(oap->oap_request);
1822                 ptlrpc_req_finished(oap->oap_request);
1823                 oap->oap_request = NULL;
1824         }
1825
1826         /* As the transfer for this page is being done, clear the flags */
1827         spin_lock(&oap->oap_lock);
1828         oap->oap_async_flags = 0;
1829         spin_unlock(&oap->oap_lock);
1830         oap->oap_interrupted = 0;
1831
1832         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1833                 spin_lock(&cli->cl_loi_list_lock);
1834                 osc_process_ar(&cli->cl_ar, xid, rc);
1835                 osc_process_ar(&loi->loi_ar, xid, rc);
1836                 spin_unlock(&cli->cl_loi_list_lock);
1837         }
1838
1839         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1840         if (rc)
1841                 CERROR("completion on oap %p obj %p returns %d.\n",
1842                        oap, osc, rc);
1843
1844         EXIT;
1845 }
1846
1847 /**
1848  * Try to add extent to one RPC. We need to think about the following things:
1849  * - # of pages must not be over max_pages_per_rpc
1850  * - extent must be compatible with previous ones
1851  */
1852 static int try_to_add_extent_for_io(struct client_obd *cli,
1853                                     struct osc_extent *ext,
1854                                     struct list_head *rpclist,
1855                                     unsigned int *pc, unsigned int *max_pages)
1856 {
1857         struct osc_extent *tmp;
1858         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1859                                                       struct osc_async_page,
1860                                                       oap_pending_item);
1861         ENTRY;
1862
1863         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1864                 ext);
1865
1866         *max_pages = max(ext->oe_mppr, *max_pages);
1867         if (*pc + ext->oe_nr_pages > *max_pages)
1868                 RETURN(0);
1869
1870         list_for_each_entry(tmp, rpclist, oe_link) {
1871                 struct osc_async_page *oap2;
1872                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1873                                         oap_pending_item);
1874                 EASSERT(tmp->oe_owner == current, tmp);
1875 #if 0
1876                 if (overlapped(tmp, ext)) {
1877                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1878                         EASSERT(0, ext);
1879                 }
1880 #endif
1881                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1882                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1883                                         " for a same RPC\n");
1884                         RETURN(0);
1885                 }
1886
1887                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1888                     !tmp->oe_grants != !ext->oe_grants)
1889                         RETURN(0);
1890
1891                 /* remove break for strict check */
1892                 break;
1893         }
1894
1895         *pc += ext->oe_nr_pages;
1896         list_move_tail(&ext->oe_link, rpclist);
1897         ext->oe_owner = current;
1898         RETURN(1);
1899 }
1900
1901 /**
1902  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1903  * get_write_extent() takes all appropriate extents in atomic.
1904  *
1905  * The following policy is used to collect extents for IO:
1906  * 1. Add as many HP extents as possible;
1907  * 2. Add the first urgent extent in urgent extent list and take it out of
1908  *    urgent list;
1909  * 3. Add subsequent extents of this urgent extent;
1910  * 4. If urgent list is not empty, goto 2;
1911  * 5. Traverse the extent tree from the 1st extent;
1912  * 6. Above steps exit if there is no space in this RPC.
1913  */
1914 static unsigned int get_write_extents(struct osc_object *obj,
1915                                       struct list_head *rpclist)
1916 {
1917         struct client_obd *cli = osc_cli(obj);
1918         struct osc_extent *ext;
1919         unsigned int page_count = 0;
1920         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1921
1922         LASSERT(osc_object_is_locked(obj));
1923         while (!list_empty(&obj->oo_hp_exts)) {
1924                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1925                                  oe_link);
1926                 LASSERT(ext->oe_state == OES_CACHE);
1927                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1928                                               &max_pages))
1929                         return page_count;
1930                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1931         }
1932         if (page_count == max_pages)
1933                 return page_count;
1934
1935         while (!list_empty(&obj->oo_urgent_exts)) {
1936                 ext = list_entry(obj->oo_urgent_exts.next,
1937                                  struct osc_extent, oe_link);
1938                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1939                                               &max_pages))
1940                         return page_count;
1941
1942                 if (!ext->oe_intree)
1943                         continue;
1944
1945                 while ((ext = next_extent(ext)) != NULL) {
1946                         if ((ext->oe_state != OES_CACHE) ||
1947                             (!list_empty(&ext->oe_link) &&
1948                              ext->oe_owner != NULL))
1949                                 continue;
1950
1951                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1952                                                       &page_count, &max_pages))
1953                                 return page_count;
1954                 }
1955         }
1956         if (page_count == max_pages)
1957                 return page_count;
1958
1959         ext = first_extent(obj);
1960         while (ext != NULL) {
1961                 if ((ext->oe_state != OES_CACHE) ||
1962                     /* this extent may be already in current rpclist */
1963                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1964                         ext = next_extent(ext);
1965                         continue;
1966                 }
1967
1968                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1969                                               &max_pages))
1970                         return page_count;
1971
1972                 ext = next_extent(ext);
1973         }
1974         return page_count;
1975 }
1976
1977 static int
1978 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1979                    struct osc_object *osc)
1980 __must_hold(osc)
1981 {
1982         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
1983         struct osc_extent *ext;
1984         struct osc_extent *tmp;
1985         struct osc_extent *first = NULL;
1986         unsigned int page_count = 0;
1987         int srvlock = 0;
1988         int rc = 0;
1989         ENTRY;
1990
1991         LASSERT(osc_object_is_locked(osc));
1992
1993         page_count = get_write_extents(osc, &rpclist);
1994         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1995
1996         if (list_empty(&rpclist))
1997                 RETURN(0);
1998
1999         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2000
2001         list_for_each_entry(ext, &rpclist, oe_link) {
2002                 LASSERT(ext->oe_state == OES_CACHE ||
2003                         ext->oe_state == OES_LOCK_DONE);
2004                 if (ext->oe_state == OES_CACHE)
2005                         osc_extent_state_set(ext, OES_LOCKING);
2006                 else
2007                         osc_extent_state_set(ext, OES_RPC);
2008         }
2009
2010         /* we're going to grab page lock, so release object lock because
2011          * lock order is page lock -> object lock. */
2012         osc_object_unlock(osc);
2013
2014         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2015                 if (ext->oe_state == OES_LOCKING) {
2016                         rc = osc_extent_make_ready(env, ext);
2017                         if (unlikely(rc < 0)) {
2018                                 list_del_init(&ext->oe_link);
2019                                 osc_extent_finish(env, ext, 0, rc);
2020                                 continue;
2021                         }
2022                 }
2023                 if (first == NULL) {
2024                         first = ext;
2025                         srvlock = ext->oe_srvlock;
2026                 } else {
2027                         LASSERT(srvlock == ext->oe_srvlock);
2028                 }
2029         }
2030
2031         if (!list_empty(&rpclist)) {
2032                 LASSERT(page_count > 0);
2033                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2034                 LASSERT(list_empty(&rpclist));
2035         }
2036
2037         osc_object_lock(osc);
2038         RETURN(rc);
2039 }
2040
2041 /**
2042  * prepare pages for ASYNC io and put pages in send queue.
2043  *
2044  * \param cmd OBD_BRW_* macroses
2045  * \param lop pending pages
2046  *
2047  * \return zero if no page added to send queue.
2048  * \return 1 if pages successfully added to send queue.
2049  * \return negative on errors.
2050  */
2051 static int
2052 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2053                   struct osc_object *osc)
2054 __must_hold(osc)
2055 {
2056         struct osc_extent *ext;
2057         struct osc_extent *next;
2058         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2059         unsigned int page_count = 0;
2060         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2061         int rc = 0;
2062         ENTRY;
2063
2064         LASSERT(osc_object_is_locked(osc));
2065         list_for_each_entry_safe(ext, next,
2066                                      &osc->oo_reading_exts, oe_link) {
2067                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2068                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2069                                               &max_pages))
2070                         break;
2071                 osc_extent_state_set(ext, OES_RPC);
2072                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2073         }
2074         LASSERT(page_count <= max_pages);
2075
2076         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2077
2078         if (!list_empty(&rpclist)) {
2079                 osc_object_unlock(osc);
2080
2081                 LASSERT(page_count > 0);
2082                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2083                 LASSERT(list_empty(&rpclist));
2084
2085                 osc_object_lock(osc);
2086         }
2087         RETURN(rc);
2088 }
2089
2090 #define list_to_obj(list, item) ({                                            \
2091         struct list_head *__tmp = (list)->next;                               \
2092         list_del_init(__tmp);                                         \
2093         list_entry(__tmp, struct osc_object, oo_##item);                      \
2094 })
2095
2096 /* This is called by osc_check_rpcs() to find which objects have pages that
2097  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2098 static struct osc_object *osc_next_obj(struct client_obd *cli)
2099 {
2100         ENTRY;
2101
2102         /* First return objects that have blocked locks so that they
2103          * will be flushed quickly and other clients can get the lock,
2104          * then objects which have pages ready to be stuffed into RPCs */
2105         if (!list_empty(&cli->cl_loi_hp_ready_list))
2106                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2107         if (!list_empty(&cli->cl_loi_ready_list))
2108                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2109
2110         /* then if we have cache waiters, return all objects with queued
2111          * writes.  This is especially important when many small files
2112          * have filled up the cache and not been fired into rpcs because
2113          * they don't pass the nr_pending/object threshhold */
2114         if (!list_empty(&cli->cl_cache_waiters) &&
2115             !list_empty(&cli->cl_loi_write_list))
2116                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2117
2118         /* then return all queued objects when we have an invalid import
2119          * so that they get flushed */
2120         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2121                 if (!list_empty(&cli->cl_loi_write_list))
2122                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2123                                            write_item));
2124                 if (!list_empty(&cli->cl_loi_read_list))
2125                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2126                                            read_item));
2127         }
2128         RETURN(NULL);
2129 }
2130
2131 /* called with the loi list lock held */
2132 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2133 __must_hold(&cli->cl_loi_list_lock)
2134 {
2135         struct osc_object *osc;
2136         int rc = 0;
2137         ENTRY;
2138
2139         while ((osc = osc_next_obj(cli)) != NULL) {
2140                 struct cl_object *obj = osc2cl(osc);
2141                 struct lu_ref_link link;
2142
2143                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2144
2145                 if (osc_max_rpc_in_flight(cli, osc)) {
2146                         __osc_list_maint(cli, osc);
2147                         break;
2148                 }
2149
2150                 cl_object_get(obj);
2151                 spin_unlock(&cli->cl_loi_list_lock);
2152                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2153
2154                 /* attempt some read/write balancing by alternating between
2155                  * reads and writes in an object.  The makes_rpc checks here
2156                  * would be redundant if we were getting read/write work items
2157                  * instead of objects.  we don't want send_oap_rpc to drain a
2158                  * partial read pending queue when we're given this object to
2159                  * do io on writes while there are cache waiters */
2160                 osc_object_lock(osc);
2161                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2162                         rc = osc_send_write_rpc(env, cli, osc);
2163                         if (rc < 0) {
2164                                 CERROR("Write request failed with %d\n", rc);
2165
2166                                 /* osc_send_write_rpc failed, mostly because of
2167                                  * memory pressure.
2168                                  *
2169                                  * It can't break here, because if:
2170                                  *  - a page was submitted by osc_io_submit, so
2171                                  *    page locked;
2172                                  *  - no request in flight
2173                                  *  - no subsequent request
2174                                  * The system will be in live-lock state,
2175                                  * because there is no chance to call
2176                                  * osc_io_unplug() and osc_check_rpcs() any
2177                                  * more. pdflush can't help in this case,
2178                                  * because it might be blocked at grabbing
2179                                  * the page lock as we mentioned.
2180                                  *
2181                                  * Anyway, continue to drain pages. */
2182                                 /* break; */
2183                         }
2184                 }
2185                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2186                         rc = osc_send_read_rpc(env, cli, osc);
2187                         if (rc < 0)
2188                                 CERROR("Read request failed with %d\n", rc);
2189                 }
2190                 osc_object_unlock(osc);
2191
2192                 osc_list_maint(cli, osc);
2193                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2194                 cl_object_put(env, obj);
2195
2196                 spin_lock(&cli->cl_loi_list_lock);
2197         }
2198 }
2199
2200 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2201                           struct osc_object *osc, int async)
2202 {
2203         int rc = 0;
2204
2205         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2206                 return 0;
2207
2208         if (!async) {
2209                 spin_lock(&cli->cl_loi_list_lock);
2210                 osc_check_rpcs(env, cli);
2211                 spin_unlock(&cli->cl_loi_list_lock);
2212         } else {
2213                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2214                 LASSERT(cli->cl_writeback_work != NULL);
2215                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2216         }
2217         return rc;
2218 }
2219
2220 static int osc_io_unplug_async(const struct lu_env *env,
2221                                 struct client_obd *cli, struct osc_object *osc)
2222 {
2223         return osc_io_unplug0(env, cli, osc, 1);
2224 }
2225
2226 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2227                    struct osc_object *osc)
2228 {
2229         (void)osc_io_unplug0(env, cli, osc, 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 & ~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         unsigned int tmp;
2272         unsigned int grants = 0;
2273         u32    brw_flags = OBD_BRW_ASYNC;
2274         int    cmd = OBD_BRW_WRITE;
2275         int    need_release = 0;
2276         int    rc = 0;
2277         ENTRY;
2278
2279         if (oap->oap_magic != OAP_MAGIC)
2280                 RETURN(-EINVAL);
2281
2282         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2283                 RETURN(-EIO);
2284
2285         if (!list_empty(&oap->oap_pending_item) ||
2286             !list_empty(&oap->oap_rpc_item))
2287                 RETURN(-EBUSY);
2288
2289         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2290         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2291         if (!client_is_remote(osc_export(osc)) &&
2292             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2293                 brw_flags |= OBD_BRW_NOQUOTA;
2294                 cmd |= OBD_BRW_NOQUOTA;
2295         }
2296
2297         /* check if the file's owner/group is over quota */
2298         if (!(cmd & OBD_BRW_NOQUOTA)) {
2299                 struct cl_object *obj;
2300                 struct cl_attr   *attr;
2301                 unsigned int qid[MAXQUOTAS];
2302
2303                 obj = cl_object_top(&osc->oo_cl);
2304                 attr = &osc_env_info(env)->oti_attr;
2305
2306                 cl_object_attr_lock(obj);
2307                 rc = cl_object_attr_get(env, obj, attr);
2308                 cl_object_attr_unlock(obj);
2309
2310                 qid[USRQUOTA] = attr->cat_uid;
2311                 qid[GRPQUOTA] = attr->cat_gid;
2312                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2313                         rc = -EDQUOT;
2314                 if (rc)
2315                         RETURN(rc);
2316         }
2317
2318         oap->oap_cmd = cmd;
2319         oap->oap_page_off = ops->ops_from;
2320         oap->oap_count = ops->ops_to - ops->ops_from;
2321         /* No need to hold a lock here,
2322          * since this page is not in any list yet. */
2323         oap->oap_async_flags = 0;
2324         oap->oap_brw_flags = brw_flags;
2325
2326         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2327                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2328
2329         index = osc_index(oap2osc(oap));
2330
2331         /* Add this page into extent by the following steps:
2332          * 1. if there exists an active extent for this IO, mostly this page
2333          *    can be added to the active extent and sometimes we need to
2334          *    expand extent to accomodate this page;
2335          * 2. otherwise, a new extent will be allocated. */
2336
2337         ext = oio->oi_active;
2338         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2339                 /* one chunk plus extent overhead must be enough to write this
2340                  * page */
2341                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2342                 if (ext->oe_end >= index)
2343                         grants = 0;
2344
2345                 /* it doesn't need any grant to dirty this page */
2346                 spin_lock(&cli->cl_loi_list_lock);
2347                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2348                 spin_unlock(&cli->cl_loi_list_lock);
2349                 if (rc == 0) { /* try failed */
2350                         grants = 0;
2351                         need_release = 1;
2352                 } else if (ext->oe_end < index) {
2353                         tmp = grants;
2354                         /* try to expand this extent */
2355                         rc = osc_extent_expand(ext, index, &tmp);
2356                         if (rc < 0) {
2357                                 need_release = 1;
2358                                 /* don't free reserved grant */
2359                         } else {
2360                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2361                                                 "expanded for %lu.\n", index);
2362                                 osc_unreserve_grant(cli, grants, tmp);
2363                                 grants = 0;
2364                         }
2365                 }
2366                 rc = 0;
2367         } else if (ext != NULL) {
2368                 /* index is located outside of active extent */
2369                 need_release = 1;
2370         }
2371         if (need_release) {
2372                 osc_extent_release(env, ext);
2373                 oio->oi_active = NULL;
2374                 ext = NULL;
2375         }
2376
2377         if (ext == NULL) {
2378                 tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2379
2380                 /* try to find new extent to cover this page */
2381                 LASSERT(oio->oi_active == NULL);
2382                 /* we may have allocated grant for this page if we failed
2383                  * to expand the previous active extent. */
2384                 LASSERT(ergo(grants > 0, grants >= tmp));
2385
2386                 rc = 0;
2387                 if (grants == 0) {
2388                         /* we haven't allocated grant for this page. */
2389                         rc = osc_enter_cache(env, cli, oap, tmp);
2390                         if (rc == 0)
2391                                 grants = tmp;
2392                 }
2393
2394                 tmp = grants;
2395                 if (rc == 0) {
2396                         ext = osc_extent_find(env, osc, index, &tmp);
2397                         if (IS_ERR(ext)) {
2398                                 LASSERT(tmp == grants);
2399                                 osc_exit_cache(cli, oap);
2400                                 rc = PTR_ERR(ext);
2401                                 ext = NULL;
2402                         } else {
2403                                 oio->oi_active = ext;
2404                         }
2405                 }
2406                 if (grants > 0)
2407                         osc_unreserve_grant(cli, grants, tmp);
2408         }
2409
2410         LASSERT(ergo(rc == 0, ext != NULL));
2411         if (ext != NULL) {
2412                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2413                          ext, "index = %lu.\n", index);
2414                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2415
2416                 osc_object_lock(osc);
2417                 if (ext->oe_nr_pages == 0)
2418                         ext->oe_srvlock = ops->ops_srvlock;
2419                 else
2420                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2421                 ++ext->oe_nr_pages;
2422                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2423                 osc_object_unlock(osc);
2424         }
2425         RETURN(rc);
2426 }
2427
2428 int osc_teardown_async_page(const struct lu_env *env,
2429                             struct osc_object *obj, struct osc_page *ops)
2430 {
2431         struct osc_async_page *oap = &ops->ops_oap;
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         if (!list_empty(&oap->oap_rpc_item)) {
2441                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2442                 rc = -EBUSY;
2443         } else if (!list_empty(&oap->oap_pending_item)) {
2444                 struct osc_extent *ext = NULL;
2445
2446                 osc_object_lock(obj);
2447                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2448                 osc_object_unlock(obj);
2449                 /* only truncated pages are allowed to be taken out.
2450                  * See osc_extent_truncate() and osc_cache_truncate_start()
2451                  * for details. */
2452                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2453                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2454                                         osc_index(oap2osc(oap)));
2455                         rc = -EBUSY;
2456                 }
2457                 if (ext != NULL)
2458                         osc_extent_put(env, ext);
2459         }
2460         RETURN(rc);
2461 }
2462
2463 /**
2464  * This is called when a page is picked up by kernel to write out.
2465  *
2466  * We should find out the corresponding extent and add the whole extent
2467  * into urgent list. The extent may be being truncated or used, handle it
2468  * carefully.
2469  */
2470 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2471                          struct osc_page *ops)
2472 {
2473         struct osc_extent *ext   = NULL;
2474         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2475         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2476         pgoff_t            index = osc_index(ops);
2477         struct osc_async_page *oap = &ops->ops_oap;
2478         bool unplug = false;
2479         int rc = 0;
2480         ENTRY;
2481
2482         osc_object_lock(obj);
2483         ext = osc_extent_lookup(obj, index);
2484         if (ext == NULL) {
2485                 osc_extent_tree_dump(D_ERROR, obj);
2486                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2487         }
2488
2489         switch (ext->oe_state) {
2490         case OES_RPC:
2491         case OES_LOCK_DONE:
2492                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2493                 LASSERT(0);
2494                 break;
2495         case OES_LOCKING:
2496                 /* If we know this extent is being written out, we should abort
2497                  * so that the writer can make this page ready. Otherwise, there
2498                  * exists a deadlock problem because other process can wait for
2499                  * page writeback bit holding page lock; and meanwhile in
2500                  * vvp_page_make_ready(), we need to grab page lock before
2501                  * really sending the RPC. */
2502         case OES_TRUNC:
2503                 /* race with truncate, page will be redirtied */
2504         case OES_ACTIVE:
2505                 /* The extent is active so we need to abort and let the caller
2506                  * re-dirty the page. If we continued on here, and we were the
2507                  * one making the extent active, we could deadlock waiting for
2508                  * the page writeback to clear but it won't because the extent
2509                  * is active and won't be written out. */
2510                 GOTO(out, rc = -EAGAIN);
2511         default:
2512                 break;
2513         }
2514
2515         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2516         if (rc)
2517                 GOTO(out, rc);
2518
2519         spin_lock(&oap->oap_lock);
2520         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2521         spin_unlock(&oap->oap_lock);
2522
2523         if (memory_pressure_get())
2524                 ext->oe_memalloc = 1;
2525
2526         ext->oe_urgent = 1;
2527         if (ext->oe_state == OES_CACHE) {
2528                 OSC_EXTENT_DUMP(D_CACHE, ext,
2529                                 "flush page %p make it urgent.\n", oap);
2530                 if (list_empty(&ext->oe_link))
2531                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2532                 unplug = true;
2533         }
2534         rc = 0;
2535         EXIT;
2536
2537 out:
2538         osc_object_unlock(obj);
2539         osc_extent_put(env, ext);
2540         if (unplug)
2541                 osc_io_unplug_async(env, osc_cli(obj), obj);
2542         return rc;
2543 }
2544
2545 /**
2546  * this is called when a sync waiter receives an interruption.  Its job is to
2547  * get the caller woken as soon as possible.  If its page hasn't been put in an
2548  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2549  * desiring interruption which will forcefully complete the rpc once the rpc
2550  * has timed out.
2551  */
2552 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2553 {
2554         struct osc_async_page *oap = &ops->ops_oap;
2555         struct osc_object     *obj = oap->oap_obj;
2556         struct client_obd     *cli = osc_cli(obj);
2557         struct osc_extent     *ext;
2558         struct osc_extent     *found = NULL;
2559         struct list_head            *plist;
2560         pgoff_t index = osc_index(ops);
2561         int     rc = -EBUSY;
2562         int     cmd;
2563         ENTRY;
2564
2565         LASSERT(!oap->oap_interrupted);
2566         oap->oap_interrupted = 1;
2567
2568         /* Find out the caching extent */
2569         osc_object_lock(obj);
2570         if (oap->oap_cmd & OBD_BRW_WRITE) {
2571                 plist = &obj->oo_urgent_exts;
2572                 cmd   = OBD_BRW_WRITE;
2573         } else {
2574                 plist = &obj->oo_reading_exts;
2575                 cmd   = OBD_BRW_READ;
2576         }
2577         list_for_each_entry(ext, plist, oe_link) {
2578                 if (ext->oe_start <= index && ext->oe_end >= index) {
2579                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2580                         /* For OES_LOCK_DONE state extent, it has already held
2581                          * a refcount for RPC. */
2582                         found = osc_extent_get(ext);
2583                         break;
2584                 }
2585         }
2586         if (found != NULL) {
2587                 list_del_init(&found->oe_link);
2588                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2589                 osc_object_unlock(obj);
2590
2591                 osc_extent_finish(env, found, 0, -EINTR);
2592                 osc_extent_put(env, found);
2593                 rc = 0;
2594         } else {
2595                 osc_object_unlock(obj);
2596                 /* ok, it's been put in an rpc. only one oap gets a request
2597                  * reference */
2598                 if (oap->oap_request != NULL) {
2599                         ptlrpc_mark_interrupted(oap->oap_request);
2600                         ptlrpcd_wake(oap->oap_request);
2601                         ptlrpc_req_finished(oap->oap_request);
2602                         oap->oap_request = NULL;
2603                 }
2604         }
2605
2606         osc_list_maint(cli, obj);
2607         RETURN(rc);
2608 }
2609
2610 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2611                          struct list_head *list, int cmd, int brw_flags)
2612 {
2613         struct client_obd     *cli = osc_cli(obj);
2614         struct osc_extent     *ext;
2615         struct osc_async_page *oap;
2616         int     page_count = 0;
2617         int     mppr       = cli->cl_max_pages_per_rpc;
2618         pgoff_t start      = CL_PAGE_EOF;
2619         pgoff_t end        = 0;
2620         ENTRY;
2621
2622         list_for_each_entry(oap, list, oap_pending_item) {
2623                 pgoff_t index = osc_index(oap2osc(oap));
2624                 if (index > end)
2625                         end = index;
2626                 if (index < start)
2627                         start = index;
2628                 ++page_count;
2629                 mppr <<= (page_count > mppr);
2630         }
2631
2632         ext = osc_extent_alloc(obj);
2633         if (ext == NULL) {
2634                 struct osc_async_page *tmp;
2635
2636                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2637                         list_del_init(&oap->oap_pending_item);
2638                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2639                 }
2640                 RETURN(-ENOMEM);
2641         }
2642
2643         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2644         ext->oe_sync = 1;
2645         ext->oe_urgent = 1;
2646         ext->oe_start = start;
2647         ext->oe_end = ext->oe_max_end = end;
2648         ext->oe_obj = obj;
2649         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2650         ext->oe_nr_pages = page_count;
2651         ext->oe_mppr = mppr;
2652         list_splice_init(list, &ext->oe_pages);
2653
2654         osc_object_lock(obj);
2655         /* Reuse the initial refcount for RPC, don't drop it */
2656         osc_extent_state_set(ext, OES_LOCK_DONE);
2657         if (cmd & OBD_BRW_WRITE) {
2658                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2659                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2660         } else {
2661                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2662                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2663         }
2664         osc_object_unlock(obj);
2665
2666         osc_io_unplug_async(env, cli, obj);
2667         RETURN(0);
2668 }
2669
2670 /**
2671  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2672  */
2673 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2674                              __u64 size, struct osc_extent **extp)
2675 {
2676         struct client_obd *cli = osc_cli(obj);
2677         struct osc_extent *ext;
2678         struct osc_extent *waiting = NULL;
2679         pgoff_t index;
2680         struct list_head list = LIST_HEAD_INIT(list);
2681         int result = 0;
2682         bool partial;
2683         ENTRY;
2684
2685         /* pages with index greater or equal to index will be truncated. */
2686         index = cl_index(osc2cl(obj), size);
2687         partial = size > cl_offset(osc2cl(obj), index);
2688
2689 again:
2690         osc_object_lock(obj);
2691         ext = osc_extent_search(obj, index);
2692         if (ext == NULL)
2693                 ext = first_extent(obj);
2694         else if (ext->oe_end < index)
2695                 ext = next_extent(ext);
2696         while (ext != NULL) {
2697                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2698
2699                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2700                         /* if ext is in urgent state, it means there must exist
2701                          * a page already having been flushed by write_page().
2702                          * We have to wait for this extent because we can't
2703                          * truncate that page. */
2704                         LASSERT(!ext->oe_hp);
2705                         OSC_EXTENT_DUMP(D_CACHE, ext,
2706                                         "waiting for busy extent\n");
2707                         waiting = osc_extent_get(ext);
2708                         break;
2709                 }
2710
2711                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2712
2713                 osc_extent_get(ext);
2714                 if (ext->oe_state == OES_ACTIVE) {
2715                         /* though we grab inode mutex for write path, but we
2716                          * release it before releasing extent(in osc_io_end()),
2717                          * so there is a race window that an extent is still
2718                          * in OES_ACTIVE when truncate starts. */
2719                         LASSERT(!ext->oe_trunc_pending);
2720                         ext->oe_trunc_pending = 1;
2721                 } else {
2722                         EASSERT(ext->oe_state == OES_CACHE, ext);
2723                         osc_extent_state_set(ext, OES_TRUNC);
2724                         osc_update_pending(obj, OBD_BRW_WRITE,
2725                                            -ext->oe_nr_pages);
2726                 }
2727                 EASSERT(list_empty(&ext->oe_link), ext);
2728                 list_add_tail(&ext->oe_link, &list);
2729
2730                 ext = next_extent(ext);
2731         }
2732         osc_object_unlock(obj);
2733
2734         osc_list_maint(cli, obj);
2735
2736         while (!list_empty(&list)) {
2737                 int rc;
2738
2739                 ext = list_entry(list.next, struct osc_extent, oe_link);
2740                 list_del_init(&ext->oe_link);
2741
2742                 /* extent may be in OES_ACTIVE state because inode mutex
2743                  * is released before osc_io_end() in file write case */
2744                 if (ext->oe_state != OES_TRUNC)
2745                         osc_extent_wait(env, ext, OES_TRUNC);
2746
2747                 rc = osc_extent_truncate(ext, index, partial);
2748                 if (rc < 0) {
2749                         if (result == 0)
2750                                 result = rc;
2751
2752                         OSC_EXTENT_DUMP(D_ERROR, ext,
2753                                         "truncate error %d\n", rc);
2754                 } else if (ext->oe_nr_pages == 0) {
2755                         osc_extent_remove(ext);
2756                 } else {
2757                         /* this must be an overlapped extent which means only
2758                          * part of pages in this extent have been truncated.
2759                          */
2760                         EASSERTF(ext->oe_start <= index, ext,
2761                                  "trunc index = %lu/%d.\n", index, partial);
2762                         /* fix index to skip this partially truncated extent */
2763                         index = ext->oe_end + 1;
2764                         partial = false;
2765
2766                         /* we need to hold this extent in OES_TRUNC state so
2767                          * that no writeback will happen. This is to avoid
2768                          * BUG 17397.
2769                          * Only partial truncate can reach here, if @size is
2770                          * not zero, the caller should provide a valid @extp. */
2771                         LASSERT(*extp == NULL);
2772                         *extp = osc_extent_get(ext);
2773                         OSC_EXTENT_DUMP(D_CACHE, ext,
2774                                         "trunc at "LPU64"\n", size);
2775                 }
2776                 osc_extent_put(env, ext);
2777         }
2778         if (waiting != NULL) {
2779                 int rc;
2780
2781                 /* ignore the result of osc_extent_wait the write initiator
2782                  * should take care of it. */
2783                 rc = osc_extent_wait(env, waiting, OES_INV);
2784                 if (rc < 0)
2785                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2786
2787                 osc_extent_put(env, waiting);
2788                 waiting = NULL;
2789                 goto again;
2790         }
2791         RETURN(result);
2792 }
2793
2794 /**
2795  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2796  */
2797 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2798 {
2799         if (ext != NULL) {
2800                 struct osc_object *obj = ext->oe_obj;
2801                 bool unplug = false;
2802
2803                 EASSERT(ext->oe_nr_pages > 0, ext);
2804                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2805                 EASSERT(!ext->oe_urgent, ext);
2806
2807                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2808                 osc_object_lock(obj);
2809                 osc_extent_state_set(ext, OES_CACHE);
2810                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2811                         ext->oe_urgent = 1;
2812                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2813                         unplug = true;
2814                 }
2815                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2816                 osc_object_unlock(obj);
2817                 osc_extent_put(env, ext);
2818
2819                 if (unplug)
2820                         osc_io_unplug_async(env, osc_cli(obj), obj);
2821         }
2822 }
2823
2824 /**
2825  * Wait for extents in a specific range to be written out.
2826  * The caller must have called osc_cache_writeback_range() to issue IO
2827  * otherwise it will take a long time for this function to finish.
2828  *
2829  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2830  * nobody else can dirty this range of file while we're waiting for
2831  * extents to be written.
2832  */
2833 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2834                          pgoff_t start, pgoff_t end)
2835 {
2836         struct osc_extent *ext;
2837         pgoff_t index = start;
2838         int     result = 0;
2839         ENTRY;
2840
2841 again:
2842         osc_object_lock(obj);
2843         ext = osc_extent_search(obj, index);
2844         if (ext == NULL)
2845                 ext = first_extent(obj);
2846         else if (ext->oe_end < index)
2847                 ext = next_extent(ext);
2848         while (ext != NULL) {
2849                 int rc;
2850
2851                 if (ext->oe_start > end)
2852                         break;
2853
2854                 if (!ext->oe_fsync_wait) {
2855                         ext = next_extent(ext);
2856                         continue;
2857                 }
2858
2859                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2860                              ext->oe_hp || ext->oe_urgent), ext);
2861                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2862                              !ext->oe_hp && ext->oe_urgent), ext);
2863
2864                 index = ext->oe_end + 1;
2865                 osc_extent_get(ext);
2866                 osc_object_unlock(obj);
2867
2868                 rc = osc_extent_wait(env, ext, OES_INV);
2869                 if (result == 0)
2870                         result = rc;
2871                 osc_extent_put(env, ext);
2872                 goto again;
2873         }
2874         osc_object_unlock(obj);
2875
2876         OSC_IO_DEBUG(obj, "sync file range.\n");
2877         RETURN(result);
2878 }
2879
2880 /**
2881  * Called to write out a range of osc object.
2882  *
2883  * @hp     : should be set this is caused by lock cancel;
2884  * @discard: is set if dirty pages should be dropped - file will be deleted or
2885  *         truncated, this implies there is no partially discarding extents.
2886  *
2887  * Return how many pages will be issued, or error code if error occurred.
2888  */
2889 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2890                               pgoff_t start, pgoff_t end, int hp, int discard)
2891 {
2892         struct osc_extent *ext;
2893         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2894         bool unplug = false;
2895         int result = 0;
2896         ENTRY;
2897
2898         osc_object_lock(obj);
2899         ext = osc_extent_search(obj, start);
2900         if (ext == NULL)
2901                 ext = first_extent(obj);
2902         else if (ext->oe_end < start)
2903                 ext = next_extent(ext);
2904         while (ext != NULL) {
2905                 if (ext->oe_start > end)
2906                         break;
2907
2908                 ext->oe_fsync_wait = 1;
2909                 switch (ext->oe_state) {
2910                 case OES_CACHE:
2911                         result += ext->oe_nr_pages;
2912                         if (!discard) {
2913                                 struct list_head *list = NULL;
2914                                 if (hp) {
2915                                         EASSERT(!ext->oe_hp, ext);
2916                                         ext->oe_hp = 1;
2917                                         list = &obj->oo_hp_exts;
2918                                 } else if (!ext->oe_urgent) {
2919                                         ext->oe_urgent = 1;
2920                                         list = &obj->oo_urgent_exts;
2921                                 }
2922                                 if (list != NULL)
2923                                         list_move_tail(&ext->oe_link, list);
2924                                 unplug = true;
2925                         } else {
2926                                 /* the only discarder is lock cancelling, so
2927                                  * [start, end] must contain this extent */
2928                                 EASSERT(ext->oe_start >= start &&
2929                                         ext->oe_max_end <= end, ext);
2930                                 osc_extent_state_set(ext, OES_LOCKING);
2931                                 ext->oe_owner = current;
2932                                 list_move_tail(&ext->oe_link,
2933                                                    &discard_list);
2934                                 osc_update_pending(obj, OBD_BRW_WRITE,
2935                                                    -ext->oe_nr_pages);
2936                         }
2937                         break;
2938                 case OES_ACTIVE:
2939                         /* It's pretty bad to wait for ACTIVE extents, because
2940                          * we don't know how long we will wait for it to be
2941                          * flushed since it may be blocked at awaiting more
2942                          * grants. We do this for the correctness of fsync. */
2943                         LASSERT(hp == 0 && discard == 0);
2944                         ext->oe_urgent = 1;
2945                         break;
2946                 case OES_TRUNC:
2947                         /* this extent is being truncated, can't do anything
2948                          * for it now. it will be set to urgent after truncate
2949                          * is finished in osc_cache_truncate_end(). */
2950                 default:
2951                         break;
2952                 }
2953                 ext = next_extent(ext);
2954         }
2955         osc_object_unlock(obj);
2956
2957         LASSERT(ergo(!discard, list_empty(&discard_list)));
2958         if (!list_empty(&discard_list)) {
2959                 struct osc_extent *tmp;
2960                 int rc;
2961
2962                 osc_list_maint(osc_cli(obj), obj);
2963                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2964                         list_del_init(&ext->oe_link);
2965                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2966
2967                         /* Discard caching pages. We don't actually write this
2968                          * extent out but we complete it as if we did. */
2969                         rc = osc_extent_make_ready(env, ext);
2970                         if (unlikely(rc < 0)) {
2971                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2972                                                 "make_ready returned %d\n", rc);
2973                                 if (result >= 0)
2974                                         result = rc;
2975                         }
2976
2977                         /* finish the extent as if the pages were sent */
2978                         osc_extent_finish(env, ext, 0, 0);
2979                 }
2980         }
2981
2982         if (unplug)
2983                 osc_io_unplug(env, osc_cli(obj), obj);
2984
2985         if (hp || discard) {
2986                 int rc;
2987                 rc = osc_cache_wait_range(env, obj, start, end);
2988                 if (result >= 0 && rc < 0)
2989                         result = rc;
2990         }
2991
2992         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2993         RETURN(result);
2994 }
2995
2996 /**
2997  * Returns a list of pages by a given [start, end] of \a obj.
2998  *
2999  * \param resched If not NULL, then we give up before hogging CPU for too
3000  * long and set *resched = 1, in that case caller should implement a retry
3001  * logic.
3002  *
3003  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3004  * crucial in the face of [offset, EOF] locks.
3005  *
3006  * Return at least one page in @queue unless there is no covered page.
3007  */
3008 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3009                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3010                         osc_page_gang_cbt cb, void *cbdata)
3011 {
3012         struct osc_page *ops;
3013         void            **pvec;
3014         pgoff_t         idx;
3015         unsigned int    nr;
3016         unsigned int    i;
3017         unsigned int    j;
3018         int             res = CLP_GANG_OKAY;
3019         bool            tree_lock = true;
3020         ENTRY;
3021
3022         idx = start;
3023         pvec = osc_env_info(env)->oti_pvec;
3024         spin_lock(&osc->oo_tree_lock);
3025         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3026                                             idx, OTI_PVEC_SIZE)) > 0) {
3027                 struct cl_page *page;
3028                 bool end_of_region = false;
3029
3030                 for (i = 0, j = 0; i < nr; ++i) {
3031                         ops = pvec[i];
3032                         pvec[i] = NULL;
3033
3034                         idx = osc_index(ops);
3035                         if (idx > end) {
3036                                 end_of_region = true;
3037                                 break;
3038                         }
3039
3040                         page = ops->ops_cl.cpl_page;
3041                         LASSERT(page->cp_type == CPT_CACHEABLE);
3042                         if (page->cp_state == CPS_FREEING)
3043                                 continue;
3044
3045                         cl_page_get(page);
3046                         lu_ref_add_atomic(&page->cp_reference,
3047                                           "gang_lookup", current);
3048                         pvec[j++] = ops;
3049                 }
3050                 ++idx;
3051
3052                 /*
3053                  * Here a delicate locking dance is performed. Current thread
3054                  * holds a reference to a page, but has to own it before it
3055                  * can be placed into queue. Owning implies waiting, so
3056                  * radix-tree lock is to be released. After a wait one has to
3057                  * check that pages weren't truncated (cl_page_own() returns
3058                  * error in the latter case).
3059                  */
3060                 spin_unlock(&osc->oo_tree_lock);
3061                 tree_lock = false;
3062
3063                 for (i = 0; i < j; ++i) {
3064                         ops = pvec[i];
3065                         if (res == CLP_GANG_OKAY)
3066                                 res = (*cb)(env, io, ops, cbdata);
3067
3068                         page = ops->ops_cl.cpl_page;
3069                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3070                         cl_page_put(env, page);
3071                 }
3072                 if (nr < OTI_PVEC_SIZE || end_of_region)
3073                         break;
3074
3075                 if (res == CLP_GANG_OKAY && need_resched())
3076                         res = CLP_GANG_RESCHED;
3077                 if (res != CLP_GANG_OKAY)
3078                         break;
3079
3080                 spin_lock(&osc->oo_tree_lock);
3081                 tree_lock = true;
3082         }
3083         if (tree_lock)
3084                 spin_unlock(&osc->oo_tree_lock);
3085         RETURN(res);
3086 }
3087
3088 /**
3089  * Check if page @page is covered by an extra lock or discard it.
3090  */
3091 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3092                                 struct osc_page *ops, void *cbdata)
3093 {
3094         struct osc_thread_info *info = osc_env_info(env);
3095         struct osc_object *osc = cbdata;
3096         pgoff_t index;
3097
3098         index = osc_index(ops);
3099         if (index >= info->oti_fn_index) {
3100                 struct ldlm_lock *tmp;
3101                 struct cl_page *page = ops->ops_cl.cpl_page;
3102
3103                 /* refresh non-overlapped index */
3104                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3105                                            OSC_DAP_FL_TEST_LOCK);
3106                 if (tmp != NULL) {
3107                         __u64 end = tmp->l_policy_data.l_extent.end;
3108                         /* Cache the first-non-overlapped index so as to skip
3109                          * all pages within [index, oti_fn_index). This is safe
3110                          * because if tmp lock is canceled, it will discard
3111                          * these pages. */
3112                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3113                         if (end == OBD_OBJECT_EOF)
3114                                 info->oti_fn_index = CL_PAGE_EOF;
3115                         LDLM_LOCK_PUT(tmp);
3116                 } else if (cl_page_own(env, io, page) == 0) {
3117                         /* discard the page */
3118                         cl_page_discard(env, io, page);
3119                         cl_page_disown(env, io, page);
3120                 } else {
3121                         LASSERT(page->cp_state == CPS_FREEING);
3122                 }
3123         }
3124
3125         info->oti_next_index = index + 1;
3126         return CLP_GANG_OKAY;
3127 }
3128
3129 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3130                       struct osc_page *ops, void *cbdata)
3131 {
3132         struct osc_thread_info *info = osc_env_info(env);
3133         struct cl_page *page = ops->ops_cl.cpl_page;
3134
3135         /* page is top page. */
3136         info->oti_next_index = osc_index(ops) + 1;
3137         if (cl_page_own(env, io, page) == 0) {
3138                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3139                           !PageDirty(cl_page_vmpage(page))))
3140                         CL_PAGE_DEBUG(D_ERROR, env, page,
3141                                         "discard dirty page?\n");
3142
3143                 /* discard the page */
3144                 cl_page_discard(env, io, page);
3145                 cl_page_disown(env, io, page);
3146         } else {
3147                 LASSERT(page->cp_state == CPS_FREEING);
3148         }
3149
3150         return CLP_GANG_OKAY;
3151 }
3152
3153 /**
3154  * Discard pages protected by the given lock. This function traverses radix
3155  * tree to find all covering pages and discard them. If a page is being covered
3156  * by other locks, it should remain in cache.
3157  *
3158  * If error happens on any step, the process continues anyway (the reasoning
3159  * behind this being that lock cancellation cannot be delayed indefinitely).
3160  */
3161 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3162                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3163 {
3164         struct osc_thread_info *info = osc_env_info(env);
3165         struct cl_io *io = &info->oti_io;
3166         osc_page_gang_cbt cb;
3167         int res;
3168         int result;
3169
3170         ENTRY;
3171
3172         io->ci_obj = cl_object_top(osc2cl(osc));
3173         io->ci_ignore_layout = 1;
3174         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3175         if (result != 0)
3176                 GOTO(out, result);
3177