4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
34 * This file is part of Lustre, http://www.lustre.org/
35 * Lustre is a trademark of Sun Microsystems, Inc.
37 * osc cache management.
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_OSC
44 #include "osc_cl_internal.h"
45 #include "osc_internal.h"
47 static int extent_debug; /* set it to be true for more debug */
49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
53 struct osc_async_page *oap, int sent, int rc);
54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
56 static int osc_refresh_count(const struct lu_env *env,
57 struct osc_async_page *oap, int cmd);
58 static int osc_io_unplug_async(const struct lu_env *env,
59 struct client_obd *cli, struct osc_object *osc);
60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
61 unsigned int lost_grant);
63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
64 const char *func, int line);
65 #define osc_extent_tree_dump(lvl, obj) \
66 osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
72 /* ------------------ osc extent ------------------ */
73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
76 *buf++ = ext->oe_rw ? 'r' : 'w';
87 if (ext->oe_trunc_pending)
89 if (ext->oe_fsync_wait)
95 static inline char list_empty_marker(struct list_head *list)
97 return list_empty(list) ? '-' : '+';
100 #define EXTSTR "[%lu -> %lu/%lu]"
101 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
102 static const char *oes_strings[] = {
103 "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
105 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do { \
106 struct osc_extent *__ext = (extent); \
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), \
118 oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \
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 ----- */ \
130 #define EASSERTF(expr, ext, fmt, args...) do { \
132 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args); \
133 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj); \
139 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
141 static inline struct osc_extent *rb_extent(struct rb_node *n)
146 return container_of(n, struct osc_extent, oe_node);
149 static inline struct osc_extent *next_extent(struct osc_extent *ext)
154 LASSERT(ext->oe_intree);
155 return rb_extent(rb_next(&ext->oe_node));
158 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
163 LASSERT(ext->oe_intree);
164 return rb_extent(rb_prev(&ext->oe_node));
167 static inline struct osc_extent *first_extent(struct osc_object *obj)
169 return rb_extent(rb_first(&obj->oo_root));
172 /* object must be locked by caller. */
173 static int osc_extent_sanity_check0(struct osc_extent *ext,
174 const char *func, const int line)
176 struct osc_object *obj = ext->oe_obj;
177 struct osc_async_page *oap;
181 if (!osc_object_is_locked(obj))
184 if (ext->oe_state >= OES_STATE_MAX)
187 if (atomic_read(&ext->oe_refc) <= 0)
190 if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
193 switch (ext->oe_state) {
195 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
200 if (atomic_read(&ext->oe_users) == 0)
204 if (ext->oe_fsync_wait && !ext->oe_urgent)
208 if (ext->oe_grants == 0)
210 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
213 if (atomic_read(&ext->oe_users) > 0)
217 if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
220 if (ext->oe_osclock == NULL && ext->oe_grants > 0)
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))
231 if (ext->oe_nr_pages > ext->oe_mppr)
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)
243 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
244 pgoff_t index = osc_index(oap2osc(oap));
246 if (index > ext->oe_end || index < ext->oe_start)
249 if (page_count != ext->oe_nr_pages)
254 OSC_EXTENT_DUMP(D_ERROR, ext,
255 "%s:%d sanity check %p failed with rc = %d\n",
256 func, line, ext, rc);
260 #define sanity_check_nolock(ext) \
261 osc_extent_sanity_check0(ext, __func__, __LINE__)
263 #define sanity_check(ext) ({ \
265 osc_object_lock((ext)->oe_obj); \
266 __res = sanity_check_nolock(ext); \
267 osc_object_unlock((ext)->oe_obj); \
273 * sanity check - to make sure there is no overlapped extent in the tree.
275 static int osc_extent_is_overlapped(struct osc_object *obj,
276 struct osc_extent *ext)
278 struct osc_extent *tmp;
280 LASSERT(osc_object_is_locked(obj));
285 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
288 if (tmp->oe_end >= ext->oe_start &&
289 tmp->oe_start <= ext->oe_end)
295 static void osc_extent_state_set(struct osc_extent *ext, int state)
297 LASSERT(osc_object_is_locked(ext->oe_obj));
298 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
300 /* Never try to sanity check a state changing extent :-) */
301 /* LASSERT(sanity_check_nolock(ext) == 0); */
303 /* TODO: validate the state machine */
304 ext->oe_state = state;
305 wake_up_all(&ext->oe_waitq);
308 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
310 struct osc_extent *ext;
312 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
316 RB_CLEAR_NODE(&ext->oe_node);
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;
329 static void osc_extent_free(struct osc_extent *ext)
331 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
334 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
336 LASSERT(atomic_read(&ext->oe_refc) >= 0);
337 atomic_inc(&ext->oe_refc);
341 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
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);
350 if (ext->oe_osclock) {
351 cl_lock_put(env, ext->oe_osclock);
352 ext->oe_osclock = NULL;
354 osc_extent_free(ext);
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 ;-).
363 static void osc_extent_put_trust(struct osc_extent *ext)
365 LASSERT(atomic_read(&ext->oe_refc) > 1);
366 LASSERT(osc_object_is_locked(ext->oe_obj));
367 atomic_dec(&ext->oe_refc);
371 * Return the extent which includes pgoff @index, or return the greatest
372 * previous extent in the tree.
374 static struct osc_extent *osc_extent_search(struct osc_object *obj,
377 struct rb_node *n = obj->oo_root.rb_node;
378 struct osc_extent *tmp, *p = NULL;
380 LASSERT(osc_object_is_locked(obj));
383 if (index < tmp->oe_start) {
385 } else if (index > tmp->oe_end) {
396 * Return the extent covering @index, otherwise return NULL.
397 * caller must have held object lock.
399 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
402 struct osc_extent *ext;
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);
410 /* caller must have held object lock. */
411 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
413 struct rb_node **n = &obj->oo_root.rb_node;
414 struct rb_node *parent = NULL;
415 struct osc_extent *tmp;
417 LASSERT(ext->oe_intree == 0);
418 LASSERT(ext->oe_obj == obj);
419 LASSERT(osc_object_is_locked(obj));
424 if (ext->oe_end < tmp->oe_start)
426 else if (ext->oe_start > tmp->oe_end)
429 EASSERTF(0, tmp, EXTSTR, EXTPARA(ext));
431 rb_link_node(&ext->oe_node, parent, n);
432 rb_insert_color(&ext->oe_node, &obj->oo_root);
437 /* caller must have held object lock. */
438 static void osc_extent_erase(struct osc_extent *ext)
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);
445 /* rbtree held a refcount */
446 osc_extent_put_trust(ext);
450 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
452 struct osc_object *obj = ext->oe_obj;
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);
460 atomic_inc(&ext->oe_users);
461 list_del_init(&ext->oe_link);
462 return osc_extent_get(ext);
465 static void __osc_extent_remove(struct osc_extent *ext)
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");
475 static void osc_extent_remove(struct osc_extent *ext)
477 struct osc_object *obj = ext->oe_obj;
479 osc_object_lock(obj);
480 __osc_extent_remove(ext);
481 osc_object_unlock(obj);
485 * This function is used to merge extents to get better performance. It checks
486 * if @cur and @victim are contiguous at chunk level.
488 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
489 struct osc_extent *victim)
491 struct osc_object *obj = cur->oe_obj;
496 LASSERT(cur->oe_state == OES_CACHE);
497 LASSERT(osc_object_is_locked(obj));
501 if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
504 if (cur->oe_max_end != victim->oe_max_end)
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)
515 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
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;
528 osc_extent_get(victim);
529 __osc_extent_remove(victim);
530 osc_extent_put(env, victim);
532 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
537 * Drop user count of osc_extent, and unplug IO asynchronously.
539 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
541 struct osc_object *obj = ext->oe_obj;
545 LASSERT(atomic_read(&ext->oe_users) > 0);
546 LASSERT(sanity_check(ext) == 0);
547 LASSERT(ext->oe_grants > 0);
549 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
550 LASSERT(ext->oe_state == OES_ACTIVE);
551 if (ext->oe_trunc_pending) {
552 /* a truncate process is waiting for this extent.
553 * This may happen due to a race, check
554 * osc_cache_truncate_start(). */
555 osc_extent_state_set(ext, OES_TRUNC);
556 ext->oe_trunc_pending = 0;
558 osc_extent_state_set(ext, OES_CACHE);
559 osc_update_pending(obj, OBD_BRW_WRITE,
562 /* try to merge the previous and next extent. */
563 osc_extent_merge(env, ext, prev_extent(ext));
564 osc_extent_merge(env, ext, next_extent(ext));
567 list_move_tail(&ext->oe_link,
568 &obj->oo_urgent_exts);
570 osc_object_unlock(obj);
572 osc_io_unplug_async(env, osc_cli(obj), obj);
574 osc_extent_put(env, ext);
578 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
580 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
584 * Find or create an extent which includes @index, core function to manage
587 struct osc_extent *osc_extent_find(const struct lu_env *env,
588 struct osc_object *obj, pgoff_t index,
592 struct client_obd *cli = osc_cli(obj);
593 struct cl_lock *lock;
594 struct osc_extent *cur;
595 struct osc_extent *ext;
596 struct osc_extent *conflict = NULL;
597 struct osc_extent *found = NULL;
600 int max_pages; /* max_pages_per_rpc */
602 int ppc_bits; /* pages per chunk bits */
607 cur = osc_extent_alloc(obj);
609 RETURN(ERR_PTR(-ENOMEM));
611 lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0);
612 LASSERT(lock != NULL);
613 LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
615 LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
616 ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
617 chunk_mask = ~((1 << ppc_bits) - 1);
618 chunksize = 1 << cli->cl_chunkbits;
619 chunk = index >> ppc_bits;
621 /* align end to rpc edge, rpc size may not be a power 2 integer. */
622 max_pages = cli->cl_max_pages_per_rpc;
623 LASSERT((max_pages & ~chunk_mask) == 0);
624 max_end = index - (index % max_pages) + max_pages - 1;
625 max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end);
627 /* initialize new extent by parameters so far */
628 cur->oe_max_end = max_end;
629 cur->oe_start = index & chunk_mask;
630 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
631 if (cur->oe_start < lock->cll_descr.cld_start)
632 cur->oe_start = lock->cll_descr.cld_start;
633 if (cur->oe_end > max_end)
634 cur->oe_end = max_end;
635 cur->oe_osclock = lock;
637 cur->oe_mppr = max_pages;
639 /* grants has been allocated by caller */
640 LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
641 "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
642 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur));
645 osc_object_lock(obj);
646 ext = osc_extent_search(obj, cur->oe_start);
648 ext = first_extent(obj);
649 while (ext != NULL) {
650 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
651 loff_t ext_chk_end = ext->oe_end >> ppc_bits;
653 LASSERT(sanity_check_nolock(ext) == 0);
654 if (chunk > ext_chk_end + 1)
657 /* if covering by different locks, no chance to match */
658 if (lock != ext->oe_osclock) {
659 EASSERTF(!overlapped(ext, cur), ext,
660 EXTSTR, EXTPARA(cur));
662 ext = next_extent(ext);
666 /* discontiguous chunks? */
667 if (chunk + 1 < ext_chk_start) {
668 ext = next_extent(ext);
672 /* ok, from now on, ext and cur have these attrs:
673 * 1. covered by the same lock
674 * 2. contiguous at chunk level or overlapping. */
676 if (overlapped(ext, cur)) {
677 /* cur is the minimum unit, so overlapping means
679 EASSERTF((ext->oe_start <= cur->oe_start &&
680 ext->oe_end >= cur->oe_end),
681 ext, EXTSTR, EXTPARA(cur));
683 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
684 /* for simplicity, we wait for this extent to
685 * finish before going forward. */
686 conflict = osc_extent_get(ext);
690 found = osc_extent_hold(ext);
694 /* non-overlapped extent */
695 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
696 /* we can't do anything for a non OES_CACHE extent, or
697 * if there is someone waiting for this extent to be
698 * flushed, try next one. */
699 ext = next_extent(ext);
703 /* check if they belong to the same rpc slot before trying to
704 * merge. the extents are not overlapped and contiguous at
705 * chunk level to get here. */
706 if (ext->oe_max_end != max_end) {
707 /* if they don't belong to the same RPC slot or
708 * max_pages_per_rpc has ever changed, do not merge. */
709 ext = next_extent(ext);
713 /* it's required that an extent must be contiguous at chunk
714 * level so that we know the whole extent is covered by grant
715 * (the pages in the extent are NOT required to be contiguous).
716 * Otherwise, it will be too much difficult to know which
717 * chunks have grants allocated. */
719 /* try to do front merge - extend ext's start */
720 if (chunk + 1 == ext_chk_start) {
721 /* ext must be chunk size aligned */
722 EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
724 /* pull ext's start back to cover cur */
725 ext->oe_start = cur->oe_start;
726 ext->oe_grants += chunksize;
727 *grants -= chunksize;
729 found = osc_extent_hold(ext);
730 } else if (chunk == ext_chk_end + 1) {
732 ext->oe_end = cur->oe_end;
733 ext->oe_grants += chunksize;
734 *grants -= chunksize;
736 /* try to merge with the next one because we just fill
738 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
739 /* we can save extent tax from next extent */
740 *grants += cli->cl_extent_tax;
742 found = osc_extent_hold(ext);
747 ext = next_extent(ext);
750 osc_extent_tree_dump(D_CACHE, obj);
752 LASSERT(conflict == NULL);
753 if (!IS_ERR(found)) {
754 LASSERT(found->oe_osclock == cur->oe_osclock);
755 OSC_EXTENT_DUMP(D_CACHE, found,
756 "found caching ext for %lu.\n", index);
758 } else if (conflict == NULL) {
759 /* create a new extent */
760 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
761 cur->oe_grants = chunksize + cli->cl_extent_tax;
762 *grants -= cur->oe_grants;
763 LASSERT(*grants >= 0);
765 cur->oe_state = OES_CACHE;
766 found = osc_extent_hold(cur);
767 osc_extent_insert(obj, cur);
768 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
769 index, lock->cll_descr.cld_end);
771 osc_object_unlock(obj);
773 if (conflict != NULL) {
774 LASSERT(found == NULL);
776 /* waiting for IO to finish. Please notice that it's impossible
777 * to be an OES_TRUNC extent. */
778 rc = osc_extent_wait(env, conflict, OES_INV);
779 osc_extent_put(env, conflict);
782 GOTO(out, found = ERR_PTR(rc));
789 osc_extent_put(env, cur);
790 LASSERT(*grants >= 0);
795 * Called when IO is finished to an extent.
797 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
800 struct client_obd *cli = osc_cli(ext->oe_obj);
801 struct osc_async_page *oap;
802 struct osc_async_page *tmp;
803 int nr_pages = ext->oe_nr_pages;
805 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
810 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
812 ext->oe_rc = rc ?: ext->oe_nr_pages;
813 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
815 osc_lru_add_batch(cli, &ext->oe_pages);
816 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
818 list_del_init(&oap->oap_rpc_item);
819 list_del_init(&oap->oap_pending_item);
820 if (last_off <= oap->oap_obj_off) {
821 last_off = oap->oap_obj_off;
822 last_count = oap->oap_count;
826 osc_ap_completion(env, cli, oap, sent, rc);
828 EASSERT(ext->oe_nr_pages == 0, ext);
831 lost_grant = ext->oe_grants;
832 } else if (blocksize < PAGE_CACHE_SIZE &&
833 last_count != PAGE_CACHE_SIZE) {
834 /* For short writes we shouldn't count parts of pages that
835 * span a whole chunk on the OST side, or our accounting goes
836 * wrong. Should match the code in filter_grant_check. */
837 int offset = last_off & ~CFS_PAGE_MASK;
838 int count = last_count + (offset & (blocksize - 1));
839 int end = (offset + last_count) & (blocksize - 1);
841 count += blocksize - end;
843 lost_grant = PAGE_CACHE_SIZE - count;
845 if (ext->oe_grants > 0)
846 osc_free_grant(cli, nr_pages, lost_grant);
848 osc_extent_remove(ext);
849 /* put the refcount for RPC */
850 osc_extent_put(env, ext);
854 static int extent_wait_cb(struct osc_extent *ext, int state)
858 osc_object_lock(ext->oe_obj);
859 ret = ext->oe_state == state;
860 osc_object_unlock(ext->oe_obj);
866 * Wait for the extent's state to become @state.
868 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
871 struct osc_object *obj = ext->oe_obj;
872 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
873 LWI_ON_SIGNAL_NOOP, NULL);
877 osc_object_lock(obj);
878 LASSERT(sanity_check_nolock(ext) == 0);
879 /* `Kick' this extent only if the caller is waiting for it to be
881 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
882 if (ext->oe_state == OES_ACTIVE) {
884 } else if (ext->oe_state == OES_CACHE) {
886 osc_extent_hold(ext);
890 osc_object_unlock(obj);
892 osc_extent_release(env, ext);
894 /* wait for the extent until its state becomes @state */
895 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
896 if (rc == -ETIMEDOUT) {
897 OSC_EXTENT_DUMP(D_ERROR, ext,
898 "%s: wait ext to %d timedout, recovery in progress?\n",
899 osc_export(obj)->exp_obd->obd_name, state);
901 lwi = LWI_INTR(NULL, NULL);
902 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
905 if (rc == 0 && ext->oe_rc < 0)
911 * Discard pages with index greater than @size. If @ext is overlapped with
912 * @size, then partial truncate happens.
914 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
917 struct cl_env_nest nest;
920 struct osc_object *obj = ext->oe_obj;
921 struct client_obd *cli = osc_cli(obj);
922 struct osc_async_page *oap;
923 struct osc_async_page *tmp;
924 int pages_in_chunk = 0;
925 int ppc_bits = cli->cl_chunkbits -
927 __u64 trunc_chunk = trunc_index >> ppc_bits;
933 LASSERT(sanity_check(ext) == 0);
934 LASSERT(ext->oe_state == OES_TRUNC);
935 LASSERT(!ext->oe_urgent);
937 /* Request new lu_env.
938 * We can't use that env from osc_cache_truncate_start() because
939 * it's from lov_io_sub and not fully initialized. */
940 env = cl_env_nested_get(&nest);
941 io = &osc_env_info(env)->oti_io;
942 io->ci_obj = cl_object_top(osc2cl(obj));
943 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
947 /* discard all pages with index greater then trunc_index */
948 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
950 pgoff_t index = osc_index(oap2osc(oap));
951 struct cl_page *page = oap2cl_page(oap);
953 LASSERT(list_empty(&oap->oap_rpc_item));
955 /* only discard the pages with their index greater than
956 * trunc_index, and ... */
957 if (index < trunc_index ||
958 (index == trunc_index && partial)) {
959 /* accounting how many pages remaining in the chunk
960 * so that we can calculate grants correctly. */
961 if (index >> ppc_bits == trunc_chunk)
966 list_del_init(&oap->oap_pending_item);
969 lu_ref_add(&page->cp_reference, "truncate", current);
971 if (cl_page_own(env, io, page) == 0) {
972 cl_page_discard(env, io, page);
973 cl_page_disown(env, io, page);
975 LASSERT(page->cp_state == CPS_FREEING);
979 lu_ref_del(&page->cp_reference, "truncate", current);
980 cl_page_put(env, page);
985 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
986 ext->oe_nr_pages == 0),
987 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
989 osc_object_lock(obj);
990 if (ext->oe_nr_pages == 0) {
991 LASSERT(pages_in_chunk == 0);
992 grants = ext->oe_grants;
994 } else { /* calculate how many grants we can free */
995 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
999 /* if there is no pages in this chunk, we can also free grants
1000 * for the last chunk */
1001 if (pages_in_chunk == 0) {
1002 /* if this is the 1st chunk and no pages in this chunk,
1003 * ext->oe_nr_pages must be zero, so we should be in
1004 * the other if-clause. */
1005 LASSERT(trunc_chunk > 0);
1010 /* this is what we can free from this extent */
1011 grants = chunks << cli->cl_chunkbits;
1012 ext->oe_grants -= grants;
1013 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1014 ext->oe_end = min(last_index, ext->oe_max_end);
1015 LASSERT(ext->oe_end >= ext->oe_start);
1016 LASSERT(ext->oe_grants > 0);
1018 osc_object_unlock(obj);
1020 if (grants > 0 || nr_pages > 0)
1021 osc_free_grant(cli, nr_pages, grants);
1024 cl_io_fini(env, io);
1025 cl_env_nested_put(&nest, env);
1030 * This function is used to make the extent prepared for transfer.
1031 * A race with flusing page - ll_writepage() has to be handled cautiously.
1033 static int osc_extent_make_ready(const struct lu_env *env,
1034 struct osc_extent *ext)
1036 struct osc_async_page *oap;
1037 struct osc_async_page *last = NULL;
1038 struct osc_object *obj = ext->oe_obj;
1043 /* we're going to grab page lock, so object lock must not be taken. */
1044 LASSERT(sanity_check(ext) == 0);
1045 /* in locking state, any process should not touch this extent. */
1046 EASSERT(ext->oe_state == OES_LOCKING, ext);
1047 EASSERT(ext->oe_owner != NULL, ext);
1049 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1051 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1053 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1056 /* checking ASYNC_READY is race safe */
1057 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1060 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1063 spin_lock(&oap->oap_lock);
1064 oap->oap_async_flags |= ASYNC_READY;
1065 spin_unlock(&oap->oap_lock);
1068 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1071 LASSERTF(0, "unknown return code: %d\n", rc);
1075 LASSERT(page_count == ext->oe_nr_pages);
1076 LASSERT(last != NULL);
1077 /* the last page is the only one we need to refresh its count by
1078 * the size of file. */
1079 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1080 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1081 LASSERT(last->oap_count > 0);
1082 LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
1083 spin_lock(&last->oap_lock);
1084 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1085 spin_unlock(&last->oap_lock);
1088 /* for the rest of pages, we don't need to call osf_refresh_count()
1089 * because it's known they are not the last page */
1090 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1091 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1092 oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1093 spin_lock(&oap->oap_lock);
1094 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1095 spin_unlock(&oap->oap_lock);
1099 osc_object_lock(obj);
1100 osc_extent_state_set(ext, OES_RPC);
1101 osc_object_unlock(obj);
1102 /* get a refcount for RPC. */
1103 osc_extent_get(ext);
1109 * Quick and simple version of osc_extent_find(). This function is frequently
1110 * called to expand the extent for the same IO. To expand the extent, the
1111 * page index must be in the same or next chunk of ext->oe_end.
1113 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1115 struct osc_object *obj = ext->oe_obj;
1116 struct client_obd *cli = osc_cli(obj);
1117 struct osc_extent *next;
1118 int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1119 pgoff_t chunk = index >> ppc_bits;
1122 int chunksize = 1 << cli->cl_chunkbits;
1126 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1127 osc_object_lock(obj);
1128 LASSERT(sanity_check_nolock(ext) == 0);
1129 end_chunk = ext->oe_end >> ppc_bits;
1130 if (chunk > end_chunk + 1)
1131 GOTO(out, rc = -ERANGE);
1133 if (end_chunk >= chunk)
1136 LASSERT(end_chunk + 1 == chunk);
1137 /* try to expand this extent to cover @index */
1138 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1140 next = next_extent(ext);
1141 if (next != NULL && next->oe_start <= end_index)
1142 /* complex mode - overlapped with the next extent,
1143 * this case will be handled by osc_extent_find() */
1144 GOTO(out, rc = -EAGAIN);
1146 ext->oe_end = end_index;
1147 ext->oe_grants += chunksize;
1148 *grants -= chunksize;
1149 LASSERT(*grants >= 0);
1150 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1151 "overlapped after expanding for %lu.\n", index);
1155 osc_object_unlock(obj);
1159 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1160 const char *func, int line)
1162 struct osc_extent *ext;
1165 CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1166 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1168 /* osc_object_lock(obj); */
1170 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1171 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1174 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1175 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1178 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1179 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1182 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1183 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1184 /* osc_object_unlock(obj); */
1187 /* ------------------ osc extent end ------------------ */
1189 static inline int osc_is_ready(struct osc_object *osc)
1191 return !list_empty(&osc->oo_ready_item) ||
1192 !list_empty(&osc->oo_hp_ready_item);
1195 #define OSC_IO_DEBUG(OSC, STR, args...) \
1196 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1197 (OSC), osc_is_ready(OSC), \
1198 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1199 list_empty_marker(&(OSC)->oo_ready_item), \
1200 atomic_read(&(OSC)->oo_nr_writes), \
1201 list_empty_marker(&(OSC)->oo_hp_exts), \
1202 list_empty_marker(&(OSC)->oo_urgent_exts), \
1203 atomic_read(&(OSC)->oo_nr_reads), \
1204 list_empty_marker(&(OSC)->oo_reading_exts), \
1207 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1210 struct osc_page *opg = oap2osc_page(oap);
1211 struct cl_page *page = oap2cl_page(oap);
1214 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1217 result = cl_page_make_ready(env, page, CRT_WRITE);
1219 opg->ops_submit_time = cfs_time_current();
1223 static int osc_refresh_count(const struct lu_env *env,
1224 struct osc_async_page *oap, int cmd)
1226 struct osc_page *opg = oap2osc_page(oap);
1227 pgoff_t index = osc_index(oap2osc(oap));
1228 struct cl_object *obj;
1229 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1234 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1235 LASSERT(!(cmd & OBD_BRW_READ));
1236 LASSERT(opg != NULL);
1237 obj = opg->ops_cl.cpl_obj;
1239 cl_object_attr_lock(obj);
1240 result = cl_object_attr_get(env, obj, attr);
1241 cl_object_attr_unlock(obj);
1244 kms = attr->cat_kms;
1245 if (cl_offset(obj, index) >= kms)
1246 /* catch race with truncate */
1248 else if (cl_offset(obj, index + 1) > kms)
1249 /* catch sub-page write at end of file */
1250 return kms % PAGE_CACHE_SIZE;
1252 return PAGE_CACHE_SIZE;
1255 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1258 struct osc_page *opg = oap2osc_page(oap);
1259 struct cl_page *page = oap2cl_page(oap);
1260 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
1261 enum cl_req_type crt;
1266 cmd &= ~OBD_BRW_NOQUOTA;
1267 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1268 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1269 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1270 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1271 LASSERT(opg->ops_transfer_pinned);
1274 * page->cp_req can be NULL if io submission failed before
1275 * cl_req was allocated.
1277 if (page->cp_req != NULL)
1278 cl_req_page_done(env, page);
1279 LASSERT(page->cp_req == NULL);
1281 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1282 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1283 opg->ops_transfer_pinned = 0;
1285 spin_lock(&obj->oo_seatbelt);
1286 LASSERT(opg->ops_submitter != NULL);
1287 LASSERT(!list_empty(&opg->ops_inflight));
1288 list_del_init(&opg->ops_inflight);
1289 opg->ops_submitter = NULL;
1290 spin_unlock(&obj->oo_seatbelt);
1292 opg->ops_submit_time = 0;
1293 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1296 if (rc == 0 && srvlock) {
1297 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1298 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1299 int bytes = oap->oap_count;
1301 if (crt == CRT_READ)
1302 stats->os_lockless_reads += bytes;
1304 stats->os_lockless_writes += bytes;
1308 * This has to be the last operation with the page, as locks are
1309 * released in cl_page_completion() and nothing except for the
1310 * reference counter protects page from concurrent reclaim.
1312 lu_ref_del(&page->cp_reference, "transfer", page);
1314 cl_page_completion(env, page, crt, rc);
1319 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do { \
1320 struct client_obd *__tmp = (cli); \
1321 CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %d/%d " \
1322 "dropped: %ld avail: %ld, reserved: %ld, flight: %d }" \
1323 "lru {in list: %d, left: %d, waiters: %d }" fmt, \
1324 __tmp->cl_import->imp_obd->obd_name, \
1325 __tmp->cl_dirty, __tmp->cl_dirty_max, \
1326 atomic_read(&obd_dirty_pages), obd_max_dirty_pages, \
1327 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1328 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1329 atomic_read(&__tmp->cl_lru_in_list), \
1330 atomic_read(&__tmp->cl_lru_busy), \
1331 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1334 /* caller must hold loi_list_lock */
1335 static void osc_consume_write_grant(struct client_obd *cli,
1336 struct brw_page *pga)
1338 assert_spin_locked(&cli->cl_loi_list_lock.lock);
1339 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1340 atomic_inc(&obd_dirty_pages);
1341 cli->cl_dirty += PAGE_CACHE_SIZE;
1342 pga->flag |= OBD_BRW_FROM_GRANT;
1343 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1344 PAGE_CACHE_SIZE, pga, pga->pg);
1345 osc_update_next_shrink(cli);
1348 /* the companion to osc_consume_write_grant, called when a brw has completed.
1349 * must be called with the loi lock held. */
1350 static void osc_release_write_grant(struct client_obd *cli,
1351 struct brw_page *pga)
1355 assert_spin_locked(&cli->cl_loi_list_lock.lock);
1356 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1361 pga->flag &= ~OBD_BRW_FROM_GRANT;
1362 atomic_dec(&obd_dirty_pages);
1363 cli->cl_dirty -= PAGE_CACHE_SIZE;
1364 if (pga->flag & OBD_BRW_NOCACHE) {
1365 pga->flag &= ~OBD_BRW_NOCACHE;
1366 atomic_dec(&obd_dirty_transit_pages);
1367 cli->cl_dirty_transit -= PAGE_CACHE_SIZE;
1373 * To avoid sleeping with object lock held, it's good for us allocate enough
1374 * grants before entering into critical section.
1376 * client_obd_list_lock held by caller
1378 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1382 if (cli->cl_avail_grant >= bytes) {
1383 cli->cl_avail_grant -= bytes;
1384 cli->cl_reserved_grant += bytes;
1390 static void __osc_unreserve_grant(struct client_obd *cli,
1391 unsigned int reserved, unsigned int unused)
1393 /* it's quite normal for us to get more grant than reserved.
1394 * Thinking about a case that two extents merged by adding a new
1395 * chunk, we can save one extent tax. If extent tax is greater than
1396 * one chunk, we can save more grant by adding a new chunk */
1397 cli->cl_reserved_grant -= reserved;
1398 if (unused > reserved) {
1399 cli->cl_avail_grant += reserved;
1400 cli->cl_lost_grant += unused - reserved;
1402 cli->cl_avail_grant += unused;
1406 void osc_unreserve_grant(struct client_obd *cli,
1407 unsigned int reserved, unsigned int unused)
1409 client_obd_list_lock(&cli->cl_loi_list_lock);
1410 __osc_unreserve_grant(cli, reserved, unused);
1412 osc_wake_cache_waiters(cli);
1413 client_obd_list_unlock(&cli->cl_loi_list_lock);
1417 * Free grant after IO is finished or canceled.
1419 * @lost_grant is used to remember how many grants we have allocated but not
1420 * used, we should return these grants to OST. There're two cases where grants
1423 * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1424 * written. In this case OST may use less chunks to serve this partial
1425 * write. OSTs don't actually know the page size on the client side. so
1426 * clients have to calculate lost grant by the blocksize on the OST.
1427 * See filter_grant_check() for details.
1429 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1430 unsigned int lost_grant)
1432 int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1434 client_obd_list_lock(&cli->cl_loi_list_lock);
1435 atomic_sub(nr_pages, &obd_dirty_pages);
1436 cli->cl_dirty -= nr_pages << PAGE_CACHE_SHIFT;
1437 cli->cl_lost_grant += lost_grant;
1438 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1439 /* borrow some grant from truncate to avoid the case that
1440 * truncate uses up all avail grant */
1441 cli->cl_lost_grant -= grant;
1442 cli->cl_avail_grant += grant;
1444 osc_wake_cache_waiters(cli);
1445 client_obd_list_unlock(&cli->cl_loi_list_lock);
1446 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1447 lost_grant, cli->cl_lost_grant,
1448 cli->cl_avail_grant, cli->cl_dirty);
1452 * The companion to osc_enter_cache(), called when @oap is no longer part of
1453 * the dirty accounting due to error.
1455 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1457 client_obd_list_lock(&cli->cl_loi_list_lock);
1458 osc_release_write_grant(cli, &oap->oap_brw_page);
1459 client_obd_list_unlock(&cli->cl_loi_list_lock);
1463 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1466 static int osc_enter_cache_try(struct client_obd *cli,
1467 struct osc_async_page *oap,
1468 int bytes, int transient)
1472 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1474 rc = osc_reserve_grant(cli, bytes);
1478 if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max &&
1479 1 + atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1480 osc_consume_write_grant(cli, &oap->oap_brw_page);
1482 cli->cl_dirty_transit += PAGE_CACHE_SIZE;
1483 atomic_inc(&obd_dirty_transit_pages);
1484 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1488 __osc_unreserve_grant(cli, bytes, bytes);
1494 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1497 client_obd_list_lock(&cli->cl_loi_list_lock);
1498 rc = list_empty(&ocw->ocw_entry);
1499 client_obd_list_unlock(&cli->cl_loi_list_lock);
1504 * The main entry to reserve dirty page accounting. Usually the grant reserved
1505 * in this function will be freed in bulk in osc_free_grant() unless it fails
1506 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1508 * The process will be put into sleep if it's already run out of grant.
1510 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1511 struct osc_async_page *oap, int bytes)
1513 struct osc_object *osc = oap->oap_obj;
1514 struct lov_oinfo *loi = osc->oo_oinfo;
1515 struct osc_cache_waiter ocw;
1516 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1517 LWI_ON_SIGNAL_NOOP, NULL);
1521 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1523 client_obd_list_lock(&cli->cl_loi_list_lock);
1525 /* force the caller to try sync io. this can jump the list
1526 * of queued writes and create a discontiguous rpc stream */
1527 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1528 cli->cl_dirty_max < PAGE_CACHE_SIZE ||
1529 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1530 GOTO(out, rc = -EDQUOT);
1532 /* Hopefully normal case - cache space and write credits available */
1533 if (osc_enter_cache_try(cli, oap, bytes, 0))
1536 /* We can get here for two reasons: too many dirty pages in cache, or
1537 * run out of grants. In both cases we should write dirty pages out.
1538 * Adding a cache waiter will trigger urgent write-out no matter what
1540 * The exiting condition is no avail grants and no dirty pages caching,
1541 * that really means there is no space on the OST. */
1542 init_waitqueue_head(&ocw.ocw_waitq);
1544 ocw.ocw_grant = bytes;
1545 while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
1546 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1548 client_obd_list_unlock(&cli->cl_loi_list_lock);
1550 osc_io_unplug_async(env, cli, NULL);
1552 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1553 cli->cl_import->imp_obd->obd_name, &ocw, oap);
1555 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1557 client_obd_list_lock(&cli->cl_loi_list_lock);
1559 /* l_wait_event is interrupted by signal, or timed out */
1563 OSC_DUMP_GRANT(D_ERROR, cli,
1564 "try to reserve %d.\n", bytes);
1565 osc_extent_tree_dump(D_ERROR, osc);
1569 /* Ensures restartability - LU-3581 */
1573 CDEBUG(D_CACHE, "%s: event for cache space @"
1574 " %p never arrived due to %d\n",
1575 cli->cl_import->imp_obd->obd_name,
1579 list_del_init(&ocw.ocw_entry);
1583 LASSERT(list_empty(&ocw.ocw_entry));
1588 if (osc_enter_cache_try(cli, oap, bytes, 0))
1593 client_obd_list_unlock(&cli->cl_loi_list_lock);
1594 OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1598 /* caller must hold loi_list_lock */
1599 void osc_wake_cache_waiters(struct client_obd *cli)
1601 struct list_head *l, *tmp;
1602 struct osc_cache_waiter *ocw;
1605 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1606 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1607 list_del_init(&ocw->ocw_entry);
1609 ocw->ocw_rc = -EDQUOT;
1610 /* we can't dirty more */
1611 if ((cli->cl_dirty + PAGE_CACHE_SIZE > cli->cl_dirty_max) ||
1612 (1 + atomic_read(&obd_dirty_pages) > obd_max_dirty_pages)) {
1613 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1614 "osc max %ld, sys max %d\n", cli->cl_dirty,
1615 cli->cl_dirty_max, obd_max_dirty_pages);
1620 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1621 ocw->ocw_rc = -EDQUOT;
1624 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1625 ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1627 wake_up(&ocw->ocw_waitq);
1633 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1635 int hprpc = !!list_empty(&osc->oo_hp_exts);
1636 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1639 /* This maintains the lists of pending pages to read/write for a given object
1640 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1641 * to quickly find objects that are ready to send an RPC. */
1642 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1645 int invalid_import = 0;
1648 /* if we have an invalid import we want to drain the queued pages
1649 * by forcing them through rpcs that immediately fail and complete
1650 * the pages. recovery relies on this to empty the queued pages
1651 * before canceling the locks and evicting down the llite pages */
1652 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1655 if (cmd & OBD_BRW_WRITE) {
1656 if (atomic_read(&osc->oo_nr_writes) == 0)
1658 if (invalid_import) {
1659 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1662 if (!list_empty(&osc->oo_hp_exts)) {
1663 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1666 if (!list_empty(&osc->oo_urgent_exts)) {
1667 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1670 /* trigger a write rpc stream as long as there are dirtiers
1671 * waiting for space. as they're waiting, they're not going to
1672 * create more pages to coalesce with what's waiting.. */
1673 if (!list_empty(&cli->cl_cache_waiters)) {
1674 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1677 if (atomic_read(&osc->oo_nr_writes) >=
1678 cli->cl_max_pages_per_rpc)
1681 if (atomic_read(&osc->oo_nr_reads) == 0)
1683 if (invalid_import) {
1684 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1687 /* all read are urgent. */
1688 if (!list_empty(&osc->oo_reading_exts))
1695 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1697 struct client_obd *cli = osc_cli(obj);
1698 if (cmd & OBD_BRW_WRITE) {
1699 atomic_add(delta, &obj->oo_nr_writes);
1700 atomic_add(delta, &cli->cl_pending_w_pages);
1701 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1703 atomic_add(delta, &obj->oo_nr_reads);
1704 atomic_add(delta, &cli->cl_pending_r_pages);
1705 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1707 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1710 static int osc_makes_hprpc(struct osc_object *obj)
1712 return !list_empty(&obj->oo_hp_exts);
1715 static void on_list(struct list_head *item, struct list_head *list,
1718 if (list_empty(item) && should_be_on)
1719 list_add_tail(item, list);
1720 else if (!list_empty(item) && !should_be_on)
1721 list_del_init(item);
1724 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1725 * can find pages to build into rpcs quickly */
1726 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1728 if (osc_makes_hprpc(osc)) {
1730 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1731 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1733 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1734 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1735 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1736 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1739 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1740 atomic_read(&osc->oo_nr_writes) > 0);
1742 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1743 atomic_read(&osc->oo_nr_reads) > 0);
1745 return osc_is_ready(osc);
1748 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1752 client_obd_list_lock(&cli->cl_loi_list_lock);
1753 is_ready = __osc_list_maint(cli, osc);
1754 client_obd_list_unlock(&cli->cl_loi_list_lock);
1759 /* this is trying to propogate async writeback errors back up to the
1760 * application. As an async write fails we record the error code for later if
1761 * the app does an fsync. As long as errors persist we force future rpcs to be
1762 * sync so that the app can get a sync error and break the cycle of queueing
1763 * pages for which writeback will fail. */
1764 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1771 ar->ar_force_sync = 1;
1772 ar->ar_min_xid = ptlrpc_sample_next_xid();
1777 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1778 ar->ar_force_sync = 0;
1781 /* this must be called holding the loi list lock to give coverage to exit_cache,
1782 * async_flag maintenance, and oap_request */
1783 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1784 struct osc_async_page *oap, int sent, int rc)
1786 struct osc_object *osc = oap->oap_obj;
1787 struct lov_oinfo *loi = osc->oo_oinfo;
1791 if (oap->oap_request != NULL) {
1792 xid = ptlrpc_req_xid(oap->oap_request);
1793 ptlrpc_req_finished(oap->oap_request);
1794 oap->oap_request = NULL;
1797 /* As the transfer for this page is being done, clear the flags */
1798 spin_lock(&oap->oap_lock);
1799 oap->oap_async_flags = 0;
1800 spin_unlock(&oap->oap_lock);
1801 oap->oap_interrupted = 0;
1803 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1804 client_obd_list_lock(&cli->cl_loi_list_lock);
1805 osc_process_ar(&cli->cl_ar, xid, rc);
1806 osc_process_ar(&loi->loi_ar, xid, rc);
1807 client_obd_list_unlock(&cli->cl_loi_list_lock);
1810 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1812 CERROR("completion on oap %p obj %p returns %d.\n",
1819 * Try to add extent to one RPC. We need to think about the following things:
1820 * - # of pages must not be over max_pages_per_rpc
1821 * - extent must be compatible with previous ones
1823 static int try_to_add_extent_for_io(struct client_obd *cli,
1824 struct osc_extent *ext,
1825 struct list_head *rpclist,
1826 int *pc, unsigned int *max_pages)
1828 struct osc_extent *tmp;
1831 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1834 *max_pages = max(ext->oe_mppr, *max_pages);
1835 if (*pc + ext->oe_nr_pages > *max_pages)
1838 list_for_each_entry(tmp, rpclist, oe_link) {
1839 EASSERT(tmp->oe_owner == current, tmp);
1841 if (overlapped(tmp, ext)) {
1842 OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1847 if (tmp->oe_srvlock != ext->oe_srvlock ||
1848 !tmp->oe_grants != !ext->oe_grants)
1851 /* remove break for strict check */
1855 *pc += ext->oe_nr_pages;
1856 list_move_tail(&ext->oe_link, rpclist);
1857 ext->oe_owner = current;
1862 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1863 * get_write_extent() takes all appropriate extents in atomic.
1865 * The following policy is used to collect extents for IO:
1866 * 1. Add as many HP extents as possible;
1867 * 2. Add the first urgent extent in urgent extent list and take it out of
1869 * 3. Add subsequent extents of this urgent extent;
1870 * 4. If urgent list is not empty, goto 2;
1871 * 5. Traverse the extent tree from the 1st extent;
1872 * 6. Above steps exit if there is no space in this RPC.
1874 static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1876 struct client_obd *cli = osc_cli(obj);
1877 struct osc_extent *ext;
1879 unsigned int max_pages = cli->cl_max_pages_per_rpc;
1881 LASSERT(osc_object_is_locked(obj));
1882 while (!list_empty(&obj->oo_hp_exts)) {
1883 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1885 LASSERT(ext->oe_state == OES_CACHE);
1886 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1889 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1891 if (page_count == max_pages)
1894 while (!list_empty(&obj->oo_urgent_exts)) {
1895 ext = list_entry(obj->oo_urgent_exts.next,
1896 struct osc_extent, oe_link);
1897 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1901 if (!ext->oe_intree)
1904 while ((ext = next_extent(ext)) != NULL) {
1905 if ((ext->oe_state != OES_CACHE) ||
1906 (!list_empty(&ext->oe_link) &&
1907 ext->oe_owner != NULL))
1910 if (!try_to_add_extent_for_io(cli, ext, rpclist,
1911 &page_count, &max_pages))
1915 if (page_count == max_pages)
1918 ext = first_extent(obj);
1919 while (ext != NULL) {
1920 if ((ext->oe_state != OES_CACHE) ||
1921 /* this extent may be already in current rpclist */
1922 (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1923 ext = next_extent(ext);
1927 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1931 ext = next_extent(ext);
1937 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1938 struct osc_object *osc, pdl_policy_t pol)
1940 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
1941 struct osc_extent *ext;
1942 struct osc_extent *tmp;
1943 struct osc_extent *first = NULL;
1944 obd_count page_count = 0;
1949 LASSERT(osc_object_is_locked(osc));
1951 page_count = get_write_extents(osc, &rpclist);
1952 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1954 if (list_empty(&rpclist))
1957 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1959 list_for_each_entry(ext, &rpclist, oe_link) {
1960 LASSERT(ext->oe_state == OES_CACHE ||
1961 ext->oe_state == OES_LOCK_DONE);
1962 if (ext->oe_state == OES_CACHE)
1963 osc_extent_state_set(ext, OES_LOCKING);
1965 osc_extent_state_set(ext, OES_RPC);
1968 /* we're going to grab page lock, so release object lock because
1969 * lock order is page lock -> object lock. */
1970 osc_object_unlock(osc);
1972 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
1973 if (ext->oe_state == OES_LOCKING) {
1974 rc = osc_extent_make_ready(env, ext);
1975 if (unlikely(rc < 0)) {
1976 list_del_init(&ext->oe_link);
1977 osc_extent_finish(env, ext, 0, rc);
1981 if (first == NULL) {
1983 srvlock = ext->oe_srvlock;
1985 LASSERT(srvlock == ext->oe_srvlock);
1989 if (!list_empty(&rpclist)) {
1990 LASSERT(page_count > 0);
1991 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
1992 LASSERT(list_empty(&rpclist));
1995 osc_object_lock(osc);
2000 * prepare pages for ASYNC io and put pages in send queue.
2002 * \param cmd OBD_BRW_* macroses
2003 * \param lop pending pages
2005 * \return zero if no page added to send queue.
2006 * \return 1 if pages successfully added to send queue.
2007 * \return negative on errors.
2010 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2011 struct osc_object *osc, pdl_policy_t pol)
2013 struct osc_extent *ext;
2014 struct osc_extent *next;
2015 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2017 unsigned int max_pages = cli->cl_max_pages_per_rpc;
2021 LASSERT(osc_object_is_locked(osc));
2022 list_for_each_entry_safe(ext, next,
2023 &osc->oo_reading_exts, oe_link) {
2024 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2025 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2028 osc_extent_state_set(ext, OES_RPC);
2029 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2031 LASSERT(page_count <= max_pages);
2033 osc_update_pending(osc, OBD_BRW_READ, -page_count);
2035 if (!list_empty(&rpclist)) {
2036 osc_object_unlock(osc);
2038 LASSERT(page_count > 0);
2039 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2040 LASSERT(list_empty(&rpclist));
2042 osc_object_lock(osc);
2047 #define list_to_obj(list, item) ({ \
2048 struct list_head *__tmp = (list)->next; \
2049 list_del_init(__tmp); \
2050 list_entry(__tmp, struct osc_object, oo_##item); \
2053 /* This is called by osc_check_rpcs() to find which objects have pages that
2054 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2055 static struct osc_object *osc_next_obj(struct client_obd *cli)
2059 /* First return objects that have blocked locks so that they
2060 * will be flushed quickly and other clients can get the lock,
2061 * then objects which have pages ready to be stuffed into RPCs */
2062 if (!list_empty(&cli->cl_loi_hp_ready_list))
2063 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2064 if (!list_empty(&cli->cl_loi_ready_list))
2065 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2067 /* then if we have cache waiters, return all objects with queued
2068 * writes. This is especially important when many small files
2069 * have filled up the cache and not been fired into rpcs because
2070 * they don't pass the nr_pending/object threshhold */
2071 if (!list_empty(&cli->cl_cache_waiters) &&
2072 !list_empty(&cli->cl_loi_write_list))
2073 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2075 /* then return all queued objects when we have an invalid import
2076 * so that they get flushed */
2077 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2078 if (!list_empty(&cli->cl_loi_write_list))
2079 RETURN(list_to_obj(&cli->cl_loi_write_list,
2081 if (!list_empty(&cli->cl_loi_read_list))
2082 RETURN(list_to_obj(&cli->cl_loi_read_list,
2088 /* called with the loi list lock held */
2089 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2092 struct osc_object *osc;
2096 while ((osc = osc_next_obj(cli)) != NULL) {
2097 struct cl_object *obj = osc2cl(osc);
2098 struct lu_ref_link link;
2100 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2102 if (osc_max_rpc_in_flight(cli, osc)) {
2103 __osc_list_maint(cli, osc);
2108 client_obd_list_unlock(&cli->cl_loi_list_lock);
2109 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2111 /* attempt some read/write balancing by alternating between
2112 * reads and writes in an object. The makes_rpc checks here
2113 * would be redundant if we were getting read/write work items
2114 * instead of objects. we don't want send_oap_rpc to drain a
2115 * partial read pending queue when we're given this object to
2116 * do io on writes while there are cache waiters */
2117 osc_object_lock(osc);
2118 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2119 rc = osc_send_write_rpc(env, cli, osc, pol);
2121 CERROR("Write request failed with %d\n", rc);
2123 /* osc_send_write_rpc failed, mostly because of
2126 * It can't break here, because if:
2127 * - a page was submitted by osc_io_submit, so
2129 * - no request in flight
2130 * - no subsequent request
2131 * The system will be in live-lock state,
2132 * because there is no chance to call
2133 * osc_io_unplug() and osc_check_rpcs() any
2134 * more. pdflush can't help in this case,
2135 * because it might be blocked at grabbing
2136 * the page lock as we mentioned.
2138 * Anyway, continue to drain pages. */
2142 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2143 rc = osc_send_read_rpc(env, cli, osc, pol);
2145 CERROR("Read request failed with %d\n", rc);
2147 osc_object_unlock(osc);
2149 osc_list_maint(cli, osc);
2150 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2151 cl_object_put(env, obj);
2153 client_obd_list_lock(&cli->cl_loi_list_lock);
2157 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2158 struct osc_object *osc, pdl_policy_t pol, int async)
2162 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2166 /* disable osc_lru_shrink() temporarily to avoid
2167 * potential stack overrun problem. LU-2859 */
2168 atomic_inc(&cli->cl_lru_shrinkers);
2169 client_obd_list_lock(&cli->cl_loi_list_lock);
2170 osc_check_rpcs(env, cli, pol);
2171 client_obd_list_unlock(&cli->cl_loi_list_lock);
2172 atomic_dec(&cli->cl_lru_shrinkers);
2174 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2175 LASSERT(cli->cl_writeback_work != NULL);
2176 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2181 static int osc_io_unplug_async(const struct lu_env *env,
2182 struct client_obd *cli, struct osc_object *osc)
2184 /* XXX: policy is no use actually. */
2185 return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2188 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2189 struct osc_object *osc, pdl_policy_t pol)
2191 (void)osc_io_unplug0(env, cli, osc, pol, 0);
2194 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2195 struct page *page, loff_t offset)
2197 struct obd_export *exp = osc_export(osc);
2198 struct osc_async_page *oap = &ops->ops_oap;
2202 return cfs_size_round(sizeof(*oap));
2204 oap->oap_magic = OAP_MAGIC;
2205 oap->oap_cli = &exp->exp_obd->u.cli;
2208 oap->oap_page = page;
2209 oap->oap_obj_off = offset;
2210 LASSERT(!(offset & ~CFS_PAGE_MASK));
2212 if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2213 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2215 INIT_LIST_HEAD(&oap->oap_pending_item);
2216 INIT_LIST_HEAD(&oap->oap_rpc_item);
2218 spin_lock_init(&oap->oap_lock);
2219 CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2220 oap, page, oap->oap_obj_off);
2224 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2225 struct osc_page *ops)
2227 struct osc_io *oio = osc_env_io(env);
2228 struct osc_extent *ext = NULL;
2229 struct osc_async_page *oap = &ops->ops_oap;
2230 struct client_obd *cli = oap->oap_cli;
2231 struct osc_object *osc = oap->oap_obj;
2234 int brw_flags = OBD_BRW_ASYNC;
2235 int cmd = OBD_BRW_WRITE;
2236 int need_release = 0;
2240 if (oap->oap_magic != OAP_MAGIC)
2243 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2246 if (!list_empty(&oap->oap_pending_item) ||
2247 !list_empty(&oap->oap_rpc_item))
2250 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2251 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2252 if (!client_is_remote(osc_export(osc)) &&
2253 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2254 brw_flags |= OBD_BRW_NOQUOTA;
2255 cmd |= OBD_BRW_NOQUOTA;
2258 /* check if the file's owner/group is over quota */
2259 if (!(cmd & OBD_BRW_NOQUOTA)) {
2260 struct cl_object *obj;
2261 struct cl_attr *attr;
2262 unsigned int qid[MAXQUOTAS];
2264 obj = cl_object_top(&osc->oo_cl);
2265 attr = &osc_env_info(env)->oti_attr;
2267 cl_object_attr_lock(obj);
2268 rc = cl_object_attr_get(env, obj, attr);
2269 cl_object_attr_unlock(obj);
2271 qid[USRQUOTA] = attr->cat_uid;
2272 qid[GRPQUOTA] = attr->cat_gid;
2273 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2280 oap->oap_page_off = ops->ops_from;
2281 oap->oap_count = ops->ops_to - ops->ops_from;
2282 /* No need to hold a lock here,
2283 * since this page is not in any list yet. */
2284 oap->oap_async_flags = 0;
2285 oap->oap_brw_flags = brw_flags;
2287 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2288 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2290 index = osc_index(oap2osc(oap));
2292 /* Add this page into extent by the following steps:
2293 * 1. if there exists an active extent for this IO, mostly this page
2294 * can be added to the active extent and sometimes we need to
2295 * expand extent to accomodate this page;
2296 * 2. otherwise, a new extent will be allocated. */
2298 ext = oio->oi_active;
2299 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2300 /* one chunk plus extent overhead must be enough to write this
2302 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2303 if (ext->oe_end >= index)
2306 /* it doesn't need any grant to dirty this page */
2307 client_obd_list_lock(&cli->cl_loi_list_lock);
2308 rc = osc_enter_cache_try(cli, oap, grants, 0);
2309 client_obd_list_unlock(&cli->cl_loi_list_lock);
2310 if (rc == 0) { /* try failed */
2313 } else if (ext->oe_end < index) {
2315 /* try to expand this extent */
2316 rc = osc_extent_expand(ext, index, &tmp);
2319 /* don't free reserved grant */
2321 OSC_EXTENT_DUMP(D_CACHE, ext,
2322 "expanded for %lu.\n", index);
2323 osc_unreserve_grant(cli, grants, tmp);
2328 } else if (ext != NULL) {
2329 /* index is located outside of active extent */
2333 osc_extent_release(env, ext);
2334 oio->oi_active = NULL;
2339 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2341 /* try to find new extent to cover this page */
2342 LASSERT(oio->oi_active == NULL);
2343 /* we may have allocated grant for this page if we failed
2344 * to expand the previous active extent. */
2345 LASSERT(ergo(grants > 0, grants >= tmp));
2349 /* we haven't allocated grant for this page. */
2350 rc = osc_enter_cache(env, cli, oap, tmp);
2357 ext = osc_extent_find(env, osc, index, &tmp);
2359 LASSERT(tmp == grants);
2360 osc_exit_cache(cli, oap);
2364 oio->oi_active = ext;
2368 osc_unreserve_grant(cli, grants, tmp);
2371 LASSERT(ergo(rc == 0, ext != NULL));
2373 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2374 ext, "index = %lu.\n", index);
2375 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2377 osc_object_lock(osc);
2378 if (ext->oe_nr_pages == 0)
2379 ext->oe_srvlock = ops->ops_srvlock;
2381 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2383 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2384 osc_object_unlock(osc);
2389 int osc_teardown_async_page(const struct lu_env *env,
2390 struct osc_object *obj, struct osc_page *ops)
2392 struct osc_async_page *oap = &ops->ops_oap;
2393 struct osc_extent *ext = NULL;
2397 LASSERT(oap->oap_magic == OAP_MAGIC);
2399 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2400 oap, ops, osc_index(oap2osc(oap)));
2402 osc_object_lock(obj);
2403 if (!list_empty(&oap->oap_rpc_item)) {
2404 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2406 } else if (!list_empty(&oap->oap_pending_item)) {
2407 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2408 /* only truncated pages are allowed to be taken out.
2409 * See osc_extent_truncate() and osc_cache_truncate_start()
2411 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2412 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2413 osc_index(oap2osc(oap)));
2417 osc_object_unlock(obj);
2419 osc_extent_put(env, ext);
2424 * This is called when a page is picked up by kernel to write out.
2426 * We should find out the corresponding extent and add the whole extent
2427 * into urgent list. The extent may be being truncated or used, handle it
2430 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2431 struct osc_page *ops)
2433 struct osc_extent *ext = NULL;
2434 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2435 struct cl_page *cp = ops->ops_cl.cpl_page;
2436 pgoff_t index = osc_index(ops);
2437 struct osc_async_page *oap = &ops->ops_oap;
2438 bool unplug = false;
2442 osc_object_lock(obj);
2443 ext = osc_extent_lookup(obj, index);
2445 osc_extent_tree_dump(D_ERROR, obj);
2446 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2449 switch (ext->oe_state) {
2452 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2456 /* If we know this extent is being written out, we should abort
2457 * so that the writer can make this page ready. Otherwise, there
2458 * exists a deadlock problem because other process can wait for
2459 * page writeback bit holding page lock; and meanwhile in
2460 * vvp_page_make_ready(), we need to grab page lock before
2461 * really sending the RPC. */
2463 /* race with truncate, page will be redirtied */
2465 /* The extent is active so we need to abort and let the caller
2466 * re-dirty the page. If we continued on here, and we were the
2467 * one making the extent active, we could deadlock waiting for
2468 * the page writeback to clear but it won't because the extent
2469 * is active and won't be written out. */
2470 GOTO(out, rc = -EAGAIN);
2475 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2479 spin_lock(&oap->oap_lock);
2480 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2481 spin_unlock(&oap->oap_lock);
2483 if (memory_pressure_get())
2484 ext->oe_memalloc = 1;
2487 if (ext->oe_state == OES_CACHE) {
2488 OSC_EXTENT_DUMP(D_CACHE, ext,
2489 "flush page %p make it urgent.\n", oap);
2490 if (list_empty(&ext->oe_link))
2491 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2498 osc_object_unlock(obj);
2499 osc_extent_put(env, ext);
2501 osc_io_unplug_async(env, osc_cli(obj), obj);
2506 * this is called when a sync waiter receives an interruption. Its job is to
2507 * get the caller woken as soon as possible. If its page hasn't been put in an
2508 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2509 * desiring interruption which will forcefully complete the rpc once the rpc
2512 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2514 struct osc_async_page *oap = &ops->ops_oap;
2515 struct osc_object *obj = oap->oap_obj;
2516 struct client_obd *cli = osc_cli(obj);
2517 struct osc_extent *ext;
2518 struct osc_extent *found = NULL;
2519 struct list_head *plist;
2520 pgoff_t index = osc_index(ops);
2525 LASSERT(!oap->oap_interrupted);
2526 oap->oap_interrupted = 1;
2528 /* Find out the caching extent */
2529 osc_object_lock(obj);
2530 if (oap->oap_cmd & OBD_BRW_WRITE) {
2531 plist = &obj->oo_urgent_exts;
2532 cmd = OBD_BRW_WRITE;
2534 plist = &obj->oo_reading_exts;
2537 list_for_each_entry(ext, plist, oe_link) {
2538 if (ext->oe_start <= index && ext->oe_end >= index) {
2539 LASSERT(ext->oe_state == OES_LOCK_DONE);
2540 /* For OES_LOCK_DONE state extent, it has already held
2541 * a refcount for RPC. */
2542 found = osc_extent_get(ext);
2546 if (found != NULL) {
2547 list_del_init(&found->oe_link);
2548 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2549 osc_object_unlock(obj);
2551 osc_extent_finish(env, found, 0, -EINTR);
2552 osc_extent_put(env, found);
2555 osc_object_unlock(obj);
2556 /* ok, it's been put in an rpc. only one oap gets a request
2558 if (oap->oap_request != NULL) {
2559 ptlrpc_mark_interrupted(oap->oap_request);
2560 ptlrpcd_wake(oap->oap_request);
2561 ptlrpc_req_finished(oap->oap_request);
2562 oap->oap_request = NULL;
2566 osc_list_maint(cli, obj);
2570 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2571 struct list_head *list, int cmd, int brw_flags)
2573 struct client_obd *cli = osc_cli(obj);
2574 struct osc_extent *ext;
2575 struct osc_async_page *oap;
2577 int mppr = cli->cl_max_pages_per_rpc;
2578 pgoff_t start = CL_PAGE_EOF;
2582 list_for_each_entry(oap, list, oap_pending_item) {
2583 pgoff_t index = osc_index(oap2osc(oap));
2589 mppr <<= (page_count > mppr);
2592 ext = osc_extent_alloc(obj);
2594 list_for_each_entry(oap, list, oap_pending_item) {
2595 list_del_init(&oap->oap_pending_item);
2596 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2601 ext->oe_rw = !!(cmd & OBD_BRW_READ);
2603 ext->oe_start = start;
2604 ext->oe_end = ext->oe_max_end = end;
2606 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2607 ext->oe_nr_pages = page_count;
2608 ext->oe_mppr = mppr;
2609 list_splice_init(list, &ext->oe_pages);
2611 osc_object_lock(obj);
2612 /* Reuse the initial refcount for RPC, don't drop it */
2613 osc_extent_state_set(ext, OES_LOCK_DONE);
2614 if (cmd & OBD_BRW_WRITE) {
2615 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2616 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2618 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2619 osc_update_pending(obj, OBD_BRW_READ, page_count);
2621 osc_object_unlock(obj);
2623 osc_io_unplug_async(env, cli, obj);
2628 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2630 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2631 struct osc_object *obj, __u64 size)
2633 struct client_obd *cli = osc_cli(obj);
2634 struct osc_extent *ext;
2635 struct osc_extent *waiting = NULL;
2637 struct list_head list = LIST_HEAD_INIT(list);
2642 /* pages with index greater or equal to index will be truncated. */
2643 index = cl_index(osc2cl(obj), size);
2644 partial = size > cl_offset(osc2cl(obj), index);
2647 osc_object_lock(obj);
2648 ext = osc_extent_search(obj, index);
2650 ext = first_extent(obj);
2651 else if (ext->oe_end < index)
2652 ext = next_extent(ext);
2653 while (ext != NULL) {
2654 EASSERT(ext->oe_state != OES_TRUNC, ext);
2656 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2657 /* if ext is in urgent state, it means there must exist
2658 * a page already having been flushed by write_page().
2659 * We have to wait for this extent because we can't
2660 * truncate that page. */
2661 LASSERT(!ext->oe_hp);
2662 OSC_EXTENT_DUMP(D_CACHE, ext,
2663 "waiting for busy extent\n");
2664 waiting = osc_extent_get(ext);
2668 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2670 osc_extent_get(ext);
2671 if (ext->oe_state == OES_ACTIVE) {
2672 /* though we grab inode mutex for write path, but we
2673 * release it before releasing extent(in osc_io_end()),
2674 * so there is a race window that an extent is still
2675 * in OES_ACTIVE when truncate starts. */
2676 LASSERT(!ext->oe_trunc_pending);
2677 ext->oe_trunc_pending = 1;
2679 EASSERT(ext->oe_state == OES_CACHE, ext);
2680 osc_extent_state_set(ext, OES_TRUNC);
2681 osc_update_pending(obj, OBD_BRW_WRITE,
2684 EASSERT(list_empty(&ext->oe_link), ext);
2685 list_add_tail(&ext->oe_link, &list);
2687 ext = next_extent(ext);
2689 osc_object_unlock(obj);
2691 osc_list_maint(cli, obj);
2693 while (!list_empty(&list)) {
2696 ext = list_entry(list.next, struct osc_extent, oe_link);
2697 list_del_init(&ext->oe_link);
2699 /* extent may be in OES_ACTIVE state because inode mutex
2700 * is released before osc_io_end() in file write case */
2701 if (ext->oe_state != OES_TRUNC)
2702 osc_extent_wait(env, ext, OES_TRUNC);
2704 rc = osc_extent_truncate(ext, index, partial);
2709 OSC_EXTENT_DUMP(D_ERROR, ext,
2710 "truncate error %d\n", rc);
2711 } else if (ext->oe_nr_pages == 0) {
2712 osc_extent_remove(ext);
2714 /* this must be an overlapped extent which means only
2715 * part of pages in this extent have been truncated.
2717 EASSERTF(ext->oe_start <= index, ext,
2718 "trunc index = %lu/%d.\n", index, partial);
2719 /* fix index to skip this partially truncated extent */
2720 index = ext->oe_end + 1;
2723 /* we need to hold this extent in OES_TRUNC state so
2724 * that no writeback will happen. This is to avoid
2726 LASSERT(oio->oi_trunc == NULL);
2727 oio->oi_trunc = osc_extent_get(ext);
2728 OSC_EXTENT_DUMP(D_CACHE, ext,
2729 "trunc at "LPU64"\n", size);
2731 osc_extent_put(env, ext);
2733 if (waiting != NULL) {
2736 /* ignore the result of osc_extent_wait the write initiator
2737 * should take care of it. */
2738 rc = osc_extent_wait(env, waiting, OES_INV);
2740 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2742 osc_extent_put(env, waiting);
2750 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2752 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2753 struct osc_object *obj)
2755 struct osc_extent *ext = oio->oi_trunc;
2757 oio->oi_trunc = NULL;
2759 bool unplug = false;
2761 EASSERT(ext->oe_nr_pages > 0, ext);
2762 EASSERT(ext->oe_state == OES_TRUNC, ext);
2763 EASSERT(!ext->oe_urgent, ext);
2765 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2766 osc_object_lock(obj);
2767 osc_extent_state_set(ext, OES_CACHE);
2768 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2770 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2773 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2774 osc_object_unlock(obj);
2775 osc_extent_put(env, ext);
2778 osc_io_unplug_async(env, osc_cli(obj), obj);
2783 * Wait for extents in a specific range to be written out.
2784 * The caller must have called osc_cache_writeback_range() to issue IO
2785 * otherwise it will take a long time for this function to finish.
2787 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2788 * nobody else can dirty this range of file while we're waiting for
2789 * extents to be written.
2791 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2792 pgoff_t start, pgoff_t end)
2794 struct osc_extent *ext;
2795 pgoff_t index = start;
2800 osc_object_lock(obj);
2801 ext = osc_extent_search(obj, index);
2803 ext = first_extent(obj);
2804 else if (ext->oe_end < index)
2805 ext = next_extent(ext);
2806 while (ext != NULL) {
2809 if (ext->oe_start > end)
2812 if (!ext->oe_fsync_wait) {
2813 ext = next_extent(ext);
2817 EASSERT(ergo(ext->oe_state == OES_CACHE,
2818 ext->oe_hp || ext->oe_urgent), ext);
2819 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2820 !ext->oe_hp && ext->oe_urgent), ext);
2822 index = ext->oe_end + 1;
2823 osc_extent_get(ext);
2824 osc_object_unlock(obj);
2826 rc = osc_extent_wait(env, ext, OES_INV);
2829 osc_extent_put(env, ext);
2832 osc_object_unlock(obj);
2834 OSC_IO_DEBUG(obj, "sync file range.\n");
2839 * Called to write out a range of osc object.
2841 * @hp : should be set this is caused by lock cancel;
2842 * @discard: is set if dirty pages should be dropped - file will be deleted or
2843 * truncated, this implies there is no partially discarding extents.
2845 * Return how many pages will be issued, or error code if error occurred.
2847 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2848 pgoff_t start, pgoff_t end, int hp, int discard)
2850 struct osc_extent *ext;
2851 struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2852 bool unplug = false;
2856 osc_object_lock(obj);
2857 ext = osc_extent_search(obj, start);
2859 ext = first_extent(obj);
2860 else if (ext->oe_end < start)
2861 ext = next_extent(ext);
2862 while (ext != NULL) {
2863 if (ext->oe_start > end)
2866 ext->oe_fsync_wait = 1;
2867 switch (ext->oe_state) {
2869 result += ext->oe_nr_pages;
2871 struct list_head *list = NULL;
2873 EASSERT(!ext->oe_hp, ext);
2875 list = &obj->oo_hp_exts;
2876 } else if (!ext->oe_urgent) {
2878 list = &obj->oo_urgent_exts;
2881 list_move_tail(&ext->oe_link, list);
2884 /* the only discarder is lock cancelling, so
2885 * [start, end] must contain this extent */
2886 EASSERT(ext->oe_start >= start &&
2887 ext->oe_max_end <= end, ext);
2888 osc_extent_state_set(ext, OES_LOCKING);
2889 ext->oe_owner = current;
2890 list_move_tail(&ext->oe_link,
2892 osc_update_pending(obj, OBD_BRW_WRITE,
2897 /* It's pretty bad to wait for ACTIVE extents, because
2898 * we don't know how long we will wait for it to be
2899 * flushed since it may be blocked at awaiting more
2900 * grants. We do this for the correctness of fsync. */
2901 LASSERT(hp == 0 && discard == 0);
2905 /* this extent is being truncated, can't do anything
2906 * for it now. it will be set to urgent after truncate
2907 * is finished in osc_cache_truncate_end(). */
2911 ext = next_extent(ext);
2913 osc_object_unlock(obj);
2915 LASSERT(ergo(!discard, list_empty(&discard_list)));
2916 if (!list_empty(&discard_list)) {
2917 struct osc_extent *tmp;
2920 osc_list_maint(osc_cli(obj), obj);
2921 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2922 list_del_init(&ext->oe_link);
2923 EASSERT(ext->oe_state == OES_LOCKING, ext);
2925 /* Discard caching pages. We don't actually write this
2926 * extent out but we complete it as if we did. */
2927 rc = osc_extent_make_ready(env, ext);
2928 if (unlikely(rc < 0)) {
2929 OSC_EXTENT_DUMP(D_ERROR, ext,
2930 "make_ready returned %d\n", rc);
2935 /* finish the extent as if the pages were sent */
2936 osc_extent_finish(env, ext, 0, 0);
2941 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2943 if (hp || discard) {
2945 rc = osc_cache_wait_range(env, obj, start, end);
2946 if (result >= 0 && rc < 0)
2950 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2955 * Returns a list of pages by a given [start, end] of \a obj.
2957 * \param resched If not NULL, then we give up before hogging CPU for too
2958 * long and set *resched = 1, in that case caller should implement a retry
2961 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
2962 * crucial in the face of [offset, EOF] locks.
2964 * Return at least one page in @queue unless there is no covered page.
2966 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
2967 struct osc_object *osc, pgoff_t start, pgoff_t end,
2968 osc_page_gang_cbt cb, void *cbdata)
2970 struct osc_page *ops;
2976 int res = CLP_GANG_OKAY;
2977 bool tree_lock = true;
2981 pvec = osc_env_info(env)->oti_pvec;
2982 spin_lock(&osc->oo_tree_lock);
2983 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
2984 idx, OTI_PVEC_SIZE)) > 0) {
2985 struct cl_page *page;
2986 bool end_of_region = false;
2988 for (i = 0, j = 0; i < nr; ++i) {
2992 idx = osc_index(ops);
2994 end_of_region = true;
2998 page = ops->ops_cl.cpl_page;
2999 LASSERT(page->cp_type == CPT_CACHEABLE);
3000 if (page->cp_state == CPS_FREEING)
3004 lu_ref_add_atomic(&page->cp_reference,
3005 "gang_lookup", current);
3011 * Here a delicate locking dance is performed. Current thread
3012 * holds a reference to a page, but has to own it before it
3013 * can be placed into queue. Owning implies waiting, so
3014 * radix-tree lock is to be released. After a wait one has to
3015 * check that pages weren't truncated (cl_page_own() returns
3016 * error in the latter case).
3018 spin_unlock(&osc->oo_tree_lock);
3021 for (i = 0; i < j; ++i) {
3023 if (res == CLP_GANG_OKAY)
3024 res = (*cb)(env, io, ops, cbdata);
3026 page = ops->ops_cl.cpl_page;
3027 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3028 cl_page_put(env, page);
3030 if (nr < OTI_PVEC_SIZE || end_of_region)
3033 if (res == CLP_GANG_OKAY && need_resched())
3034 res = CLP_GANG_RESCHED;
3035 if (res != CLP_GANG_OKAY)
3038 spin_lock(&osc->oo_tree_lock);
3042 spin_unlock(&osc->oo_tree_lock);
3047 * Check if page @page is covered by an extra lock or discard it.
3049 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3050 struct osc_page *ops, void *cbdata)
3052 struct osc_thread_info *info = osc_env_info(env);
3053 struct cl_lock *lock = cbdata;
3056 index = osc_index(ops);
3057 if (index >= info->oti_fn_index) {
3058 struct cl_lock *tmp;
3059 struct cl_page *page = ops->ops_cl.cpl_page;
3061 /* refresh non-overlapped index */
3062 tmp = cl_lock_at_pgoff(env, lock->cll_descr.cld_obj, index,
3065 /* Cache the first-non-overlapped index so as to skip
3066 * all pages within [index, oti_fn_index). This
3067 * is safe because if tmp lock is canceled, it will
3068 * discard these pages. */
3069 info->oti_fn_index = tmp->cll_descr.cld_end + 1;
3070 if (tmp->cll_descr.cld_end == CL_PAGE_EOF)
3071 info->oti_fn_index = CL_PAGE_EOF;
3072 cl_lock_put(env, tmp);
3073 } else if (cl_page_own(env, io, page) == 0) {
3074 /* discard the page */
3075 cl_page_discard(env, io, page);
3076 cl_page_disown(env, io, page);
3078 LASSERT(page->cp_state == CPS_FREEING);
3082 info->oti_next_index = index + 1;
3083 return CLP_GANG_OKAY;
3086 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3087 struct osc_page *ops, void *cbdata)
3089 struct osc_thread_info *info = osc_env_info(env);
3090 struct cl_lock *lock = cbdata;
3091 struct cl_page *page = ops->ops_cl.cpl_page;
3093 LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
3095 /* page is top page. */
3096 info->oti_next_index = osc_index(ops) + 1;
3097 if (cl_page_own(env, io, page) == 0) {
3098 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3099 !PageDirty(cl_page_vmpage(page))));
3101 /* discard the page */
3102 cl_page_discard(env, io, page);
3103 cl_page_disown(env, io, page);
3105 LASSERT(page->cp_state == CPS_FREEING);
3108 return CLP_GANG_OKAY;
3112 * Discard pages protected by the given lock. This function traverses radix
3113 * tree to find all covering pages and discard them. If a page is being covered
3114 * by other locks, it should remain in cache.
3116 * If error happens on any step, the process continues anyway (the reasoning
3117 * behind this being that lock cancellation cannot be delayed indefinitely).
3119 int osc_lock_discard_pages(const struct lu_env *env, struct osc_lock *ols)
3121 struct osc_thread_info *info = osc_env_info(env);
3122 struct cl_io *io = &info->oti_io;
3123 struct cl_object *osc = ols->ols_cl.cls_obj;
3124 struct cl_lock *lock = ols->ols_cl.cls_lock;
3125 struct cl_lock_descr *descr = &lock->cll_descr;
3126 osc_page_gang_cbt cb;
3132 io->ci_obj = cl_object_top(osc);
3133 io->ci_ignore_layout = 1;
3134 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3138 cb = descr->cld_mode == CLM_READ ? check_and_discard_cb : discard_cb;
3139 info->oti_fn_index = info->oti_next_index = descr->cld_start;
3141 res = osc_page_gang_lookup(env, io, cl2osc(osc),
3142 info->oti_next_index, descr->cld_end,
3144 if (info->oti_next_index > descr->cld_end)
3147 if (res == CLP_GANG_RESCHED)
3149 } while (res != CLP_GANG_OKAY);
3151 cl_io_fini(env, io);