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(cfs_list_t *list)
97 return cfs_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 || !cfs_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 cfs_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 CFS_INIT_LIST_HEAD(&ext->oe_link);
321 ext->oe_state = OES_INV;
322 CFS_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(cfs_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 cfs_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(cfs_list_empty(&ext->oe_pages));
469 osc_extent_erase(ext);
470 cfs_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 cfs_list_splice_init(&victim->oe_pages, &cur->oe_pages);
525 cfs_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 cfs_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 cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
818 cfs_list_del_init(&oap->oap_rpc_item);
819 cfs_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 cfs_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(cfs_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 cfs_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 cfs_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 cfs_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 cfs_list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1175 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1178 cfs_list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1179 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1182 cfs_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 !cfs_list_empty(&osc->oo_ready_item) ||
1192 !cfs_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(!cfs_list_empty(&opg->ops_inflight));
1288 cfs_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 "unstable_pages: %d/%d dropped: %ld avail: %ld, " \
1323 "reserved: %ld, flight: %d } lru {in list: %d, " \
1324 "left: %d, waiters: %d }" fmt, \
1325 __tmp->cl_import->imp_obd->obd_name, \
1326 __tmp->cl_dirty, __tmp->cl_dirty_max, \
1327 atomic_read(&obd_dirty_pages), obd_max_dirty_pages, \
1328 atomic_read(&obd_unstable_pages), obd_max_dirty_pages, \
1329 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1330 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1331 atomic_read(&__tmp->cl_lru_in_list), \
1332 atomic_read(&__tmp->cl_lru_busy), \
1333 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1336 /* caller must hold loi_list_lock */
1337 static void osc_consume_write_grant(struct client_obd *cli,
1338 struct brw_page *pga)
1340 LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock));
1341 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1342 atomic_inc(&obd_dirty_pages);
1343 cli->cl_dirty += PAGE_CACHE_SIZE;
1344 pga->flag |= OBD_BRW_FROM_GRANT;
1345 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1346 PAGE_CACHE_SIZE, pga, pga->pg);
1347 osc_update_next_shrink(cli);
1350 /* the companion to osc_consume_write_grant, called when a brw has completed.
1351 * must be called with the loi lock held. */
1352 static void osc_release_write_grant(struct client_obd *cli,
1353 struct brw_page *pga)
1357 LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock));
1358 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1363 pga->flag &= ~OBD_BRW_FROM_GRANT;
1364 atomic_dec(&obd_dirty_pages);
1365 cli->cl_dirty -= PAGE_CACHE_SIZE;
1366 if (pga->flag & OBD_BRW_NOCACHE) {
1367 pga->flag &= ~OBD_BRW_NOCACHE;
1368 atomic_dec(&obd_dirty_transit_pages);
1369 cli->cl_dirty_transit -= PAGE_CACHE_SIZE;
1375 * To avoid sleeping with object lock held, it's good for us allocate enough
1376 * grants before entering into critical section.
1378 * client_obd_list_lock held by caller
1380 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1384 if (cli->cl_avail_grant >= bytes) {
1385 cli->cl_avail_grant -= bytes;
1386 cli->cl_reserved_grant += bytes;
1392 static void __osc_unreserve_grant(struct client_obd *cli,
1393 unsigned int reserved, unsigned int unused)
1395 /* it's quite normal for us to get more grant than reserved.
1396 * Thinking about a case that two extents merged by adding a new
1397 * chunk, we can save one extent tax. If extent tax is greater than
1398 * one chunk, we can save more grant by adding a new chunk */
1399 cli->cl_reserved_grant -= reserved;
1400 if (unused > reserved) {
1401 cli->cl_avail_grant += reserved;
1402 cli->cl_lost_grant += unused - reserved;
1404 cli->cl_avail_grant += unused;
1408 void osc_unreserve_grant(struct client_obd *cli,
1409 unsigned int reserved, unsigned int unused)
1411 client_obd_list_lock(&cli->cl_loi_list_lock);
1412 __osc_unreserve_grant(cli, reserved, unused);
1414 osc_wake_cache_waiters(cli);
1415 client_obd_list_unlock(&cli->cl_loi_list_lock);
1419 * Free grant after IO is finished or canceled.
1421 * @lost_grant is used to remember how many grants we have allocated but not
1422 * used, we should return these grants to OST. There're two cases where grants
1425 * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1426 * written. In this case OST may use less chunks to serve this partial
1427 * write. OSTs don't actually know the page size on the client side. so
1428 * clients have to calculate lost grant by the blocksize on the OST.
1429 * See filter_grant_check() for details.
1431 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1432 unsigned int lost_grant)
1434 int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1436 client_obd_list_lock(&cli->cl_loi_list_lock);
1437 atomic_sub(nr_pages, &obd_dirty_pages);
1438 cli->cl_dirty -= nr_pages << PAGE_CACHE_SHIFT;
1439 cli->cl_lost_grant += lost_grant;
1440 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1441 /* borrow some grant from truncate to avoid the case that
1442 * truncate uses up all avail grant */
1443 cli->cl_lost_grant -= grant;
1444 cli->cl_avail_grant += grant;
1446 osc_wake_cache_waiters(cli);
1447 client_obd_list_unlock(&cli->cl_loi_list_lock);
1448 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1449 lost_grant, cli->cl_lost_grant,
1450 cli->cl_avail_grant, cli->cl_dirty);
1454 * The companion to osc_enter_cache(), called when @oap is no longer part of
1455 * the dirty accounting due to error.
1457 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1459 client_obd_list_lock(&cli->cl_loi_list_lock);
1460 osc_release_write_grant(cli, &oap->oap_brw_page);
1461 client_obd_list_unlock(&cli->cl_loi_list_lock);
1465 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1468 static int osc_enter_cache_try(struct client_obd *cli,
1469 struct osc_async_page *oap,
1470 int bytes, int transient)
1474 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1476 rc = osc_reserve_grant(cli, bytes);
1480 if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max &&
1481 atomic_read(&obd_unstable_pages) + 1 +
1482 atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1483 osc_consume_write_grant(cli, &oap->oap_brw_page);
1485 cli->cl_dirty_transit += PAGE_CACHE_SIZE;
1486 atomic_inc(&obd_dirty_transit_pages);
1487 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1491 __osc_unreserve_grant(cli, bytes, bytes);
1497 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1500 client_obd_list_lock(&cli->cl_loi_list_lock);
1501 rc = cfs_list_empty(&ocw->ocw_entry);
1502 client_obd_list_unlock(&cli->cl_loi_list_lock);
1507 * The main entry to reserve dirty page accounting. Usually the grant reserved
1508 * in this function will be freed in bulk in osc_free_grant() unless it fails
1509 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1511 * The process will be put into sleep if it's already run out of grant.
1513 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1514 struct osc_async_page *oap, int bytes)
1516 struct osc_object *osc = oap->oap_obj;
1517 struct lov_oinfo *loi = osc->oo_oinfo;
1518 struct osc_cache_waiter ocw;
1519 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1520 LWI_ON_SIGNAL_NOOP, NULL);
1524 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1526 client_obd_list_lock(&cli->cl_loi_list_lock);
1528 /* force the caller to try sync io. this can jump the list
1529 * of queued writes and create a discontiguous rpc stream */
1530 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1531 cli->cl_dirty_max < PAGE_CACHE_SIZE ||
1532 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1533 GOTO(out, rc = -EDQUOT);
1535 /* Hopefully normal case - cache space and write credits available */
1536 if (osc_enter_cache_try(cli, oap, bytes, 0))
1539 /* We can get here for two reasons: too many dirty pages in cache, or
1540 * run out of grants. In both cases we should write dirty pages out.
1541 * Adding a cache waiter will trigger urgent write-out no matter what
1543 * The exiting condition is no avail grants and no dirty pages caching,
1544 * that really means there is no space on the OST. */
1545 init_waitqueue_head(&ocw.ocw_waitq);
1547 ocw.ocw_grant = bytes;
1548 while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
1549 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1551 client_obd_list_unlock(&cli->cl_loi_list_lock);
1553 osc_io_unplug_async(env, cli, NULL);
1555 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1556 cli->cl_import->imp_obd->obd_name, &ocw, oap);
1558 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1560 client_obd_list_lock(&cli->cl_loi_list_lock);
1562 /* l_wait_event is interrupted by signal, or timed out */
1566 OSC_DUMP_GRANT(D_ERROR, cli,
1567 "try to reserve %d.\n", bytes);
1568 osc_extent_tree_dump(D_ERROR, osc);
1572 /* Ensures restartability - LU-3581 */
1576 CDEBUG(D_CACHE, "%s: event for cache space @"
1577 " %p never arrived due to %d\n",
1578 cli->cl_import->imp_obd->obd_name,
1582 cfs_list_del_init(&ocw.ocw_entry);
1586 LASSERT(cfs_list_empty(&ocw.ocw_entry));
1591 if (osc_enter_cache_try(cli, oap, bytes, 0))
1596 client_obd_list_unlock(&cli->cl_loi_list_lock);
1597 OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1601 /* caller must hold loi_list_lock */
1602 void osc_wake_cache_waiters(struct client_obd *cli)
1604 cfs_list_t *l, *tmp;
1605 struct osc_cache_waiter *ocw;
1608 cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1609 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
1610 cfs_list_del_init(&ocw->ocw_entry);
1612 ocw->ocw_rc = -EDQUOT;
1613 /* we can't dirty more */
1614 if ((cli->cl_dirty + PAGE_CACHE_SIZE > cli->cl_dirty_max) ||
1615 (atomic_read(&obd_unstable_pages) + 1 +
1616 atomic_read(&obd_dirty_pages) > obd_max_dirty_pages)) {
1617 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1618 "osc max %ld, sys max %d\n", cli->cl_dirty,
1619 cli->cl_dirty_max, obd_max_dirty_pages);
1624 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1625 ocw->ocw_rc = -EDQUOT;
1628 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1629 ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1631 wake_up(&ocw->ocw_waitq);
1637 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1639 int hprpc = !!cfs_list_empty(&osc->oo_hp_exts);
1640 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1643 /* This maintains the lists of pending pages to read/write for a given object
1644 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1645 * to quickly find objects that are ready to send an RPC. */
1646 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1649 int invalid_import = 0;
1652 /* if we have an invalid import we want to drain the queued pages
1653 * by forcing them through rpcs that immediately fail and complete
1654 * the pages. recovery relies on this to empty the queued pages
1655 * before canceling the locks and evicting down the llite pages */
1656 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1659 if (cmd & OBD_BRW_WRITE) {
1660 if (atomic_read(&osc->oo_nr_writes) == 0)
1662 if (invalid_import) {
1663 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1666 if (!cfs_list_empty(&osc->oo_hp_exts)) {
1667 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1670 if (!cfs_list_empty(&osc->oo_urgent_exts)) {
1671 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1674 /* trigger a write rpc stream as long as there are dirtiers
1675 * waiting for space. as they're waiting, they're not going to
1676 * create more pages to coalesce with what's waiting.. */
1677 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1678 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1681 if (atomic_read(&osc->oo_nr_writes) >=
1682 cli->cl_max_pages_per_rpc)
1685 if (atomic_read(&osc->oo_nr_reads) == 0)
1687 if (invalid_import) {
1688 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1691 /* all read are urgent. */
1692 if (!cfs_list_empty(&osc->oo_reading_exts))
1699 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1701 struct client_obd *cli = osc_cli(obj);
1702 if (cmd & OBD_BRW_WRITE) {
1703 atomic_add(delta, &obj->oo_nr_writes);
1704 atomic_add(delta, &cli->cl_pending_w_pages);
1705 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1707 atomic_add(delta, &obj->oo_nr_reads);
1708 atomic_add(delta, &cli->cl_pending_r_pages);
1709 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1711 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1714 static int osc_makes_hprpc(struct osc_object *obj)
1716 return !cfs_list_empty(&obj->oo_hp_exts);
1719 static void on_list(cfs_list_t *item, cfs_list_t *list, int should_be_on)
1721 if (cfs_list_empty(item) && should_be_on)
1722 cfs_list_add_tail(item, list);
1723 else if (!cfs_list_empty(item) && !should_be_on)
1724 cfs_list_del_init(item);
1727 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1728 * can find pages to build into rpcs quickly */
1729 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1731 if (osc_makes_hprpc(osc)) {
1733 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1734 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1736 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1737 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1738 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1739 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1742 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1743 atomic_read(&osc->oo_nr_writes) > 0);
1745 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1746 atomic_read(&osc->oo_nr_reads) > 0);
1748 return osc_is_ready(osc);
1751 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1755 client_obd_list_lock(&cli->cl_loi_list_lock);
1756 is_ready = __osc_list_maint(cli, osc);
1757 client_obd_list_unlock(&cli->cl_loi_list_lock);
1762 /* this is trying to propogate async writeback errors back up to the
1763 * application. As an async write fails we record the error code for later if
1764 * the app does an fsync. As long as errors persist we force future rpcs to be
1765 * sync so that the app can get a sync error and break the cycle of queueing
1766 * pages for which writeback will fail. */
1767 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1774 ar->ar_force_sync = 1;
1775 ar->ar_min_xid = ptlrpc_sample_next_xid();
1780 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1781 ar->ar_force_sync = 0;
1784 /* Performs "unstable" page accounting. This function balances the
1785 * increment operations performed in osc_inc_unstable_pages. It is
1786 * registered as the RPC request callback, and is executed when the
1787 * bulk RPC is committed on the server. Thus at this point, the pages
1788 * involved in the bulk transfer are no longer considered unstable. */
1789 void osc_dec_unstable_pages(struct ptlrpc_request *req)
1791 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1792 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1793 obd_count page_count = desc->bd_iov_count;
1796 /* No unstable page tracking */
1797 if (cli->cl_cache == NULL)
1800 LASSERT(page_count >= 0);
1802 for (i = 0; i < page_count; i++)
1803 dec_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1805 atomic_sub(page_count, &cli->cl_cache->ccc_unstable_nr);
1806 LASSERT(atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1808 atomic_sub(page_count, &cli->cl_unstable_count);
1809 LASSERT(atomic_read(&cli->cl_unstable_count) >= 0);
1811 atomic_sub(page_count, &obd_unstable_pages);
1812 LASSERT(atomic_read(&obd_unstable_pages) >= 0);
1814 spin_lock(&req->rq_lock);
1815 req->rq_committed = 1;
1816 req->rq_unstable = 0;
1817 spin_unlock(&req->rq_lock);
1819 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
1822 /* "unstable" page accounting. See: osc_dec_unstable_pages. */
1823 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1825 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1826 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
1827 obd_count page_count = desc->bd_iov_count;
1830 /* No unstable page tracking */
1831 if (cli->cl_cache == NULL)
1834 LASSERT(page_count >= 0);
1836 for (i = 0; i < page_count; i++)
1837 inc_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1839 LASSERT(atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1840 atomic_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1842 LASSERT(atomic_read(&cli->cl_unstable_count) >= 0);
1843 atomic_add(page_count, &cli->cl_unstable_count);
1845 LASSERT(atomic_read(&obd_unstable_pages) >= 0);
1846 atomic_add(page_count, &obd_unstable_pages);
1848 spin_lock(&req->rq_lock);
1850 /* If the request has already been committed (i.e. brw_commit
1851 * called via rq_commit_cb), we need to undo the unstable page
1852 * increments we just performed because rq_commit_cb wont be
1853 * called again. Otherwise, just set the commit callback so the
1854 * unstable page accounting is properly updated when the request
1856 if (req->rq_committed) {
1857 /* Drop lock before calling osc_dec_unstable_pages */
1858 spin_unlock(&req->rq_lock);
1859 osc_dec_unstable_pages(req);
1860 spin_lock(&req->rq_lock);
1862 req->rq_unstable = 1;
1863 req->rq_commit_cb = osc_dec_unstable_pages;
1866 spin_unlock(&req->rq_lock);
1869 /* this must be called holding the loi list lock to give coverage to exit_cache,
1870 * async_flag maintenance, and oap_request */
1871 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1872 struct osc_async_page *oap, int sent, int rc)
1874 struct osc_object *osc = oap->oap_obj;
1875 struct lov_oinfo *loi = osc->oo_oinfo;
1879 if (oap->oap_request != NULL) {
1881 osc_inc_unstable_pages(oap->oap_request);
1883 xid = ptlrpc_req_xid(oap->oap_request);
1884 ptlrpc_req_finished(oap->oap_request);
1885 oap->oap_request = NULL;
1888 /* As the transfer for this page is being done, clear the flags */
1889 spin_lock(&oap->oap_lock);
1890 oap->oap_async_flags = 0;
1891 spin_unlock(&oap->oap_lock);
1892 oap->oap_interrupted = 0;
1894 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1895 client_obd_list_lock(&cli->cl_loi_list_lock);
1896 osc_process_ar(&cli->cl_ar, xid, rc);
1897 osc_process_ar(&loi->loi_ar, xid, rc);
1898 client_obd_list_unlock(&cli->cl_loi_list_lock);
1901 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1903 CERROR("completion on oap %p obj %p returns %d.\n",
1910 * Try to add extent to one RPC. We need to think about the following things:
1911 * - # of pages must not be over max_pages_per_rpc
1912 * - extent must be compatible with previous ones
1914 static int try_to_add_extent_for_io(struct client_obd *cli,
1915 struct osc_extent *ext, cfs_list_t *rpclist,
1916 int *pc, unsigned int *max_pages)
1918 struct osc_extent *tmp;
1921 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1924 *max_pages = max(ext->oe_mppr, *max_pages);
1925 if (*pc + ext->oe_nr_pages > *max_pages)
1928 cfs_list_for_each_entry(tmp, rpclist, oe_link) {
1929 EASSERT(tmp->oe_owner == current, tmp);
1931 if (overlapped(tmp, ext)) {
1932 OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1937 if (tmp->oe_srvlock != ext->oe_srvlock ||
1938 !tmp->oe_grants != !ext->oe_grants)
1941 /* remove break for strict check */
1945 *pc += ext->oe_nr_pages;
1946 cfs_list_move_tail(&ext->oe_link, rpclist);
1947 ext->oe_owner = current;
1952 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1953 * get_write_extent() takes all appropriate extents in atomic.
1955 * The following policy is used to collect extents for IO:
1956 * 1. Add as many HP extents as possible;
1957 * 2. Add the first urgent extent in urgent extent list and take it out of
1959 * 3. Add subsequent extents of this urgent extent;
1960 * 4. If urgent list is not empty, goto 2;
1961 * 5. Traverse the extent tree from the 1st extent;
1962 * 6. Above steps exit if there is no space in this RPC.
1964 static int get_write_extents(struct osc_object *obj, cfs_list_t *rpclist)
1966 struct client_obd *cli = osc_cli(obj);
1967 struct osc_extent *ext;
1969 unsigned int max_pages = cli->cl_max_pages_per_rpc;
1971 LASSERT(osc_object_is_locked(obj));
1972 while (!cfs_list_empty(&obj->oo_hp_exts)) {
1973 ext = cfs_list_entry(obj->oo_hp_exts.next, struct osc_extent,
1975 LASSERT(ext->oe_state == OES_CACHE);
1976 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1979 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1981 if (page_count == max_pages)
1984 while (!cfs_list_empty(&obj->oo_urgent_exts)) {
1985 ext = cfs_list_entry(obj->oo_urgent_exts.next,
1986 struct osc_extent, oe_link);
1987 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1991 if (!ext->oe_intree)
1994 while ((ext = next_extent(ext)) != NULL) {
1995 if ((ext->oe_state != OES_CACHE) ||
1996 (!cfs_list_empty(&ext->oe_link) &&
1997 ext->oe_owner != NULL))
2000 if (!try_to_add_extent_for_io(cli, ext, rpclist,
2001 &page_count, &max_pages))
2005 if (page_count == max_pages)
2008 ext = first_extent(obj);
2009 while (ext != NULL) {
2010 if ((ext->oe_state != OES_CACHE) ||
2011 /* this extent may be already in current rpclist */
2012 (!cfs_list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2013 ext = next_extent(ext);
2017 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
2021 ext = next_extent(ext);
2027 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2028 struct osc_object *osc, pdl_policy_t pol)
2030 CFS_LIST_HEAD(rpclist);
2031 struct osc_extent *ext;
2032 struct osc_extent *tmp;
2033 struct osc_extent *first = NULL;
2034 obd_count page_count = 0;
2039 LASSERT(osc_object_is_locked(osc));
2041 page_count = get_write_extents(osc, &rpclist);
2042 LASSERT(equi(page_count == 0, cfs_list_empty(&rpclist)));
2044 if (cfs_list_empty(&rpclist))
2047 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2049 cfs_list_for_each_entry(ext, &rpclist, oe_link) {
2050 LASSERT(ext->oe_state == OES_CACHE ||
2051 ext->oe_state == OES_LOCK_DONE);
2052 if (ext->oe_state == OES_CACHE)
2053 osc_extent_state_set(ext, OES_LOCKING);
2055 osc_extent_state_set(ext, OES_RPC);
2058 /* we're going to grab page lock, so release object lock because
2059 * lock order is page lock -> object lock. */
2060 osc_object_unlock(osc);
2062 cfs_list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2063 if (ext->oe_state == OES_LOCKING) {
2064 rc = osc_extent_make_ready(env, ext);
2065 if (unlikely(rc < 0)) {
2066 cfs_list_del_init(&ext->oe_link);
2067 osc_extent_finish(env, ext, 0, rc);
2071 if (first == NULL) {
2073 srvlock = ext->oe_srvlock;
2075 LASSERT(srvlock == ext->oe_srvlock);
2079 if (!cfs_list_empty(&rpclist)) {
2080 LASSERT(page_count > 0);
2081 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2082 LASSERT(cfs_list_empty(&rpclist));
2085 osc_object_lock(osc);
2090 * prepare pages for ASYNC io and put pages in send queue.
2092 * \param cmd OBD_BRW_* macroses
2093 * \param lop pending pages
2095 * \return zero if no page added to send queue.
2096 * \return 1 if pages successfully added to send queue.
2097 * \return negative on errors.
2100 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2101 struct osc_object *osc, pdl_policy_t pol)
2103 struct osc_extent *ext;
2104 struct osc_extent *next;
2105 CFS_LIST_HEAD(rpclist);
2107 unsigned int max_pages = cli->cl_max_pages_per_rpc;
2111 LASSERT(osc_object_is_locked(osc));
2112 cfs_list_for_each_entry_safe(ext, next,
2113 &osc->oo_reading_exts, oe_link) {
2114 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2115 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2118 osc_extent_state_set(ext, OES_RPC);
2119 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2121 LASSERT(page_count <= max_pages);
2123 osc_update_pending(osc, OBD_BRW_READ, -page_count);
2125 if (!cfs_list_empty(&rpclist)) {
2126 osc_object_unlock(osc);
2128 LASSERT(page_count > 0);
2129 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2130 LASSERT(cfs_list_empty(&rpclist));
2132 osc_object_lock(osc);
2137 #define list_to_obj(list, item) ({ \
2138 cfs_list_t *__tmp = (list)->next; \
2139 cfs_list_del_init(__tmp); \
2140 cfs_list_entry(__tmp, struct osc_object, oo_##item); \
2143 /* This is called by osc_check_rpcs() to find which objects have pages that
2144 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2145 static struct osc_object *osc_next_obj(struct client_obd *cli)
2149 /* First return objects that have blocked locks so that they
2150 * will be flushed quickly and other clients can get the lock,
2151 * then objects which have pages ready to be stuffed into RPCs */
2152 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2153 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2154 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2155 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2157 /* then if we have cache waiters, return all objects with queued
2158 * writes. This is especially important when many small files
2159 * have filled up the cache and not been fired into rpcs because
2160 * they don't pass the nr_pending/object threshhold */
2161 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2162 !cfs_list_empty(&cli->cl_loi_write_list))
2163 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2165 /* then return all queued objects when we have an invalid import
2166 * so that they get flushed */
2167 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2168 if (!cfs_list_empty(&cli->cl_loi_write_list))
2169 RETURN(list_to_obj(&cli->cl_loi_write_list,
2171 if (!cfs_list_empty(&cli->cl_loi_read_list))
2172 RETURN(list_to_obj(&cli->cl_loi_read_list,
2178 /* called with the loi list lock held */
2179 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2182 struct osc_object *osc;
2186 while ((osc = osc_next_obj(cli)) != NULL) {
2187 struct cl_object *obj = osc2cl(osc);
2188 struct lu_ref_link link;
2190 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2192 if (osc_max_rpc_in_flight(cli, osc)) {
2193 __osc_list_maint(cli, osc);
2198 client_obd_list_unlock(&cli->cl_loi_list_lock);
2199 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2201 /* attempt some read/write balancing by alternating between
2202 * reads and writes in an object. The makes_rpc checks here
2203 * would be redundant if we were getting read/write work items
2204 * instead of objects. we don't want send_oap_rpc to drain a
2205 * partial read pending queue when we're given this object to
2206 * do io on writes while there are cache waiters */
2207 osc_object_lock(osc);
2208 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2209 rc = osc_send_write_rpc(env, cli, osc, pol);
2211 CERROR("Write request failed with %d\n", rc);
2213 /* osc_send_write_rpc failed, mostly because of
2216 * It can't break here, because if:
2217 * - a page was submitted by osc_io_submit, so
2219 * - no request in flight
2220 * - no subsequent request
2221 * The system will be in live-lock state,
2222 * because there is no chance to call
2223 * osc_io_unplug() and osc_check_rpcs() any
2224 * more. pdflush can't help in this case,
2225 * because it might be blocked at grabbing
2226 * the page lock as we mentioned.
2228 * Anyway, continue to drain pages. */
2232 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2233 rc = osc_send_read_rpc(env, cli, osc, pol);
2235 CERROR("Read request failed with %d\n", rc);
2237 osc_object_unlock(osc);
2239 osc_list_maint(cli, osc);
2240 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2241 cl_object_put(env, obj);
2243 client_obd_list_lock(&cli->cl_loi_list_lock);
2247 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2248 struct osc_object *osc, pdl_policy_t pol, int async)
2252 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2256 /* disable osc_lru_shrink() temporarily to avoid
2257 * potential stack overrun problem. LU-2859 */
2258 atomic_inc(&cli->cl_lru_shrinkers);
2259 client_obd_list_lock(&cli->cl_loi_list_lock);
2260 osc_check_rpcs(env, cli, pol);
2261 client_obd_list_unlock(&cli->cl_loi_list_lock);
2262 atomic_dec(&cli->cl_lru_shrinkers);
2264 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2265 LASSERT(cli->cl_writeback_work != NULL);
2266 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2271 static int osc_io_unplug_async(const struct lu_env *env,
2272 struct client_obd *cli, struct osc_object *osc)
2274 /* XXX: policy is no use actually. */
2275 return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2278 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2279 struct osc_object *osc, pdl_policy_t pol)
2281 (void)osc_io_unplug0(env, cli, osc, pol, 0);
2284 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2285 struct page *page, loff_t offset)
2287 struct obd_export *exp = osc_export(osc);
2288 struct osc_async_page *oap = &ops->ops_oap;
2292 return cfs_size_round(sizeof(*oap));
2294 oap->oap_magic = OAP_MAGIC;
2295 oap->oap_cli = &exp->exp_obd->u.cli;
2298 oap->oap_page = page;
2299 oap->oap_obj_off = offset;
2300 LASSERT(!(offset & ~CFS_PAGE_MASK));
2302 if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2303 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2305 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2306 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2308 spin_lock_init(&oap->oap_lock);
2309 CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2310 oap, page, oap->oap_obj_off);
2314 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2315 struct osc_page *ops)
2317 struct osc_io *oio = osc_env_io(env);
2318 struct osc_extent *ext = NULL;
2319 struct osc_async_page *oap = &ops->ops_oap;
2320 struct client_obd *cli = oap->oap_cli;
2321 struct osc_object *osc = oap->oap_obj;
2324 int brw_flags = OBD_BRW_ASYNC;
2325 int cmd = OBD_BRW_WRITE;
2326 int need_release = 0;
2330 if (oap->oap_magic != OAP_MAGIC)
2333 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2336 if (!cfs_list_empty(&oap->oap_pending_item) ||
2337 !cfs_list_empty(&oap->oap_rpc_item))
2340 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2341 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2342 if (!client_is_remote(osc_export(osc)) &&
2343 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2344 brw_flags |= OBD_BRW_NOQUOTA;
2345 cmd |= OBD_BRW_NOQUOTA;
2348 /* check if the file's owner/group is over quota */
2349 if (!(cmd & OBD_BRW_NOQUOTA)) {
2350 struct cl_object *obj;
2351 struct cl_attr *attr;
2352 unsigned int qid[MAXQUOTAS];
2354 obj = cl_object_top(&osc->oo_cl);
2355 attr = &osc_env_info(env)->oti_attr;
2357 cl_object_attr_lock(obj);
2358 rc = cl_object_attr_get(env, obj, attr);
2359 cl_object_attr_unlock(obj);
2361 qid[USRQUOTA] = attr->cat_uid;
2362 qid[GRPQUOTA] = attr->cat_gid;
2363 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2369 if (osc_over_unstable_soft_limit(cli))
2370 brw_flags |= OBD_BRW_SOFT_SYNC;
2373 oap->oap_page_off = ops->ops_from;
2374 oap->oap_count = ops->ops_to - ops->ops_from;
2375 /* No need to hold a lock here,
2376 * since this page is not in any list yet. */
2377 oap->oap_async_flags = 0;
2378 oap->oap_brw_flags = brw_flags;
2380 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2381 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2383 index = osc_index(oap2osc(oap));
2385 /* Add this page into extent by the following steps:
2386 * 1. if there exists an active extent for this IO, mostly this page
2387 * can be added to the active extent and sometimes we need to
2388 * expand extent to accomodate this page;
2389 * 2. otherwise, a new extent will be allocated. */
2391 ext = oio->oi_active;
2392 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2393 /* one chunk plus extent overhead must be enough to write this
2395 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2396 if (ext->oe_end >= index)
2399 /* it doesn't need any grant to dirty this page */
2400 client_obd_list_lock(&cli->cl_loi_list_lock);
2401 rc = osc_enter_cache_try(cli, oap, grants, 0);
2402 client_obd_list_unlock(&cli->cl_loi_list_lock);
2403 if (rc == 0) { /* try failed */
2406 } else if (ext->oe_end < index) {
2408 /* try to expand this extent */
2409 rc = osc_extent_expand(ext, index, &tmp);
2412 /* don't free reserved grant */
2414 OSC_EXTENT_DUMP(D_CACHE, ext,
2415 "expanded for %lu.\n", index);
2416 osc_unreserve_grant(cli, grants, tmp);
2421 } else if (ext != NULL) {
2422 /* index is located outside of active extent */
2426 osc_extent_release(env, ext);
2427 oio->oi_active = NULL;
2432 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2434 /* try to find new extent to cover this page */
2435 LASSERT(oio->oi_active == NULL);
2436 /* we may have allocated grant for this page if we failed
2437 * to expand the previous active extent. */
2438 LASSERT(ergo(grants > 0, grants >= tmp));
2442 /* we haven't allocated grant for this page. */
2443 rc = osc_enter_cache(env, cli, oap, tmp);
2450 ext = osc_extent_find(env, osc, index, &tmp);
2452 LASSERT(tmp == grants);
2453 osc_exit_cache(cli, oap);
2457 oio->oi_active = ext;
2461 osc_unreserve_grant(cli, grants, tmp);
2464 LASSERT(ergo(rc == 0, ext != NULL));
2466 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2467 ext, "index = %lu.\n", index);
2468 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2470 osc_object_lock(osc);
2471 if (ext->oe_nr_pages == 0)
2472 ext->oe_srvlock = ops->ops_srvlock;
2474 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2476 cfs_list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2477 osc_object_unlock(osc);
2482 int osc_teardown_async_page(const struct lu_env *env,
2483 struct osc_object *obj, struct osc_page *ops)
2485 struct osc_async_page *oap = &ops->ops_oap;
2486 struct osc_extent *ext = NULL;
2490 LASSERT(oap->oap_magic == OAP_MAGIC);
2492 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2493 oap, ops, osc_index(oap2osc(oap)));
2495 osc_object_lock(obj);
2496 if (!cfs_list_empty(&oap->oap_rpc_item)) {
2497 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2499 } else if (!cfs_list_empty(&oap->oap_pending_item)) {
2500 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2501 /* only truncated pages are allowed to be taken out.
2502 * See osc_extent_truncate() and osc_cache_truncate_start()
2504 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2505 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2506 osc_index(oap2osc(oap)));
2510 osc_object_unlock(obj);
2512 osc_extent_put(env, ext);
2517 * This is called when a page is picked up by kernel to write out.
2519 * We should find out the corresponding extent and add the whole extent
2520 * into urgent list. The extent may be being truncated or used, handle it
2523 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2524 struct osc_page *ops)
2526 struct osc_extent *ext = NULL;
2527 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2528 struct cl_page *cp = ops->ops_cl.cpl_page;
2529 pgoff_t index = osc_index(ops);
2530 struct osc_async_page *oap = &ops->ops_oap;
2531 bool unplug = false;
2535 osc_object_lock(obj);
2536 ext = osc_extent_lookup(obj, index);
2538 osc_extent_tree_dump(D_ERROR, obj);
2539 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2542 switch (ext->oe_state) {
2545 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2549 /* If we know this extent is being written out, we should abort
2550 * so that the writer can make this page ready. Otherwise, there
2551 * exists a deadlock problem because other process can wait for
2552 * page writeback bit holding page lock; and meanwhile in
2553 * vvp_page_make_ready(), we need to grab page lock before
2554 * really sending the RPC. */
2556 /* race with truncate, page will be redirtied */
2558 /* The extent is active so we need to abort and let the caller
2559 * re-dirty the page. If we continued on here, and we were the
2560 * one making the extent active, we could deadlock waiting for
2561 * the page writeback to clear but it won't because the extent
2562 * is active and won't be written out. */
2563 GOTO(out, rc = -EAGAIN);
2568 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2572 spin_lock(&oap->oap_lock);
2573 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2574 spin_unlock(&oap->oap_lock);
2576 if (memory_pressure_get())
2577 ext->oe_memalloc = 1;
2580 if (ext->oe_state == OES_CACHE) {
2581 OSC_EXTENT_DUMP(D_CACHE, ext,
2582 "flush page %p make it urgent.\n", oap);
2583 if (cfs_list_empty(&ext->oe_link))
2584 cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2591 osc_object_unlock(obj);
2592 osc_extent_put(env, ext);
2594 osc_io_unplug_async(env, osc_cli(obj), obj);
2599 * this is called when a sync waiter receives an interruption. Its job is to
2600 * get the caller woken as soon as possible. If its page hasn't been put in an
2601 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2602 * desiring interruption which will forcefully complete the rpc once the rpc
2605 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2607 struct osc_async_page *oap = &ops->ops_oap;
2608 struct osc_object *obj = oap->oap_obj;
2609 struct client_obd *cli = osc_cli(obj);
2610 struct osc_extent *ext;
2611 struct osc_extent *found = NULL;
2613 pgoff_t index = osc_index(ops);
2618 LASSERT(!oap->oap_interrupted);
2619 oap->oap_interrupted = 1;
2621 /* Find out the caching extent */
2622 osc_object_lock(obj);
2623 if (oap->oap_cmd & OBD_BRW_WRITE) {
2624 plist = &obj->oo_urgent_exts;
2625 cmd = OBD_BRW_WRITE;
2627 plist = &obj->oo_reading_exts;
2630 cfs_list_for_each_entry(ext, plist, oe_link) {
2631 if (ext->oe_start <= index && ext->oe_end >= index) {
2632 LASSERT(ext->oe_state == OES_LOCK_DONE);
2633 /* For OES_LOCK_DONE state extent, it has already held
2634 * a refcount for RPC. */
2635 found = osc_extent_get(ext);
2639 if (found != NULL) {
2640 cfs_list_del_init(&found->oe_link);
2641 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2642 osc_object_unlock(obj);
2644 osc_extent_finish(env, found, 0, -EINTR);
2645 osc_extent_put(env, found);
2648 osc_object_unlock(obj);
2649 /* ok, it's been put in an rpc. only one oap gets a request
2651 if (oap->oap_request != NULL) {
2652 ptlrpc_mark_interrupted(oap->oap_request);
2653 ptlrpcd_wake(oap->oap_request);
2654 ptlrpc_req_finished(oap->oap_request);
2655 oap->oap_request = NULL;
2659 osc_list_maint(cli, obj);
2663 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2664 cfs_list_t *list, int cmd, int brw_flags)
2666 struct client_obd *cli = osc_cli(obj);
2667 struct osc_extent *ext;
2668 struct osc_async_page *oap;
2670 int mppr = cli->cl_max_pages_per_rpc;
2671 pgoff_t start = CL_PAGE_EOF;
2675 cfs_list_for_each_entry(oap, list, oap_pending_item) {
2676 pgoff_t index = osc_index(oap2osc(oap));
2682 mppr <<= (page_count > mppr);
2685 ext = osc_extent_alloc(obj);
2687 cfs_list_for_each_entry(oap, list, oap_pending_item) {
2688 cfs_list_del_init(&oap->oap_pending_item);
2689 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2694 ext->oe_rw = !!(cmd & OBD_BRW_READ);
2696 ext->oe_start = start;
2697 ext->oe_end = ext->oe_max_end = end;
2699 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2700 ext->oe_nr_pages = page_count;
2701 ext->oe_mppr = mppr;
2702 cfs_list_splice_init(list, &ext->oe_pages);
2704 osc_object_lock(obj);
2705 /* Reuse the initial refcount for RPC, don't drop it */
2706 osc_extent_state_set(ext, OES_LOCK_DONE);
2707 if (cmd & OBD_BRW_WRITE) {
2708 cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2709 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2711 cfs_list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2712 osc_update_pending(obj, OBD_BRW_READ, page_count);
2714 osc_object_unlock(obj);
2716 osc_io_unplug(env, cli, obj, PDL_POLICY_ROUND);
2721 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2723 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2724 struct osc_object *obj, __u64 size)
2726 struct client_obd *cli = osc_cli(obj);
2727 struct osc_extent *ext;
2728 struct osc_extent *waiting = NULL;
2730 CFS_LIST_HEAD(list);
2735 /* pages with index greater or equal to index will be truncated. */
2736 index = cl_index(osc2cl(obj), size);
2737 partial = size > cl_offset(osc2cl(obj), index);
2740 osc_object_lock(obj);
2741 ext = osc_extent_search(obj, index);
2743 ext = first_extent(obj);
2744 else if (ext->oe_end < index)
2745 ext = next_extent(ext);
2746 while (ext != NULL) {
2747 EASSERT(ext->oe_state != OES_TRUNC, ext);
2749 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2750 /* if ext is in urgent state, it means there must exist
2751 * a page already having been flushed by write_page().
2752 * We have to wait for this extent because we can't
2753 * truncate that page. */
2754 LASSERT(!ext->oe_hp);
2755 OSC_EXTENT_DUMP(D_CACHE, ext,
2756 "waiting for busy extent\n");
2757 waiting = osc_extent_get(ext);
2761 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2763 osc_extent_get(ext);
2764 if (ext->oe_state == OES_ACTIVE) {
2765 /* though we grab inode mutex for write path, but we
2766 * release it before releasing extent(in osc_io_end()),
2767 * so there is a race window that an extent is still
2768 * in OES_ACTIVE when truncate starts. */
2769 LASSERT(!ext->oe_trunc_pending);
2770 ext->oe_trunc_pending = 1;
2772 EASSERT(ext->oe_state == OES_CACHE, ext);
2773 osc_extent_state_set(ext, OES_TRUNC);
2774 osc_update_pending(obj, OBD_BRW_WRITE,
2777 EASSERT(cfs_list_empty(&ext->oe_link), ext);
2778 cfs_list_add_tail(&ext->oe_link, &list);
2780 ext = next_extent(ext);
2782 osc_object_unlock(obj);
2784 osc_list_maint(cli, obj);
2786 while (!cfs_list_empty(&list)) {
2789 ext = cfs_list_entry(list.next, struct osc_extent, oe_link);
2790 cfs_list_del_init(&ext->oe_link);
2792 /* extent may be in OES_ACTIVE state because inode mutex
2793 * is released before osc_io_end() in file write case */
2794 if (ext->oe_state != OES_TRUNC)
2795 osc_extent_wait(env, ext, OES_TRUNC);
2797 rc = osc_extent_truncate(ext, index, partial);
2802 OSC_EXTENT_DUMP(D_ERROR, ext,
2803 "truncate error %d\n", rc);
2804 } else if (ext->oe_nr_pages == 0) {
2805 osc_extent_remove(ext);
2807 /* this must be an overlapped extent which means only
2808 * part of pages in this extent have been truncated.
2810 EASSERTF(ext->oe_start <= index, ext,
2811 "trunc index = %lu/%d.\n", index, partial);
2812 /* fix index to skip this partially truncated extent */
2813 index = ext->oe_end + 1;
2816 /* we need to hold this extent in OES_TRUNC state so
2817 * that no writeback will happen. This is to avoid
2819 LASSERT(oio->oi_trunc == NULL);
2820 oio->oi_trunc = osc_extent_get(ext);
2821 OSC_EXTENT_DUMP(D_CACHE, ext,
2822 "trunc at "LPU64"\n", size);
2824 osc_extent_put(env, ext);
2826 if (waiting != NULL) {
2829 /* ignore the result of osc_extent_wait the write initiator
2830 * should take care of it. */
2831 rc = osc_extent_wait(env, waiting, OES_INV);
2833 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2835 osc_extent_put(env, waiting);
2843 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2845 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2846 struct osc_object *obj)
2848 struct osc_extent *ext = oio->oi_trunc;
2850 oio->oi_trunc = NULL;
2852 bool unplug = false;
2854 EASSERT(ext->oe_nr_pages > 0, ext);
2855 EASSERT(ext->oe_state == OES_TRUNC, ext);
2856 EASSERT(!ext->oe_urgent, ext);
2858 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2859 osc_object_lock(obj);
2860 osc_extent_state_set(ext, OES_CACHE);
2861 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2863 cfs_list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2866 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2867 osc_object_unlock(obj);
2868 osc_extent_put(env, ext);
2871 osc_io_unplug_async(env, osc_cli(obj), obj);
2876 * Wait for extents in a specific range to be written out.
2877 * The caller must have called osc_cache_writeback_range() to issue IO
2878 * otherwise it will take a long time for this function to finish.
2880 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2881 * nobody else can dirty this range of file while we're waiting for
2882 * extents to be written.
2884 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2885 pgoff_t start, pgoff_t end)
2887 struct osc_extent *ext;
2888 pgoff_t index = start;
2893 osc_object_lock(obj);
2894 ext = osc_extent_search(obj, index);
2896 ext = first_extent(obj);
2897 else if (ext->oe_end < index)
2898 ext = next_extent(ext);
2899 while (ext != NULL) {
2902 if (ext->oe_start > end)
2905 if (!ext->oe_fsync_wait) {
2906 ext = next_extent(ext);
2910 EASSERT(ergo(ext->oe_state == OES_CACHE,
2911 ext->oe_hp || ext->oe_urgent), ext);
2912 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2913 !ext->oe_hp && ext->oe_urgent), ext);
2915 index = ext->oe_end + 1;
2916 osc_extent_get(ext);
2917 osc_object_unlock(obj);
2919 rc = osc_extent_wait(env, ext, OES_INV);
2922 osc_extent_put(env, ext);
2925 osc_object_unlock(obj);
2927 OSC_IO_DEBUG(obj, "sync file range.\n");
2932 * Called to write out a range of osc object.
2934 * @hp : should be set this is caused by lock cancel;
2935 * @discard: is set if dirty pages should be dropped - file will be deleted or
2936 * truncated, this implies there is no partially discarding extents.
2938 * Return how many pages will be issued, or error code if error occurred.
2940 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2941 pgoff_t start, pgoff_t end, int hp, int discard)
2943 struct osc_extent *ext;
2944 CFS_LIST_HEAD(discard_list);
2945 bool unplug = false;
2949 osc_object_lock(obj);
2950 ext = osc_extent_search(obj, start);
2952 ext = first_extent(obj);
2953 else if (ext->oe_end < start)
2954 ext = next_extent(ext);
2955 while (ext != NULL) {
2956 if (ext->oe_start > end)
2959 ext->oe_fsync_wait = 1;
2960 switch (ext->oe_state) {
2962 result += ext->oe_nr_pages;
2964 cfs_list_t *list = NULL;
2966 EASSERT(!ext->oe_hp, ext);
2968 list = &obj->oo_hp_exts;
2969 } else if (!ext->oe_urgent) {
2971 list = &obj->oo_urgent_exts;
2974 cfs_list_move_tail(&ext->oe_link, list);
2977 /* the only discarder is lock cancelling, so
2978 * [start, end] must contain this extent */
2979 EASSERT(ext->oe_start >= start &&
2980 ext->oe_max_end <= end, ext);
2981 osc_extent_state_set(ext, OES_LOCKING);
2982 ext->oe_owner = current;
2983 cfs_list_move_tail(&ext->oe_link,
2985 osc_update_pending(obj, OBD_BRW_WRITE,
2990 /* It's pretty bad to wait for ACTIVE extents, because
2991 * we don't know how long we will wait for it to be
2992 * flushed since it may be blocked at awaiting more
2993 * grants. We do this for the correctness of fsync. */
2994 LASSERT(hp == 0 && discard == 0);
2998 /* this extent is being truncated, can't do anything
2999 * for it now. it will be set to urgent after truncate
3000 * is finished in osc_cache_truncate_end(). */
3004 ext = next_extent(ext);
3006 osc_object_unlock(obj);
3008 LASSERT(ergo(!discard, cfs_list_empty(&discard_list)));
3009 if (!cfs_list_empty(&discard_list)) {
3010 struct osc_extent *tmp;
3013 osc_list_maint(osc_cli(obj), obj);
3014 cfs_list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3015 cfs_list_del_init(&ext->oe_link);
3016 EASSERT(ext->oe_state == OES_LOCKING, ext);
3018 /* Discard caching pages. We don't actually write this
3019 * extent out but we complete it as if we did. */
3020 rc = osc_extent_make_ready(env, ext);
3021 if (unlikely(rc < 0)) {
3022 OSC_EXTENT_DUMP(D_ERROR, ext,
3023 "make_ready returned %d\n", rc);
3028 /* finish the extent as if the pages were sent */
3029 osc_extent_finish(env, ext, 0, 0);
3034 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
3036 if (hp || discard) {
3038 rc = osc_cache_wait_range(env, obj, start, end);
3039 if (result >= 0 && rc < 0)
3043 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3048 * Returns a list of pages by a given [start, end] of \a obj.
3050 * \param resched If not NULL, then we give up before hogging CPU for too
3051 * long and set *resched = 1, in that case caller should implement a retry
3054 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3055 * crucial in the face of [offset, EOF] locks.
3057 * Return at least one page in @queue unless there is no covered page.
3059 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3060 struct osc_object *osc, pgoff_t start, pgoff_t end,
3061 osc_page_gang_cbt cb, void *cbdata)
3063 struct osc_page *ops;
3069 int res = CLP_GANG_OKAY;
3070 bool tree_lock = true;
3074 pvec = osc_env_info(env)->oti_pvec;
3075 spin_lock(&osc->oo_tree_lock);
3076 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3077 idx, OTI_PVEC_SIZE)) > 0) {
3078 struct cl_page *page;
3079 bool end_of_region = false;
3081 for (i = 0, j = 0; i < nr; ++i) {
3085 idx = osc_index(ops);
3087 end_of_region = true;
3091 page = ops->ops_cl.cpl_page;
3092 LASSERT(page->cp_type == CPT_CACHEABLE);
3093 if (page->cp_state == CPS_FREEING)
3097 lu_ref_add_atomic(&page->cp_reference,
3098 "gang_lookup", current);
3104 * Here a delicate locking dance is performed. Current thread
3105 * holds a reference to a page, but has to own it before it
3106 * can be placed into queue. Owning implies waiting, so
3107 * radix-tree lock is to be released. After a wait one has to
3108 * check that pages weren't truncated (cl_page_own() returns
3109 * error in the latter case).
3111 spin_unlock(&osc->oo_tree_lock);
3114 for (i = 0; i < j; ++i) {
3116 if (res == CLP_GANG_OKAY)
3117 res = (*cb)(env, io, ops, cbdata);
3119 page = ops->ops_cl.cpl_page;
3120 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3121 cl_page_put(env, page);
3123 if (nr < OTI_PVEC_SIZE || end_of_region)
3126 if (res == CLP_GANG_OKAY && need_resched())
3127 res = CLP_GANG_RESCHED;
3128 if (res != CLP_GANG_OKAY)
3131 spin_lock(&osc->oo_tree_lock);
3135 spin_unlock(&osc->oo_tree_lock);
3140 * Check if page @page is covered by an extra lock or discard it.
3142 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3143 struct osc_page *ops, void *cbdata)
3145 struct osc_thread_info *info = osc_env_info(env);
3146 struct cl_lock *lock = cbdata;
3149 index = osc_index(ops);
3150 if (index >= info->oti_fn_index) {
3151 struct cl_lock *tmp;
3152 struct cl_page *page = ops->ops_cl.cpl_page;
3154 /* refresh non-overlapped index */
3155 tmp = cl_lock_at_pgoff(env, lock->cll_descr.cld_obj, index,
3158 /* Cache the first-non-overlapped index so as to skip
3159 * all pages within [index, oti_fn_index). This
3160 * is safe because if tmp lock is canceled, it will
3161 * discard these pages. */
3162 info->oti_fn_index = tmp->cll_descr.cld_end + 1;
3163 if (tmp->cll_descr.cld_end == CL_PAGE_EOF)
3164 info->oti_fn_index = CL_PAGE_EOF;
3165 cl_lock_put(env, tmp);
3166 } else if (cl_page_own(env, io, page) == 0) {
3167 /* discard the page */
3168 cl_page_discard(env, io, page);
3169 cl_page_disown(env, io, page);
3171 LASSERT(page->cp_state == CPS_FREEING);
3175 info->oti_next_index = index + 1;
3176 return CLP_GANG_OKAY;
3179 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3180 struct osc_page *ops, void *cbdata)
3182 struct osc_thread_info *info = osc_env_info(env);
3183 struct cl_lock *lock = cbdata;
3184 struct cl_page *page = ops->ops_cl.cpl_page;
3186 LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
3188 /* page is top page. */
3189 info->oti_next_index = osc_index(ops) + 1;
3190 if (cl_page_own(env, io, page) == 0) {
3191 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3192 !PageDirty(cl_page_vmpage(page))));
3194 /* discard the page */
3195 cl_page_discard(env, io, page);
3196 cl_page_disown(env, io, page);
3198 LASSERT(page->cp_state == CPS_FREEING);
3201 return CLP_GANG_OKAY;
3205 * Discard pages protected by the given lock. This function traverses radix
3206 * tree to find all covering pages and discard them. If a page is being covered
3207 * by other locks, it should remain in cache.
3209 * If error happens on any step, the process continues anyway (the reasoning
3210 * behind this being that lock cancellation cannot be delayed indefinitely).
3212 int osc_lock_discard_pages(const struct lu_env *env, struct osc_lock *ols)
3214 struct osc_thread_info *info = osc_env_info(env);
3215 struct cl_io *io = &info->oti_io;
3216 struct cl_object *osc = ols->ols_cl.cls_obj;
3217 struct cl_lock *lock = ols->ols_cl.cls_lock;
3218 struct cl_lock_descr *descr = &lock->cll_descr;
3219 osc_page_gang_cbt cb;
3225 io->ci_obj = cl_object_top(osc);
3226 io->ci_ignore_layout = 1;
3227 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3231 cb = descr->cld_mode == CLM_READ ? check_and_discard_cb : discard_cb;
3232 info->oti_fn_index = info->oti_next_index = descr->cld_start;
3234 res = osc_page_gang_lookup(env, io, cl2osc(osc),
3235 info->oti_next_index, descr->cld_end,
3237 if (info->oti_next_index > descr->cld_end)
3240 if (res == CLP_GANG_RESCHED)
3242 } while (res != CLP_GANG_OKAY);
3244 cl_io_fini(env, io);