]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blame - drivers/staging/lustre/lustre/osc/osc_cache.c
projects / mirror_ubuntu-hirsute-kernel.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
staging/lustre: get rid of obd_* typedefs
[mirror_ubuntu-hirsute-kernel.git] / drivers / staging / lustre / lustre / osc / osc_cache.c
CommitLineData
d7e09d03
PT		 1/*
2 * GPL HEADER START
3 *
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19 *
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26/*
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2012, Intel Corporation.
31 *
32 */
33/*
34 * This file is part of Lustre, http://www.lustre.org/
35 * Lustre is a trademark of Sun Microsystems, Inc.
36 *
37 * osc cache management.
38 *
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
40 */
41
42#define DEBUG_SUBSYSTEM S_OSC
43
44#include "osc_cl_internal.h"
45#include "osc_internal.h"
46
47static int extent_debug; /* set it to be true for more debug */
48
49static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
50static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
51 int state);
52static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
53 struct osc_async_page *oap, int sent, int rc);
54static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
55 int cmd);
56static int osc_refresh_count(const struct lu_env *env,
57 struct osc_async_page *oap, int cmd);
58static int osc_io_unplug_async(const struct lu_env *env,
59 struct client_obd *cli, struct osc_object *osc);
60static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
61 unsigned int lost_grant);
62
63static 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 ------------------ */
73static inline char *ext_flags(struct osc_extent *ext, char *flags)
74{
75 char *buf = flags;
76 *buf++ = ext->oe_rw ? 'r' : 'w';
77 if (ext->oe_intree)
78 *buf++ = 'i';
79 if (ext->oe_srvlock)
80 *buf++ = 's';
81 if (ext->oe_hp)
82 *buf++ = 'h';
83 if (ext->oe_urgent)
84 *buf++ = 'u';
85 if (ext->oe_memalloc)
86 *buf++ = 'm';
87 if (ext->oe_trunc_pending)
88 *buf++ = 't';
89 if (ext->oe_fsync_wait)
90 *buf++ = 'Y';
91 *buf = 0;
92 return flags;
93}
94
95static inline char list_empty_marker(struct list_head *list)
96{
97 return list_empty(list) ? '-' : '+';
98}
99
100#define EXTSTR "[%lu -> %lu/%lu]"
101#define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
cad6fafa
BJ		 102static const char *oes_strings[] = {
103 "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
d7e09d03
PT		 104
105#define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do { \
106 struct osc_extent *__ext = (extent); \
d7e09d03
PT		 107 char __buf[16]; \
108 \
109 CDEBUG(lvl, \
110 "extent %p@{" EXTSTR ", " \
111 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt, \
112 /* ----- extent part 0 ----- */ \
113 __ext, EXTPARA(__ext), \
114 /* ----- part 1 ----- */ \
115 atomic_read(&__ext->oe_refc), \
116 atomic_read(&__ext->oe_users), \
117 list_empty_marker(&__ext->oe_link), \
cad6fafa 118 oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \
d7e09d03
PT		 119 __ext->oe_obj, \
120 /* ----- part 2 ----- */ \
121 __ext->oe_grants, __ext->oe_nr_pages, \
122 list_empty_marker(&__ext->oe_pages), \
123 waitqueue_active(&__ext->oe_waitq) ? '+' : '-', \
124 __ext->oe_osclock, __ext->oe_mppr, __ext->oe_owner, \
125 /* ----- part 4 ----- */ \
126 ## __VA_ARGS__); \
127} while (0)
128
129#undef EASSERTF
130#define EASSERTF(expr, ext, fmt, args...) do { \
131 if (!(expr)) { \
cad6fafa
BJ		 132 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args); \
133 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj); \
d7e09d03 134 LASSERT(expr); \
cad6fafa 135 } \
d7e09d03
PT		 136} while (0)
137
138#undef EASSERT
139#define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
140
141static inline struct osc_extent *rb_extent(struct rb_node *n)
142{
143 if (n == NULL)
144 return NULL;
145
146 return container_of(n, struct osc_extent, oe_node);
147}
148
149static inline struct osc_extent *next_extent(struct osc_extent *ext)
150{
151 if (ext == NULL)
152 return NULL;
153
154 LASSERT(ext->oe_intree);
155 return rb_extent(rb_next(&ext->oe_node));
156}
157
158static inline struct osc_extent *prev_extent(struct osc_extent *ext)
159{
160 if (ext == NULL)
161 return NULL;
162
163 LASSERT(ext->oe_intree);
164 return rb_extent(rb_prev(&ext->oe_node));
165}
166
167static inline struct osc_extent *first_extent(struct osc_object *obj)
168{
169 return rb_extent(rb_first(&obj->oo_root));
170}
171
172/* object must be locked by caller. */
173static int osc_extent_sanity_check0(struct osc_extent *ext,
174 const char *func, const int line)
175{
176 struct osc_object *obj = ext->oe_obj;
177 struct osc_async_page *oap;
178 int page_count;
179 int rc = 0;
180
181 if (!osc_object_is_locked(obj))
182 GOTO(out, rc = 9);
183
184 if (ext->oe_state >= OES_STATE_MAX)
185 GOTO(out, rc = 10);
186
187 if (atomic_read(&ext->oe_refc) <= 0)
188 GOTO(out, rc = 20);
189
190 if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
191 GOTO(out, rc = 30);
192
193 switch (ext->oe_state) {
194 case OES_INV:
195 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
196 GOTO(out, rc = 35);
197 GOTO(out, rc = 0);
198 break;
199 case OES_ACTIVE:
200 if (atomic_read(&ext->oe_users) == 0)
201 GOTO(out, rc = 40);
202 if (ext->oe_hp)
203 GOTO(out, rc = 50);
204 if (ext->oe_fsync_wait && !ext->oe_urgent)
205 GOTO(out, rc = 55);
206 break;
207 case OES_CACHE:
208 if (ext->oe_grants == 0)
209 GOTO(out, rc = 60);
210 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
211 GOTO(out, rc = 65);
212 default:
213 if (atomic_read(&ext->oe_users) > 0)
214 GOTO(out, rc = 70);
215 }
216
217 if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
218 GOTO(out, rc = 80);
219
220 if (ext->oe_osclock == NULL && ext->oe_grants > 0)
221 GOTO(out, rc = 90);
222
223 if (ext->oe_osclock) {
224 struct cl_lock_descr *descr;
225 descr = &ext->oe_osclock->cll_descr;
226 if (!(descr->cld_start <= ext->oe_start &&
227 descr->cld_end >= ext->oe_max_end))
228 GOTO(out, rc = 100);
229 }
230
231 if (ext->oe_nr_pages > ext->oe_mppr)
232 GOTO(out, rc = 105);
233
234 /* Do not verify page list if extent is in RPC. This is because an
235 * in-RPC extent is supposed to be exclusively accessible w/o lock. */
236 if (ext->oe_state > OES_CACHE)
237 GOTO(out, rc = 0);
238
239 if (!extent_debug)
240 GOTO(out, rc = 0);
241
242 page_count = 0;
243 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
244 pgoff_t index = oap2cl_page(oap)->cp_index;
245 ++page_count;
246 if (index > ext->oe_end || index < ext->oe_start)
247 GOTO(out, rc = 110);
248 }
249 if (page_count != ext->oe_nr_pages)
250 GOTO(out, rc = 120);
251
252out:
253 if (rc != 0)
254 OSC_EXTENT_DUMP(D_ERROR, ext,
255 "%s:%d sanity check %p failed with rc = %d\n",
256 func, line, ext, rc);
257 return rc;
258}
259
260#define sanity_check_nolock(ext) \
261 osc_extent_sanity_check0(ext, __func__, __LINE__)
262
263#define sanity_check(ext) ({ \
264 int __res; \
265 osc_object_lock((ext)->oe_obj); \
266 __res = sanity_check_nolock(ext); \
267 osc_object_unlock((ext)->oe_obj); \
268 __res; \
269})
270
271
272/**
273 * sanity check - to make sure there is no overlapped extent in the tree.
274 */
275static int osc_extent_is_overlapped(struct osc_object *obj,
276 struct osc_extent *ext)
277{
278 struct osc_extent *tmp;
279
280 LASSERT(osc_object_is_locked(obj));
281
282 if (!extent_debug)
283 return 0;
284
285 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
286 if (tmp == ext)
287 continue;
288 if (tmp->oe_end >= ext->oe_start &&
289 tmp->oe_start <= ext->oe_end)
290 return 1;
291 }
292 return 0;
293}
294
295static void osc_extent_state_set(struct osc_extent *ext, int state)
296{
297 LASSERT(osc_object_is_locked(ext->oe_obj));
298 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
299
300 /* Never try to sanity check a state changing extent :-) */
301 /* LASSERT(sanity_check_nolock(ext) == 0); */
302
303 /* TODO: validate the state machine */
304 ext->oe_state = state;
305 wake_up_all(&ext->oe_waitq);
306}
307
308static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
309{
310 struct osc_extent *ext;
311
312 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
313 if (ext == NULL)
314 return NULL;
315
316 RB_CLEAR_NODE(&ext->oe_node);
317 ext->oe_obj = obj;
318 atomic_set(&ext->oe_refc, 1);
319 atomic_set(&ext->oe_users, 0);
320 INIT_LIST_HEAD(&ext->oe_link);
321 ext->oe_state = OES_INV;
322 INIT_LIST_HEAD(&ext->oe_pages);
323 init_waitqueue_head(&ext->oe_waitq);
324 ext->oe_osclock = NULL;
325
326 return ext;
327}
328
329static void osc_extent_free(struct osc_extent *ext)
330{
331 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
332}
333
334static struct osc_extent *osc_extent_get(struct osc_extent *ext)
335{
336 LASSERT(atomic_read(&ext->oe_refc) >= 0);
337 atomic_inc(&ext->oe_refc);
338 return ext;
339}
340
341static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
342{
343 LASSERT(atomic_read(&ext->oe_refc) > 0);
344 if (atomic_dec_and_test(&ext->oe_refc)) {
345 LASSERT(list_empty(&ext->oe_link));
346 LASSERT(atomic_read(&ext->oe_users) == 0);
347 LASSERT(ext->oe_state == OES_INV);
348 LASSERT(!ext->oe_intree);
349
350 if (ext->oe_osclock) {
351 cl_lock_put(env, ext->oe_osclock);
352 ext->oe_osclock = NULL;
353 }
354 osc_extent_free(ext);
355 }
356}
357
358/**
359 * osc_extent_put_trust() is a special version of osc_extent_put() when
360 * it's known that the caller is not the last user. This is to address the
361 * problem of lacking of lu_env ;-).
362 */
363static void osc_extent_put_trust(struct osc_extent *ext)
364{
365 LASSERT(atomic_read(&ext->oe_refc) > 1);
366 LASSERT(osc_object_is_locked(ext->oe_obj));
367 atomic_dec(&ext->oe_refc);
368}
369
370/**
371 * Return the extent which includes pgoff @index, or return the greatest
372 * previous extent in the tree.
373 */
374static struct osc_extent *osc_extent_search(struct osc_object *obj,
375 pgoff_t index)
376{
377 struct rb_node *n = obj->oo_root.rb_node;
378 struct osc_extent *tmp, *p = NULL;
379
380 LASSERT(osc_object_is_locked(obj));
381 while (n != NULL) {
382 tmp = rb_extent(n);
383 if (index < tmp->oe_start) {
384 n = n->rb_left;
385 } else if (index > tmp->oe_end) {
386 p = rb_extent(n);
387 n = n->rb_right;
388 } else {
389 return tmp;
390 }
391 }
392 return p;
393}
394
395/*
396 * Return the extent covering @index, otherwise return NULL.
397 * caller must have held object lock.
398 */
399static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
400 pgoff_t index)
401{
402 struct osc_extent *ext;
403
404 ext = osc_extent_search(obj, index);
405 if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
406 return osc_extent_get(ext);
407 return NULL;
408}
409
410/* caller must have held object lock. */
411static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
412{
413 struct rb_node **n = &obj->oo_root.rb_node;
414 struct rb_node *parent = NULL;
415 struct osc_extent *tmp;
416
417 LASSERT(ext->oe_intree == 0);
418 LASSERT(ext->oe_obj == obj);
419 LASSERT(osc_object_is_locked(obj));
420 while (*n != NULL) {
421 tmp = rb_extent(*n);
422 parent = *n;
423
424 if (ext->oe_end < tmp->oe_start)
425 n = &(*n)->rb_left;
426 else if (ext->oe_start > tmp->oe_end)
427 n = &(*n)->rb_right;
428 else
429 EASSERTF(0, tmp, EXTSTR, EXTPARA(ext));
430 }
431 rb_link_node(&ext->oe_node, parent, n);
432 rb_insert_color(&ext->oe_node, &obj->oo_root);
433 osc_extent_get(ext);
434 ext->oe_intree = 1;
435}
436
437/* caller must have held object lock. */
438static void osc_extent_erase(struct osc_extent *ext)
439{
440 struct osc_object *obj = ext->oe_obj;
441 LASSERT(osc_object_is_locked(obj));
442 if (ext->oe_intree) {
443 rb_erase(&ext->oe_node, &obj->oo_root);
444 ext->oe_intree = 0;
445 /* rbtree held a refcount */
446 osc_extent_put_trust(ext);
447 }
448}
449
450static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
451{
452 struct osc_object *obj = ext->oe_obj;
453
454 LASSERT(osc_object_is_locked(obj));
455 LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
456 if (ext->oe_state == OES_CACHE) {
457 osc_extent_state_set(ext, OES_ACTIVE);
458 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
459 }
460 atomic_inc(&ext->oe_users);
461 list_del_init(&ext->oe_link);
462 return osc_extent_get(ext);
463}
464
465static void __osc_extent_remove(struct osc_extent *ext)
466{
467 LASSERT(osc_object_is_locked(ext->oe_obj));
468 LASSERT(list_empty(&ext->oe_pages));
469 osc_extent_erase(ext);
470 list_del_init(&ext->oe_link);
471 osc_extent_state_set(ext, OES_INV);
472 OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
473}
474
475static void osc_extent_remove(struct osc_extent *ext)
476{
477 struct osc_object *obj = ext->oe_obj;
478
479 osc_object_lock(obj);
480 __osc_extent_remove(ext);
481 osc_object_unlock(obj);
482}
483
484/**
485 * This function is used to merge extents to get better performance. It checks
486 * if @cur and @victim are contiguous at chunk level.
487 */
488static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
489 struct osc_extent *victim)
490{
491 struct osc_object *obj = cur->oe_obj;
492 pgoff_t chunk_start;
493 pgoff_t chunk_end;
494 int ppc_bits;
495
496 LASSERT(cur->oe_state == OES_CACHE);
497 LASSERT(osc_object_is_locked(obj));
498 if (victim == NULL)
499 return -EINVAL;
500
501 if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
502 return -EBUSY;
503
504 if (cur->oe_max_end != victim->oe_max_end)
505 return -ERANGE;
506
507 LASSERT(cur->oe_osclock == victim->oe_osclock);
508 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
509 chunk_start = cur->oe_start >> ppc_bits;
510 chunk_end = cur->oe_end >> ppc_bits;
511 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
512 chunk_end + 1 != victim->oe_start >> ppc_bits)
513 return -ERANGE;
514
515 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
516
517 cur->oe_start = min(cur->oe_start, victim->oe_start);
518 cur->oe_end = max(cur->oe_end, victim->oe_end);
519 cur->oe_grants += victim->oe_grants;
520 cur->oe_nr_pages += victim->oe_nr_pages;
521 /* only the following bits are needed to merge */
522 cur->oe_urgent |= victim->oe_urgent;
523 cur->oe_memalloc |= victim->oe_memalloc;
524 list_splice_init(&victim->oe_pages, &cur->oe_pages);
525 list_del_init(&victim->oe_link);
526 victim->oe_nr_pages = 0;
527
528 osc_extent_get(victim);
529 __osc_extent_remove(victim);
530 osc_extent_put(env, victim);
531
532 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
533 return 0;
534}
535
536/**
537 * Drop user count of osc_extent, and unplug IO asynchronously.
538 */
539int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
540{
541 struct osc_object *obj = ext->oe_obj;
542 int rc = 0;
d7e09d03
PT		 543
544 LASSERT(atomic_read(&ext->oe_users) > 0);
545 LASSERT(sanity_check(ext) == 0);
546 LASSERT(ext->oe_grants > 0);
547
548 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
549 LASSERT(ext->oe_state == OES_ACTIVE);
550 if (ext->oe_trunc_pending) {
551 /* a truncate process is waiting for this extent.
552 * This may happen due to a race, check
553 * osc_cache_truncate_start(). */
554 osc_extent_state_set(ext, OES_TRUNC);
555 ext->oe_trunc_pending = 0;
556 } else {
557 osc_extent_state_set(ext, OES_CACHE);
558 osc_update_pending(obj, OBD_BRW_WRITE,
559 ext->oe_nr_pages);
560
561 /* try to merge the previous and next extent. */
562 osc_extent_merge(env, ext, prev_extent(ext));
563 osc_extent_merge(env, ext, next_extent(ext));
564
565 if (ext->oe_urgent)
566 list_move_tail(&ext->oe_link,
567 &obj->oo_urgent_exts);
568 }
569 osc_object_unlock(obj);
570
571 osc_io_unplug_async(env, osc_cli(obj), obj);
572 }
573 osc_extent_put(env, ext);
0a3bdb00 574 return rc;
d7e09d03
PT		 575}
576
577static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
578{
579 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
580}
581
582/**
583 * Find or create an extent which includes @index, core function to manage
584 * extent tree.
585 */
586struct osc_extent *osc_extent_find(const struct lu_env *env,
587 struct osc_object *obj, pgoff_t index,
588 int *grants)
589
590{
591 struct client_obd *cli = osc_cli(obj);
592 struct cl_lock *lock;
593 struct osc_extent *cur;
594 struct osc_extent *ext;
595 struct osc_extent *conflict = NULL;
596 struct osc_extent *found = NULL;
597 pgoff_t chunk;
598 pgoff_t max_end;
599 int max_pages; /* max_pages_per_rpc */
600 int chunksize;
601 int ppc_bits; /* pages per chunk bits */
602 int chunk_mask;
603 int rc;
d7e09d03
PT		 604
605 cur = osc_extent_alloc(obj);
606 if (cur == NULL)
0a3bdb00 607 return ERR_PTR(-ENOMEM);
d7e09d03
PT		 608
609 lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0);
610 LASSERT(lock != NULL);
611 LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
612
613 LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
614 ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
615 chunk_mask = ~((1 << ppc_bits) - 1);
616 chunksize = 1 << cli->cl_chunkbits;
617 chunk = index >> ppc_bits;
618
619 /* align end to rpc edge, rpc size may not be a power 2 integer. */
620 max_pages = cli->cl_max_pages_per_rpc;
621 LASSERT((max_pages & ~chunk_mask) == 0);
622 max_end = index - (index % max_pages) + max_pages - 1;
623 max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end);
624
625 /* initialize new extent by parameters so far */
626 cur->oe_max_end = max_end;
627 cur->oe_start = index & chunk_mask;
628 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
629 if (cur->oe_start < lock->cll_descr.cld_start)
630 cur->oe_start = lock->cll_descr.cld_start;
631 if (cur->oe_end > max_end)
632 cur->oe_end = max_end;
633 cur->oe_osclock = lock;
634 cur->oe_grants = 0;
635 cur->oe_mppr = max_pages;
636
637 /* grants has been allocated by caller */
638 LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
639 "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
640 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur));
641
642restart:
643 osc_object_lock(obj);
644 ext = osc_extent_search(obj, cur->oe_start);
645 if (ext == NULL)
646 ext = first_extent(obj);
647 while (ext != NULL) {
648 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
649 loff_t ext_chk_end = ext->oe_end >> ppc_bits;
650
651 LASSERT(sanity_check_nolock(ext) == 0);
652 if (chunk > ext_chk_end + 1)
653 break;
654
655 /* if covering by different locks, no chance to match */
656 if (lock != ext->oe_osclock) {
657 EASSERTF(!overlapped(ext, cur), ext,
658 EXTSTR, EXTPARA(cur));
659
660 ext = next_extent(ext);
661 continue;
662 }
663
664 /* discontiguous chunks? */
665 if (chunk + 1 < ext_chk_start) {
666 ext = next_extent(ext);
667 continue;
668 }
669
670 /* ok, from now on, ext and cur have these attrs:
671 * 1. covered by the same lock
672 * 2. contiguous at chunk level or overlapping. */
673
674 if (overlapped(ext, cur)) {
675 /* cur is the minimum unit, so overlapping means
676 * full contain. */
677 EASSERTF((ext->oe_start <= cur->oe_start &&
678 ext->oe_end >= cur->oe_end),
679 ext, EXTSTR, EXTPARA(cur));
680
681 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
682 /* for simplicity, we wait for this extent to
683 * finish before going forward. */
684 conflict = osc_extent_get(ext);
685 break;
686 }
687
688 found = osc_extent_hold(ext);
689 break;
690 }
691
692 /* non-overlapped extent */
693 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
694 /* we can't do anything for a non OES_CACHE extent, or
695 * if there is someone waiting for this extent to be
696 * flushed, try next one. */
697 ext = next_extent(ext);
698 continue;
699 }
700
701 /* check if they belong to the same rpc slot before trying to
702 * merge. the extents are not overlapped and contiguous at
703 * chunk level to get here. */
704 if (ext->oe_max_end != max_end) {
705 /* if they don't belong to the same RPC slot or
706 * max_pages_per_rpc has ever changed, do not merge. */
707 ext = next_extent(ext);
708 continue;
709 }
710
711 /* it's required that an extent must be contiguous at chunk
712 * level so that we know the whole extent is covered by grant
713 * (the pages in the extent are NOT required to be contiguous).
714 * Otherwise, it will be too much difficult to know which
715 * chunks have grants allocated. */
716
717 /* try to do front merge - extend ext's start */
718 if (chunk + 1 == ext_chk_start) {
719 /* ext must be chunk size aligned */
720 EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
721
722 /* pull ext's start back to cover cur */
723 ext->oe_start = cur->oe_start;
724 ext->oe_grants += chunksize;
725 *grants -= chunksize;
726
727 found = osc_extent_hold(ext);
728 } else if (chunk == ext_chk_end + 1) {
729 /* rear merge */
730 ext->oe_end = cur->oe_end;
731 ext->oe_grants += chunksize;
732 *grants -= chunksize;
733
734 /* try to merge with the next one because we just fill
735 * in a gap */
736 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
737 /* we can save extent tax from next extent */
738 *grants += cli->cl_extent_tax;
739
740 found = osc_extent_hold(ext);
741 }
742 if (found != NULL)
743 break;
744
745 ext = next_extent(ext);
746 }
747
748 osc_extent_tree_dump(D_CACHE, obj);
749 if (found != NULL) {
750 LASSERT(conflict == NULL);
751 if (!IS_ERR(found)) {
752 LASSERT(found->oe_osclock == cur->oe_osclock);
753 OSC_EXTENT_DUMP(D_CACHE, found,
754 "found caching ext for %lu.\n", index);
755 }
756 } else if (conflict == NULL) {
757 /* create a new extent */
758 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
759 cur->oe_grants = chunksize + cli->cl_extent_tax;
760 *grants -= cur->oe_grants;
761 LASSERT(*grants >= 0);
762
763 cur->oe_state = OES_CACHE;
764 found = osc_extent_hold(cur);
765 osc_extent_insert(obj, cur);
766 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
767 index, lock->cll_descr.cld_end);
768 }
769 osc_object_unlock(obj);
770
771 if (conflict != NULL) {
772 LASSERT(found == NULL);
773
774 /* waiting for IO to finish. Please notice that it's impossible
775 * to be an OES_TRUNC extent. */
776 rc = osc_extent_wait(env, conflict, OES_INV);
777 osc_extent_put(env, conflict);
778 conflict = NULL;
779 if (rc < 0)
780 GOTO(out, found = ERR_PTR(rc));
781
782 goto restart;
783 }
d7e09d03
PT		 784
785out:
786 osc_extent_put(env, cur);
787 LASSERT(*grants >= 0);
788 return found;
789}
790
791/**
792 * Called when IO is finished to an extent.
793 */
794int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
795 int sent, int rc)
796{
797 struct client_obd *cli = osc_cli(ext->oe_obj);
798 struct osc_async_page *oap;
799 struct osc_async_page *tmp;
800 int nr_pages = ext->oe_nr_pages;
801 int lost_grant = 0;
802 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
803 __u64 last_off = 0;
804 int last_count = -1;
d7e09d03
PT		 805
806 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
807
808 ext->oe_rc = rc ?: ext->oe_nr_pages;
809 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
810 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
811 oap_pending_item) {
812 list_del_init(&oap->oap_rpc_item);
813 list_del_init(&oap->oap_pending_item);
814 if (last_off <= oap->oap_obj_off) {
815 last_off = oap->oap_obj_off;
816 last_count = oap->oap_count;
817 }
818
819 --ext->oe_nr_pages;
820 osc_ap_completion(env, cli, oap, sent, rc);
821 }
822 EASSERT(ext->oe_nr_pages == 0, ext);
823
824 if (!sent) {
825 lost_grant = ext->oe_grants;
826 } else if (blocksize < PAGE_CACHE_SIZE &&
827 last_count != PAGE_CACHE_SIZE) {
828 /* For short writes we shouldn't count parts of pages that
829 * span a whole chunk on the OST side, or our accounting goes
830 * wrong. Should match the code in filter_grant_check. */
831 int offset = oap->oap_page_off & ~CFS_PAGE_MASK;
832 int count = oap->oap_count + (offset & (blocksize - 1));
833 int end = (offset + oap->oap_count) & (blocksize - 1);
834 if (end)
835 count += blocksize - end;
836
837 lost_grant = PAGE_CACHE_SIZE - count;
838 }
839 if (ext->oe_grants > 0)
840 osc_free_grant(cli, nr_pages, lost_grant);
841
842 osc_extent_remove(ext);
843 /* put the refcount for RPC */
844 osc_extent_put(env, ext);
0a3bdb00 845 return 0;
d7e09d03
PT		 846}
847
848static int extent_wait_cb(struct osc_extent *ext, int state)
849{
850 int ret;
851
852 osc_object_lock(ext->oe_obj);
853 ret = ext->oe_state == state;
854 osc_object_unlock(ext->oe_obj);
855
856 return ret;
857}
858
859/**
860 * Wait for the extent's state to become @state.
861 */
862static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
863 int state)
864{
865 struct osc_object *obj = ext->oe_obj;
866 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
867 LWI_ON_SIGNAL_NOOP, NULL);
868 int rc = 0;
d7e09d03
PT		 869
870 osc_object_lock(obj);
871 LASSERT(sanity_check_nolock(ext) == 0);
872 /* `Kick' this extent only if the caller is waiting for it to be
873 * written out. */
ce248d59
AS		 874 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp &&
875 !ext->oe_trunc_pending) {
d7e09d03
PT		 876 if (ext->oe_state == OES_ACTIVE) {
877 ext->oe_urgent = 1;
878 } else if (ext->oe_state == OES_CACHE) {
879 ext->oe_urgent = 1;
880 osc_extent_hold(ext);
881 rc = 1;
882 }
883 }
884 osc_object_unlock(obj);
885 if (rc == 1)
886 osc_extent_release(env, ext);
887
888 /* wait for the extent until its state becomes @state */
889 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
890 if (rc == -ETIMEDOUT) {
891 OSC_EXTENT_DUMP(D_ERROR, ext,
892 "%s: wait ext to %d timedout, recovery in progress?\n",
893 osc_export(obj)->exp_obd->obd_name, state);
894
895 lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
896 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
897 &lwi);
898 }
899 if (rc == 0 && ext->oe_rc < 0)
900 rc = ext->oe_rc;
0a3bdb00 901 return rc;
d7e09d03
PT		 902}
903
904/**
905 * Discard pages with index greater than @size. If @ext is overlapped with
906 * @size, then partial truncate happens.
907 */
908static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
909 bool partial)
910{
911 struct cl_env_nest nest;
912 struct lu_env *env;
913 struct cl_io *io;
914 struct osc_object *obj = ext->oe_obj;
915 struct client_obd *cli = osc_cli(obj);
916 struct osc_async_page *oap;
917 struct osc_async_page *tmp;
918 int pages_in_chunk = 0;
919 int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
920 __u64 trunc_chunk = trunc_index >> ppc_bits;
921 int grants = 0;
922 int nr_pages = 0;
923 int rc = 0;
d7e09d03
PT		 924
925 LASSERT(sanity_check(ext) == 0);
ce248d59
AS		 926 EASSERT(ext->oe_state == OES_TRUNC, ext);
927 EASSERT(!ext->oe_urgent, ext);
d7e09d03
PT		 928
929 /* Request new lu_env.
930 * We can't use that env from osc_cache_truncate_start() because
931 * it's from lov_io_sub and not fully initialized. */
932 env = cl_env_nested_get(&nest);
933 io = &osc_env_info(env)->oti_io;
934 io->ci_obj = cl_object_top(osc2cl(obj));
935 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
936 if (rc < 0)
937 GOTO(out, rc);
938
939 /* discard all pages with index greater then trunc_index */
940 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
941 oap_pending_item) {
942 struct cl_page *sub = oap2cl_page(oap);
943 struct cl_page *page = cl_page_top(sub);
944
945 LASSERT(list_empty(&oap->oap_rpc_item));
946
947 /* only discard the pages with their index greater than
948 * trunc_index, and ... */
949 if (sub->cp_index < trunc_index ||
950 (sub->cp_index == trunc_index && partial)) {
951 /* accounting how many pages remaining in the chunk
952 * so that we can calculate grants correctly. */
953 if (sub->cp_index >> ppc_bits == trunc_chunk)
954 ++pages_in_chunk;
955 continue;
956 }
957
958 list_del_init(&oap->oap_pending_item);
959
960 cl_page_get(page);
961 lu_ref_add(&page->cp_reference, "truncate", current);
962
963 if (cl_page_own(env, io, page) == 0) {
964 cl_page_unmap(env, io, page);
965 cl_page_discard(env, io, page);
966 cl_page_disown(env, io, page);
967 } else {
968 LASSERT(page->cp_state == CPS_FREEING);
969 LASSERT(0);
970 }
971
972 lu_ref_del(&page->cp_reference, "truncate", current);
973 cl_page_put(env, page);
974
975 --ext->oe_nr_pages;
976 ++nr_pages;
977 }
978 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
979 ext->oe_nr_pages == 0),
980 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
981
982 osc_object_lock(obj);
983 if (ext->oe_nr_pages == 0) {
984 LASSERT(pages_in_chunk == 0);
985 grants = ext->oe_grants;
986 ext->oe_grants = 0;
987 } else { /* calculate how many grants we can free */
988 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
989 pgoff_t last_index;
990
991
992 /* if there is no pages in this chunk, we can also free grants
993 * for the last chunk */
994 if (pages_in_chunk == 0) {
995 /* if this is the 1st chunk and no pages in this chunk,
996 * ext->oe_nr_pages must be zero, so we should be in
997 * the other if-clause. */
998 LASSERT(trunc_chunk > 0);
999 --trunc_chunk;
1000 ++chunks;
1001 }
1002
1003 /* this is what we can free from this extent */
1004 grants = chunks << cli->cl_chunkbits;
1005 ext->oe_grants -= grants;
1006 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1007 ext->oe_end = min(last_index, ext->oe_max_end);
1008 LASSERT(ext->oe_end >= ext->oe_start);
1009 LASSERT(ext->oe_grants > 0);
1010 }
1011 osc_object_unlock(obj);
1012
1013 if (grants > 0 || nr_pages > 0)
1014 osc_free_grant(cli, nr_pages, grants);
1015
1016out:
1017 cl_io_fini(env, io);
1018 cl_env_nested_put(&nest, env);
0a3bdb00 1019 return rc;
d7e09d03
PT		 1020}
1021
1022/**
1023 * This function is used to make the extent prepared for transfer.
34ca8748 1024 * A race with flushing page - ll_writepage() has to be handled cautiously.
d7e09d03
PT		 1025 */
1026static int osc_extent_make_ready(const struct lu_env *env,
1027 struct osc_extent *ext)
1028{
1029 struct osc_async_page *oap;
1030 struct osc_async_page *last = NULL;
1031 struct osc_object *obj = ext->oe_obj;
1032 int page_count = 0;
1033 int rc;
d7e09d03
PT		 1034
1035 /* we're going to grab page lock, so object lock must not be taken. */
1036 LASSERT(sanity_check(ext) == 0);
1037 /* in locking state, any process should not touch this extent. */
1038 EASSERT(ext->oe_state == OES_LOCKING, ext);
1039 EASSERT(ext->oe_owner != NULL, ext);
1040
1041 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1042
1043 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1044 ++page_count;
1045 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1046 last = oap;
1047
1048 /* checking ASYNC_READY is race safe */
1049 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1050 continue;
1051
1052 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1053 switch (rc) {
1054 case 0:
1055 spin_lock(&oap->oap_lock);
1056 oap->oap_async_flags |= ASYNC_READY;
1057 spin_unlock(&oap->oap_lock);
1058 break;
1059 case -EALREADY:
1060 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1061 break;
1062 default:
1063 LASSERTF(0, "unknown return code: %d\n", rc);
1064 }
1065 }
1066
1067 LASSERT(page_count == ext->oe_nr_pages);
1068 LASSERT(last != NULL);
1069 /* the last page is the only one we need to refresh its count by
1070 * the size of file. */
1071 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1072 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1073 LASSERT(last->oap_count > 0);
1074 LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
1075 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1076 }
1077
1078 /* for the rest of pages, we don't need to call osf_refresh_count()
1079 * because it's known they are not the last page */
1080 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1081 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1082 oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1083 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1084 }
1085 }
1086
1087 osc_object_lock(obj);
1088 osc_extent_state_set(ext, OES_RPC);
1089 osc_object_unlock(obj);
1090 /* get a refcount for RPC. */
1091 osc_extent_get(ext);
1092
0a3bdb00 1093 return 0;
d7e09d03
PT		 1094}
1095
1096/**
1097 * Quick and simple version of osc_extent_find(). This function is frequently
1098 * called to expand the extent for the same IO. To expand the extent, the
1099 * page index must be in the same or next chunk of ext->oe_end.
1100 */
1101static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1102{
1103 struct osc_object *obj = ext->oe_obj;
1104 struct client_obd *cli = osc_cli(obj);
1105 struct osc_extent *next;
1106 int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1107 pgoff_t chunk = index >> ppc_bits;
1108 pgoff_t end_chunk;
1109 pgoff_t end_index;
1110 int chunksize = 1 << cli->cl_chunkbits;
1111 int rc = 0;
d7e09d03
PT		 1112
1113 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1114 osc_object_lock(obj);
1115 LASSERT(sanity_check_nolock(ext) == 0);
1116 end_chunk = ext->oe_end >> ppc_bits;
1117 if (chunk > end_chunk + 1)
1118 GOTO(out, rc = -ERANGE);
1119
1120 if (end_chunk >= chunk)
1121 GOTO(out, rc = 0);
1122
1123 LASSERT(end_chunk + 1 == chunk);
1124 /* try to expand this extent to cover @index */
1125 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1126
1127 next = next_extent(ext);
1128 if (next != NULL && next->oe_start <= end_index)
1129 /* complex mode - overlapped with the next extent,
1130 * this case will be handled by osc_extent_find() */
1131 GOTO(out, rc = -EAGAIN);
1132
1133 ext->oe_end = end_index;
1134 ext->oe_grants += chunksize;
1135 *grants -= chunksize;
1136 LASSERT(*grants >= 0);
1137 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1138 "overlapped after expanding for %lu.\n", index);
d7e09d03
PT		 1139
1140out:
1141 osc_object_unlock(obj);
0a3bdb00 1142 return rc;
d7e09d03
PT		 1143}
1144
1145static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1146 const char *func, int line)
1147{
1148 struct osc_extent *ext;
1149 int cnt;
1150
1151 CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1152 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1153
1154 /* osc_object_lock(obj); */
1155 cnt = 1;
1156 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1157 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1158
1159 cnt = 1;
1160 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1161 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1162
1163 cnt = 1;
1164 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1165 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1166
1167 cnt = 1;
1168 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1169 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1170 /* osc_object_unlock(obj); */
1171}
1172
1173/* ------------------ osc extent end ------------------ */
1174
1175static inline int osc_is_ready(struct osc_object *osc)
1176{
1177 return !list_empty(&osc->oo_ready_item) ||
1178 !list_empty(&osc->oo_hp_ready_item);
1179}
1180
1181#define OSC_IO_DEBUG(OSC, STR, args...) \
1182 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1183 (OSC), osc_is_ready(OSC), \
1184 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1185 list_empty_marker(&(OSC)->oo_ready_item), \
1186 atomic_read(&(OSC)->oo_nr_writes), \
1187 list_empty_marker(&(OSC)->oo_hp_exts), \
1188 list_empty_marker(&(OSC)->oo_urgent_exts), \
1189 atomic_read(&(OSC)->oo_nr_reads), \
1190 list_empty_marker(&(OSC)->oo_reading_exts), \
1191 ##args)
1192
1193static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1194 int cmd)
1195{
1196 struct osc_page *opg = oap2osc_page(oap);
1197 struct cl_page *page = cl_page_top(oap2cl_page(oap));
1198 int result;
1199
1200 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1201
d7e09d03
PT		 1202 result = cl_page_make_ready(env, page, CRT_WRITE);
1203 if (result == 0)
1204 opg->ops_submit_time = cfs_time_current();
0a3bdb00 1205 return result;
d7e09d03
PT		 1206}
1207
1208static int osc_refresh_count(const struct lu_env *env,
1209 struct osc_async_page *oap, int cmd)
1210{
1211 struct osc_page *opg = oap2osc_page(oap);
1212 struct cl_page *page = oap2cl_page(oap);
1213 struct cl_object *obj;
1214 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1215
1216 int result;
1217 loff_t kms;
1218
1219 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1220 LASSERT(!(cmd & OBD_BRW_READ));
1221 LASSERT(opg != NULL);
1222 obj = opg->ops_cl.cpl_obj;
1223
1224 cl_object_attr_lock(obj);
1225 result = cl_object_attr_get(env, obj, attr);
1226 cl_object_attr_unlock(obj);
1227 if (result < 0)
1228 return result;
1229 kms = attr->cat_kms;
1230 if (cl_offset(obj, page->cp_index) >= kms)
1231 /* catch race with truncate */
1232 return 0;
1233 else if (cl_offset(obj, page->cp_index + 1) > kms)
1234 /* catch sub-page write at end of file */
1235 return kms % PAGE_CACHE_SIZE;
1236 else
1237 return PAGE_CACHE_SIZE;
1238}
1239
1240static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1241 int cmd, int rc)
1242{
1243 struct osc_page *opg = oap2osc_page(oap);
1244 struct cl_page *page = cl_page_top(oap2cl_page(oap));
1245 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
1246 enum cl_req_type crt;
1247 int srvlock;
1248
d7e09d03
PT		 1249 cmd &= ~OBD_BRW_NOQUOTA;
1250 LASSERT(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ));
1251 LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
1252 LASSERT(opg->ops_transfer_pinned);
1253
1254 /*
1255 * page->cp_req can be NULL if io submission failed before
1256 * cl_req was allocated.
1257 */
1258 if (page->cp_req != NULL)
1259 cl_req_page_done(env, page);
1260 LASSERT(page->cp_req == NULL);
1261
1262 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1263 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1264 opg->ops_transfer_pinned = 0;
1265
1266 spin_lock(&obj->oo_seatbelt);
1267 LASSERT(opg->ops_submitter != NULL);
1268 LASSERT(!list_empty(&opg->ops_inflight));
1269 list_del_init(&opg->ops_inflight);
1270 opg->ops_submitter = NULL;
1271 spin_unlock(&obj->oo_seatbelt);
1272
1273 opg->ops_submit_time = 0;
1274 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1275
1276 /* statistic */
1277 if (rc == 0 && srvlock) {
1278 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1279 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1280 int bytes = oap->oap_count;
1281
1282 if (crt == CRT_READ)
1283 stats->os_lockless_reads += bytes;
1284 else
1285 stats->os_lockless_writes += bytes;
1286 }
1287
1288 /*
1289 * This has to be the last operation with the page, as locks are
1290 * released in cl_page_completion() and nothing except for the
1291 * reference counter protects page from concurrent reclaim.
1292 */
1293 lu_ref_del(&page->cp_reference, "transfer", page);
1294
1295 cl_page_completion(env, page, crt, rc);
1296
0a3bdb00 1297 return 0;
d7e09d03
PT		 1298}
1299
1300#define OSC_DUMP_GRANT(cli, fmt, args...) do { \
1301 struct client_obd *__tmp = (cli); \
1302 CDEBUG(D_CACHE, "%s: { dirty: %ld/%ld dirty_pages: %d/%d " \
c52f69c5 1303 "dropped: %ld avail: %ld, reserved: %ld, flight: %d } " fmt, \
d7e09d03
PT		 1304 __tmp->cl_import->imp_obd->obd_name, \
1305 __tmp->cl_dirty, __tmp->cl_dirty_max, \
1306 atomic_read(&obd_dirty_pages), obd_max_dirty_pages, \
d7e09d03
PT		 1307 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1308 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, ##args); \
1309} while (0)
1310
1311/* caller must hold loi_list_lock */
1312static void osc_consume_write_grant(struct client_obd *cli,
1313 struct brw_page *pga)
1314{
5e42bc9d 1315 assert_spin_locked(&cli->cl_loi_list_lock.lock);
d7e09d03
PT		 1316 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1317 atomic_inc(&obd_dirty_pages);
1318 cli->cl_dirty += PAGE_CACHE_SIZE;
1319 pga->flag |= OBD_BRW_FROM_GRANT;
1320 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1321 PAGE_CACHE_SIZE, pga, pga->pg);
1322 osc_update_next_shrink(cli);
1323}
1324
1325/* the companion to osc_consume_write_grant, called when a brw has completed.
1326 * must be called with the loi lock held. */
1327static void osc_release_write_grant(struct client_obd *cli,
1328 struct brw_page *pga)
1329{
5e42bc9d 1330 assert_spin_locked(&cli->cl_loi_list_lock.lock);
d7e09d03 1331 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
d7e09d03
PT		 1332 return;
1333 }
1334
1335 pga->flag &= ~OBD_BRW_FROM_GRANT;
1336 atomic_dec(&obd_dirty_pages);
1337 cli->cl_dirty -= PAGE_CACHE_SIZE;
1338 if (pga->flag & OBD_BRW_NOCACHE) {
1339 pga->flag &= ~OBD_BRW_NOCACHE;
1340 atomic_dec(&obd_dirty_transit_pages);
1341 cli->cl_dirty_transit -= PAGE_CACHE_SIZE;
1342 }
d7e09d03
PT		 1343}
1344
1345/**
1346 * To avoid sleeping with object lock held, it's good for us allocate enough
1347 * grants before entering into critical section.
1348 *
1349 * client_obd_list_lock held by caller
1350 */
1351static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1352{
1353 int rc = -EDQUOT;
1354
1355 if (cli->cl_avail_grant >= bytes) {
1356 cli->cl_avail_grant -= bytes;
1357 cli->cl_reserved_grant += bytes;
1358 rc = 0;
1359 }
1360 return rc;
1361}
1362
1363static void __osc_unreserve_grant(struct client_obd *cli,
1364 unsigned int reserved, unsigned int unused)
1365{
1366 /* it's quite normal for us to get more grant than reserved.
1367 * Thinking about a case that two extents merged by adding a new
1368 * chunk, we can save one extent tax. If extent tax is greater than
1369 * one chunk, we can save more grant by adding a new chunk */
1370 cli->cl_reserved_grant -= reserved;
1371 if (unused > reserved) {
1372 cli->cl_avail_grant += reserved;
1373 cli->cl_lost_grant += unused - reserved;
1374 } else {
1375 cli->cl_avail_grant += unused;
1376 }
1377}
1378
1379void osc_unreserve_grant(struct client_obd *cli,
1380 unsigned int reserved, unsigned int unused)
1381{
1382 client_obd_list_lock(&cli->cl_loi_list_lock);
1383 __osc_unreserve_grant(cli, reserved, unused);
1384 if (unused > 0)
1385 osc_wake_cache_waiters(cli);
1386 client_obd_list_unlock(&cli->cl_loi_list_lock);
1387}
1388
1389/**
1390 * Free grant after IO is finished or canceled.
1391 *
1392 * @lost_grant is used to remember how many grants we have allocated but not
1393 * used, we should return these grants to OST. There're two cases where grants
1394 * can be lost:
1395 * 1. truncate;
1396 * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1397 * written. In this case OST may use less chunks to serve this partial
1398 * write. OSTs don't actually know the page size on the client side. so
1399 * clients have to calculate lost grant by the blocksize on the OST.
1400 * See filter_grant_check() for details.
1401 */
1402static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1403 unsigned int lost_grant)
1404{
1405 int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1406
1407 client_obd_list_lock(&cli->cl_loi_list_lock);
1408 atomic_sub(nr_pages, &obd_dirty_pages);
1409 cli->cl_dirty -= nr_pages << PAGE_CACHE_SHIFT;
1410 cli->cl_lost_grant += lost_grant;
1411 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1412 /* borrow some grant from truncate to avoid the case that
1413 * truncate uses up all avail grant */
1414 cli->cl_lost_grant -= grant;
1415 cli->cl_avail_grant += grant;
1416 }
1417 osc_wake_cache_waiters(cli);
1418 client_obd_list_unlock(&cli->cl_loi_list_lock);
1419 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1420 lost_grant, cli->cl_lost_grant,
1421 cli->cl_avail_grant, cli->cl_dirty);
1422}
1423
1424/**
1425 * The companion to osc_enter_cache(), called when @oap is no longer part of
1426 * the dirty accounting due to error.
1427 */
1428static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1429{
1430 client_obd_list_lock(&cli->cl_loi_list_lock);
1431 osc_release_write_grant(cli, &oap->oap_brw_page);
1432 client_obd_list_unlock(&cli->cl_loi_list_lock);
1433}
1434
1435/**
1436 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1437 * is available.
1438 */
1439static int osc_enter_cache_try(struct client_obd *cli,
1440 struct osc_async_page *oap,
1441 int bytes, int transient)
1442{
1443 int rc;
1444
1445 OSC_DUMP_GRANT(cli, "need:%d.\n", bytes);
1446
1447 rc = osc_reserve_grant(cli, bytes);
1448 if (rc < 0)
1449 return 0;
1450
1451 if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max &&
c52f69c5 1452 atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) {
d7e09d03
PT		 1453 osc_consume_write_grant(cli, &oap->oap_brw_page);
1454 if (transient) {
1455 cli->cl_dirty_transit += PAGE_CACHE_SIZE;
1456 atomic_inc(&obd_dirty_transit_pages);
1457 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1458 }
1459 rc = 1;
1460 } else {
1461 __osc_unreserve_grant(cli, bytes, bytes);
1462 rc = 0;
1463 }
1464 return rc;
1465}
1466
1467static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1468{
1469 int rc;
1470 client_obd_list_lock(&cli->cl_loi_list_lock);
1471 rc = list_empty(&ocw->ocw_entry);
1472 client_obd_list_unlock(&cli->cl_loi_list_lock);
1473 return rc;
1474}
1475
1476/**
1477 * The main entry to reserve dirty page accounting. Usually the grant reserved
1478 * in this function will be freed in bulk in osc_free_grant() unless it fails
1479 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1480 *
1481 * The process will be put into sleep if it's already run out of grant.
1482 */
1483static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1484 struct osc_async_page *oap, int bytes)
1485{
1486 struct osc_object *osc = oap->oap_obj;
1487 struct lov_oinfo *loi = osc->oo_oinfo;
1488 struct osc_cache_waiter ocw;
1489 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1490 int rc = -EDQUOT;
d7e09d03
PT		 1491
1492 OSC_DUMP_GRANT(cli, "need:%d.\n", bytes);
1493
1494 client_obd_list_lock(&cli->cl_loi_list_lock);
1495
1496 /* force the caller to try sync io. this can jump the list
1497 * of queued writes and create a discontiguous rpc stream */
1498 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1499 cli->cl_dirty_max < PAGE_CACHE_SIZE ||
1500 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1501 GOTO(out, rc = -EDQUOT);
1502
1503 /* Hopefully normal case - cache space and write credits available */
1504 if (osc_enter_cache_try(cli, oap, bytes, 0))
1505 GOTO(out, rc = 0);
1506
1507 /* We can get here for two reasons: too many dirty pages in cache, or
1508 * run out of grants. In both cases we should write dirty pages out.
1509 * Adding a cache waiter will trigger urgent write-out no matter what
1510 * RPC size will be.
1511 * The exiting condition is no avail grants and no dirty pages caching,
1512 * that really means there is no space on the OST. */
1513 init_waitqueue_head(&ocw.ocw_waitq);
1514 ocw.ocw_oap = oap;
1515 ocw.ocw_grant = bytes;
1516 while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
1517 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1518 ocw.ocw_rc = 0;
1519 client_obd_list_unlock(&cli->cl_loi_list_lock);
1520
1521 osc_io_unplug_async(env, cli, NULL);
1522
1523 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1524 cli->cl_import->imp_obd->obd_name, &ocw, oap);
1525
1526 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1527
1528 client_obd_list_lock(&cli->cl_loi_list_lock);
1529
1530 /* l_wait_event is interrupted by signal */
1531 if (rc < 0) {
1532 list_del_init(&ocw.ocw_entry);
1533 GOTO(out, rc);
1534 }
1535
1536 LASSERT(list_empty(&ocw.ocw_entry));
1537 rc = ocw.ocw_rc;
1538
1539 if (rc != -EDQUOT)
1540 GOTO(out, rc);
1541 if (osc_enter_cache_try(cli, oap, bytes, 0))
1542 GOTO(out, rc = 0);
1543 }
d7e09d03
PT		 1544out:
1545 client_obd_list_unlock(&cli->cl_loi_list_lock);
1546 OSC_DUMP_GRANT(cli, "returned %d.\n", rc);
0a3bdb00 1547 return rc;
d7e09d03
PT		 1548}
1549
1550/* caller must hold loi_list_lock */
1551void osc_wake_cache_waiters(struct client_obd *cli)
1552{
1553 struct list_head *l, *tmp;
1554 struct osc_cache_waiter *ocw;
1555
d7e09d03
PT		 1556 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1557 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1558 list_del_init(&ocw->ocw_entry);
1559
1560 ocw->ocw_rc = -EDQUOT;
1561 /* we can't dirty more */
c52f69c5
OD		 1562 if ((cli->cl_dirty + PAGE_CACHE_SIZE > cli->cl_dirty_max) ||
1563 (atomic_read(&obd_dirty_pages) + 1 >
1564 obd_max_dirty_pages)) {
d7e09d03
PT		 1565 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1566 "osc max %ld, sys max %d\n", cli->cl_dirty,
1567 cli->cl_dirty_max, obd_max_dirty_pages);
1568 goto wakeup;
1569 }
1570
1571 ocw->ocw_rc = 0;
1572 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1573 ocw->ocw_rc = -EDQUOT;
1574
1575wakeup:
1576 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1577 ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1578
1579 wake_up(&ocw->ocw_waitq);
1580 }
d7e09d03
PT		 1581}
1582
1583static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1584{
1585 int hprpc = !!list_empty(&osc->oo_hp_exts);
1586 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1587}
1588
1589/* This maintains the lists of pending pages to read/write for a given object
1590 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1591 * to quickly find objects that are ready to send an RPC. */
1592static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1593 int cmd)
1594{
1595 int invalid_import = 0;
d7e09d03
PT		 1596
1597 /* if we have an invalid import we want to drain the queued pages
1598 * by forcing them through rpcs that immediately fail and complete
1599 * the pages. recovery relies on this to empty the queued pages
1600 * before canceling the locks and evicting down the llite pages */
1601 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1602 invalid_import = 1;
1603
1604 if (cmd & OBD_BRW_WRITE) {
1605 if (atomic_read(&osc->oo_nr_writes) == 0)
0a3bdb00 1606 return 0;
d7e09d03
PT		 1607 if (invalid_import) {
1608 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
0a3bdb00 1609 return 1;
d7e09d03
PT		 1610 }
1611 if (!list_empty(&osc->oo_hp_exts)) {
1612 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
0a3bdb00 1613 return 1;
d7e09d03
PT		 1614 }
1615 if (!list_empty(&osc->oo_urgent_exts)) {
1616 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
0a3bdb00 1617 return 1;
d7e09d03
PT		 1618 }
1619 /* trigger a write rpc stream as long as there are dirtiers
1620 * waiting for space. as they're waiting, they're not going to
1621 * create more pages to coalesce with what's waiting.. */
1622 if (!list_empty(&cli->cl_cache_waiters)) {
1623 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
0a3bdb00 1624 return 1;
d7e09d03
PT		 1625 }
1626 if (atomic_read(&osc->oo_nr_writes) >=
1627 cli->cl_max_pages_per_rpc)
0a3bdb00 1628 return 1;
d7e09d03
PT		 1629 } else {
1630 if (atomic_read(&osc->oo_nr_reads) == 0)
0a3bdb00 1631 return 0;
d7e09d03
PT		 1632 if (invalid_import) {
1633 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
0a3bdb00 1634 return 1;
d7e09d03
PT		 1635 }
1636 /* all read are urgent. */
1637 if (!list_empty(&osc->oo_reading_exts))
0a3bdb00 1638 return 1;
d7e09d03
PT		 1639 }
1640
0a3bdb00 1641 return 0;
d7e09d03
PT		 1642}
1643
1644static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1645{
1646 struct client_obd *cli = osc_cli(obj);
1647 if (cmd & OBD_BRW_WRITE) {
1648 atomic_add(delta, &obj->oo_nr_writes);
1649 atomic_add(delta, &cli->cl_pending_w_pages);
1650 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1651 } else {
1652 atomic_add(delta, &obj->oo_nr_reads);
1653 atomic_add(delta, &cli->cl_pending_r_pages);
1654 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1655 }
1656 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1657}
1658
1659static int osc_makes_hprpc(struct osc_object *obj)
1660{
1661 return !list_empty(&obj->oo_hp_exts);
1662}
1663
1664static void on_list(struct list_head *item, struct list_head *list, int should_be_on)
1665{
1666 if (list_empty(item) && should_be_on)
1667 list_add_tail(item, list);
1668 else if (!list_empty(item) && !should_be_on)
1669 list_del_init(item);
1670}
1671
1672/* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1673 * can find pages to build into rpcs quickly */
1674static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1675{
1676 if (osc_makes_hprpc(osc)) {
1677 /* HP rpc */
1678 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1679 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1680 } else {
1681 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1682 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1683 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1684 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1685 }
1686
1687 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1688 atomic_read(&osc->oo_nr_writes) > 0);
1689
1690 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1691 atomic_read(&osc->oo_nr_reads) > 0);
1692
1693 return osc_is_ready(osc);
1694}
1695
1696static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1697{
1698 int is_ready;
1699
1700 client_obd_list_lock(&cli->cl_loi_list_lock);
1701 is_ready = __osc_list_maint(cli, osc);
1702 client_obd_list_unlock(&cli->cl_loi_list_lock);
1703
1704 return is_ready;
1705}
1706
11d66e89 1707/* this is trying to propagate async writeback errors back up to the
d7e09d03
PT		 1708 * application. As an async write fails we record the error code for later if
1709 * the app does an fsync. As long as errors persist we force future rpcs to be
1710 * sync so that the app can get a sync error and break the cycle of queueing
1711 * pages for which writeback will fail. */
1712static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1713 int rc)
1714{
1715 if (rc) {
1716 if (!ar->ar_rc)
1717 ar->ar_rc = rc;
1718
1719 ar->ar_force_sync = 1;
1720 ar->ar_min_xid = ptlrpc_sample_next_xid();
1721 return;
1722
1723 }
1724
1725 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1726 ar->ar_force_sync = 0;
1727}
1728
d7e09d03
PT		 1729
1730/* this must be called holding the loi list lock to give coverage to exit_cache,
1731 * async_flag maintenance, and oap_request */
1732static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1733 struct osc_async_page *oap, int sent, int rc)
1734{
1735 struct osc_object *osc = oap->oap_obj;
1736 struct lov_oinfo *loi = osc->oo_oinfo;
1737 __u64 xid = 0;
1738
d7e09d03 1739 if (oap->oap_request != NULL) {
d7e09d03
PT		 1740 xid = ptlrpc_req_xid(oap->oap_request);
1741 ptlrpc_req_finished(oap->oap_request);
1742 oap->oap_request = NULL;
1743 }
1744
1745 /* As the transfer for this page is being done, clear the flags */
1746 spin_lock(&oap->oap_lock);
1747 oap->oap_async_flags = 0;
1748 spin_unlock(&oap->oap_lock);
1749 oap->oap_interrupted = 0;
1750
1751 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1752 client_obd_list_lock(&cli->cl_loi_list_lock);
1753 osc_process_ar(&cli->cl_ar, xid, rc);
1754 osc_process_ar(&loi->loi_ar, xid, rc);
1755 client_obd_list_unlock(&cli->cl_loi_list_lock);
1756 }
1757
1758 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1759 if (rc)
1760 CERROR("completion on oap %p obj %p returns %d.\n",
1761 oap, osc, rc);
d7e09d03
PT		 1762}
1763
1764/**
1765 * Try to add extent to one RPC. We need to think about the following things:
1766 * - # of pages must not be over max_pages_per_rpc
1767 * - extent must be compatible with previous ones
1768 */
1769static int try_to_add_extent_for_io(struct client_obd *cli,
1770 struct osc_extent *ext, struct list_head *rpclist,
1771 int *pc, unsigned int *max_pages)
1772{
1773 struct osc_extent *tmp;
d7e09d03
PT		 1774
1775 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1776 ext);
1777
1778 *max_pages = max(ext->oe_mppr, *max_pages);
1779 if (*pc + ext->oe_nr_pages > *max_pages)
0a3bdb00 1780 return 0;
d7e09d03
PT		 1781
1782 list_for_each_entry(tmp, rpclist, oe_link) {
1783 EASSERT(tmp->oe_owner == current, tmp);
1784#if 0
1785 if (overlapped(tmp, ext)) {
1786 OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1787 EASSERT(0, ext);
1788 }
1789#endif
1790
1791 if (tmp->oe_srvlock != ext->oe_srvlock ||
1792 !tmp->oe_grants != !ext->oe_grants)
0a3bdb00 1793 return 0;
d7e09d03
PT		 1794
1795 /* remove break for strict check */
1796 break;
1797 }
1798
1799 *pc += ext->oe_nr_pages;
1800 list_move_tail(&ext->oe_link, rpclist);
1801 ext->oe_owner = current;
0a3bdb00 1802 return 1;
d7e09d03
PT		 1803}
1804
1805/**
1806 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1807 * get_write_extent() takes all appropriate extents in atomic.
1808 *
1809 * The following policy is used to collect extents for IO:
1810 * 1. Add as many HP extents as possible;
1811 * 2. Add the first urgent extent in urgent extent list and take it out of
1812 * urgent list;
1813 * 3. Add subsequent extents of this urgent extent;
1814 * 4. If urgent list is not empty, goto 2;
1815 * 5. Traverse the extent tree from the 1st extent;
1816 * 6. Above steps exit if there is no space in this RPC.
1817 */
1818static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1819{
1820 struct client_obd *cli = osc_cli(obj);
1821 struct osc_extent *ext;
1822 int page_count = 0;
1823 unsigned int max_pages = cli->cl_max_pages_per_rpc;
1824
1825 LASSERT(osc_object_is_locked(obj));
1826 while (!list_empty(&obj->oo_hp_exts)) {
1827 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1828 oe_link);
1829 LASSERT(ext->oe_state == OES_CACHE);
1830 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1831 &max_pages))
1832 return page_count;
1833 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1834 }
1835 if (page_count == max_pages)
1836 return page_count;
1837
1838 while (!list_empty(&obj->oo_urgent_exts)) {
1839 ext = list_entry(obj->oo_urgent_exts.next,
1840 struct osc_extent, oe_link);
1841 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1842 &max_pages))
1843 return page_count;
1844
1845 if (!ext->oe_intree)
1846 continue;
1847
1848 while ((ext = next_extent(ext)) != NULL) {
1849 if ((ext->oe_state != OES_CACHE) ||
1850 (!list_empty(&ext->oe_link) &&
1851 ext->oe_owner != NULL))
1852 continue;
1853
1854 if (!try_to_add_extent_for_io(cli, ext, rpclist,
1855 &page_count, &max_pages))
1856 return page_count;
1857 }
1858 }
1859 if (page_count == max_pages)
1860 return page_count;
1861
1862 ext = first_extent(obj);
1863 while (ext != NULL) {
1864 if ((ext->oe_state != OES_CACHE) ||
1865 /* this extent may be already in current rpclist */
1866 (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1867 ext = next_extent(ext);
1868 continue;
1869 }
1870
1871 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1872 &max_pages))
1873 return page_count;
1874
1875 ext = next_extent(ext);
1876 }
1877 return page_count;
1878}
1879
1880static int
1881osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1882 struct osc_object *osc, pdl_policy_t pol)
1883{
1884 LIST_HEAD(rpclist);
1885 struct osc_extent *ext;
1886 struct osc_extent *tmp;
1887 struct osc_extent *first = NULL;
21aef7d9 1888 u32 page_count = 0;
d7e09d03
PT		 1889 int srvlock = 0;
1890 int rc = 0;
d7e09d03
PT		 1891
1892 LASSERT(osc_object_is_locked(osc));
1893
1894 page_count = get_write_extents(osc, &rpclist);
1895 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1896
1897 if (list_empty(&rpclist))
0a3bdb00 1898 return 0;
d7e09d03
PT		 1899
1900 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1901
1902 list_for_each_entry(ext, &rpclist, oe_link) {
1903 LASSERT(ext->oe_state == OES_CACHE ||
1904 ext->oe_state == OES_LOCK_DONE);
1905 if (ext->oe_state == OES_CACHE)
1906 osc_extent_state_set(ext, OES_LOCKING);
1907 else
1908 osc_extent_state_set(ext, OES_RPC);
1909 }
1910
1911 /* we're going to grab page lock, so release object lock because
1912 * lock order is page lock -> object lock. */
1913 osc_object_unlock(osc);
1914
1915 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
1916 if (ext->oe_state == OES_LOCKING) {
1917 rc = osc_extent_make_ready(env, ext);
1918 if (unlikely(rc < 0)) {
1919 list_del_init(&ext->oe_link);
1920 osc_extent_finish(env, ext, 0, rc);
1921 continue;
1922 }
1923 }
1924 if (first == NULL) {
1925 first = ext;
1926 srvlock = ext->oe_srvlock;
1927 } else {
1928 LASSERT(srvlock == ext->oe_srvlock);
1929 }
1930 }
1931
1932 if (!list_empty(&rpclist)) {
1933 LASSERT(page_count > 0);
1934 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
1935 LASSERT(list_empty(&rpclist));
1936 }
1937
1938 osc_object_lock(osc);
0a3bdb00 1939 return rc;
d7e09d03
PT		 1940}
1941
1942/**
1943 * prepare pages for ASYNC io and put pages in send queue.
1944 *
1945 * \param cmd OBD_BRW_* macroses
1946 * \param lop pending pages
1947 *
1948 * \return zero if no page added to send queue.
1949 * \return 1 if pages successfully added to send queue.
1950 * \return negative on errors.
1951 */
1952static int
1953osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
1954 struct osc_object *osc, pdl_policy_t pol)
1955{
1956 struct osc_extent *ext;
1957 struct osc_extent *next;
1958 LIST_HEAD(rpclist);
1959 int page_count = 0;
1960 unsigned int max_pages = cli->cl_max_pages_per_rpc;
1961 int rc = 0;
d7e09d03
PT		 1962
1963 LASSERT(osc_object_is_locked(osc));
1964 list_for_each_entry_safe(ext, next,
1965 &osc->oo_reading_exts, oe_link) {
1966 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
1967 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
1968 &max_pages))
1969 break;
1970 osc_extent_state_set(ext, OES_RPC);
1971 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1972 }
1973 LASSERT(page_count <= max_pages);
1974
1975 osc_update_pending(osc, OBD_BRW_READ, -page_count);
1976
1977 if (!list_empty(&rpclist)) {
1978 osc_object_unlock(osc);
1979
1980 LASSERT(page_count > 0);
1981 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
1982 LASSERT(list_empty(&rpclist));
1983
1984 osc_object_lock(osc);
1985 }
0a3bdb00 1986 return rc;
d7e09d03
PT		 1987}
1988
1989#define list_to_obj(list, item) ({ \
1990 struct list_head *__tmp = (list)->next; \
1991 list_del_init(__tmp); \
1992 list_entry(__tmp, struct osc_object, oo_##item); \
1993})
1994
1995/* This is called by osc_check_rpcs() to find which objects have pages that
1996 * we could be sending. These lists are maintained by osc_makes_rpc(). */
1997static struct osc_object *osc_next_obj(struct client_obd *cli)
1998{
d7e09d03
PT		 1999 /* First return objects that have blocked locks so that they
2000 * will be flushed quickly and other clients can get the lock,
2001 * then objects which have pages ready to be stuffed into RPCs */
2002 if (!list_empty(&cli->cl_loi_hp_ready_list))
0a3bdb00 2003 return list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item);
d7e09d03 2004 if (!list_empty(&cli->cl_loi_ready_list))
0a3bdb00 2005 return list_to_obj(&cli->cl_loi_ready_list, ready_item);
d7e09d03
PT		 2006
2007 /* then if we have cache waiters, return all objects with queued
2008 * writes. This is especially important when many small files
2009 * have filled up the cache and not been fired into rpcs because
11d66e89 2010 * they don't pass the nr_pending/object threshold */
d7e09d03
PT		 2011 if (!list_empty(&cli->cl_cache_waiters) &&
2012 !list_empty(&cli->cl_loi_write_list))
0a3bdb00 2013 return list_to_obj(&cli->cl_loi_write_list, write_item);
d7e09d03
PT		 2014
2015 /* then return all queued objects when we have an invalid import
2016 * so that they get flushed */
2017 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2018 if (!list_empty(&cli->cl_loi_write_list))
0a3bdb00 2019 return list_to_obj(&cli->cl_loi_write_list, write_item);
d7e09d03 2020 if (!list_empty(&cli->cl_loi_read_list))
0a3bdb00 2021 return list_to_obj(&cli->cl_loi_read_list, read_item);
d7e09d03 2022 }
0a3bdb00 2023 return NULL;
d7e09d03
PT		 2024}
2025
2026/* called with the loi list lock held */
2027static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2028 pdl_policy_t pol)
2029{
2030 struct osc_object *osc;
2031 int rc = 0;
d7e09d03
PT		 2032
2033 while ((osc = osc_next_obj(cli)) != NULL) {
2034 struct cl_object *obj = osc2cl(osc);
631abc6e 2035 struct lu_ref_link link;
d7e09d03
PT		 2036
2037 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2038
2039 if (osc_max_rpc_in_flight(cli, osc)) {
2040 __osc_list_maint(cli, osc);
2041 break;
2042 }
2043
2044 cl_object_get(obj);
2045 client_obd_list_unlock(&cli->cl_loi_list_lock);
631abc6e
JH		 2046 lu_object_ref_add_at(&obj->co_lu, &link, "check",
2047 current);
d7e09d03
PT		 2048
2049 /* attempt some read/write balancing by alternating between
2050 * reads and writes in an object. The makes_rpc checks here
2051 * would be redundant if we were getting read/write work items
2052 * instead of objects. we don't want send_oap_rpc to drain a
2053 * partial read pending queue when we're given this object to
2054 * do io on writes while there are cache waiters */
2055 osc_object_lock(osc);
2056 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2057 rc = osc_send_write_rpc(env, cli, osc, pol);
2058 if (rc < 0) {
2059 CERROR("Write request failed with %d\n", rc);
2060
2061 /* osc_send_write_rpc failed, mostly because of
2062 * memory pressure.
2063 *
2064 * It can't break here, because if:
2065 * - a page was submitted by osc_io_submit, so
2066 * page locked;
2067 * - no request in flight
2068 * - no subsequent request
2069 * The system will be in live-lock state,
2070 * because there is no chance to call
2071 * osc_io_unplug() and osc_check_rpcs() any
2072 * more. pdflush can't help in this case,
2073 * because it might be blocked at grabbing
2074 * the page lock as we mentioned.
2075 *
2076 * Anyway, continue to drain pages. */
2077 /* break; */
2078 }
2079 }
2080 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2081 rc = osc_send_read_rpc(env, cli, osc, pol);
2082 if (rc < 0)
2083 CERROR("Read request failed with %d\n", rc);
2084 }
2085 osc_object_unlock(osc);
2086
2087 osc_list_maint(cli, osc);
631abc6e
JH		 2088 lu_object_ref_del_at(&obj->co_lu, &link, "check",
2089 current);
d7e09d03
PT		 2090 cl_object_put(env, obj);
2091
2092 client_obd_list_lock(&cli->cl_loi_list_lock);
2093 }
2094}
2095
2096static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2097 struct osc_object *osc, pdl_policy_t pol, int async)
2098{
d7e09d03
PT		 2099 int rc = 0;
2100
cad6fafa
BJ		 2101 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2102 return 0;
2103
2104 if (!async) {
2105 /* disable osc_lru_shrink() temporarily to avoid
2106 * potential stack overrun problem. LU-2859 */
2107 atomic_inc(&cli->cl_lru_shrinkers);
2108 client_obd_list_lock(&cli->cl_loi_list_lock);
2109 osc_check_rpcs(env, cli, pol);
2110 client_obd_list_unlock(&cli->cl_loi_list_lock);
2111 atomic_dec(&cli->cl_lru_shrinkers);
2112 } else {
2113 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2114 LASSERT(cli->cl_writeback_work != NULL);
2115 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
d7e09d03 2116 }
d7e09d03
PT		 2117 return rc;
2118}
2119
2120static int osc_io_unplug_async(const struct lu_env *env,
2121 struct client_obd *cli, struct osc_object *osc)
2122{
2123 /* XXX: policy is no use actually. */
2124 return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2125}
2126
2127void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2128 struct osc_object *osc, pdl_policy_t pol)
2129{
2130 (void)osc_io_unplug0(env, cli, osc, pol, 0);
2131}
2132
2133int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2134 struct page *page, loff_t offset)
2135{
2136 struct obd_export *exp = osc_export(osc);
2137 struct osc_async_page *oap = &ops->ops_oap;
d7e09d03
PT		 2138
2139 if (!page)
2140 return cfs_size_round(sizeof(*oap));
2141
2142 oap->oap_magic = OAP_MAGIC;
2143 oap->oap_cli = &exp->exp_obd->u.cli;
2144 oap->oap_obj = osc;
2145
2146 oap->oap_page = page;
2147 oap->oap_obj_off = offset;
2148 LASSERT(!(offset & ~CFS_PAGE_MASK));
2149
2eb90a75 2150 if (!client_is_remote(exp) && capable(CFS_CAP_SYS_RESOURCE))
d7e09d03
PT		 2151 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2152
2153 INIT_LIST_HEAD(&oap->oap_pending_item);
2154 INIT_LIST_HEAD(&oap->oap_rpc_item);
2155
2156 spin_lock_init(&oap->oap_lock);
b0f5aad5 2157 CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
d7e09d03 2158 oap, page, oap->oap_obj_off);
0a3bdb00 2159 return 0;
d7e09d03
PT		 2160}
2161
2162int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2163 struct osc_page *ops)
2164{
2165 struct osc_io *oio = osc_env_io(env);
2166 struct osc_extent *ext = NULL;
2167 struct osc_async_page *oap = &ops->ops_oap;
2168 struct client_obd *cli = oap->oap_cli;
2169 struct osc_object *osc = oap->oap_obj;
2170 pgoff_t index;
2171 int grants = 0;
2172 int brw_flags = OBD_BRW_ASYNC;
2173 int cmd = OBD_BRW_WRITE;
2174 int need_release = 0;
2175 int rc = 0;
d7e09d03
PT		 2176
2177 if (oap->oap_magic != OAP_MAGIC)
0a3bdb00 2178 return -EINVAL;
d7e09d03
PT		 2179
2180 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
0a3bdb00 2181 return -EIO;
d7e09d03
PT		 2182
2183 if (!list_empty(&oap->oap_pending_item) ||
2184 !list_empty(&oap->oap_rpc_item))
0a3bdb00 2185 return -EBUSY;
d7e09d03
PT		 2186
2187 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2188 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2189 if (!client_is_remote(osc_export(osc)) &&
2eb90a75 2190 capable(CFS_CAP_SYS_RESOURCE)) {
d7e09d03
PT		 2191 brw_flags |= OBD_BRW_NOQUOTA;
2192 cmd |= OBD_BRW_NOQUOTA;
2193 }
2194
2195 /* check if the file's owner/group is over quota */
2196 if (!(cmd & OBD_BRW_NOQUOTA)) {
2197 struct cl_object *obj;
2198 struct cl_attr *attr;
2199 unsigned int qid[MAXQUOTAS];
2200
2201 obj = cl_object_top(&osc->oo_cl);
2202 attr = &osc_env_info(env)->oti_attr;
2203
2204 cl_object_attr_lock(obj);
2205 rc = cl_object_attr_get(env, obj, attr);
2206 cl_object_attr_unlock(obj);
2207
2208 qid[USRQUOTA] = attr->cat_uid;
2209 qid[GRPQUOTA] = attr->cat_gid;
2210 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2211 rc = -EDQUOT;
2212 if (rc)
0a3bdb00 2213 return rc;
d7e09d03
PT		 2214 }
2215
2216 oap->oap_cmd = cmd;
2217 oap->oap_page_off = ops->ops_from;
2218 oap->oap_count = ops->ops_to - ops->ops_from;
2219 oap->oap_async_flags = 0;
2220 oap->oap_brw_flags = brw_flags;
2221
2222 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2223 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2224
2225 index = oap2cl_page(oap)->cp_index;
2226
2227 /* Add this page into extent by the following steps:
2228 * 1. if there exists an active extent for this IO, mostly this page
2229 * can be added to the active extent and sometimes we need to
11d66e89 2230 * expand extent to accommodate this page;
d7e09d03
PT		 2231 * 2. otherwise, a new extent will be allocated. */
2232
2233 ext = oio->oi_active;
2234 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2235 /* one chunk plus extent overhead must be enough to write this
2236 * page */
2237 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2238 if (ext->oe_end >= index)
2239 grants = 0;
2240
2241 /* it doesn't need any grant to dirty this page */
2242 client_obd_list_lock(&cli->cl_loi_list_lock);
2243 rc = osc_enter_cache_try(cli, oap, grants, 0);
2244 client_obd_list_unlock(&cli->cl_loi_list_lock);
2245 if (rc == 0) { /* try failed */
2246 grants = 0;
2247 need_release = 1;
2248 } else if (ext->oe_end < index) {
2249 int tmp = grants;
2250 /* try to expand this extent */
2251 rc = osc_extent_expand(ext, index, &tmp);
2252 if (rc < 0) {
2253 need_release = 1;
2254 /* don't free reserved grant */
2255 } else {
2256 OSC_EXTENT_DUMP(D_CACHE, ext,
2257 "expanded for %lu.\n", index);
2258 osc_unreserve_grant(cli, grants, tmp);
2259 grants = 0;
2260 }
2261 }
2262 rc = 0;
2263 } else if (ext != NULL) {
2264 /* index is located outside of active extent */
2265 need_release = 1;
2266 }
2267 if (need_release) {
2268 osc_extent_release(env, ext);
2269 oio->oi_active = NULL;
2270 ext = NULL;
2271 }
2272
2273 if (ext == NULL) {
2274 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2275
2276 /* try to find new extent to cover this page */
2277 LASSERT(oio->oi_active == NULL);
2278 /* we may have allocated grant for this page if we failed
2279 * to expand the previous active extent. */
2280 LASSERT(ergo(grants > 0, grants >= tmp));
2281
2282 rc = 0;
2283 if (grants == 0) {
2284 /* we haven't allocated grant for this page. */
2285 rc = osc_enter_cache(env, cli, oap, tmp);
2286 if (rc == 0)
2287 grants = tmp;
2288 }
2289
2290 tmp = grants;
2291 if (rc == 0) {
2292 ext = osc_extent_find(env, osc, index, &tmp);
2293 if (IS_ERR(ext)) {
2294 LASSERT(tmp == grants);
2295 osc_exit_cache(cli, oap);
2296 rc = PTR_ERR(ext);
2297 ext = NULL;
2298 } else {
2299 oio->oi_active = ext;
2300 }
2301 }
2302 if (grants > 0)
2303 osc_unreserve_grant(cli, grants, tmp);
2304 }
2305
2306 LASSERT(ergo(rc == 0, ext != NULL));
2307 if (ext != NULL) {
2308 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2309 ext, "index = %lu.\n", index);
2310 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2311
2312 osc_object_lock(osc);
2313 if (ext->oe_nr_pages == 0)
2314 ext->oe_srvlock = ops->ops_srvlock;
2315 else
2316 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2317 ++ext->oe_nr_pages;
2318 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2319 osc_object_unlock(osc);
2320 }
0a3bdb00 2321 return rc;
d7e09d03
PT		 2322}
2323
2324int osc_teardown_async_page(const struct lu_env *env,
2325 struct osc_object *obj, struct osc_page *ops)
2326{
2327 struct osc_async_page *oap = &ops->ops_oap;
2328 struct osc_extent *ext = NULL;
2329 int rc = 0;
d7e09d03
PT		 2330
2331 LASSERT(oap->oap_magic == OAP_MAGIC);
2332
2333 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2334 oap, ops, oap2cl_page(oap)->cp_index);
2335
2336 osc_object_lock(obj);
2337 if (!list_empty(&oap->oap_rpc_item)) {
2338 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2339 rc = -EBUSY;
2340 } else if (!list_empty(&oap->oap_pending_item)) {
2341 ext = osc_extent_lookup(obj, oap2cl_page(oap)->cp_index);
2342 /* only truncated pages are allowed to be taken out.
2343 * See osc_extent_truncate() and osc_cache_truncate_start()
2344 * for details. */
2345 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2346 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2347 oap2cl_page(oap)->cp_index);
2348 rc = -EBUSY;
2349 }
2350 }
2351 osc_object_unlock(obj);
2352 if (ext != NULL)
2353 osc_extent_put(env, ext);
0a3bdb00 2354 return rc;
d7e09d03
PT		 2355}
2356
2357/**
2358 * This is called when a page is picked up by kernel to write out.
2359 *
2360 * We should find out the corresponding extent and add the whole extent
2361 * into urgent list. The extent may be being truncated or used, handle it
2362 * carefully.
2363 */
2364int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2365 struct osc_page *ops)
2366{
2367 struct osc_extent *ext = NULL;
2368 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2369 struct cl_page *cp = ops->ops_cl.cpl_page;
2370 pgoff_t index = cp->cp_index;
2371 struct osc_async_page *oap = &ops->ops_oap;
2372 bool unplug = false;
2373 int rc = 0;
d7e09d03
PT		 2374
2375 osc_object_lock(obj);
2376 ext = osc_extent_lookup(obj, index);
2377 if (ext == NULL) {
2378 osc_extent_tree_dump(D_ERROR, obj);
2379 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2380 }
2381
2382 switch (ext->oe_state) {
2383 case OES_RPC:
2384 case OES_LOCK_DONE:
2385 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(cp),
2386 "flush an in-rpc page?\n");
2387 LASSERT(0);
2388 break;
2389 case OES_LOCKING:
2390 /* If we know this extent is being written out, we should abort
2391 * so that the writer can make this page ready. Otherwise, there
2392 * exists a deadlock problem because other process can wait for
2393 * page writeback bit holding page lock; and meanwhile in
2394 * vvp_page_make_ready(), we need to grab page lock before
2395 * really sending the RPC. */
2396 case OES_TRUNC:
2397 /* race with truncate, page will be redirtied */
15f13cde
AK		 2398 case OES_ACTIVE:
2399 /* The extent is active so we need to abort and let the caller
2400 * re-dirty the page. If we continued on here, and we were the
2401 * one making the extent active, we could deadlock waiting for
2402 * the page writeback to clear but it won't because the extent
2403 * is active and won't be written out. */
d7e09d03
PT		 2404 GOTO(out, rc = -EAGAIN);
2405 default:
2406 break;
2407 }
2408
2409 rc = cl_page_prep(env, io, cl_page_top(cp), CRT_WRITE);
2410 if (rc)
2411 GOTO(out, rc);
2412
2413 spin_lock(&oap->oap_lock);
2414 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2415 spin_unlock(&oap->oap_lock);
2416
2417 if (memory_pressure_get())
2418 ext->oe_memalloc = 1;
2419
2420 ext->oe_urgent = 1;
2421 if (ext->oe_state == OES_CACHE) {
2422 OSC_EXTENT_DUMP(D_CACHE, ext,
2423 "flush page %p make it urgent.\n", oap);
2424 if (list_empty(&ext->oe_link))
2425 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2426 unplug = true;
2427 }
2428 rc = 0;
d7e09d03
PT		 2429
2430out:
2431 osc_object_unlock(obj);
2432 osc_extent_put(env, ext);
2433 if (unplug)
2434 osc_io_unplug_async(env, osc_cli(obj), obj);
2435 return rc;
2436}
2437
2438/**
2439 * this is called when a sync waiter receives an interruption. Its job is to
2440 * get the caller woken as soon as possible. If its page hasn't been put in an
2441 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2442 * desiring interruption which will forcefully complete the rpc once the rpc
2443 * has timed out.
2444 */
2445int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2446{
2447 struct osc_async_page *oap = &ops->ops_oap;
2448 struct osc_object *obj = oap->oap_obj;
2449 struct client_obd *cli = osc_cli(obj);
2450 struct osc_extent *ext;
2451 struct osc_extent *found = NULL;
2452 struct list_head *plist;
2453 pgoff_t index = oap2cl_page(oap)->cp_index;
2454 int rc = -EBUSY;
2455 int cmd;
d7e09d03
PT		 2456
2457 LASSERT(!oap->oap_interrupted);
2458 oap->oap_interrupted = 1;
2459
2460 /* Find out the caching extent */
2461 osc_object_lock(obj);
2462 if (oap->oap_cmd & OBD_BRW_WRITE) {
2463 plist = &obj->oo_urgent_exts;
2464 cmd = OBD_BRW_WRITE;
2465 } else {
2466 plist = &obj->oo_reading_exts;
2467 cmd = OBD_BRW_READ;
2468 }
2469 list_for_each_entry(ext, plist, oe_link) {
2470 if (ext->oe_start <= index && ext->oe_end >= index) {
2471 LASSERT(ext->oe_state == OES_LOCK_DONE);
2472 /* For OES_LOCK_DONE state extent, it has already held
2473 * a refcount for RPC. */
2474 found = osc_extent_get(ext);
2475 break;
2476 }
2477 }
2478 if (found != NULL) {
2479 list_del_init(&found->oe_link);
2480 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2481 osc_object_unlock(obj);
2482
2483 osc_extent_finish(env, found, 0, -EINTR);
2484 osc_extent_put(env, found);
2485 rc = 0;
2486 } else {
2487 osc_object_unlock(obj);
2488 /* ok, it's been put in an rpc. only one oap gets a request
2489 * reference */
2490 if (oap->oap_request != NULL) {
2491 ptlrpc_mark_interrupted(oap->oap_request);
2492 ptlrpcd_wake(oap->oap_request);
2493 ptlrpc_req_finished(oap->oap_request);
2494 oap->oap_request = NULL;
2495 }
2496 }
2497
2498 osc_list_maint(cli, obj);
0a3bdb00 2499 return rc;
d7e09d03
PT		 2500}
2501
2502int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2503 struct list_head *list, int cmd, int brw_flags)
2504{
2505 struct client_obd *cli = osc_cli(obj);
2506 struct osc_extent *ext;
f13ab92e 2507 struct osc_async_page *oap, *tmp;
d7e09d03
PT		 2508 int page_count = 0;
2509 int mppr = cli->cl_max_pages_per_rpc;
2510 pgoff_t start = CL_PAGE_EOF;
2511 pgoff_t end = 0;
d7e09d03
PT		 2512
2513 list_for_each_entry(oap, list, oap_pending_item) {
2514 struct cl_page *cp = oap2cl_page(oap);
2515 if (cp->cp_index > end)
2516 end = cp->cp_index;
2517 if (cp->cp_index < start)
2518 start = cp->cp_index;
2519 ++page_count;
2520 mppr <<= (page_count > mppr);
2521 }
2522
2523 ext = osc_extent_alloc(obj);
2524 if (ext == NULL) {
f13ab92e 2525 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
d7e09d03
PT		 2526 list_del_init(&oap->oap_pending_item);
2527 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2528 }
0a3bdb00 2529 return -ENOMEM;
d7e09d03
PT		 2530 }
2531
2532 ext->oe_rw = !!(cmd & OBD_BRW_READ);
2533 ext->oe_urgent = 1;
2534 ext->oe_start = start;
2535 ext->oe_end = ext->oe_max_end = end;
2536 ext->oe_obj = obj;
2537 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2538 ext->oe_nr_pages = page_count;
2539 ext->oe_mppr = mppr;
2540 list_splice_init(list, &ext->oe_pages);
2541
2542 osc_object_lock(obj);
2543 /* Reuse the initial refcount for RPC, don't drop it */
2544 osc_extent_state_set(ext, OES_LOCK_DONE);
2545 if (cmd & OBD_BRW_WRITE) {
2546 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2547 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2548 } else {
2549 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2550 osc_update_pending(obj, OBD_BRW_READ, page_count);
2551 }
2552 osc_object_unlock(obj);
2553
2554 osc_io_unplug(env, cli, obj, PDL_POLICY_ROUND);
0a3bdb00 2555 return 0;
d7e09d03
PT		 2556}
2557
2558/**
2559 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2560 */
2561int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2562 struct osc_object *obj, __u64 size)
2563{
2564 struct client_obd *cli = osc_cli(obj);
2565 struct osc_extent *ext;
2566 struct osc_extent *waiting = NULL;
2567 pgoff_t index;
2568 LIST_HEAD(list);
2569 int result = 0;
2570 bool partial;
d7e09d03
PT		 2571
2572 /* pages with index greater or equal to index will be truncated. */
2573 index = cl_index(osc2cl(obj), size);
2574 partial = size > cl_offset(osc2cl(obj), index);
2575
2576again:
2577 osc_object_lock(obj);
2578 ext = osc_extent_search(obj, index);
2579 if (ext == NULL)
2580 ext = first_extent(obj);
2581 else if (ext->oe_end < index)
2582 ext = next_extent(ext);
2583 while (ext != NULL) {
2584 EASSERT(ext->oe_state != OES_TRUNC, ext);
2585
2586 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2587 /* if ext is in urgent state, it means there must exist
2588 * a page already having been flushed by write_page().
2589 * We have to wait for this extent because we can't
2590 * truncate that page. */
2591 LASSERT(!ext->oe_hp);
2592 OSC_EXTENT_DUMP(D_CACHE, ext,
2593 "waiting for busy extent\n");
2594 waiting = osc_extent_get(ext);
2595 break;
2596 }
2597
b0f5aad5 2598 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
d7e09d03
PT		 2599
2600 osc_extent_get(ext);
2601 if (ext->oe_state == OES_ACTIVE) {
2602 /* though we grab inode mutex for write path, but we
2603 * release it before releasing extent(in osc_io_end()),
2604 * so there is a race window that an extent is still
2605 * in OES_ACTIVE when truncate starts. */
2606 LASSERT(!ext->oe_trunc_pending);
2607 ext->oe_trunc_pending = 1;
2608 } else {
2609 EASSERT(ext->oe_state == OES_CACHE, ext);
2610 osc_extent_state_set(ext, OES_TRUNC);
2611 osc_update_pending(obj, OBD_BRW_WRITE,
2612 -ext->oe_nr_pages);
2613 }
2614 EASSERT(list_empty(&ext->oe_link), ext);
2615 list_add_tail(&ext->oe_link, &list);
2616
2617 ext = next_extent(ext);
2618 }
2619 osc_object_unlock(obj);
2620
2621 osc_list_maint(cli, obj);
2622
2623 while (!list_empty(&list)) {
2624 int rc;
2625
2626 ext = list_entry(list.next, struct osc_extent, oe_link);
2627 list_del_init(&ext->oe_link);
2628
2629 /* extent may be in OES_ACTIVE state because inode mutex
2630 * is released before osc_io_end() in file write case */
2631 if (ext->oe_state != OES_TRUNC)
2632 osc_extent_wait(env, ext, OES_TRUNC);
2633
2634 rc = osc_extent_truncate(ext, index, partial);
2635 if (rc < 0) {
2636 if (result == 0)
2637 result = rc;
2638
2639 OSC_EXTENT_DUMP(D_ERROR, ext,
2640 "truncate error %d\n", rc);
2641 } else if (ext->oe_nr_pages == 0) {
2642 osc_extent_remove(ext);
2643 } else {
2644 /* this must be an overlapped extent which means only
2645 * part of pages in this extent have been truncated.
2646 */
2647 EASSERTF(ext->oe_start <= index, ext,
2648 "trunc index = %lu/%d.\n", index, partial);
2649 /* fix index to skip this partially truncated extent */
2650 index = ext->oe_end + 1;
2651 partial = false;
2652
2653 /* we need to hold this extent in OES_TRUNC state so
2654 * that no writeback will happen. This is to avoid
2655 * BUG 17397. */
2656 LASSERT(oio->oi_trunc == NULL);
2657 oio->oi_trunc = osc_extent_get(ext);
2658 OSC_EXTENT_DUMP(D_CACHE, ext,
b0f5aad5 2659 "trunc at %llu\n", size);
d7e09d03
PT		 2660 }
2661 osc_extent_put(env, ext);
2662 }
2663 if (waiting != NULL) {
2664 int rc;
2665
2666 /* ignore the result of osc_extent_wait the write initiator
2667 * should take care of it. */
2668 rc = osc_extent_wait(env, waiting, OES_INV);
2669 if (rc < 0)
451721cc 2670 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
d7e09d03
PT		 2671
2672 osc_extent_put(env, waiting);
2673 waiting = NULL;
2674 goto again;
2675 }
0a3bdb00 2676 return result;
d7e09d03
PT		 2677}
2678
2679/**
2680 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2681 */
2682void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2683 struct osc_object *obj)
2684{
2685 struct osc_extent *ext = oio->oi_trunc;
2686
2687 oio->oi_trunc = NULL;
2688 if (ext != NULL) {
2689 bool unplug = false;
2690
2691 EASSERT(ext->oe_nr_pages > 0, ext);
2692 EASSERT(ext->oe_state == OES_TRUNC, ext);
2693 EASSERT(!ext->oe_urgent, ext);
2694
2695 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2696 osc_object_lock(obj);
2697 osc_extent_state_set(ext, OES_CACHE);
2698 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2699 ext->oe_urgent = 1;
2700 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2701 unplug = true;
2702 }
2703 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2704 osc_object_unlock(obj);
2705 osc_extent_put(env, ext);
2706
2707 if (unplug)
2708 osc_io_unplug_async(env, osc_cli(obj), obj);
2709 }
2710}
2711
2712/**
2713 * Wait for extents in a specific range to be written out.
2714 * The caller must have called osc_cache_writeback_range() to issue IO
2715 * otherwise it will take a long time for this function to finish.
2716 *
2717 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2718 * nobody else can dirty this range of file while we're waiting for
2719 * extents to be written.
2720 */
2721int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2722 pgoff_t start, pgoff_t end)
2723{
2724 struct osc_extent *ext;
2725 pgoff_t index = start;
2726 int result = 0;
d7e09d03
PT		 2727
2728again:
2729 osc_object_lock(obj);
2730 ext = osc_extent_search(obj, index);
2731 if (ext == NULL)
2732 ext = first_extent(obj);
2733 else if (ext->oe_end < index)
2734 ext = next_extent(ext);
2735 while (ext != NULL) {
2736 int rc;
2737
2738 if (ext->oe_start > end)
2739 break;
2740
2741 if (!ext->oe_fsync_wait) {
2742 ext = next_extent(ext);
2743 continue;
2744 }
2745
2746 EASSERT(ergo(ext->oe_state == OES_CACHE,
2747 ext->oe_hp || ext->oe_urgent), ext);
2748 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2749 !ext->oe_hp && ext->oe_urgent), ext);
2750
2751 index = ext->oe_end + 1;
2752 osc_extent_get(ext);
2753 osc_object_unlock(obj);
2754
2755 rc = osc_extent_wait(env, ext, OES_INV);
2756 if (result == 0)
2757 result = rc;
2758 osc_extent_put(env, ext);
2759 goto again;
2760 }
2761 osc_object_unlock(obj);
2762
2763 OSC_IO_DEBUG(obj, "sync file range.\n");
0a3bdb00 2764 return result;
d7e09d03
PT		 2765}
2766
2767/**
2768 * Called to write out a range of osc object.
2769 *
2770 * @hp : should be set this is caused by lock cancel;
2771 * @discard: is set if dirty pages should be dropped - file will be deleted or
2772 * truncated, this implies there is no partially discarding extents.
2773 *
2774 * Return how many pages will be issued, or error code if error occurred.
2775 */
2776int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2777 pgoff_t start, pgoff_t end, int hp, int discard)
2778{
2779 struct osc_extent *ext;
2780 LIST_HEAD(discard_list);
2781 bool unplug = false;
2782 int result = 0;
d7e09d03
PT		 2783
2784 osc_object_lock(obj);
2785 ext = osc_extent_search(obj, start);
2786 if (ext == NULL)
2787 ext = first_extent(obj);
2788 else if (ext->oe_end < start)
2789 ext = next_extent(ext);
2790 while (ext != NULL) {
2791 if (ext->oe_start > end)
2792 break;
2793
2794 ext->oe_fsync_wait = 1;
2795 switch (ext->oe_state) {
2796 case OES_CACHE:
2797 result += ext->oe_nr_pages;
2798 if (!discard) {
2799 struct list_head *list = NULL;
2800 if (hp) {
2801 EASSERT(!ext->oe_hp, ext);
2802 ext->oe_hp = 1;
2803 list = &obj->oo_hp_exts;
2804 } else if (!ext->oe_urgent) {
2805 ext->oe_urgent = 1;
2806 list = &obj->oo_urgent_exts;
2807 }
2808 if (list != NULL)
2809 list_move_tail(&ext->oe_link, list);
2810 unplug = true;
2811 } else {
2812 /* the only discarder is lock cancelling, so
2813 * [start, end] must contain this extent */
2814 EASSERT(ext->oe_start >= start &&
2815 ext->oe_max_end <= end, ext);
2816 osc_extent_state_set(ext, OES_LOCKING);
2817 ext->oe_owner = current;
2818 list_move_tail(&ext->oe_link,
2819 &discard_list);
2820 osc_update_pending(obj, OBD_BRW_WRITE,
2821 -ext->oe_nr_pages);
2822 }
2823 break;
2824 case OES_ACTIVE:
2825 /* It's pretty bad to wait for ACTIVE extents, because
2826 * we don't know how long we will wait for it to be
2827 * flushed since it may be blocked at awaiting more
2828 * grants. We do this for the correctness of fsync. */
2829 LASSERT(hp == 0 && discard == 0);
2830 ext->oe_urgent = 1;
2831 break;
2832 case OES_TRUNC:
2833 /* this extent is being truncated, can't do anything
2834 * for it now. it will be set to urgent after truncate
2835 * is finished in osc_cache_truncate_end(). */
2836 default:
2837 break;
2838 }
2839 ext = next_extent(ext);
2840 }
2841 osc_object_unlock(obj);
2842
2843 LASSERT(ergo(!discard, list_empty(&discard_list)));
2844 if (!list_empty(&discard_list)) {
2845 struct osc_extent *tmp;
2846 int rc;
2847
2848 osc_list_maint(osc_cli(obj), obj);
2849 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2850 list_del_init(&ext->oe_link);
2851 EASSERT(ext->oe_state == OES_LOCKING, ext);
2852
2853 /* Discard caching pages. We don't actually write this
2854 * extent out but we complete it as if we did. */
2855 rc = osc_extent_make_ready(env, ext);
2856 if (unlikely(rc < 0)) {
2857 OSC_EXTENT_DUMP(D_ERROR, ext,
2858 "make_ready returned %d\n", rc);
2859 if (result >= 0)
2860 result = rc;
2861 }
2862
2863 /* finish the extent as if the pages were sent */
2864 osc_extent_finish(env, ext, 0, 0);
2865 }
2866 }
2867
2868 if (unplug)
2869 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2870
2871 if (hp || discard) {
2872 int rc;
2873 rc = osc_cache_wait_range(env, obj, start, end);
2874 if (result >= 0 && rc < 0)
2875 result = rc;
2876 }
2877
2878 OSC_IO_DEBUG(obj, "cache page out.\n");
0a3bdb00 2879 return result;
d7e09d03
PT		 2880}
2881
2882/** @} osc */
Ubuntu Hirsute Linux kernel mirror