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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
30 * This file is part of Lustre, http://www.lustre.org/
31 * Lustre is a trademark of Sun Microsystems, Inc.
33 * osc cache management.
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_OSC
40 #include <lustre_osc.h>
41 #include <lustre_dlm.h>
43 #include "osc_internal.h"
45 static int extent_debug; /* set it to be true for more debug */
47 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
48 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
49 enum osc_extent_state state);
50 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
51 struct osc_async_page *oap, int sent, int rc);
52 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
54 static int osc_refresh_count(const struct lu_env *env,
55 struct osc_async_page *oap, int cmd);
56 static int osc_io_unplug_async(const struct lu_env *env,
57 struct client_obd *cli, struct osc_object *osc);
58 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
59 unsigned int lost_grant, unsigned int dirty_grant);
61 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
62 const char *func, int line);
63 #define osc_extent_tree_dump(mask, obj) \
64 osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
66 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
73 /* ------------------ osc extent ------------------ */
74 static inline char *ext_flags(struct osc_extent *ext, char *flags)
77 *buf++ = ext->oe_rw ? 'r' : 'w';
78 if (!RB_EMPTY_NODE(&ext->oe_node))
90 if (ext->oe_trunc_pending)
92 if (ext->oe_fsync_wait)
98 #define EXTSTR "[%lu -> %lu/%lu]"
99 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
100 static const char *oes_strings[] = {
101 "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
103 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
104 static struct cfs_debug_limit_state cdls; \
105 struct osc_extent *__ext = (extent); \
108 __CDEBUG_WITH_LOC(file, func, line, mask, &cdls, \
109 "extent %p@{" EXTSTR ", " \
110 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt, \
111 /* ----- extent part 0 ----- */ \
112 __ext, EXTPARA(__ext), \
113 /* ----- part 1 ----- */ \
114 kref_read(&__ext->oe_refc), \
115 atomic_read(&__ext->oe_users), \
116 list_empty_marker(&__ext->oe_link), \
117 oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \
119 /* ----- part 2 ----- */ \
120 __ext->oe_grants, __ext->oe_nr_pages, \
121 list_empty_marker(&__ext->oe_pages), \
122 waitqueue_active(&__ext->oe_waitq) ? '+' : '-', \
123 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner, \
124 /* ----- part 4 ----- */ \
126 if (mask == D_ERROR && __ext->oe_dlmlock != NULL) \
127 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext); \
129 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext); \
132 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...) \
133 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__, \
134 mask, ext, fmt, ## __VA_ARGS__)
137 #define EASSERTF(expr, ext, fmt, args...) do { \
139 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args); \
140 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj); \
146 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
148 static inline struct osc_extent *rb_extent(struct rb_node *n)
150 return rb_entry_safe(n, struct osc_extent, oe_node);
153 static inline struct osc_extent *next_extent(struct osc_extent *ext)
158 LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
159 return rb_extent(rb_next(&ext->oe_node));
162 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
167 LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
168 return rb_extent(rb_prev(&ext->oe_node));
171 static inline struct osc_extent *first_extent(struct osc_object *obj)
173 return rb_extent(rb_first(&obj->oo_root));
176 /* object must be locked by caller. */
177 static int osc_extent_sanity_check0(struct osc_extent *ext,
178 const char *func, const int line)
180 struct osc_object *obj = ext->oe_obj;
181 struct osc_async_page *oap;
185 assert_osc_object_is_locked(obj);
187 if (ext->oe_state >= OES_STATE_MAX)
190 if (kref_read(&ext->oe_refc) <= 0)
193 if (kref_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
196 switch (ext->oe_state) {
198 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
203 if (atomic_read(&ext->oe_users) == 0)
207 if (ext->oe_fsync_wait && !ext->oe_urgent)
211 if (ext->oe_grants == 0)
213 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
217 if (atomic_read(&ext->oe_users) > 0)
221 if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
224 if (ext->oe_sync && ext->oe_grants > 0)
227 if (ext->oe_dlmlock != NULL &&
228 ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
229 !ldlm_is_failed(ext->oe_dlmlock)) {
230 struct ldlm_extent *extent;
232 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
233 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
234 extent->end >= cl_offset(osc2cl(obj), ext->oe_max_end)))
237 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
241 if (ext->oe_nr_pages > ext->oe_mppr)
244 /* Do not verify page list if extent is in RPC. This is because an
245 * in-RPC extent is supposed to be exclusively accessible w/o lock. */
246 if (ext->oe_state > OES_CACHE)
253 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
254 pgoff_t index = osc_index(oap2osc(oap));
256 if (index > ext->oe_end || index < ext->oe_start)
259 if (page_count != ext->oe_nr_pages)
264 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
265 "sanity check %p failed: rc = %d\n",
270 #define sanity_check_nolock(ext) \
271 osc_extent_sanity_check0(ext, __func__, __LINE__)
273 #define sanity_check(ext) ({ \
275 osc_object_lock((ext)->oe_obj); \
276 __res = sanity_check_nolock(ext); \
277 osc_object_unlock((ext)->oe_obj); \
282 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
284 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
288 * sanity check - to make sure there is no overlapped extent in the tree.
290 static int osc_extent_is_overlapped(struct osc_object *obj,
291 struct osc_extent *ext)
293 struct osc_extent *tmp;
295 assert_osc_object_is_locked(obj);
300 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
303 if (overlapped(tmp, ext))
309 static void osc_extent_state_set(struct osc_extent *ext, int state)
311 assert_osc_object_is_locked(ext->oe_obj);
312 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
314 /* Never try to sanity check a state changing extent :-) */
315 /* LASSERT(sanity_check_nolock(ext) == 0); */
317 /* TODO: validate the state machine */
318 smp_store_release(&ext->oe_state, state);
319 wake_up(&ext->oe_waitq);
322 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
324 struct osc_extent *ext;
326 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
330 RB_CLEAR_NODE(&ext->oe_node);
332 cl_object_get(osc2cl(obj));
333 kref_init(&ext->oe_refc);
334 atomic_set(&ext->oe_users, 0);
335 INIT_LIST_HEAD(&ext->oe_link);
336 ext->oe_state = OES_INV;
337 INIT_LIST_HEAD(&ext->oe_pages);
338 init_waitqueue_head(&ext->oe_waitq);
339 ext->oe_dlmlock = NULL;
344 static void osc_extent_free(struct kref *kref)
346 struct osc_extent *ext = container_of(kref, struct osc_extent,
349 LASSERT(list_empty(&ext->oe_link));
350 LASSERT(atomic_read(&ext->oe_users) == 0);
351 LASSERT(ext->oe_state == OES_INV);
352 LASSERT(RB_EMPTY_NODE(&ext->oe_node));
354 if (ext->oe_dlmlock) {
355 lu_ref_del(&ext->oe_dlmlock->l_reference,
357 LDLM_LOCK_PUT(ext->oe_dlmlock);
358 ext->oe_dlmlock = NULL;
361 /* If/When cl_object_put drops the need for 'env',
362 * this code can be enabled, and matching code in
363 * osc_extent_put removed.
365 cl_object_put(osc2cl(ext->oe_obj));
367 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
371 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
373 LASSERT(kref_read(&ext->oe_refc) >= 0);
374 kref_get(&ext->oe_refc);
378 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
380 LASSERT(kref_read(&ext->oe_refc) > 0);
381 if (kref_put(&ext->oe_refc, osc_extent_free)) {
382 /* This should be in osc_extent_free(), but
383 * while we need to pass 'env' it cannot be.
385 cl_object_put(env, osc2cl(ext->oe_obj));
387 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
392 * osc_extent_put_trust() is a special version of osc_extent_put() when
393 * it's known that the caller is not the last user. This is to address the
394 * problem of lacking of lu_env ;-).
396 static void osc_extent_put_trust(struct osc_extent *ext)
398 LASSERT(kref_read(&ext->oe_refc) > 1);
399 assert_osc_object_is_locked(ext->oe_obj);
400 osc_extent_put(NULL, ext);
404 * Return the extent which includes pgoff @index, or return the greatest
405 * previous extent in the tree.
407 static struct osc_extent *osc_extent_search(struct osc_object *obj,
410 struct rb_node *n = obj->oo_root.rb_node;
411 struct osc_extent *tmp, *p = NULL;
413 assert_osc_object_is_locked(obj);
416 if (index < tmp->oe_start) {
418 } else if (index > tmp->oe_end) {
429 * Return the extent covering @index, otherwise return NULL.
430 * caller must have held object lock.
432 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
435 struct osc_extent *ext;
437 ext = osc_extent_search(obj, index);
438 if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
439 return osc_extent_get(ext);
443 /* caller must have held object lock. */
444 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
446 struct rb_node **n = &obj->oo_root.rb_node;
447 struct rb_node *parent = NULL;
448 struct osc_extent *tmp;
450 LASSERT(RB_EMPTY_NODE(&ext->oe_node));
451 LASSERT(ext->oe_obj == obj);
452 assert_osc_object_is_locked(obj);
457 if (ext->oe_end < tmp->oe_start)
459 else if (ext->oe_start > tmp->oe_end)
462 EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
464 rb_link_node(&ext->oe_node, parent, n);
465 rb_insert_color(&ext->oe_node, &obj->oo_root);
469 /* caller must have held object lock. */
470 static void osc_extent_erase(struct osc_extent *ext)
472 struct osc_object *obj = ext->oe_obj;
473 assert_osc_object_is_locked(obj);
474 if (!RB_EMPTY_NODE(&ext->oe_node)) {
475 rb_erase(&ext->oe_node, &obj->oo_root);
476 RB_CLEAR_NODE(&ext->oe_node);
477 /* rbtree held a refcount */
478 osc_extent_put_trust(ext);
482 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
484 struct osc_object *obj = ext->oe_obj;
486 assert_osc_object_is_locked(obj);
487 LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
488 if (ext->oe_state == OES_CACHE) {
489 osc_extent_state_set(ext, OES_ACTIVE);
490 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
492 atomic_inc(&ext->oe_users);
493 list_del_init(&ext->oe_link);
494 return osc_extent_get(ext);
497 static void __osc_extent_remove(struct osc_extent *ext)
499 assert_osc_object_is_locked(ext->oe_obj);
500 LASSERT(list_empty(&ext->oe_pages));
501 osc_extent_erase(ext);
502 list_del_init(&ext->oe_link);
503 osc_extent_state_set(ext, OES_INV);
504 OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
507 static void osc_extent_remove(struct osc_extent *ext)
509 struct osc_object *obj = ext->oe_obj;
511 osc_object_lock(obj);
512 __osc_extent_remove(ext);
513 osc_object_unlock(obj);
517 * This function is used to merge extents to get better performance. It checks
518 * if @cur and @victim are contiguous at block level.
520 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
521 struct osc_extent *victim)
523 struct osc_object *obj = cur->oe_obj;
524 struct client_obd *cli = osc_cli(obj);
529 LASSERT(cur->oe_state == OES_CACHE);
530 assert_osc_object_is_locked(obj);
534 if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
537 if (cur->oe_max_end != victim->oe_max_end)
540 LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
541 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
542 chunk_start = cur->oe_start >> ppc_bits;
543 chunk_end = cur->oe_end >> ppc_bits;
544 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
545 chunk_end + 1 != victim->oe_start >> ppc_bits)
548 /* overall extent size should not exceed the max supported limit
549 * reported by the server */
550 if (cur->oe_end - cur->oe_start + 1 +
551 victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
554 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
556 cur->oe_start = min(cur->oe_start, victim->oe_start);
557 cur->oe_end = max(cur->oe_end, victim->oe_end);
558 /* per-extent tax should be accounted only once for the whole extent */
559 cur->oe_grants += victim->oe_grants - cli->cl_grant_extent_tax;
560 cur->oe_nr_pages += victim->oe_nr_pages;
561 /* only the following bits are needed to merge */
562 cur->oe_urgent |= victim->oe_urgent;
563 cur->oe_memalloc |= victim->oe_memalloc;
564 list_splice_init(&victim->oe_pages, &cur->oe_pages);
565 list_del_init(&victim->oe_link);
566 victim->oe_nr_pages = 0;
568 osc_extent_get(victim);
569 __osc_extent_remove(victim);
570 osc_extent_put(env, victim);
572 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
577 * Drop user count of osc_extent, and unplug IO asynchronously.
579 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
581 struct osc_object *obj = ext->oe_obj;
582 struct client_obd *cli = osc_cli(obj);
586 LASSERT(atomic_read(&ext->oe_users) > 0);
587 LASSERT(sanity_check(ext) == 0);
588 LASSERT(ext->oe_grants > 0);
590 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
591 LASSERT(ext->oe_state == OES_ACTIVE);
592 if (ext->oe_trunc_pending) {
593 /* a truncate process is waiting for this extent.
594 * This may happen due to a race, check
595 * osc_cache_truncate_start(). */
596 osc_extent_state_set(ext, OES_TRUNC);
597 ext->oe_trunc_pending = 0;
598 osc_object_unlock(obj);
602 osc_extent_state_set(ext, OES_CACHE);
603 osc_update_pending(obj, OBD_BRW_WRITE,
606 /* try to merge the previous and next extent. */
607 if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
608 grant += cli->cl_grant_extent_tax;
609 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
610 grant += cli->cl_grant_extent_tax;
613 list_move_tail(&ext->oe_link,
615 else if (ext->oe_urgent)
616 list_move_tail(&ext->oe_link,
617 &obj->oo_urgent_exts);
618 else if (ext->oe_nr_pages == ext->oe_mppr) {
619 list_move_tail(&ext->oe_link,
622 osc_object_unlock(obj);
624 osc_unreserve_grant(cli, 0, grant);
627 osc_io_unplug_async(env, cli, obj);
629 osc_extent_put(env, ext);
634 * Find or create an extent which includes @index, core function to manage
637 static struct osc_extent *osc_extent_find(const struct lu_env *env,
638 struct osc_object *obj, pgoff_t index,
639 unsigned int *grants)
641 struct client_obd *cli = osc_cli(obj);
642 struct osc_lock *olck;
643 struct cl_lock_descr *descr;
644 struct osc_extent *cur;
645 struct osc_extent *ext;
646 struct osc_extent *conflict = NULL;
647 struct osc_extent *found = NULL;
650 unsigned int max_pages; /* max_pages_per_rpc */
651 unsigned int chunksize;
652 int ppc_bits; /* pages per chunk bits */
657 cur = osc_extent_alloc(obj);
659 RETURN(ERR_PTR(-ENOMEM));
661 olck = osc_env_io(env)->oi_write_osclock;
662 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
663 LASSERT(olck->ols_state == OLS_GRANTED);
665 descr = &olck->ols_cl.cls_lock->cll_descr;
666 LASSERT(descr->cld_mode >= CLM_WRITE);
668 LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
669 "chunkbits: %u\n", cli->cl_chunkbits);
670 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
671 chunk_mask = ~((1 << ppc_bits) - 1);
672 chunksize = 1 << cli->cl_chunkbits;
673 chunk = index >> ppc_bits;
675 /* align end to RPC edge. */
676 max_pages = cli->cl_max_pages_per_rpc;
677 if ((max_pages & ~chunk_mask) != 0) {
678 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
679 max_pages, cli->cl_chunkbits, chunk_mask);
680 RETURN(ERR_PTR(-EINVAL));
682 max_end = index - (index % max_pages) + max_pages - 1;
683 max_end = min_t(pgoff_t, max_end, descr->cld_end);
685 /* initialize new extent by parameters so far */
686 cur->oe_max_end = max_end;
687 cur->oe_start = index & chunk_mask;
688 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
689 if (cur->oe_start < descr->cld_start)
690 cur->oe_start = descr->cld_start;
691 if (cur->oe_end > max_end)
692 cur->oe_end = max_end;
694 cur->oe_mppr = max_pages;
695 if (olck->ols_dlmlock != NULL) {
696 LASSERT(olck->ols_hold);
697 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
698 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
701 /* grants has been allocated by caller */
702 LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
703 "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
704 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
708 osc_object_lock(obj);
709 ext = osc_extent_search(obj, cur->oe_start);
711 ext = first_extent(obj);
712 for (; ext; ext = next_extent(ext)) {
713 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
714 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
716 LASSERT(sanity_check_nolock(ext) == 0);
717 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
720 /* if covering by different locks, no chance to match */
721 if (olck->ols_dlmlock != ext->oe_dlmlock) {
722 EASSERTF(!overlapped(ext, cur), ext,
723 EXTSTR"\n", EXTPARA(cur));
728 /* discontiguous chunks? */
729 if (chunk + 1 < ext_chk_start)
732 /* ok, from now on, ext and cur have these attrs:
733 * 1. covered by the same lock
734 * 2. contiguous at chunk level or overlapping. */
736 if (overlapped(ext, cur)) {
737 /* cur is the minimum unit, so overlapping means
739 EASSERTF((ext->oe_start <= cur->oe_start &&
740 ext->oe_end >= cur->oe_end),
741 ext, EXTSTR"\n", EXTPARA(cur));
743 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
744 /* for simplicity, we wait for this extent to
745 * finish before going forward. */
746 conflict = osc_extent_get(ext);
750 found = osc_extent_hold(ext);
754 /* non-overlapped extent */
755 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
756 /* we can't do anything for a non OES_CACHE extent, or
757 * if there is someone waiting for this extent to be
758 * flushed, try next one. */
761 /* check if they belong to the same rpc slot before trying to
762 * merge. the extents are not overlapped and contiguous at
763 * chunk level to get here. */
764 if (ext->oe_max_end != max_end)
765 /* if they don't belong to the same RPC slot or
766 * max_pages_per_rpc has ever changed, do not merge. */
769 /* check whether maximum extent size will be hit */
770 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
771 cli->cl_max_extent_pages)
774 /* it's required that an extent must be contiguous at chunk
775 * level so that we know the whole extent is covered by grant
776 * (the pages in the extent are NOT required to be contiguous).
777 * Otherwise, it will be too much difficult to know which
778 * chunks have grants allocated. */
780 /* try to do front merge - extend ext's start */
781 if (chunk + 1 == ext_chk_start) {
782 /* ext must be chunk size aligned */
783 EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
785 /* pull ext's start back to cover cur */
786 ext->oe_start = cur->oe_start;
787 ext->oe_grants += chunksize;
788 LASSERT(*grants >= chunksize);
789 *grants -= chunksize;
791 found = osc_extent_hold(ext);
792 } else if (chunk == ext_chk_end + 1) {
794 ext->oe_end = cur->oe_end;
795 ext->oe_grants += chunksize;
796 LASSERT(*grants >= chunksize);
797 *grants -= chunksize;
799 /* try to merge with the next one because we just fill
801 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
802 /* we can save extent tax from next extent */
803 *grants += cli->cl_grant_extent_tax;
805 found = osc_extent_hold(ext);
811 osc_extent_tree_dump(D_CACHE, obj);
813 LASSERT(conflict == NULL);
814 if (!IS_ERR(found)) {
815 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
816 OSC_EXTENT_DUMP(D_CACHE, found,
817 "found caching ext for %lu.\n", index);
819 } else if (conflict == NULL) {
820 /* create a new extent */
821 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
822 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
823 LASSERT(*grants >= cur->oe_grants);
824 *grants -= cur->oe_grants;
826 cur->oe_state = OES_CACHE;
827 found = osc_extent_hold(cur);
828 osc_extent_insert(obj, cur);
829 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
830 index, descr->cld_end);
832 osc_object_unlock(obj);
834 if (conflict != NULL) {
835 LASSERT(found == NULL);
837 /* waiting for IO to finish. Please notice that it's impossible
838 * to be an OES_TRUNC extent. */
839 rc = osc_extent_wait(env, conflict, OES_INV);
840 osc_extent_put(env, conflict);
843 GOTO(out, found = ERR_PTR(rc));
850 osc_extent_put(env, cur);
855 * Called when IO is finished to an extent.
857 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
860 struct client_obd *cli = osc_cli(ext->oe_obj);
861 struct osc_async_page *oap;
862 struct osc_async_page *tmp;
863 int nr_pages = ext->oe_nr_pages;
865 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
870 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
872 ext->oe_rc = rc ?: ext->oe_nr_pages;
873 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
875 osc_lru_add_batch(cli, &ext->oe_pages);
876 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
878 list_del_init(&oap->oap_rpc_item);
879 list_del_init(&oap->oap_pending_item);
880 if (last_off <= oap->oap_obj_off) {
881 last_off = oap->oap_obj_off;
882 last_count = oap->oap_count;
886 osc_ap_completion(env, cli, oap, sent, rc);
888 EASSERT(ext->oe_nr_pages == 0, ext);
891 lost_grant = ext->oe_grants;
892 } else if (blocksize < PAGE_SIZE &&
893 last_count != PAGE_SIZE) {
894 /* For short writes we shouldn't count parts of pages that
895 * span a whole chunk on the OST side, or our accounting goes
896 * wrong. Should match the code in filter_grant_check. */
897 int offset = last_off & ~PAGE_MASK;
898 int count = last_count + (offset & (blocksize - 1));
899 int end = (offset + last_count) & (blocksize - 1);
901 count += blocksize - end;
903 lost_grant = PAGE_SIZE - count;
905 if (ext->oe_grants > 0)
906 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
908 osc_extent_remove(ext);
909 /* put the refcount for RPC */
910 osc_extent_put(env, ext);
915 * Wait for the extent's state to become @state.
917 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
918 enum osc_extent_state state)
920 struct osc_object *obj = ext->oe_obj;
924 osc_object_lock(obj);
925 LASSERT(sanity_check_nolock(ext) == 0);
926 /* `Kick' this extent only if the caller is waiting for it to be
928 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
929 if (ext->oe_state == OES_ACTIVE) {
931 } else if (ext->oe_state == OES_CACHE) {
933 osc_extent_hold(ext);
937 osc_object_unlock(obj);
939 osc_extent_release(env, ext);
941 /* wait for the extent until its state becomes @state */
942 rc = wait_event_idle_timeout(ext->oe_waitq,
943 smp_load_acquire(&ext->oe_state) == state,
944 cfs_time_seconds(600));
946 OSC_EXTENT_DUMP(D_ERROR, ext,
947 "%s: wait ext to %u timedout, recovery in progress?\n",
948 cli_name(osc_cli(obj)), state);
950 wait_event_idle(ext->oe_waitq,
951 smp_load_acquire(&ext->oe_state) == state);
961 * Discard pages with index greater than @size. If @ext is overlapped with
962 * @size, then partial truncate happens.
964 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
969 struct osc_object *obj = ext->oe_obj;
970 struct client_obd *cli = osc_cli(obj);
971 struct osc_async_page *oap;
972 struct osc_async_page *tmp;
973 struct pagevec *pvec;
974 int pages_in_chunk = 0;
975 int ppc_bits = cli->cl_chunkbits -
977 __u64 trunc_chunk = trunc_index >> ppc_bits;
984 LASSERT(sanity_check(ext) == 0);
985 LASSERT(ext->oe_state == OES_TRUNC);
986 LASSERT(!ext->oe_urgent);
988 /* Request new lu_env.
989 * We can't use that env from osc_cache_truncate_start() because
990 * it's from lov_io_sub and not fully initialized. */
991 env = cl_env_get(&refcheck);
993 RETURN(PTR_ERR(env));
995 io = osc_env_thread_io(env);
996 io->ci_obj = cl_object_top(osc2cl(obj));
997 io->ci_ignore_layout = 1;
998 pvec = &osc_env_info(env)->oti_pagevec;
999 ll_pagevec_init(pvec, 0);
1000 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1004 /* discard all pages with index greater than trunc_index */
1005 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1007 pgoff_t index = osc_index(oap2osc(oap));
1008 struct cl_page *page = oap2cl_page(oap);
1010 LASSERT(list_empty(&oap->oap_rpc_item));
1012 /* only discard the pages with their index greater than
1013 * trunc_index, and ... */
1014 if (index < trunc_index ||
1015 (index == trunc_index && partial)) {
1016 /* accounting how many pages remaining in the chunk
1017 * so that we can calculate grants correctly. */
1018 if (index >> ppc_bits == trunc_chunk)
1023 list_del_init(&oap->oap_pending_item);
1026 lu_ref_add(&page->cp_reference, "truncate", current);
1028 if (cl_page_own(env, io, page) == 0) {
1029 cl_page_discard(env, io, page);
1030 cl_page_disown(env, io, page);
1032 LASSERT(page->cp_state == CPS_FREEING);
1036 lu_ref_del(&page->cp_reference, "truncate", current);
1037 cl_pagevec_put(env, page, pvec);
1042 pagevec_release(pvec);
1044 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1045 ext->oe_nr_pages == 0),
1046 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1048 osc_object_lock(obj);
1049 if (ext->oe_nr_pages == 0) {
1050 LASSERT(pages_in_chunk == 0);
1051 grants = ext->oe_grants;
1053 } else { /* calculate how many grants we can free */
1054 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1058 /* if there is no pages in this chunk, we can also free grants
1059 * for the last chunk */
1060 if (pages_in_chunk == 0) {
1061 /* if this is the 1st chunk and no pages in this chunk,
1062 * ext->oe_nr_pages must be zero, so we should be in
1063 * the other if-clause. */
1064 LASSERT(trunc_chunk > 0);
1069 /* this is what we can free from this extent */
1070 grants = chunks << cli->cl_chunkbits;
1071 ext->oe_grants -= grants;
1072 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1073 ext->oe_end = min(last_index, ext->oe_max_end);
1074 LASSERT(ext->oe_end >= ext->oe_start);
1075 LASSERT(ext->oe_grants > 0);
1077 osc_object_unlock(obj);
1079 if (grants > 0 || nr_pages > 0)
1080 osc_free_grant(cli, nr_pages, grants, grants);
1083 cl_io_fini(env, io);
1084 cl_env_put(env, &refcheck);
1089 * This function is used to make the extent prepared for transfer.
1090 * A race with flusing page - ll_writepage() has to be handled cautiously.
1092 static int osc_extent_make_ready(const struct lu_env *env,
1093 struct osc_extent *ext)
1095 struct osc_async_page *oap;
1096 struct osc_async_page *last = NULL;
1097 struct osc_object *obj = ext->oe_obj;
1098 unsigned int page_count = 0;
1102 /* we're going to grab page lock, so object lock must not be taken. */
1103 LASSERT(sanity_check(ext) == 0);
1104 /* in locking state, any process should not touch this extent. */
1105 EASSERT(ext->oe_state == OES_LOCKING, ext);
1106 EASSERT(ext->oe_owner != NULL, ext);
1108 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1110 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1112 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1115 /* checking ASYNC_READY is race safe */
1116 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1119 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1122 spin_lock(&oap->oap_lock);
1123 oap->oap_async_flags |= ASYNC_READY;
1124 spin_unlock(&oap->oap_lock);
1127 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1130 LASSERTF(0, "unknown return code: %d\n", rc);
1134 LASSERT(page_count == ext->oe_nr_pages);
1135 LASSERT(last != NULL);
1136 /* the last page is the only one we need to refresh its count by
1137 * the size of file. */
1138 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1139 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1140 LASSERT(last_oap_count > 0);
1141 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1142 last->oap_count = last_oap_count;
1143 spin_lock(&last->oap_lock);
1144 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1145 spin_unlock(&last->oap_lock);
1148 /* for the rest of pages, we don't need to call osf_refresh_count()
1149 * because it's known they are not the last page */
1150 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1151 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1152 oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1153 spin_lock(&oap->oap_lock);
1154 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1155 spin_unlock(&oap->oap_lock);
1159 osc_object_lock(obj);
1160 osc_extent_state_set(ext, OES_RPC);
1161 osc_object_unlock(obj);
1162 /* get a refcount for RPC. */
1163 osc_extent_get(ext);
1169 * Quick and simple version of osc_extent_find(). This function is frequently
1170 * called to expand the extent for the same IO. To expand the extent, the
1171 * page index must be in the same or next chunk of ext->oe_end.
1173 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1174 unsigned int *grants)
1176 struct osc_object *obj = ext->oe_obj;
1177 struct client_obd *cli = osc_cli(obj);
1178 struct osc_extent *next;
1179 int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1180 pgoff_t chunk = index >> ppc_bits;
1183 unsigned int chunksize = 1 << cli->cl_chunkbits;
1187 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1188 osc_object_lock(obj);
1189 LASSERT(sanity_check_nolock(ext) == 0);
1190 end_chunk = ext->oe_end >> ppc_bits;
1191 if (chunk > end_chunk + 1)
1192 GOTO(out, rc = -ERANGE);
1194 if (end_chunk >= chunk)
1197 LASSERT(end_chunk + 1 == chunk);
1199 /* try to expand this extent to cover @index */
1200 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1202 /* don't go over the maximum extent size reported by server */
1203 if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1204 GOTO(out, rc = -ERANGE);
1206 next = next_extent(ext);
1207 if (next != NULL && next->oe_start <= end_index)
1208 /* complex mode - overlapped with the next extent,
1209 * this case will be handled by osc_extent_find() */
1210 GOTO(out, rc = -EAGAIN);
1212 ext->oe_end = end_index;
1213 ext->oe_grants += chunksize;
1214 LASSERT(*grants >= chunksize);
1215 *grants -= chunksize;
1216 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1217 "overlapped after expanding for %lu.\n", index);
1221 osc_object_unlock(obj);
1225 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1226 const char *func, int line)
1228 struct osc_extent *ext;
1231 if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1234 CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1235 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1237 /* osc_object_lock(obj); */
1239 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1240 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1243 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1244 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1247 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1248 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1251 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1252 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1253 /* osc_object_unlock(obj); */
1256 /* ------------------ osc extent end ------------------ */
1258 static inline int osc_is_ready(struct osc_object *osc)
1260 return !list_empty(&osc->oo_ready_item) ||
1261 !list_empty(&osc->oo_hp_ready_item);
1264 #define OSC_IO_DEBUG(OSC, STR, args...) \
1265 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1266 (OSC), osc_is_ready(OSC), \
1267 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1268 list_empty_marker(&(OSC)->oo_ready_item), \
1269 atomic_read(&(OSC)->oo_nr_writes), \
1270 list_empty_marker(&(OSC)->oo_hp_exts), \
1271 list_empty_marker(&(OSC)->oo_urgent_exts), \
1272 atomic_read(&(OSC)->oo_nr_reads), \
1273 list_empty_marker(&(OSC)->oo_reading_exts), \
1276 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1279 struct osc_page *opg = oap2osc_page(oap);
1280 struct cl_page *page = oap2cl_page(oap);
1283 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1286 result = cl_page_make_ready(env, page, CRT_WRITE);
1288 opg->ops_submit_time = ktime_get();
1292 static int osc_refresh_count(const struct lu_env *env,
1293 struct osc_async_page *oap, int cmd)
1295 struct osc_page *opg = oap2osc_page(oap);
1296 pgoff_t index = osc_index(oap2osc(oap));
1297 struct cl_object *obj;
1298 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1302 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1303 LASSERT(!(cmd & OBD_BRW_READ));
1304 LASSERT(opg != NULL);
1305 obj = opg->ops_cl.cpl_obj;
1307 cl_object_attr_lock(obj);
1308 result = cl_object_attr_get(env, obj, attr);
1309 cl_object_attr_unlock(obj);
1312 kms = attr->cat_kms;
1313 if (cl_offset(obj, index) >= kms)
1314 /* catch race with truncate */
1316 else if (cl_offset(obj, index + 1) > kms)
1317 /* catch sub-page write at end of file */
1318 return kms & ~PAGE_MASK;
1323 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1326 struct osc_page *opg = oap2osc_page(oap);
1327 struct cl_page *page = oap2cl_page(oap);
1328 enum cl_req_type crt;
1333 cmd &= ~OBD_BRW_NOQUOTA;
1334 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1335 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1336 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1337 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1338 LASSERT(opg->ops_transfer_pinned);
1340 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1341 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1342 opg->ops_transfer_pinned = 0;
1344 opg->ops_submit_time = ktime_set(0, 0);
1345 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1348 if (rc == 0 && srvlock) {
1349 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1350 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1351 size_t bytes = oap->oap_count;
1353 if (crt == CRT_READ)
1354 stats->os_lockless_reads += bytes;
1356 stats->os_lockless_writes += bytes;
1360 * This has to be the last operation with the page, as locks are
1361 * released in cl_page_completion() and nothing except for the
1362 * reference counter protects page from concurrent reclaim.
1364 lu_ref_del(&page->cp_reference, "transfer", page);
1366 cl_page_completion(env, page, crt, rc);
1367 cl_page_put(env, page);
1372 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do { \
1373 struct client_obd *__tmp = (cli); \
1374 CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1375 "dropped: %ld avail: %ld, dirty_grant: %ld, " \
1376 "reserved: %ld, flight: %d } lru {in list: %ld, " \
1377 "left: %ld, waiters: %d }" fmt "\n", \
1379 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1380 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1381 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1382 __tmp->cl_dirty_grant, \
1383 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1384 atomic_long_read(&__tmp->cl_lru_in_list), \
1385 atomic_long_read(&__tmp->cl_lru_busy), \
1386 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1389 /* caller must hold loi_list_lock */
1390 static void osc_consume_write_grant(struct client_obd *cli,
1391 struct brw_page *pga)
1393 assert_spin_locked(&cli->cl_loi_list_lock);
1394 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1395 cli->cl_dirty_pages++;
1396 pga->flag |= OBD_BRW_FROM_GRANT;
1397 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1398 PAGE_SIZE, pga, pga->pg);
1399 osc_update_next_shrink(cli);
1402 /* the companion to osc_consume_write_grant, called when a brw has completed.
1403 * must be called with the loi lock held. */
1404 static void osc_release_write_grant(struct client_obd *cli,
1405 struct brw_page *pga)
1409 assert_spin_locked(&cli->cl_loi_list_lock);
1410 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1415 pga->flag &= ~OBD_BRW_FROM_GRANT;
1416 atomic_long_dec(&obd_dirty_pages);
1417 cli->cl_dirty_pages--;
1422 * To avoid sleeping with object lock held, it's good for us allocate enough
1423 * grants before entering into critical section.
1425 * client_obd_list_lock held by caller
1427 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1431 if (cli->cl_avail_grant >= bytes) {
1432 cli->cl_avail_grant -= bytes;
1433 cli->cl_reserved_grant += bytes;
1439 static void __osc_unreserve_grant(struct client_obd *cli,
1440 unsigned int reserved, unsigned int unused)
1442 /* it's quite normal for us to get more grant than reserved.
1443 * Thinking about a case that two extents merged by adding a new
1444 * chunk, we can save one extent tax. If extent tax is greater than
1445 * one chunk, we can save more grant by adding a new chunk */
1446 cli->cl_reserved_grant -= reserved;
1447 if (unused > reserved) {
1448 cli->cl_avail_grant += reserved;
1449 cli->cl_lost_grant += unused - reserved;
1450 cli->cl_dirty_grant -= unused - reserved;
1452 cli->cl_avail_grant += unused;
1453 cli->cl_dirty_grant += reserved - unused;
1457 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1458 unsigned int reserved,
1459 unsigned int unused)
1461 __osc_unreserve_grant(cli, reserved, unused);
1463 osc_wake_cache_waiters(cli);
1466 static void osc_unreserve_grant(struct client_obd *cli,
1467 unsigned int reserved, unsigned int unused)
1469 spin_lock(&cli->cl_loi_list_lock);
1470 osc_unreserve_grant_nolock(cli, reserved, unused);
1471 spin_unlock(&cli->cl_loi_list_lock);
1475 * Free grant after IO is finished or canceled.
1477 * @lost_grant is used to remember how many grants we have allocated but not
1478 * used, we should return these grants to OST. There're two cases where grants
1481 * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1482 * written. In this case OST may use less chunks to serve this partial
1483 * write. OSTs don't actually know the page size on the client side. so
1484 * clients have to calculate lost grant by the blocksize on the OST.
1485 * See filter_grant_check() for details.
1487 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1488 unsigned int lost_grant, unsigned int dirty_grant)
1490 unsigned long grant;
1492 grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1494 spin_lock(&cli->cl_loi_list_lock);
1495 atomic_long_sub(nr_pages, &obd_dirty_pages);
1496 cli->cl_dirty_pages -= nr_pages;
1497 cli->cl_lost_grant += lost_grant;
1498 cli->cl_dirty_grant -= dirty_grant;
1499 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1500 /* borrow some grant from truncate to avoid the case that
1501 * truncate uses up all avail grant */
1502 cli->cl_lost_grant -= grant;
1503 cli->cl_avail_grant += grant;
1505 osc_wake_cache_waiters(cli);
1506 spin_unlock(&cli->cl_loi_list_lock);
1507 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1508 lost_grant, cli->cl_lost_grant,
1509 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1510 cli->cl_dirty_grant);
1514 * The companion to osc_enter_cache(), called when @oap is no longer part of
1515 * the dirty accounting due to error.
1517 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1519 spin_lock(&cli->cl_loi_list_lock);
1520 osc_release_write_grant(cli, &oap->oap_brw_page);
1521 spin_unlock(&cli->cl_loi_list_lock);
1525 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1528 static int osc_enter_cache_try(struct client_obd *cli,
1529 struct osc_async_page *oap,
1534 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1536 rc = osc_reserve_grant(cli, bytes);
1540 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1541 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1542 obd_max_dirty_pages) {
1543 osc_consume_write_grant(cli, &oap->oap_brw_page);
1547 atomic_long_dec(&obd_dirty_pages);
1549 __osc_unreserve_grant(cli, bytes, bytes);
1555 /* Following two inlines exist to pass code fragments
1556 * to wait_event_idle_exclusive_timeout_cmd(). Passing
1557 * code fragments as macro args can look confusing, so
1558 * we provide inlines to encapsulate them.
1560 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1561 struct client_obd *cli,
1562 struct osc_async_page *oap)
1564 spin_unlock(&cli->cl_loi_list_lock);
1565 osc_io_unplug_async(env, cli, NULL);
1567 "%s: sleeping for cache space for %p\n",
1568 cli_name(cli), oap);
1571 static inline void cli_lock_after_unplug(struct client_obd *cli)
1573 spin_lock(&cli->cl_loi_list_lock);
1576 * The main entry to reserve dirty page accounting. Usually the grant reserved
1577 * in this function will be freed in bulk in osc_free_grant() unless it fails
1578 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1580 * The process will be put into sleep if it's already run out of grant.
1582 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1583 struct osc_async_page *oap, int bytes)
1585 struct osc_object *osc = oap->oap_obj;
1586 struct lov_oinfo *loi = osc->oo_oinfo;
1589 bool entered = false;
1590 /* We cannot wait for a long time here since we are holding ldlm lock
1591 * across the actual IO. If no requests complete fast (e.g. due to
1592 * overloaded OST that takes a long time to process everything, we'd
1593 * get evicted if we wait for a normal obd_timeout or some such.
1594 * So we try to wait half the time it would take the client to be
1595 * evicted by server which is half obd_timeout when AT is off
1596 * or at least ldlm_enqueue_min with AT on.
1598 unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1599 ldlm_enqueue_min / 2);
1603 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1605 spin_lock(&cli->cl_loi_list_lock);
1607 /* force the caller to try sync io. this can jump the list
1608 * of queued writes and create a discontiguous rpc stream */
1609 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1610 cli->cl_dirty_max_pages == 0 ||
1611 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1612 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1613 GOTO(out, rc = -EDQUOT);
1617 * We can wait here for two reasons: too many dirty pages in cache, or
1618 * run out of grants. In both cases we should write dirty pages out.
1619 * Adding a cache waiter will trigger urgent write-out no matter what
1621 * The exiting condition (other than success) is no avail grants
1622 * and no dirty pages caching, that really means there is no space
1625 remain = wait_event_idle_exclusive_timeout_cmd(
1626 cli->cl_cache_waiters,
1627 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1628 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1630 cli_unlock_and_unplug(env, cli, oap),
1631 cli_lock_after_unplug(cli));
1634 if (remain == timeout)
1635 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1637 OSC_DUMP_GRANT(D_CACHE, cli,
1638 "finally got grant space\n");
1639 wake_up(&cli->cl_cache_waiters);
1641 } else if (remain == 0) {
1642 OSC_DUMP_GRANT(D_CACHE, cli,
1643 "timeout, fall back to sync i/o\n");
1644 osc_extent_tree_dump(D_CACHE, osc);
1645 /* fall back to synchronous I/O */
1647 OSC_DUMP_GRANT(D_CACHE, cli,
1648 "no grant space, fall back to sync i/o\n");
1649 wake_up_all(&cli->cl_cache_waiters);
1653 spin_unlock(&cli->cl_loi_list_lock);
1657 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1659 int hprpc = !!list_empty(&osc->oo_hp_exts);
1660 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1663 /* This maintains the lists of pending pages to read/write for a given object
1664 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1665 * to quickly find objects that are ready to send an RPC. */
1666 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1669 int invalid_import = 0;
1672 /* if we have an invalid import we want to drain the queued pages
1673 * by forcing them through rpcs that immediately fail and complete
1674 * the pages. recovery relies on this to empty the queued pages
1675 * before canceling the locks and evicting down the llite pages */
1676 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1679 if (cmd & OBD_BRW_WRITE) {
1680 if (atomic_read(&osc->oo_nr_writes) == 0)
1682 if (invalid_import) {
1683 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1686 if (!list_empty(&osc->oo_hp_exts)) {
1687 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1690 if (!list_empty(&osc->oo_urgent_exts)) {
1691 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1694 /* trigger a write rpc stream as long as there are dirtiers
1695 * waiting for space. as they're waiting, they're not going to
1696 * create more pages to coalesce with what's waiting..
1698 if (waitqueue_active(&cli->cl_cache_waiters)) {
1699 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1702 if (!list_empty(&osc->oo_full_exts)) {
1703 CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1707 if (atomic_read(&osc->oo_nr_reads) == 0)
1709 if (invalid_import) {
1710 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1713 /* all read are urgent. */
1714 if (!list_empty(&osc->oo_reading_exts))
1721 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1723 struct client_obd *cli = osc_cli(obj);
1724 if (cmd & OBD_BRW_WRITE) {
1725 atomic_add(delta, &obj->oo_nr_writes);
1726 atomic_add(delta, &cli->cl_pending_w_pages);
1727 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1729 atomic_add(delta, &obj->oo_nr_reads);
1730 atomic_add(delta, &cli->cl_pending_r_pages);
1731 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1733 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1736 static int osc_makes_hprpc(struct osc_object *obj)
1738 return !list_empty(&obj->oo_hp_exts);
1741 static void on_list(struct list_head *item, struct list_head *list,
1744 if (list_empty(item) && should_be_on)
1745 list_add_tail(item, list);
1746 else if (!list_empty(item) && !should_be_on)
1747 list_del_init(item);
1750 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1751 * can find pages to build into rpcs quickly */
1752 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1754 if (osc_makes_hprpc(osc)) {
1756 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1757 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1759 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1760 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1761 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1762 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1765 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1766 atomic_read(&osc->oo_nr_writes) > 0);
1768 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1769 atomic_read(&osc->oo_nr_reads) > 0);
1771 return osc_is_ready(osc);
1774 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1778 spin_lock(&cli->cl_loi_list_lock);
1779 is_ready = __osc_list_maint(cli, osc);
1780 spin_unlock(&cli->cl_loi_list_lock);
1785 /* this is trying to propogate async writeback errors back up to the
1786 * application. As an async write fails we record the error code for later if
1787 * the app does an fsync. As long as errors persist we force future rpcs to be
1788 * sync so that the app can get a sync error and break the cycle of queueing
1789 * pages for which writeback will fail. */
1790 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1797 ar->ar_force_sync = 1;
1798 ar->ar_min_xid = ptlrpc_sample_next_xid();
1803 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1804 ar->ar_force_sync = 0;
1807 /* this must be called holding the loi list lock to give coverage to exit_cache,
1808 * async_flag maintenance, and oap_request */
1809 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1810 struct osc_async_page *oap, int sent, int rc)
1812 struct osc_object *osc = oap->oap_obj;
1813 struct lov_oinfo *loi = osc->oo_oinfo;
1817 if (oap->oap_request != NULL) {
1818 xid = ptlrpc_req_xid(oap->oap_request);
1819 ptlrpc_req_finished(oap->oap_request);
1820 oap->oap_request = NULL;
1823 /* As the transfer for this page is being done, clear the flags */
1824 spin_lock(&oap->oap_lock);
1825 oap->oap_async_flags = 0;
1826 spin_unlock(&oap->oap_lock);
1828 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1829 spin_lock(&cli->cl_loi_list_lock);
1830 osc_process_ar(&cli->cl_ar, xid, rc);
1831 osc_process_ar(&loi->loi_ar, xid, rc);
1832 spin_unlock(&cli->cl_loi_list_lock);
1835 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1837 CERROR("completion on oap %p obj %p returns %d.\n",
1843 struct extent_rpc_data {
1844 struct list_head *erd_rpc_list;
1845 unsigned int erd_page_count;
1846 unsigned int erd_max_pages;
1847 unsigned int erd_max_chunks;
1848 unsigned int erd_max_extents;
1851 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1853 struct client_obd *cli = osc_cli(ext->oe_obj);
1854 unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1856 return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1860 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1862 if (ext->oe_no_merge || in_rpc->oe_no_merge)
1865 if (ext->oe_srvlock != in_rpc->oe_srvlock)
1868 if (ext->oe_ndelay != in_rpc->oe_ndelay)
1871 if (!ext->oe_grants != !in_rpc->oe_grants)
1874 if (ext->oe_dio != in_rpc->oe_dio)
1877 /* It's possible to have overlap on DIO */
1878 if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1885 * Try to add extent to one RPC. We need to think about the following things:
1886 * - # of pages must not be over max_pages_per_rpc
1887 * - extent must be compatible with previous ones
1889 static int try_to_add_extent_for_io(struct client_obd *cli,
1890 struct osc_extent *ext,
1891 struct extent_rpc_data *data)
1893 struct osc_extent *tmp;
1894 unsigned int chunk_count;
1897 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1899 OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1901 if (data->erd_max_extents == 0)
1904 chunk_count = osc_extent_chunks(ext);
1905 EASSERTF(data->erd_page_count != 0 ||
1906 chunk_count <= data->erd_max_chunks, ext,
1907 "The first extent to be fit in a RPC contains %u chunks, "
1908 "which is over the limit %u.\n", chunk_count,
1909 data->erd_max_chunks);
1910 if (chunk_count > data->erd_max_chunks)
1913 data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1914 EASSERTF(data->erd_page_count != 0 ||
1915 ext->oe_nr_pages <= data->erd_max_pages, ext,
1916 "The first extent to be fit in a RPC contains %u pages, "
1917 "which is over the limit %u.\n", ext->oe_nr_pages,
1918 data->erd_max_pages);
1919 if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1922 list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1923 EASSERT(tmp->oe_owner == current, tmp);
1925 if (!can_merge(ext, tmp))
1929 data->erd_max_extents--;
1930 data->erd_max_chunks -= chunk_count;
1931 data->erd_page_count += ext->oe_nr_pages;
1932 list_move_tail(&ext->oe_link, data->erd_rpc_list);
1933 ext->oe_owner = current;
1938 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1939 * get_write_extent() takes all appropriate extents in atomic.
1941 * The following policy is used to collect extents for IO:
1942 * 1. Add as many HP extents as possible;
1943 * 2. Add the first urgent extent in urgent extent list and take it out of
1945 * 3. Add subsequent extents of this urgent extent;
1946 * 4. If urgent list is not empty, goto 2;
1947 * 5. Traverse the extent tree from the 1st extent;
1948 * 6. Above steps exit if there is no space in this RPC.
1950 static unsigned int get_write_extents(struct osc_object *obj,
1951 struct list_head *rpclist)
1953 struct client_obd *cli = osc_cli(obj);
1954 struct osc_extent *ext;
1955 struct extent_rpc_data data = {
1956 .erd_rpc_list = rpclist,
1957 .erd_page_count = 0,
1958 .erd_max_pages = cli->cl_max_pages_per_rpc,
1959 .erd_max_chunks = osc_max_write_chunks(cli),
1960 .erd_max_extents = 256,
1963 assert_osc_object_is_locked(obj);
1964 while (!list_empty(&obj->oo_hp_exts)) {
1965 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1967 if (!try_to_add_extent_for_io(cli, ext, &data))
1968 return data.erd_page_count;
1969 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1971 if (data.erd_page_count == data.erd_max_pages)
1972 return data.erd_page_count;
1974 while (!list_empty(&obj->oo_urgent_exts)) {
1975 ext = list_entry(obj->oo_urgent_exts.next,
1976 struct osc_extent, oe_link);
1977 if (!try_to_add_extent_for_io(cli, ext, &data))
1978 return data.erd_page_count;
1980 if (data.erd_page_count == data.erd_max_pages)
1981 return data.erd_page_count;
1983 /* One key difference between full extents and other extents: full
1984 * extents can usually only be added if the rpclist was empty, so if we
1985 * can't add one, we continue on to trying to add normal extents. This
1986 * is so we don't miss adding extra extents to an RPC containing high
1987 * priority or urgent extents. */
1988 while (!list_empty(&obj->oo_full_exts)) {
1989 ext = list_entry(obj->oo_full_exts.next,
1990 struct osc_extent, oe_link);
1991 if (!try_to_add_extent_for_io(cli, ext, &data))
1994 if (data.erd_page_count == data.erd_max_pages)
1995 return data.erd_page_count;
1997 for (ext = first_extent(obj);
1999 ext = next_extent(ext)) {
2000 if ((ext->oe_state != OES_CACHE) ||
2001 /* this extent may be already in current rpclist */
2002 (!list_empty(&ext->oe_link) && ext->oe_owner))
2005 if (!try_to_add_extent_for_io(cli, ext, &data))
2006 return data.erd_page_count;
2008 return data.erd_page_count;
2012 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2013 struct osc_object *osc)
2017 struct osc_extent *ext;
2018 struct osc_extent *tmp;
2019 struct osc_extent *first = NULL;
2020 unsigned int page_count = 0;
2025 assert_osc_object_is_locked(osc);
2027 page_count = get_write_extents(osc, &rpclist);
2028 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2030 if (list_empty(&rpclist))
2033 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2035 list_for_each_entry(ext, &rpclist, oe_link) {
2036 LASSERT(ext->oe_state == OES_CACHE ||
2037 ext->oe_state == OES_LOCK_DONE);
2038 if (ext->oe_state == OES_CACHE)
2039 osc_extent_state_set(ext, OES_LOCKING);
2041 osc_extent_state_set(ext, OES_RPC);
2044 /* we're going to grab page lock, so release object lock because
2045 * lock order is page lock -> object lock. */
2046 osc_object_unlock(osc);
2048 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2049 if (ext->oe_state == OES_LOCKING) {
2050 rc = osc_extent_make_ready(env, ext);
2051 if (unlikely(rc < 0)) {
2052 list_del_init(&ext->oe_link);
2053 osc_extent_finish(env, ext, 0, rc);
2057 if (first == NULL) {
2059 srvlock = ext->oe_srvlock;
2061 LASSERT(srvlock == ext->oe_srvlock);
2065 if (!list_empty(&rpclist)) {
2066 LASSERT(page_count > 0);
2067 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2068 LASSERT(list_empty(&rpclist));
2071 osc_object_lock(osc);
2076 * prepare pages for ASYNC io and put pages in send queue.
2078 * \param cmd OBD_BRW_* macroses
2079 * \param lop pending pages
2081 * \return zero if no page added to send queue.
2082 * \return 1 if pages successfully added to send queue.
2083 * \return negative on errors.
2086 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2087 struct osc_object *osc)
2090 struct osc_extent *ext;
2091 struct osc_extent *next;
2093 struct extent_rpc_data data = {
2094 .erd_rpc_list = &rpclist,
2095 .erd_page_count = 0,
2096 .erd_max_pages = cli->cl_max_pages_per_rpc,
2097 .erd_max_chunks = UINT_MAX,
2098 .erd_max_extents = UINT_MAX,
2103 assert_osc_object_is_locked(osc);
2104 list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2105 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2106 if (!try_to_add_extent_for_io(cli, ext, &data))
2108 osc_extent_state_set(ext, OES_RPC);
2109 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2111 LASSERT(data.erd_page_count <= data.erd_max_pages);
2113 osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2115 if (!list_empty(&rpclist)) {
2116 osc_object_unlock(osc);
2118 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2119 LASSERT(list_empty(&rpclist));
2121 osc_object_lock(osc);
2126 #define list_to_obj(list, item) ({ \
2127 struct list_head *__tmp = (list)->next; \
2128 list_del_init(__tmp); \
2129 list_entry(__tmp, struct osc_object, oo_##item); \
2132 /* This is called by osc_check_rpcs() to find which objects have pages that
2133 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2134 static struct osc_object *osc_next_obj(struct client_obd *cli)
2138 /* First return objects that have blocked locks so that they
2139 * will be flushed quickly and other clients can get the lock,
2140 * then objects which have pages ready to be stuffed into RPCs */
2141 if (!list_empty(&cli->cl_loi_hp_ready_list))
2142 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2143 if (!list_empty(&cli->cl_loi_ready_list))
2144 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2146 /* then if we have cache waiters, return all objects with queued
2147 * writes. This is especially important when many small files
2148 * have filled up the cache and not been fired into rpcs because
2149 * they don't pass the nr_pending/object threshhold
2151 if (waitqueue_active(&cli->cl_cache_waiters) &&
2152 !list_empty(&cli->cl_loi_write_list))
2153 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2155 /* then return all queued objects when we have an invalid import
2156 * so that they get flushed */
2157 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2158 if (!list_empty(&cli->cl_loi_write_list))
2159 RETURN(list_to_obj(&cli->cl_loi_write_list,
2161 if (!list_empty(&cli->cl_loi_read_list))
2162 RETURN(list_to_obj(&cli->cl_loi_read_list,
2168 /* called with the loi list lock held */
2169 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2170 __must_hold(&cli->cl_loi_list_lock)
2172 struct osc_object *osc;
2176 while ((osc = osc_next_obj(cli)) != NULL) {
2177 struct cl_object *obj = osc2cl(osc);
2178 struct lu_ref_link link;
2180 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2182 /* even if we have reached our max in flight RPCs, we still
2183 * allow all high-priority RPCs through to prevent their
2184 * starvation and leading to server evicting us for not
2185 * writing out pages in a timely manner LU-13131 */
2186 if (osc_max_rpc_in_flight(cli, osc) &&
2187 list_empty(&osc->oo_hp_exts)) {
2188 __osc_list_maint(cli, osc);
2193 spin_unlock(&cli->cl_loi_list_lock);
2194 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2196 /* attempt some read/write balancing by alternating between
2197 * reads and writes in an object. The makes_rpc checks here
2198 * would be redundant if we were getting read/write work items
2199 * instead of objects. we don't want send_oap_rpc to drain a
2200 * partial read pending queue when we're given this object to
2201 * do io on writes while there are cache waiters */
2202 osc_object_lock(osc);
2203 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2204 rc = osc_send_write_rpc(env, cli, osc);
2206 CERROR("Write request failed with %d\n", rc);
2208 /* osc_send_write_rpc failed, mostly because of
2211 * It can't break here, because if:
2212 * - a page was submitted by osc_io_submit, so
2214 * - no request in flight
2215 * - no subsequent request
2216 * The system will be in live-lock state,
2217 * because there is no chance to call
2218 * osc_io_unplug() and osc_check_rpcs() any
2219 * more. pdflush can't help in this case,
2220 * because it might be blocked at grabbing
2221 * the page lock as we mentioned.
2223 * Anyway, continue to drain pages. */
2227 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2228 rc = osc_send_read_rpc(env, cli, osc);
2230 CERROR("Read request failed with %d\n", rc);
2232 osc_object_unlock(osc);
2234 osc_list_maint(cli, osc);
2235 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2236 cl_object_put(env, obj);
2238 spin_lock(&cli->cl_loi_list_lock);
2242 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2243 struct osc_object *osc, int async)
2247 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2251 spin_lock(&cli->cl_loi_list_lock);
2252 osc_check_rpcs(env, cli);
2253 spin_unlock(&cli->cl_loi_list_lock);
2255 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2256 LASSERT(cli->cl_writeback_work != NULL);
2257 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2261 EXPORT_SYMBOL(osc_io_unplug0);
2263 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2264 struct page *page, loff_t offset)
2266 struct obd_export *exp = osc_export(osc);
2267 struct osc_async_page *oap = &ops->ops_oap;
2271 return cfs_size_round(sizeof(*oap));
2273 oap->oap_magic = OAP_MAGIC;
2274 oap->oap_cli = &exp->exp_obd->u.cli;
2277 oap->oap_page = page;
2278 oap->oap_obj_off = offset;
2279 LASSERT(!(offset & ~PAGE_MASK));
2281 INIT_LIST_HEAD(&oap->oap_pending_item);
2282 INIT_LIST_HEAD(&oap->oap_rpc_item);
2284 spin_lock_init(&oap->oap_lock);
2285 CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2286 oap, page, oap->oap_obj_off);
2289 EXPORT_SYMBOL(osc_prep_async_page);
2291 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2292 struct osc_page *ops, cl_commit_cbt cb)
2294 struct osc_io *oio = osc_env_io(env);
2295 struct osc_extent *ext = NULL;
2296 struct osc_async_page *oap = &ops->ops_oap;
2297 struct client_obd *cli = oap->oap_cli;
2298 struct osc_object *osc = oap->oap_obj;
2299 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
2302 unsigned int grants = 0;
2303 u32 brw_flags = OBD_BRW_ASYNC;
2304 int cmd = OBD_BRW_WRITE;
2305 int need_release = 0;
2309 if (oap->oap_magic != OAP_MAGIC)
2312 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2315 if (!list_empty(&oap->oap_pending_item) ||
2316 !list_empty(&oap->oap_rpc_item))
2319 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2320 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2321 if (oio->oi_cap_sys_resource || io->ci_noquota) {
2322 brw_flags |= OBD_BRW_NOQUOTA;
2323 cmd |= OBD_BRW_NOQUOTA;
2326 /* check if the file's owner/group is over quota */
2327 if (!(cmd & OBD_BRW_NOQUOTA)) {
2328 struct cl_object *obj;
2329 struct cl_attr *attr;
2330 unsigned int qid[LL_MAXQUOTAS];
2332 obj = cl_object_top(&osc->oo_cl);
2333 attr = &osc_env_info(env)->oti_attr;
2335 cl_object_attr_lock(obj);
2336 rc = cl_object_attr_get(env, obj, attr);
2337 cl_object_attr_unlock(obj);
2339 qid[USRQUOTA] = attr->cat_uid;
2340 qid[GRPQUOTA] = attr->cat_gid;
2341 qid[PRJQUOTA] = attr->cat_projid;
2342 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2349 oap->oap_page_off = ops->ops_from;
2350 oap->oap_count = ops->ops_to - ops->ops_from;
2351 /* No need to hold a lock here,
2352 * since this page is not in any list yet. */
2353 oap->oap_async_flags = 0;
2354 oap->oap_brw_flags = brw_flags;
2356 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2357 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2359 index = osc_index(oap2osc(oap));
2361 /* Add this page into extent by the following steps:
2362 * 1. if there exists an active extent for this IO, mostly this page
2363 * can be added to the active extent and sometimes we need to
2364 * expand extent to accomodate this page;
2365 * 2. otherwise, a new extent will be allocated. */
2367 ext = oio->oi_active;
2368 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2369 /* one chunk plus extent overhead must be enough to write this
2371 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2372 if (ext->oe_end >= index)
2375 /* it doesn't need any grant to dirty this page */
2376 spin_lock(&cli->cl_loi_list_lock);
2377 rc = osc_enter_cache_try(cli, oap, grants);
2378 if (rc == 0) { /* try failed */
2381 } else if (ext->oe_end < index) {
2383 /* try to expand this extent */
2384 rc = osc_extent_expand(ext, index, &tmp);
2387 /* don't free reserved grant */
2389 OSC_EXTENT_DUMP(D_CACHE, ext,
2390 "expanded for %lu.\n", index);
2391 osc_unreserve_grant_nolock(cli, grants, tmp);
2395 spin_unlock(&cli->cl_loi_list_lock);
2397 } else if (ext != NULL) {
2398 /* index is located outside of active extent */
2402 osc_extent_release(env, ext);
2403 oio->oi_active = NULL;
2408 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2410 /* try to find new extent to cover this page */
2411 LASSERT(oio->oi_active == NULL);
2412 /* we may have allocated grant for this page if we failed
2413 * to expand the previous active extent. */
2414 LASSERT(ergo(grants > 0, grants >= tmp));
2418 /* We haven't allocated grant for this page, and we
2419 * must not hold a page lock while we do enter_cache,
2420 * so we must mark dirty & unlock any pages in the
2421 * write commit pagevec. */
2422 if (pagevec_count(pvec)) {
2424 pagevec_reinit(pvec);
2426 rc = osc_enter_cache(env, cli, oap, tmp);
2433 ext = osc_extent_find(env, osc, index, &tmp);
2435 LASSERT(tmp == grants);
2436 osc_exit_cache(cli, oap);
2440 oio->oi_active = ext;
2444 osc_unreserve_grant(cli, grants, tmp);
2447 LASSERT(ergo(rc == 0, ext != NULL));
2449 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2450 ext, "index = %lu.\n", index);
2451 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2453 osc_object_lock(osc);
2454 if (ext->oe_nr_pages == 0)
2455 ext->oe_srvlock = ops->ops_srvlock;
2457 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2459 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2460 osc_object_unlock(osc);
2462 if (!ext->oe_layout_version)
2463 ext->oe_layout_version = io->ci_layout_version;
2469 int osc_teardown_async_page(const struct lu_env *env,
2470 struct osc_object *obj, struct osc_page *ops)
2472 struct osc_async_page *oap = &ops->ops_oap;
2476 LASSERT(oap->oap_magic == OAP_MAGIC);
2478 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2479 oap, ops, osc_index(oap2osc(oap)));
2481 if (!list_empty(&oap->oap_rpc_item)) {
2482 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2484 } else if (!list_empty(&oap->oap_pending_item)) {
2485 struct osc_extent *ext = NULL;
2487 osc_object_lock(obj);
2488 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2489 osc_object_unlock(obj);
2490 /* only truncated pages are allowed to be taken out.
2491 * See osc_extent_truncate() and osc_cache_truncate_start()
2493 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2494 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2495 osc_index(oap2osc(oap)));
2499 osc_extent_put(env, ext);
2505 * This is called when a page is picked up by kernel to write out.
2507 * We should find out the corresponding extent and add the whole extent
2508 * into urgent list. The extent may be being truncated or used, handle it
2511 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2512 struct osc_page *ops)
2514 struct osc_extent *ext = NULL;
2515 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2516 struct cl_page *cp = ops->ops_cl.cpl_page;
2517 pgoff_t index = osc_index(ops);
2518 struct osc_async_page *oap = &ops->ops_oap;
2519 bool unplug = false;
2523 osc_object_lock(obj);
2524 ext = osc_extent_lookup(obj, index);
2526 osc_extent_tree_dump(D_ERROR, obj);
2527 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2530 switch (ext->oe_state) {
2533 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2537 /* If we know this extent is being written out, we should abort
2538 * so that the writer can make this page ready. Otherwise, there
2539 * exists a deadlock problem because other process can wait for
2540 * page writeback bit holding page lock; and meanwhile in
2541 * vvp_page_make_ready(), we need to grab page lock before
2542 * really sending the RPC. */
2544 /* race with truncate, page will be redirtied */
2546 /* The extent is active so we need to abort and let the caller
2547 * re-dirty the page. If we continued on here, and we were the
2548 * one making the extent active, we could deadlock waiting for
2549 * the page writeback to clear but it won't because the extent
2550 * is active and won't be written out. */
2551 GOTO(out, rc = -EAGAIN);
2556 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2560 spin_lock(&oap->oap_lock);
2561 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2562 spin_unlock(&oap->oap_lock);
2564 if (current->flags & PF_MEMALLOC)
2565 ext->oe_memalloc = 1;
2568 if (ext->oe_state == OES_CACHE) {
2569 OSC_EXTENT_DUMP(D_CACHE, ext,
2570 "flush page %p make it urgent.\n", oap);
2571 if (list_empty(&ext->oe_link))
2572 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2579 osc_object_unlock(obj);
2580 osc_extent_put(env, ext);
2582 osc_io_unplug_async(env, osc_cli(obj), obj);
2586 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2587 struct osc_object *obj, struct list_head *list,
2590 struct client_obd *cli = osc_cli(obj);
2591 struct osc_extent *ext;
2592 struct osc_async_page *oap;
2594 int mppr = cli->cl_max_pages_per_rpc;
2595 bool can_merge = true;
2596 pgoff_t start = CL_PAGE_EOF;
2600 list_for_each_entry(oap, list, oap_pending_item) {
2601 struct osc_page *opg = oap2osc_page(oap);
2602 pgoff_t index = osc_index(opg);
2609 mppr <<= (page_count > mppr);
2611 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2615 ext = osc_extent_alloc(obj);
2617 struct osc_async_page *tmp;
2619 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2620 list_del_init(&oap->oap_pending_item);
2621 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2626 ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2628 ext->oe_no_merge = !can_merge;
2630 ext->oe_start = start;
2631 ext->oe_end = ext->oe_max_end = end;
2633 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2634 ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2635 ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2636 ext->oe_nr_pages = page_count;
2637 ext->oe_mppr = mppr;
2638 list_splice_init(list, &ext->oe_pages);
2639 ext->oe_layout_version = io->ci_layout_version;
2641 osc_object_lock(obj);
2642 /* Reuse the initial refcount for RPC, don't drop it */
2643 osc_extent_state_set(ext, OES_LOCK_DONE);
2644 if (!ext->oe_rw) { /* write */
2645 if (!ext->oe_srvlock && !ext->oe_dio) {
2646 /* The most likely case here is from lack of grants
2647 * so we are either out of quota or out of space.
2648 * Since this means we are holding locks across
2649 * potentially multi-striped IO, we must send out
2650 * everything out instantly to avoid prolonged
2651 * waits resulting in lock eviction (likely since
2652 * the extended wait in osc_cache_enter() did not
2653 * yield any additional grant due to a timeout.
2656 list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2658 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2660 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2662 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2663 osc_update_pending(obj, OBD_BRW_READ, page_count);
2665 osc_object_unlock(obj);
2667 osc_io_unplug_async(env, cli, obj);
2672 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2674 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2675 __u64 size, struct osc_extent **extp)
2677 struct client_obd *cli = osc_cli(obj);
2678 struct osc_extent *ext;
2679 struct osc_extent *waiting = NULL;
2686 /* pages with index greater or equal to index will be truncated. */
2687 index = cl_index(osc2cl(obj), size);
2688 partial = size > cl_offset(osc2cl(obj), index);
2691 osc_object_lock(obj);
2692 ext = osc_extent_search(obj, index);
2694 ext = first_extent(obj);
2695 else if (ext->oe_end < index)
2696 ext = next_extent(ext);
2697 while (ext != NULL) {
2698 EASSERT(ext->oe_state != OES_TRUNC, ext);
2700 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2701 /* if ext is in urgent state, it means there must exist
2702 * a page already having been flushed by write_page().
2703 * We have to wait for this extent because we can't
2704 * truncate that page. */
2705 OSC_EXTENT_DUMP(D_CACHE, ext,
2706 "waiting for busy extent\n");
2707 waiting = osc_extent_get(ext);
2711 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2713 osc_extent_get(ext);
2714 if (ext->oe_state == OES_ACTIVE) {
2715 /* though we grab inode mutex for write path, but we
2716 * release it before releasing extent(in osc_io_end()),
2717 * so there is a race window that an extent is still
2718 * in OES_ACTIVE when truncate starts. */
2719 LASSERT(!ext->oe_trunc_pending);
2720 ext->oe_trunc_pending = 1;
2722 EASSERT(ext->oe_state == OES_CACHE, ext);
2723 osc_extent_state_set(ext, OES_TRUNC);
2724 osc_update_pending(obj, OBD_BRW_WRITE,
2727 /* This extent could be on the full extents list, that's OK */
2728 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2729 if (!list_empty(&ext->oe_link))
2730 list_move_tail(&ext->oe_link, &list);
2732 list_add_tail(&ext->oe_link, &list);
2734 ext = next_extent(ext);
2736 osc_object_unlock(obj);
2738 osc_list_maint(cli, obj);
2740 while (!list_empty(&list)) {
2743 ext = list_entry(list.next, struct osc_extent, oe_link);
2744 list_del_init(&ext->oe_link);
2746 /* extent may be in OES_ACTIVE state because inode mutex
2747 * is released before osc_io_end() in file write case */
2748 if (ext->oe_state != OES_TRUNC)
2749 osc_extent_wait(env, ext, OES_TRUNC);
2751 rc = osc_extent_truncate(ext, index, partial);
2756 OSC_EXTENT_DUMP(D_ERROR, ext,
2757 "truncate error %d\n", rc);
2758 } else if (ext->oe_nr_pages == 0) {
2759 osc_extent_remove(ext);
2761 /* this must be an overlapped extent which means only
2762 * part of pages in this extent have been truncated.
2764 EASSERTF(ext->oe_start <= index, ext,
2765 "trunc index = %lu/%d.\n", index, partial);
2766 /* fix index to skip this partially truncated extent */
2767 index = ext->oe_end + 1;
2770 /* we need to hold this extent in OES_TRUNC state so
2771 * that no writeback will happen. This is to avoid
2773 * Only partial truncate can reach here, if @size is
2774 * not zero, the caller should provide a valid @extp. */
2775 LASSERT(*extp == NULL);
2776 *extp = osc_extent_get(ext);
2777 OSC_EXTENT_DUMP(D_CACHE, ext,
2778 "trunc at %llu\n", size);
2780 osc_extent_put(env, ext);
2782 if (waiting != NULL) {
2785 /* ignore the result of osc_extent_wait the write initiator
2786 * should take care of it. */
2787 rc = osc_extent_wait(env, waiting, OES_INV);
2789 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2791 osc_extent_put(env, waiting);
2797 EXPORT_SYMBOL(osc_cache_truncate_start);
2800 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2802 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2805 struct osc_object *obj = ext->oe_obj;
2806 bool unplug = false;
2808 EASSERT(ext->oe_nr_pages > 0, ext);
2809 EASSERT(ext->oe_state == OES_TRUNC, ext);
2810 EASSERT(!ext->oe_urgent, ext);
2812 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2813 osc_object_lock(obj);
2814 osc_extent_state_set(ext, OES_CACHE);
2815 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2817 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2820 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2821 osc_object_unlock(obj);
2822 osc_extent_put(env, ext);
2825 osc_io_unplug_async(env, osc_cli(obj), obj);
2830 * Wait for extents in a specific range to be written out.
2831 * The caller must have called osc_cache_writeback_range() to issue IO
2832 * otherwise it will take a long time for this function to finish.
2834 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2835 * nobody else can dirty this range of file while we're waiting for
2836 * extents to be written.
2838 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2839 pgoff_t start, pgoff_t end)
2841 struct osc_extent *ext;
2842 pgoff_t index = start;
2847 osc_object_lock(obj);
2848 ext = osc_extent_search(obj, index);
2850 ext = first_extent(obj);
2851 else if (ext->oe_end < index)
2852 ext = next_extent(ext);
2853 while (ext != NULL) {
2856 if (ext->oe_start > end)
2859 if (!ext->oe_fsync_wait) {
2860 ext = next_extent(ext);
2864 EASSERT(ergo(ext->oe_state == OES_CACHE,
2865 ext->oe_hp || ext->oe_urgent), ext);
2866 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2867 !ext->oe_hp && ext->oe_urgent), ext);
2869 index = ext->oe_end + 1;
2870 osc_extent_get(ext);
2871 osc_object_unlock(obj);
2873 rc = osc_extent_wait(env, ext, OES_INV);
2876 osc_extent_put(env, ext);
2879 osc_object_unlock(obj);
2881 OSC_IO_DEBUG(obj, "sync file range.\n");
2884 EXPORT_SYMBOL(osc_cache_wait_range);
2887 * Called to write out a range of osc object.
2889 * @hp : should be set this is caused by lock cancel;
2890 * @discard: is set if dirty pages should be dropped - file will be deleted or
2891 * truncated, this implies there is no partially discarding extents.
2893 * Return how many pages will be issued, or error code if error occurred.
2895 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2896 pgoff_t start, pgoff_t end, int hp, int discard)
2898 struct osc_extent *ext;
2899 LIST_HEAD(discard_list);
2900 bool unplug = false;
2904 osc_object_lock(obj);
2905 ext = osc_extent_search(obj, start);
2907 ext = first_extent(obj);
2908 else if (ext->oe_end < start)
2909 ext = next_extent(ext);
2910 while (ext != NULL) {
2911 if (ext->oe_start > end)
2914 ext->oe_fsync_wait = 1;
2915 switch (ext->oe_state) {
2917 result += ext->oe_nr_pages;
2919 struct list_head *list = NULL;
2921 EASSERT(!ext->oe_hp, ext);
2923 list = &obj->oo_hp_exts;
2924 } else if (!ext->oe_urgent && !ext->oe_hp) {
2926 list = &obj->oo_urgent_exts;
2929 list_move_tail(&ext->oe_link, list);
2932 struct client_obd *cli = osc_cli(obj);
2933 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2934 pgoff_t align_by = (1 << pcc_bits);
2935 pgoff_t a_start = round_down(start, align_by);
2936 pgoff_t a_end = round_up(end, align_by);
2940 a_end = CL_PAGE_EOF;
2941 /* the only discarder is lock cancelling, so
2942 * [start, end], aligned by chunk size, must
2943 * contain this extent */
2944 LASSERTF(ext->oe_start >= a_start &&
2945 ext->oe_end <= a_end,
2946 "ext [%lu, %lu] reg [%lu, %lu] "
2947 "orig [%lu %lu] align %lu bits "
2948 "%d\n", ext->oe_start, ext->oe_end,
2949 a_start, a_end, start, end,
2950 align_by, pcc_bits);
2951 osc_extent_state_set(ext, OES_LOCKING);
2952 ext->oe_owner = current;
2953 list_move_tail(&ext->oe_link,
2955 osc_update_pending(obj, OBD_BRW_WRITE,
2960 /* It's pretty bad to wait for ACTIVE extents, because
2961 * we don't know how long we will wait for it to be
2962 * flushed since it may be blocked at awaiting more
2963 * grants. We do this for the correctness of fsync. */
2964 LASSERT(hp == 0 && discard == 0);
2968 /* this extent is being truncated, can't do anything
2969 * for it now. it will be set to urgent after truncate
2970 * is finished in osc_cache_truncate_end(). */
2974 ext = next_extent(ext);
2976 osc_object_unlock(obj);
2978 LASSERT(ergo(!discard, list_empty(&discard_list)));
2979 if (!list_empty(&discard_list)) {
2980 struct osc_extent *tmp;
2983 osc_list_maint(osc_cli(obj), obj);
2984 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2985 list_del_init(&ext->oe_link);
2986 EASSERT(ext->oe_state == OES_LOCKING, ext);
2988 /* Discard caching pages. We don't actually write this
2989 * extent out but we complete it as if we did. */
2990 rc = osc_extent_make_ready(env, ext);
2991 if (unlikely(rc < 0)) {
2992 OSC_EXTENT_DUMP(D_ERROR, ext,
2993 "make_ready returned %d\n", rc);
2998 /* finish the extent as if the pages were sent */
2999 osc_extent_finish(env, ext, 0, 0);
3004 osc_io_unplug(env, osc_cli(obj), obj);
3006 if (hp || discard) {
3008 rc = osc_cache_wait_range(env, obj, start, end);
3009 if (result >= 0 && rc < 0)
3013 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3016 EXPORT_SYMBOL(osc_cache_writeback_range);
3019 * Returns a list of pages by a given [start, end] of \a obj.
3021 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3022 * crucial in the face of [offset, EOF] locks.
3024 * Return at least one page in @queue unless there is no covered page.
3026 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3027 struct osc_object *osc, pgoff_t start, pgoff_t end,
3028 osc_page_gang_cbt cb, void *cbdata)
3030 struct osc_page *ops;
3031 struct pagevec *pagevec;
3038 bool tree_lock = true;
3042 pvec = osc_env_info(env)->oti_pvec;
3043 pagevec = &osc_env_info(env)->oti_pagevec;
3044 ll_pagevec_init(pagevec, 0);
3045 spin_lock(&osc->oo_tree_lock);
3046 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3047 idx, OTI_PVEC_SIZE)) > 0) {
3048 struct cl_page *page;
3049 bool end_of_region = false;
3051 for (i = 0, j = 0; i < nr; ++i) {
3055 idx = osc_index(ops);
3057 end_of_region = true;
3061 page = ops->ops_cl.cpl_page;
3062 LASSERT(page->cp_type == CPT_CACHEABLE);
3063 if (page->cp_state == CPS_FREEING)
3067 lu_ref_add_atomic(&page->cp_reference,
3068 "gang_lookup", current);
3074 * Here a delicate locking dance is performed. Current thread
3075 * holds a reference to a page, but has to own it before it
3076 * can be placed into queue. Owning implies waiting, so
3077 * radix-tree lock is to be released. After a wait one has to
3078 * check that pages weren't truncated (cl_page_own() returns
3079 * error in the latter case).
3081 spin_unlock(&osc->oo_tree_lock);
3084 for (i = 0; i < j; ++i) {
3087 res = (*cb)(env, io, ops, cbdata);
3089 page = ops->ops_cl.cpl_page;
3090 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3091 cl_pagevec_put(env, page, pagevec);
3093 pagevec_release(pagevec);
3095 if (nr < OTI_PVEC_SIZE || end_of_region)
3103 spin_lock(&osc->oo_tree_lock);
3107 spin_unlock(&osc->oo_tree_lock);
3110 EXPORT_SYMBOL(osc_page_gang_lookup);
3113 * Check if page @page is covered by an extra lock or discard it.
3115 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3116 struct osc_page *ops, void *cbdata)
3118 struct osc_thread_info *info = osc_env_info(env);
3119 struct osc_object *osc = cbdata;
3122 index = osc_index(ops);
3123 if (index >= info->oti_fn_index) {
3124 struct ldlm_lock *tmp;
3125 struct cl_page *page = ops->ops_cl.cpl_page;
3127 /* refresh non-overlapped index */
3128 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3129 OSC_DAP_FL_TEST_LOCK);
3131 __u64 end = tmp->l_policy_data.l_extent.end;
3132 /* Cache the first-non-overlapped index so as to skip
3133 * all pages within [index, oti_fn_index). This is safe
3134 * because if tmp lock is canceled, it will discard
3136 info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3137 if (end == OBD_OBJECT_EOF)
3138 info->oti_fn_index = CL_PAGE_EOF;
3140 } else if (cl_page_own(env, io, page) == 0) {
3141 /* discard the page */
3142 cl_page_discard(env, io, page);
3143 cl_page_disown(env, io, page);
3145 LASSERT(page->cp_state == CPS_FREEING);
3149 info->oti_next_index = index + 1;
3153 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3154 struct osc_page *ops, void *cbdata)
3156 struct osc_thread_info *info = osc_env_info(env);
3157 struct cl_page *page = ops->ops_cl.cpl_page;
3159 /* page is top page. */
3160 info->oti_next_index = osc_index(ops) + 1;
3161 if (cl_page_own(env, io, page) == 0) {
3162 if (!ergo(page->cp_type == CPT_CACHEABLE,
3163 !PageDirty(cl_page_vmpage(page))))
3164 CL_PAGE_DEBUG(D_ERROR, env, page,
3165 "discard dirty page?\n");
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);
3176 EXPORT_SYMBOL(osc_discard_cb);
3179 * Discard pages protected by the given lock. This function traverses radix
3180 * tree to find all covering pages and discard them. If a page is being covered
3181 * by other locks, it should remain in cache.
3183 * If error happens on any step, the process continues anyway (the reasoning
3184 * behind this being that lock cancellation cannot be delayed indefinitely).
3186 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3187 pgoff_t start, pgoff_t end, bool discard)
3189 struct osc_thread_info *info = osc_env_info(env);
3190 struct cl_io *io = osc_env_thread_io(env);
3191 osc_page_gang_cbt cb;
3196 io->ci_obj = cl_object_top(osc2cl(osc));
3197 io->ci_ignore_layout = 1;
3198 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3202 cb = discard ? osc_discard_cb : check_and_discard_cb;
3203 info->oti_fn_index = info->oti_next_index = start;
3205 osc_page_gang_lookup(env, io, osc,
3206 info->oti_next_index, end, cb, osc);
3208 cl_io_fini(env, io);