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)
541 * In the rare case max_pages_per_rpc (mppr) is changed, don't
542 * merge extents until after old ones have been sent, or the
543 * "extents are aligned to RPCs" checks are unhappy.
545 if (cur->oe_mppr != victim->oe_mppr)
548 LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
549 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
550 chunk_start = cur->oe_start >> ppc_bits;
551 chunk_end = cur->oe_end >> ppc_bits;
552 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
553 chunk_end + 1 != victim->oe_start >> ppc_bits)
556 /* overall extent size should not exceed the max supported limit
557 * reported by the server */
558 if (cur->oe_end - cur->oe_start + 1 +
559 victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
562 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
564 cur->oe_start = min(cur->oe_start, victim->oe_start);
565 cur->oe_end = max(cur->oe_end, victim->oe_end);
566 /* per-extent tax should be accounted only once for the whole extent */
567 cur->oe_grants += victim->oe_grants - cli->cl_grant_extent_tax;
568 cur->oe_nr_pages += victim->oe_nr_pages;
569 /* only the following bits are needed to merge */
570 cur->oe_urgent |= victim->oe_urgent;
571 cur->oe_memalloc |= victim->oe_memalloc;
572 list_splice_init(&victim->oe_pages, &cur->oe_pages);
573 victim->oe_nr_pages = 0;
575 osc_extent_get(victim);
576 __osc_extent_remove(victim);
577 osc_extent_put(env, victim);
579 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
584 * Drop user count of osc_extent, and unplug IO asynchronously.
586 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
588 struct osc_object *obj = ext->oe_obj;
589 struct client_obd *cli = osc_cli(obj);
593 LASSERT(atomic_read(&ext->oe_users) > 0);
594 LASSERT(sanity_check(ext) == 0);
595 LASSERT(ext->oe_grants > 0);
597 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
598 LASSERT(ext->oe_state == OES_ACTIVE);
599 if (ext->oe_trunc_pending) {
600 /* a truncate process is waiting for this extent.
601 * This may happen due to a race, check
602 * osc_cache_truncate_start(). */
603 osc_extent_state_set(ext, OES_TRUNC);
604 ext->oe_trunc_pending = 0;
605 osc_object_unlock(obj);
609 osc_extent_state_set(ext, OES_CACHE);
610 osc_update_pending(obj, OBD_BRW_WRITE,
613 /* try to merge the previous and next extent. */
614 if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
615 grant += cli->cl_grant_extent_tax;
616 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
617 grant += cli->cl_grant_extent_tax;
620 list_move_tail(&ext->oe_link,
622 else if (ext->oe_urgent)
623 list_move_tail(&ext->oe_link,
624 &obj->oo_urgent_exts);
625 else if (ext->oe_nr_pages == ext->oe_mppr) {
626 list_move_tail(&ext->oe_link,
629 osc_object_unlock(obj);
631 osc_unreserve_grant(cli, 0, grant);
634 osc_io_unplug_async(env, cli, obj);
636 osc_extent_put(env, ext);
641 * Find or create an extent which includes @index, core function to manage
644 static struct osc_extent *osc_extent_find(const struct lu_env *env,
645 struct osc_object *obj, pgoff_t index,
646 unsigned int *grants)
648 struct client_obd *cli = osc_cli(obj);
649 struct osc_lock *olck;
650 struct cl_lock_descr *descr;
651 struct osc_extent *cur;
652 struct osc_extent *ext;
653 struct osc_extent *conflict = NULL;
654 struct osc_extent *found = NULL;
657 unsigned int max_pages; /* max_pages_per_rpc */
658 unsigned int chunksize;
659 int ppc_bits; /* pages per chunk bits */
664 cur = osc_extent_alloc(obj);
666 RETURN(ERR_PTR(-ENOMEM));
668 olck = osc_env_io(env)->oi_write_osclock;
669 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
670 LASSERT(olck->ols_state == OLS_GRANTED);
672 descr = &olck->ols_cl.cls_lock->cll_descr;
673 LASSERT(descr->cld_mode >= CLM_WRITE);
675 LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
676 "chunkbits: %u\n", cli->cl_chunkbits);
677 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
678 chunk_mask = ~((1 << ppc_bits) - 1);
679 chunksize = 1 << cli->cl_chunkbits;
680 chunk = index >> ppc_bits;
682 /* align end to RPC edge. */
683 max_pages = cli->cl_max_pages_per_rpc;
684 if ((max_pages & ~chunk_mask) != 0) {
685 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
686 max_pages, cli->cl_chunkbits, chunk_mask);
687 RETURN(ERR_PTR(-EINVAL));
689 max_end = index - (index % max_pages) + max_pages - 1;
690 max_end = min_t(pgoff_t, max_end, descr->cld_end);
692 /* initialize new extent by parameters so far */
693 cur->oe_max_end = max_end;
694 cur->oe_start = index & chunk_mask;
695 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
696 if (cur->oe_start < descr->cld_start)
697 cur->oe_start = descr->cld_start;
698 if (cur->oe_end > max_end)
699 cur->oe_end = max_end;
700 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
701 cur->oe_mppr = max_pages;
702 if (olck->ols_dlmlock != NULL) {
703 LASSERT(olck->ols_hold);
704 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
705 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
708 /* grants has been allocated by caller */
709 LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
710 "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
711 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
715 osc_object_lock(obj);
716 ext = osc_extent_search(obj, cur->oe_start);
718 ext = first_extent(obj);
719 for (; ext; ext = next_extent(ext)) {
720 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
721 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
723 LASSERT(sanity_check_nolock(ext) == 0);
724 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
727 /* if covering by different locks, no chance to match */
728 if (olck->ols_dlmlock != ext->oe_dlmlock) {
729 EASSERTF(!overlapped(ext, cur), ext,
730 EXTSTR"\n", EXTPARA(cur));
735 /* discontiguous chunks? */
736 if (chunk + 1 < ext_chk_start)
739 /* ok, from now on, ext and cur have these attrs:
740 * 1. covered by the same lock
741 * 2. contiguous at chunk level or overlapping. */
743 if (overlapped(ext, cur)) {
744 /* cur is the minimum unit, so overlapping means
746 EASSERTF((ext->oe_start <= cur->oe_start &&
747 ext->oe_end >= cur->oe_end),
748 ext, EXTSTR"\n", EXTPARA(cur));
750 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
751 /* for simplicity, we wait for this extent to
752 * finish before going forward. */
753 conflict = osc_extent_get(ext);
757 found = osc_extent_hold(ext);
761 /* non-overlapped extent */
762 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
763 /* we can't do anything for a non OES_CACHE extent, or
764 * if there is someone waiting for this extent to be
765 * flushed, try next one. */
768 if (osc_extent_merge(env, ext, cur) == 0) {
769 LASSERT(*grants >= chunksize);
770 *grants -= chunksize;
771 found = osc_extent_hold(ext);
774 * Try to merge with the next one too because we
775 * might have just filled in a gap.
777 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
778 /* we can save extent tax from next extent */
779 *grants += cli->cl_grant_extent_tax;
785 osc_extent_tree_dump(D_CACHE, obj);
787 LASSERT(conflict == NULL);
788 if (!IS_ERR(found)) {
789 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
790 OSC_EXTENT_DUMP(D_CACHE, found,
791 "found caching ext for %lu.\n", index);
793 } else if (conflict == NULL) {
794 /* create a new extent */
795 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
796 LASSERT(*grants >= cur->oe_grants);
797 *grants -= cur->oe_grants;
799 cur->oe_state = OES_CACHE;
800 found = osc_extent_hold(cur);
801 osc_extent_insert(obj, cur);
802 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
803 index, descr->cld_end);
805 osc_object_unlock(obj);
807 if (conflict != NULL) {
808 LASSERT(found == NULL);
810 /* waiting for IO to finish. Please notice that it's impossible
811 * to be an OES_TRUNC extent. */
812 rc = osc_extent_wait(env, conflict, OES_INV);
813 osc_extent_put(env, conflict);
816 GOTO(out, found = ERR_PTR(rc));
823 osc_extent_put(env, cur);
828 * Called when IO is finished to an extent.
830 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
833 struct client_obd *cli = osc_cli(ext->oe_obj);
834 struct osc_async_page *oap;
835 struct osc_async_page *tmp;
836 int nr_pages = ext->oe_nr_pages;
838 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
843 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
845 ext->oe_rc = rc ?: ext->oe_nr_pages;
846 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
848 osc_lru_add_batch(cli, &ext->oe_pages);
849 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
851 list_del_init(&oap->oap_rpc_item);
852 list_del_init(&oap->oap_pending_item);
853 if (last_off <= oap->oap_obj_off) {
854 last_off = oap->oap_obj_off;
855 last_count = oap->oap_count;
859 osc_ap_completion(env, cli, oap, sent, rc);
861 EASSERT(ext->oe_nr_pages == 0, ext);
864 lost_grant = ext->oe_grants;
865 } else if (blocksize < PAGE_SIZE &&
866 last_count != PAGE_SIZE) {
867 /* For short writes we shouldn't count parts of pages that
868 * span a whole chunk on the OST side, or our accounting goes
869 * wrong. Should match the code in filter_grant_check. */
870 int offset = last_off & ~PAGE_MASK;
871 int count = last_count + (offset & (blocksize - 1));
872 int end = (offset + last_count) & (blocksize - 1);
874 count += blocksize - end;
876 lost_grant = PAGE_SIZE - count;
878 if (ext->oe_grants > 0)
879 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
881 osc_extent_remove(ext);
882 /* put the refcount for RPC */
883 osc_extent_put(env, ext);
888 * Wait for the extent's state to become @state.
890 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
891 enum osc_extent_state state)
893 struct osc_object *obj = ext->oe_obj;
897 osc_object_lock(obj);
898 LASSERT(sanity_check_nolock(ext) == 0);
899 /* `Kick' this extent only if the caller is waiting for it to be
901 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
902 if (ext->oe_state == OES_ACTIVE) {
904 } else if (ext->oe_state == OES_CACHE) {
906 osc_extent_hold(ext);
910 osc_object_unlock(obj);
912 osc_extent_release(env, ext);
914 /* wait for the extent until its state becomes @state */
915 rc = wait_event_idle_timeout(ext->oe_waitq,
916 smp_load_acquire(&ext->oe_state) == state,
917 cfs_time_seconds(600));
919 OSC_EXTENT_DUMP(D_ERROR, ext,
920 "%s: wait ext to %u timedout, recovery in progress?\n",
921 cli_name(osc_cli(obj)), state);
923 wait_event_idle(ext->oe_waitq,
924 smp_load_acquire(&ext->oe_state) == state);
934 * Discard pages with index greater than @size. If @ext is overlapped with
935 * @size, then partial truncate happens.
937 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
942 struct osc_object *obj = ext->oe_obj;
943 struct client_obd *cli = osc_cli(obj);
944 struct osc_async_page *oap;
945 struct osc_async_page *tmp;
946 struct pagevec *pvec;
947 int pages_in_chunk = 0;
948 int ppc_bits = cli->cl_chunkbits -
950 __u64 trunc_chunk = trunc_index >> ppc_bits;
957 LASSERT(sanity_check(ext) == 0);
958 LASSERT(ext->oe_state == OES_TRUNC);
959 LASSERT(!ext->oe_urgent);
961 /* Request new lu_env.
962 * We can't use that env from osc_cache_truncate_start() because
963 * it's from lov_io_sub and not fully initialized. */
964 env = cl_env_get(&refcheck);
966 RETURN(PTR_ERR(env));
968 io = osc_env_thread_io(env);
969 io->ci_obj = cl_object_top(osc2cl(obj));
970 io->ci_ignore_layout = 1;
971 pvec = &osc_env_info(env)->oti_pagevec;
972 ll_pagevec_init(pvec, 0);
973 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
977 /* discard all pages with index greater than trunc_index */
978 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
980 pgoff_t index = osc_index(oap2osc(oap));
981 struct cl_page *page = oap2cl_page(oap);
983 LASSERT(list_empty(&oap->oap_rpc_item));
985 /* only discard the pages with their index greater than
986 * trunc_index, and ... */
987 if (index < trunc_index ||
988 (index == trunc_index && partial)) {
989 /* accounting how many pages remaining in the chunk
990 * so that we can calculate grants correctly. */
991 if (index >> ppc_bits == trunc_chunk)
996 list_del_init(&oap->oap_pending_item);
999 lu_ref_add(&page->cp_reference, "truncate", current);
1001 if (cl_page_own(env, io, page) == 0) {
1002 cl_page_discard(env, io, page);
1003 cl_page_disown(env, io, page);
1005 LASSERT(page->cp_state == CPS_FREEING);
1009 lu_ref_del(&page->cp_reference, "truncate", current);
1010 cl_pagevec_put(env, page, pvec);
1015 pagevec_release(pvec);
1017 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1018 ext->oe_nr_pages == 0),
1019 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1021 osc_object_lock(obj);
1022 if (ext->oe_nr_pages == 0) {
1023 LASSERT(pages_in_chunk == 0);
1024 grants = ext->oe_grants;
1026 } else { /* calculate how many grants we can free */
1027 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1031 /* if there is no pages in this chunk, we can also free grants
1032 * for the last chunk */
1033 if (pages_in_chunk == 0) {
1034 /* if this is the 1st chunk and no pages in this chunk,
1035 * ext->oe_nr_pages must be zero, so we should be in
1036 * the other if-clause. */
1037 LASSERT(trunc_chunk > 0);
1042 /* this is what we can free from this extent */
1043 grants = chunks << cli->cl_chunkbits;
1044 ext->oe_grants -= grants;
1045 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1046 ext->oe_end = min(last_index, ext->oe_max_end);
1047 LASSERT(ext->oe_end >= ext->oe_start);
1048 LASSERT(ext->oe_grants > 0);
1050 osc_object_unlock(obj);
1052 if (grants > 0 || nr_pages > 0)
1053 osc_free_grant(cli, nr_pages, grants, grants);
1056 cl_io_fini(env, io);
1057 cl_env_put(env, &refcheck);
1062 * This function is used to make the extent prepared for transfer.
1063 * A race with flusing page - ll_writepage() has to be handled cautiously.
1065 static int osc_extent_make_ready(const struct lu_env *env,
1066 struct osc_extent *ext)
1068 struct osc_async_page *oap;
1069 struct osc_async_page *last = NULL;
1070 struct osc_object *obj = ext->oe_obj;
1071 unsigned int page_count = 0;
1075 /* we're going to grab page lock, so object lock must not be taken. */
1076 LASSERT(sanity_check(ext) == 0);
1077 /* in locking state, any process should not touch this extent. */
1078 EASSERT(ext->oe_state == OES_LOCKING, ext);
1079 EASSERT(ext->oe_owner != NULL, ext);
1081 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1083 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1085 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1088 /* checking ASYNC_READY is race safe */
1089 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1092 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1095 spin_lock(&oap->oap_lock);
1096 oap->oap_async_flags |= ASYNC_READY;
1097 spin_unlock(&oap->oap_lock);
1100 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1103 LASSERTF(0, "unknown return code: %d\n", rc);
1107 LASSERT(page_count == ext->oe_nr_pages);
1108 LASSERT(last != NULL);
1109 /* the last page is the only one we need to refresh its count by
1110 * the size of file. */
1111 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1112 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1113 LASSERT(last_oap_count > 0);
1114 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1115 last->oap_count = last_oap_count;
1116 spin_lock(&last->oap_lock);
1117 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1118 spin_unlock(&last->oap_lock);
1121 /* for the rest of pages, we don't need to call osf_refresh_count()
1122 * because it's known they are not the last page */
1123 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1124 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1125 oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1126 spin_lock(&oap->oap_lock);
1127 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1128 spin_unlock(&oap->oap_lock);
1132 osc_object_lock(obj);
1133 osc_extent_state_set(ext, OES_RPC);
1134 osc_object_unlock(obj);
1135 /* get a refcount for RPC. */
1136 osc_extent_get(ext);
1142 * Quick and simple version of osc_extent_find(). This function is frequently
1143 * called to expand the extent for the same IO. To expand the extent, the
1144 * page index must be in the same or next chunk of ext->oe_end.
1146 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1147 unsigned int *grants)
1149 struct osc_object *obj = ext->oe_obj;
1150 struct client_obd *cli = osc_cli(obj);
1151 struct osc_extent *next;
1152 int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1153 pgoff_t chunk = index >> ppc_bits;
1156 unsigned int chunksize = 1 << cli->cl_chunkbits;
1160 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1161 osc_object_lock(obj);
1162 LASSERT(sanity_check_nolock(ext) == 0);
1163 end_chunk = ext->oe_end >> ppc_bits;
1164 if (chunk > end_chunk + 1)
1165 GOTO(out, rc = -ERANGE);
1167 if (end_chunk >= chunk)
1170 LASSERT(end_chunk + 1 == chunk);
1172 /* try to expand this extent to cover @index */
1173 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1175 /* don't go over the maximum extent size reported by server */
1176 if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1177 GOTO(out, rc = -ERANGE);
1179 next = next_extent(ext);
1180 if (next != NULL && next->oe_start <= end_index)
1181 /* complex mode - overlapped with the next extent,
1182 * this case will be handled by osc_extent_find() */
1183 GOTO(out, rc = -EAGAIN);
1185 ext->oe_end = end_index;
1186 ext->oe_grants += chunksize;
1187 LASSERT(*grants >= chunksize);
1188 *grants -= chunksize;
1189 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1190 "overlapped after expanding for %lu.\n", index);
1194 osc_object_unlock(obj);
1198 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1199 const char *func, int line)
1201 struct osc_extent *ext;
1204 if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1207 CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1208 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1210 /* osc_object_lock(obj); */
1212 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1213 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1216 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1217 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1220 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1221 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1224 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1225 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1226 /* osc_object_unlock(obj); */
1229 /* ------------------ osc extent end ------------------ */
1231 static inline int osc_is_ready(struct osc_object *osc)
1233 return !list_empty(&osc->oo_ready_item) ||
1234 !list_empty(&osc->oo_hp_ready_item);
1237 #define OSC_IO_DEBUG(OSC, STR, args...) \
1238 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1239 (OSC), osc_is_ready(OSC), \
1240 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1241 list_empty_marker(&(OSC)->oo_ready_item), \
1242 atomic_read(&(OSC)->oo_nr_writes), \
1243 list_empty_marker(&(OSC)->oo_hp_exts), \
1244 list_empty_marker(&(OSC)->oo_urgent_exts), \
1245 atomic_read(&(OSC)->oo_nr_reads), \
1246 list_empty_marker(&(OSC)->oo_reading_exts), \
1249 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1252 struct osc_page *opg = oap2osc_page(oap);
1253 struct cl_page *page = oap2cl_page(oap);
1256 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1259 result = cl_page_make_ready(env, page, CRT_WRITE);
1261 opg->ops_submit_time = ktime_get();
1265 static int osc_refresh_count(const struct lu_env *env,
1266 struct osc_async_page *oap, int cmd)
1268 struct osc_page *opg = oap2osc_page(oap);
1269 pgoff_t index = osc_index(oap2osc(oap));
1270 struct cl_object *obj;
1271 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1275 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1276 LASSERT(!(cmd & OBD_BRW_READ));
1277 LASSERT(opg != NULL);
1278 obj = opg->ops_cl.cpl_obj;
1280 cl_object_attr_lock(obj);
1281 result = cl_object_attr_get(env, obj, attr);
1282 cl_object_attr_unlock(obj);
1285 kms = attr->cat_kms;
1286 if (cl_offset(obj, index) >= kms)
1287 /* catch race with truncate */
1289 else if (cl_offset(obj, index + 1) > kms)
1290 /* catch sub-page write at end of file */
1291 return kms & ~PAGE_MASK;
1296 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1299 struct osc_page *opg = oap2osc_page(oap);
1300 struct cl_page *page = oap2cl_page(oap);
1301 enum cl_req_type crt;
1306 cmd &= ~OBD_BRW_NOQUOTA;
1307 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1308 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1309 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1310 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1311 LASSERT(opg->ops_transfer_pinned);
1313 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1314 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1315 opg->ops_transfer_pinned = 0;
1317 opg->ops_submit_time = ktime_set(0, 0);
1318 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1321 if (rc == 0 && srvlock) {
1322 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1323 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1324 size_t bytes = oap->oap_count;
1326 if (crt == CRT_READ)
1327 stats->os_lockless_reads += bytes;
1329 stats->os_lockless_writes += bytes;
1333 * This has to be the last operation with the page, as locks are
1334 * released in cl_page_completion() and nothing except for the
1335 * reference counter protects page from concurrent reclaim.
1337 lu_ref_del(&page->cp_reference, "transfer", page);
1339 cl_page_completion(env, page, crt, rc);
1340 cl_page_put(env, page);
1345 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do { \
1346 struct client_obd *__tmp = (cli); \
1347 CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1348 "dropped: %ld avail: %ld, dirty_grant: %ld, " \
1349 "reserved: %ld, flight: %d } lru {in list: %ld, " \
1350 "left: %ld, waiters: %d }" fmt "\n", \
1352 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1353 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1354 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1355 __tmp->cl_dirty_grant, \
1356 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1357 atomic_long_read(&__tmp->cl_lru_in_list), \
1358 atomic_long_read(&__tmp->cl_lru_busy), \
1359 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1362 /* caller must hold loi_list_lock */
1363 static void osc_consume_write_grant(struct client_obd *cli,
1364 struct brw_page *pga)
1366 assert_spin_locked(&cli->cl_loi_list_lock);
1367 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1368 cli->cl_dirty_pages++;
1369 pga->flag |= OBD_BRW_FROM_GRANT;
1370 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1371 PAGE_SIZE, pga, pga->pg);
1372 osc_update_next_shrink(cli);
1375 /* the companion to osc_consume_write_grant, called when a brw has completed.
1376 * must be called with the loi lock held. */
1377 static void osc_release_write_grant(struct client_obd *cli,
1378 struct brw_page *pga)
1382 assert_spin_locked(&cli->cl_loi_list_lock);
1383 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1388 pga->flag &= ~OBD_BRW_FROM_GRANT;
1389 atomic_long_dec(&obd_dirty_pages);
1390 cli->cl_dirty_pages--;
1395 * To avoid sleeping with object lock held, it's good for us allocate enough
1396 * grants before entering into critical section.
1398 * client_obd_list_lock held by caller
1400 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1404 if (cli->cl_avail_grant >= bytes) {
1405 cli->cl_avail_grant -= bytes;
1406 cli->cl_reserved_grant += bytes;
1412 static void __osc_unreserve_grant(struct client_obd *cli,
1413 unsigned int reserved, unsigned int unused)
1415 /* it's quite normal for us to get more grant than reserved.
1416 * Thinking about a case that two extents merged by adding a new
1417 * chunk, we can save one extent tax. If extent tax is greater than
1418 * one chunk, we can save more grant by adding a new chunk */
1419 cli->cl_reserved_grant -= reserved;
1420 if (unused > reserved) {
1421 cli->cl_avail_grant += reserved;
1422 cli->cl_lost_grant += unused - reserved;
1423 cli->cl_dirty_grant -= unused - reserved;
1425 cli->cl_avail_grant += unused;
1426 cli->cl_dirty_grant += reserved - unused;
1430 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1431 unsigned int reserved,
1432 unsigned int unused)
1434 __osc_unreserve_grant(cli, reserved, unused);
1436 osc_wake_cache_waiters(cli);
1439 static void osc_unreserve_grant(struct client_obd *cli,
1440 unsigned int reserved, unsigned int unused)
1442 spin_lock(&cli->cl_loi_list_lock);
1443 osc_unreserve_grant_nolock(cli, reserved, unused);
1444 spin_unlock(&cli->cl_loi_list_lock);
1448 * Free grant after IO is finished or canceled.
1450 * @lost_grant is used to remember how many grants we have allocated but not
1451 * used, we should return these grants to OST. There're two cases where grants
1454 * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1455 * written. In this case OST may use less chunks to serve this partial
1456 * write. OSTs don't actually know the page size on the client side. so
1457 * clients have to calculate lost grant by the blocksize on the OST.
1458 * See filter_grant_check() for details.
1460 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1461 unsigned int lost_grant, unsigned int dirty_grant)
1463 unsigned long grant;
1465 grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1467 spin_lock(&cli->cl_loi_list_lock);
1468 atomic_long_sub(nr_pages, &obd_dirty_pages);
1469 cli->cl_dirty_pages -= nr_pages;
1470 cli->cl_lost_grant += lost_grant;
1471 cli->cl_dirty_grant -= dirty_grant;
1472 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1473 /* borrow some grant from truncate to avoid the case that
1474 * truncate uses up all avail grant */
1475 cli->cl_lost_grant -= grant;
1476 cli->cl_avail_grant += grant;
1478 osc_wake_cache_waiters(cli);
1479 spin_unlock(&cli->cl_loi_list_lock);
1480 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1481 lost_grant, cli->cl_lost_grant,
1482 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1483 cli->cl_dirty_grant);
1487 * The companion to osc_enter_cache(), called when @oap is no longer part of
1488 * the dirty accounting due to error.
1490 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1492 spin_lock(&cli->cl_loi_list_lock);
1493 osc_release_write_grant(cli, &oap->oap_brw_page);
1494 spin_unlock(&cli->cl_loi_list_lock);
1498 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1501 static int osc_enter_cache_try(struct client_obd *cli,
1502 struct osc_async_page *oap,
1507 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1509 rc = osc_reserve_grant(cli, bytes);
1513 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1514 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1515 obd_max_dirty_pages) {
1516 osc_consume_write_grant(cli, &oap->oap_brw_page);
1520 atomic_long_dec(&obd_dirty_pages);
1522 __osc_unreserve_grant(cli, bytes, bytes);
1528 /* Following two inlines exist to pass code fragments
1529 * to wait_event_idle_exclusive_timeout_cmd(). Passing
1530 * code fragments as macro args can look confusing, so
1531 * we provide inlines to encapsulate them.
1533 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1534 struct client_obd *cli,
1535 struct osc_async_page *oap)
1537 spin_unlock(&cli->cl_loi_list_lock);
1538 osc_io_unplug_async(env, cli, NULL);
1540 "%s: sleeping for cache space for %p\n",
1541 cli_name(cli), oap);
1544 static inline void cli_lock_after_unplug(struct client_obd *cli)
1546 spin_lock(&cli->cl_loi_list_lock);
1549 * The main entry to reserve dirty page accounting. Usually the grant reserved
1550 * in this function will be freed in bulk in osc_free_grant() unless it fails
1551 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1553 * The process will be put into sleep if it's already run out of grant.
1555 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1556 struct osc_async_page *oap, int bytes)
1558 struct osc_object *osc = oap->oap_obj;
1559 struct lov_oinfo *loi = osc->oo_oinfo;
1562 bool entered = false;
1563 /* We cannot wait for a long time here since we are holding ldlm lock
1564 * across the actual IO. If no requests complete fast (e.g. due to
1565 * overloaded OST that takes a long time to process everything, we'd
1566 * get evicted if we wait for a normal obd_timeout or some such.
1567 * So we try to wait half the time it would take the client to be
1568 * evicted by server which is half obd_timeout when AT is off
1569 * or at least ldlm_enqueue_min with AT on.
1571 unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1572 ldlm_enqueue_min / 2);
1576 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1578 spin_lock(&cli->cl_loi_list_lock);
1580 /* force the caller to try sync io. this can jump the list
1581 * of queued writes and create a discontiguous rpc stream */
1582 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1583 cli->cl_dirty_max_pages == 0 ||
1584 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1585 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1586 GOTO(out, rc = -EDQUOT);
1590 * We can wait here for two reasons: too many dirty pages in cache, or
1591 * run out of grants. In both cases we should write dirty pages out.
1592 * Adding a cache waiter will trigger urgent write-out no matter what
1594 * The exiting condition (other than success) is no avail grants
1595 * and no dirty pages caching, that really means there is no space
1598 remain = wait_event_idle_exclusive_timeout_cmd(
1599 cli->cl_cache_waiters,
1600 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1601 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1603 cli_unlock_and_unplug(env, cli, oap),
1604 cli_lock_after_unplug(cli));
1607 if (remain == timeout)
1608 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1610 OSC_DUMP_GRANT(D_CACHE, cli,
1611 "finally got grant space\n");
1612 wake_up(&cli->cl_cache_waiters);
1614 } else if (remain == 0) {
1615 OSC_DUMP_GRANT(D_CACHE, cli,
1616 "timeout, fall back to sync i/o\n");
1617 osc_extent_tree_dump(D_CACHE, osc);
1618 /* fall back to synchronous I/O */
1620 OSC_DUMP_GRANT(D_CACHE, cli,
1621 "no grant space, fall back to sync i/o\n");
1622 wake_up_all(&cli->cl_cache_waiters);
1626 spin_unlock(&cli->cl_loi_list_lock);
1630 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1632 int hprpc = !!list_empty(&osc->oo_hp_exts);
1633 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1636 /* This maintains the lists of pending pages to read/write for a given object
1637 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1638 * to quickly find objects that are ready to send an RPC. */
1639 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1642 int invalid_import = 0;
1645 /* if we have an invalid import we want to drain the queued pages
1646 * by forcing them through rpcs that immediately fail and complete
1647 * the pages. recovery relies on this to empty the queued pages
1648 * before canceling the locks and evicting down the llite pages */
1649 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1652 if (cmd & OBD_BRW_WRITE) {
1653 if (atomic_read(&osc->oo_nr_writes) == 0)
1655 if (invalid_import) {
1656 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1659 if (!list_empty(&osc->oo_hp_exts)) {
1660 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1663 if (!list_empty(&osc->oo_urgent_exts)) {
1664 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1667 /* trigger a write rpc stream as long as there are dirtiers
1668 * waiting for space. as they're waiting, they're not going to
1669 * create more pages to coalesce with what's waiting..
1671 if (waitqueue_active(&cli->cl_cache_waiters)) {
1672 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1675 if (!list_empty(&osc->oo_full_exts)) {
1676 CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1680 if (atomic_read(&osc->oo_nr_reads) == 0)
1682 if (invalid_import) {
1683 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1686 /* all read are urgent. */
1687 if (!list_empty(&osc->oo_reading_exts))
1694 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1696 struct client_obd *cli = osc_cli(obj);
1697 if (cmd & OBD_BRW_WRITE) {
1698 atomic_add(delta, &obj->oo_nr_writes);
1699 atomic_add(delta, &cli->cl_pending_w_pages);
1700 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1702 atomic_add(delta, &obj->oo_nr_reads);
1703 atomic_add(delta, &cli->cl_pending_r_pages);
1704 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1706 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1709 static int osc_makes_hprpc(struct osc_object *obj)
1711 return !list_empty(&obj->oo_hp_exts);
1714 static void on_list(struct list_head *item, struct list_head *list,
1717 if (list_empty(item) && should_be_on)
1718 list_add_tail(item, list);
1719 else if (!list_empty(item) && !should_be_on)
1720 list_del_init(item);
1723 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1724 * can find pages to build into rpcs quickly */
1725 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1727 if (osc_makes_hprpc(osc)) {
1729 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1730 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1732 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1733 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1734 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1735 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1738 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1739 atomic_read(&osc->oo_nr_writes) > 0);
1741 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1742 atomic_read(&osc->oo_nr_reads) > 0);
1744 return osc_is_ready(osc);
1747 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1751 spin_lock(&cli->cl_loi_list_lock);
1752 is_ready = __osc_list_maint(cli, osc);
1753 spin_unlock(&cli->cl_loi_list_lock);
1758 /* this is trying to propogate async writeback errors back up to the
1759 * application. As an async write fails we record the error code for later if
1760 * the app does an fsync. As long as errors persist we force future rpcs to be
1761 * sync so that the app can get a sync error and break the cycle of queueing
1762 * pages for which writeback will fail. */
1763 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1770 ar->ar_force_sync = 1;
1771 ar->ar_min_xid = ptlrpc_sample_next_xid();
1776 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1777 ar->ar_force_sync = 0;
1780 /* this must be called holding the loi list lock to give coverage to exit_cache,
1781 * async_flag maintenance, and oap_request */
1782 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1783 struct osc_async_page *oap, int sent, int rc)
1785 struct osc_object *osc = oap->oap_obj;
1786 struct lov_oinfo *loi = osc->oo_oinfo;
1790 if (oap->oap_request != NULL) {
1791 xid = ptlrpc_req_xid(oap->oap_request);
1792 ptlrpc_req_finished(oap->oap_request);
1793 oap->oap_request = NULL;
1796 /* As the transfer for this page is being done, clear the flags */
1797 spin_lock(&oap->oap_lock);
1798 oap->oap_async_flags = 0;
1799 spin_unlock(&oap->oap_lock);
1801 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1802 spin_lock(&cli->cl_loi_list_lock);
1803 osc_process_ar(&cli->cl_ar, xid, rc);
1804 osc_process_ar(&loi->loi_ar, xid, rc);
1805 spin_unlock(&cli->cl_loi_list_lock);
1808 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1810 CERROR("completion on oap %p obj %p returns %d.\n",
1816 struct extent_rpc_data {
1817 struct list_head *erd_rpc_list;
1818 unsigned int erd_page_count;
1819 unsigned int erd_max_pages;
1820 unsigned int erd_max_chunks;
1821 unsigned int erd_max_extents;
1824 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1826 struct client_obd *cli = osc_cli(ext->oe_obj);
1827 unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1829 return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1833 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1835 if (ext->oe_no_merge || in_rpc->oe_no_merge)
1838 if (ext->oe_srvlock != in_rpc->oe_srvlock)
1841 if (ext->oe_ndelay != in_rpc->oe_ndelay)
1844 if (!ext->oe_grants != !in_rpc->oe_grants)
1847 if (ext->oe_dio != in_rpc->oe_dio)
1850 /* It's possible to have overlap on DIO */
1851 if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1854 if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1861 * Try to add extent to one RPC. We need to think about the following things:
1862 * - # of pages must not be over max_pages_per_rpc
1863 * - extent must be compatible with previous ones
1865 static int try_to_add_extent_for_io(struct client_obd *cli,
1866 struct osc_extent *ext,
1867 struct extent_rpc_data *data)
1869 struct osc_extent *tmp;
1870 unsigned int chunk_count;
1873 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1875 OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1877 if (data->erd_max_extents == 0)
1880 chunk_count = osc_extent_chunks(ext);
1881 EASSERTF(data->erd_page_count != 0 ||
1882 chunk_count <= data->erd_max_chunks, ext,
1883 "The first extent to be fit in a RPC contains %u chunks, "
1884 "which is over the limit %u.\n", chunk_count,
1885 data->erd_max_chunks);
1886 if (chunk_count > data->erd_max_chunks)
1889 data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1890 EASSERTF(data->erd_page_count != 0 ||
1891 ext->oe_nr_pages <= data->erd_max_pages, ext,
1892 "The first extent to be fit in a RPC contains %u pages, "
1893 "which is over the limit %u.\n", ext->oe_nr_pages,
1894 data->erd_max_pages);
1895 if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1898 list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1899 EASSERT(tmp->oe_owner == current, tmp);
1901 if (!can_merge(ext, tmp))
1905 data->erd_max_extents--;
1906 data->erd_max_chunks -= chunk_count;
1907 data->erd_page_count += ext->oe_nr_pages;
1908 list_move_tail(&ext->oe_link, data->erd_rpc_list);
1909 ext->oe_owner = current;
1914 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1915 * get_write_extent() takes all appropriate extents in atomic.
1917 * The following policy is used to collect extents for IO:
1918 * 1. Add as many HP extents as possible;
1919 * 2. Add the first urgent extent in urgent extent list and take it out of
1921 * 3. Add subsequent extents of this urgent extent;
1922 * 4. If urgent list is not empty, goto 2;
1923 * 5. Traverse the extent tree from the 1st extent;
1924 * 6. Above steps exit if there is no space in this RPC.
1926 static unsigned int get_write_extents(struct osc_object *obj,
1927 struct list_head *rpclist)
1929 struct client_obd *cli = osc_cli(obj);
1930 struct osc_extent *ext;
1931 struct extent_rpc_data data = {
1932 .erd_rpc_list = rpclist,
1933 .erd_page_count = 0,
1934 .erd_max_pages = cli->cl_max_pages_per_rpc,
1935 .erd_max_chunks = osc_max_write_chunks(cli),
1936 .erd_max_extents = 256,
1939 assert_osc_object_is_locked(obj);
1940 while (!list_empty(&obj->oo_hp_exts)) {
1941 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1943 if (!try_to_add_extent_for_io(cli, ext, &data))
1944 return data.erd_page_count;
1945 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1947 if (data.erd_page_count == data.erd_max_pages)
1948 return data.erd_page_count;
1950 while (!list_empty(&obj->oo_urgent_exts)) {
1951 ext = list_entry(obj->oo_urgent_exts.next,
1952 struct osc_extent, oe_link);
1953 if (!try_to_add_extent_for_io(cli, ext, &data))
1954 return data.erd_page_count;
1956 if (data.erd_page_count == data.erd_max_pages)
1957 return data.erd_page_count;
1959 /* One key difference between full extents and other extents: full
1960 * extents can usually only be added if the rpclist was empty, so if we
1961 * can't add one, we continue on to trying to add normal extents. This
1962 * is so we don't miss adding extra extents to an RPC containing high
1963 * priority or urgent extents. */
1964 while (!list_empty(&obj->oo_full_exts)) {
1965 ext = list_entry(obj->oo_full_exts.next,
1966 struct osc_extent, oe_link);
1967 if (!try_to_add_extent_for_io(cli, ext, &data))
1970 if (data.erd_page_count == data.erd_max_pages)
1971 return data.erd_page_count;
1973 for (ext = first_extent(obj);
1975 ext = next_extent(ext)) {
1976 if ((ext->oe_state != OES_CACHE) ||
1977 /* this extent may be already in current rpclist */
1978 (!list_empty(&ext->oe_link) && ext->oe_owner))
1981 if (!try_to_add_extent_for_io(cli, ext, &data))
1982 return data.erd_page_count;
1984 return data.erd_page_count;
1988 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1989 struct osc_object *osc)
1993 struct osc_extent *ext;
1994 struct osc_extent *tmp;
1995 struct osc_extent *first = NULL;
1996 unsigned int page_count = 0;
2001 assert_osc_object_is_locked(osc);
2003 page_count = get_write_extents(osc, &rpclist);
2004 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2006 if (list_empty(&rpclist))
2009 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2011 list_for_each_entry(ext, &rpclist, oe_link) {
2012 LASSERT(ext->oe_state == OES_CACHE ||
2013 ext->oe_state == OES_LOCK_DONE);
2014 if (ext->oe_state == OES_CACHE)
2015 osc_extent_state_set(ext, OES_LOCKING);
2017 osc_extent_state_set(ext, OES_RPC);
2020 /* we're going to grab page lock, so release object lock because
2021 * lock order is page lock -> object lock. */
2022 osc_object_unlock(osc);
2024 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2025 if (ext->oe_state == OES_LOCKING) {
2026 rc = osc_extent_make_ready(env, ext);
2027 if (unlikely(rc < 0)) {
2028 list_del_init(&ext->oe_link);
2029 osc_extent_finish(env, ext, 0, rc);
2033 if (first == NULL) {
2035 srvlock = ext->oe_srvlock;
2037 LASSERT(srvlock == ext->oe_srvlock);
2041 if (!list_empty(&rpclist)) {
2042 LASSERT(page_count > 0);
2043 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2044 LASSERT(list_empty(&rpclist));
2047 osc_object_lock(osc);
2052 * prepare pages for ASYNC io and put pages in send queue.
2054 * \param cmd OBD_BRW_* macroses
2055 * \param lop pending pages
2057 * \return zero if no page added to send queue.
2058 * \return 1 if pages successfully added to send queue.
2059 * \return negative on errors.
2062 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2063 struct osc_object *osc)
2066 struct osc_extent *ext;
2067 struct osc_extent *next;
2069 struct extent_rpc_data data = {
2070 .erd_rpc_list = &rpclist,
2071 .erd_page_count = 0,
2072 .erd_max_pages = cli->cl_max_pages_per_rpc,
2073 .erd_max_chunks = UINT_MAX,
2074 .erd_max_extents = UINT_MAX,
2079 assert_osc_object_is_locked(osc);
2080 list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2081 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2082 if (!try_to_add_extent_for_io(cli, ext, &data))
2084 osc_extent_state_set(ext, OES_RPC);
2085 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2087 LASSERT(data.erd_page_count <= data.erd_max_pages);
2089 osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2091 if (!list_empty(&rpclist)) {
2092 osc_object_unlock(osc);
2094 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2095 LASSERT(list_empty(&rpclist));
2097 osc_object_lock(osc);
2102 #define list_to_obj(list, item) ({ \
2103 struct list_head *__tmp = (list)->next; \
2104 list_del_init(__tmp); \
2105 list_entry(__tmp, struct osc_object, oo_##item); \
2108 /* This is called by osc_check_rpcs() to find which objects have pages that
2109 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2110 static struct osc_object *osc_next_obj(struct client_obd *cli)
2114 /* First return objects that have blocked locks so that they
2115 * will be flushed quickly and other clients can get the lock,
2116 * then objects which have pages ready to be stuffed into RPCs */
2117 if (!list_empty(&cli->cl_loi_hp_ready_list))
2118 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2119 if (!list_empty(&cli->cl_loi_ready_list))
2120 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2122 /* then if we have cache waiters, return all objects with queued
2123 * writes. This is especially important when many small files
2124 * have filled up the cache and not been fired into rpcs because
2125 * they don't pass the nr_pending/object threshhold
2127 if (waitqueue_active(&cli->cl_cache_waiters) &&
2128 !list_empty(&cli->cl_loi_write_list))
2129 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2131 /* then return all queued objects when we have an invalid import
2132 * so that they get flushed */
2133 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2134 if (!list_empty(&cli->cl_loi_write_list))
2135 RETURN(list_to_obj(&cli->cl_loi_write_list,
2137 if (!list_empty(&cli->cl_loi_read_list))
2138 RETURN(list_to_obj(&cli->cl_loi_read_list,
2144 /* called with the loi list lock held */
2145 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2146 __must_hold(&cli->cl_loi_list_lock)
2148 struct osc_object *osc;
2152 while ((osc = osc_next_obj(cli)) != NULL) {
2153 struct cl_object *obj = osc2cl(osc);
2154 struct lu_ref_link link;
2156 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2158 /* even if we have reached our max in flight RPCs, we still
2159 * allow all high-priority RPCs through to prevent their
2160 * starvation and leading to server evicting us for not
2161 * writing out pages in a timely manner LU-13131 */
2162 if (osc_max_rpc_in_flight(cli, osc) &&
2163 list_empty(&osc->oo_hp_exts)) {
2164 __osc_list_maint(cli, osc);
2169 spin_unlock(&cli->cl_loi_list_lock);
2170 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2172 /* attempt some read/write balancing by alternating between
2173 * reads and writes in an object. The makes_rpc checks here
2174 * would be redundant if we were getting read/write work items
2175 * instead of objects. we don't want send_oap_rpc to drain a
2176 * partial read pending queue when we're given this object to
2177 * do io on writes while there are cache waiters */
2178 osc_object_lock(osc);
2179 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2180 rc = osc_send_write_rpc(env, cli, osc);
2182 CERROR("Write request failed with %d\n", rc);
2184 /* osc_send_write_rpc failed, mostly because of
2187 * It can't break here, because if:
2188 * - a page was submitted by osc_io_submit, so
2190 * - no request in flight
2191 * - no subsequent request
2192 * The system will be in live-lock state,
2193 * because there is no chance to call
2194 * osc_io_unplug() and osc_check_rpcs() any
2195 * more. pdflush can't help in this case,
2196 * because it might be blocked at grabbing
2197 * the page lock as we mentioned.
2199 * Anyway, continue to drain pages. */
2203 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2204 rc = osc_send_read_rpc(env, cli, osc);
2206 CERROR("Read request failed with %d\n", rc);
2208 osc_object_unlock(osc);
2210 osc_list_maint(cli, osc);
2211 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2212 cl_object_put(env, obj);
2214 spin_lock(&cli->cl_loi_list_lock);
2218 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2219 struct osc_object *osc, int async)
2223 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2227 spin_lock(&cli->cl_loi_list_lock);
2228 osc_check_rpcs(env, cli);
2229 spin_unlock(&cli->cl_loi_list_lock);
2231 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2232 LASSERT(cli->cl_writeback_work != NULL);
2233 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2237 EXPORT_SYMBOL(osc_io_unplug0);
2239 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2240 struct page *page, loff_t offset)
2242 struct obd_export *exp = osc_export(osc);
2243 struct osc_async_page *oap = &ops->ops_oap;
2247 return cfs_size_round(sizeof(*oap));
2249 oap->oap_magic = OAP_MAGIC;
2250 oap->oap_cli = &exp->exp_obd->u.cli;
2253 oap->oap_page = page;
2254 oap->oap_obj_off = offset;
2255 LASSERT(!(offset & ~PAGE_MASK));
2257 INIT_LIST_HEAD(&oap->oap_pending_item);
2258 INIT_LIST_HEAD(&oap->oap_rpc_item);
2260 spin_lock_init(&oap->oap_lock);
2261 CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2262 oap, page, oap->oap_obj_off);
2265 EXPORT_SYMBOL(osc_prep_async_page);
2267 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2268 struct osc_page *ops, cl_commit_cbt cb)
2270 struct osc_io *oio = osc_env_io(env);
2271 struct osc_extent *ext = NULL;
2272 struct osc_async_page *oap = &ops->ops_oap;
2273 struct client_obd *cli = oap->oap_cli;
2274 struct osc_object *osc = oap->oap_obj;
2275 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
2278 unsigned int grants = 0;
2279 u32 brw_flags = OBD_BRW_ASYNC;
2280 int cmd = OBD_BRW_WRITE;
2281 int need_release = 0;
2285 if (oap->oap_magic != OAP_MAGIC)
2288 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2291 if (!list_empty(&oap->oap_pending_item) ||
2292 !list_empty(&oap->oap_rpc_item))
2295 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2296 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2297 if (oio->oi_cap_sys_resource || io->ci_noquota) {
2298 brw_flags |= OBD_BRW_NOQUOTA;
2299 cmd |= OBD_BRW_NOQUOTA;
2302 /* check if the file's owner/group is over quota */
2303 if (!(cmd & OBD_BRW_NOQUOTA)) {
2304 struct cl_object *obj;
2305 struct cl_attr *attr;
2306 unsigned int qid[LL_MAXQUOTAS];
2308 obj = cl_object_top(&osc->oo_cl);
2309 attr = &osc_env_info(env)->oti_attr;
2311 cl_object_attr_lock(obj);
2312 rc = cl_object_attr_get(env, obj, attr);
2313 cl_object_attr_unlock(obj);
2315 qid[USRQUOTA] = attr->cat_uid;
2316 qid[GRPQUOTA] = attr->cat_gid;
2317 qid[PRJQUOTA] = attr->cat_projid;
2318 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2325 oap->oap_page_off = ops->ops_from;
2326 oap->oap_count = ops->ops_to - ops->ops_from + 1;
2327 /* No need to hold a lock here,
2328 * since this page is not in any list yet. */
2329 oap->oap_async_flags = 0;
2330 oap->oap_brw_flags = brw_flags;
2332 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2333 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2335 index = osc_index(oap2osc(oap));
2337 /* Add this page into extent by the following steps:
2338 * 1. if there exists an active extent for this IO, mostly this page
2339 * can be added to the active extent and sometimes we need to
2340 * expand extent to accomodate this page;
2341 * 2. otherwise, a new extent will be allocated. */
2343 ext = oio->oi_active;
2344 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2345 /* one chunk plus extent overhead must be enough to write this
2347 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2348 if (ext->oe_end >= index)
2351 /* it doesn't need any grant to dirty this page */
2352 spin_lock(&cli->cl_loi_list_lock);
2353 rc = osc_enter_cache_try(cli, oap, grants);
2354 if (rc == 0) { /* try failed */
2357 } else if (ext->oe_end < index) {
2359 /* try to expand this extent */
2360 rc = osc_extent_expand(ext, index, &tmp);
2363 /* don't free reserved grant */
2365 OSC_EXTENT_DUMP(D_CACHE, ext,
2366 "expanded for %lu.\n", index);
2367 osc_unreserve_grant_nolock(cli, grants, tmp);
2371 spin_unlock(&cli->cl_loi_list_lock);
2373 } else if (ext != NULL) {
2374 /* index is located outside of active extent */
2378 osc_extent_release(env, ext);
2379 oio->oi_active = NULL;
2384 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2386 /* try to find new extent to cover this page */
2387 LASSERT(oio->oi_active == NULL);
2388 /* we may have allocated grant for this page if we failed
2389 * to expand the previous active extent. */
2390 LASSERT(ergo(grants > 0, grants >= tmp));
2394 /* We haven't allocated grant for this page, and we
2395 * must not hold a page lock while we do enter_cache,
2396 * so we must mark dirty & unlock any pages in the
2397 * write commit pagevec. */
2398 if (pagevec_count(pvec)) {
2400 pagevec_reinit(pvec);
2402 rc = osc_enter_cache(env, cli, oap, tmp);
2409 ext = osc_extent_find(env, osc, index, &tmp);
2411 LASSERT(tmp == grants);
2412 osc_exit_cache(cli, oap);
2416 oio->oi_active = ext;
2420 osc_unreserve_grant(cli, grants, tmp);
2423 LASSERT(ergo(rc == 0, ext != NULL));
2425 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2426 ext, "index = %lu.\n", index);
2427 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2429 osc_object_lock(osc);
2430 if (ext->oe_nr_pages == 0)
2431 ext->oe_srvlock = ops->ops_srvlock;
2433 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2435 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2436 osc_object_unlock(osc);
2438 if (!ext->oe_layout_version)
2439 ext->oe_layout_version = io->ci_layout_version;
2445 int osc_teardown_async_page(const struct lu_env *env,
2446 struct osc_object *obj, struct osc_page *ops)
2448 struct osc_async_page *oap = &ops->ops_oap;
2452 LASSERT(oap->oap_magic == OAP_MAGIC);
2454 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2455 oap, ops, osc_index(oap2osc(oap)));
2457 if (!list_empty(&oap->oap_rpc_item)) {
2458 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2460 } else if (!list_empty(&oap->oap_pending_item)) {
2461 struct osc_extent *ext = NULL;
2463 osc_object_lock(obj);
2464 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2465 osc_object_unlock(obj);
2466 /* only truncated pages are allowed to be taken out.
2467 * See osc_extent_truncate() and osc_cache_truncate_start()
2469 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2470 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2471 osc_index(oap2osc(oap)));
2475 osc_extent_put(env, ext);
2481 * This is called when a page is picked up by kernel to write out.
2483 * We should find out the corresponding extent and add the whole extent
2484 * into urgent list. The extent may be being truncated or used, handle it
2487 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2488 struct osc_page *ops)
2490 struct osc_extent *ext = NULL;
2491 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2492 struct cl_page *cp = ops->ops_cl.cpl_page;
2493 pgoff_t index = osc_index(ops);
2494 struct osc_async_page *oap = &ops->ops_oap;
2495 bool unplug = false;
2499 osc_object_lock(obj);
2500 ext = osc_extent_lookup(obj, index);
2502 osc_extent_tree_dump(D_ERROR, obj);
2503 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2506 switch (ext->oe_state) {
2509 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2513 /* If we know this extent is being written out, we should abort
2514 * so that the writer can make this page ready. Otherwise, there
2515 * exists a deadlock problem because other process can wait for
2516 * page writeback bit holding page lock; and meanwhile in
2517 * vvp_page_make_ready(), we need to grab page lock before
2518 * really sending the RPC. */
2520 /* race with truncate, page will be redirtied */
2522 /* The extent is active so we need to abort and let the caller
2523 * re-dirty the page. If we continued on here, and we were the
2524 * one making the extent active, we could deadlock waiting for
2525 * the page writeback to clear but it won't because the extent
2526 * is active and won't be written out. */
2527 GOTO(out, rc = -EAGAIN);
2532 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2536 spin_lock(&oap->oap_lock);
2537 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2538 spin_unlock(&oap->oap_lock);
2540 if (current->flags & PF_MEMALLOC)
2541 ext->oe_memalloc = 1;
2544 if (ext->oe_state == OES_CACHE) {
2545 OSC_EXTENT_DUMP(D_CACHE, ext,
2546 "flush page %p make it urgent.\n", oap);
2547 if (list_empty(&ext->oe_link))
2548 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2555 osc_object_unlock(obj);
2556 osc_extent_put(env, ext);
2558 osc_io_unplug_async(env, osc_cli(obj), obj);
2562 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2563 struct osc_object *obj, struct list_head *list,
2566 struct client_obd *cli = osc_cli(obj);
2567 struct osc_extent *ext;
2568 struct osc_async_page *oap;
2570 int mppr = cli->cl_max_pages_per_rpc;
2571 bool can_merge = true;
2572 pgoff_t start = CL_PAGE_EOF;
2576 list_for_each_entry(oap, list, oap_pending_item) {
2577 struct osc_page *opg = oap2osc_page(oap);
2578 pgoff_t index = osc_index(opg);
2585 mppr <<= (page_count > mppr);
2587 if (unlikely(opg->ops_from > 0 ||
2588 opg->ops_to < PAGE_SIZE - 1))
2592 ext = osc_extent_alloc(obj);
2594 struct osc_async_page *tmp;
2596 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2597 list_del_init(&oap->oap_pending_item);
2598 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2603 ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2605 ext->oe_no_merge = !can_merge;
2607 ext->oe_start = start;
2608 ext->oe_end = ext->oe_max_end = end;
2610 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2611 ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2612 ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2613 if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2617 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2618 grants = cli->cl_grant_extent_tax;
2619 grants += (1 << cli->cl_chunkbits) *
2620 ((page_count + ppc - 1) / ppc);
2622 spin_lock(&cli->cl_loi_list_lock);
2623 if (osc_reserve_grant(cli, grants) == 0) {
2624 list_for_each_entry(oap, list, oap_pending_item) {
2625 osc_consume_write_grant(cli,
2626 &oap->oap_brw_page);
2627 atomic_long_inc(&obd_dirty_pages);
2629 osc_unreserve_grant_nolock(cli, grants, 0);
2630 ext->oe_grants = grants;
2632 spin_unlock(&cli->cl_loi_list_lock);
2635 ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
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;
3120 struct cl_page *page = ops->ops_cl.cpl_page;
3122 bool discard = false;
3124 index = osc_index(ops);
3126 /* negative lock caching */
3127 if (index < info->oti_ng_index) {
3129 } else if (index >= info->oti_fn_index) {
3130 struct ldlm_lock *tmp;
3131 /* refresh non-overlapped index */
3132 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3133 OSC_DAP_FL_TEST_LOCK |
3134 OSC_DAP_FL_AST | OSC_DAP_FL_RIGHT);
3136 __u64 end = tmp->l_policy_data.l_extent.end;
3137 __u64 start = tmp->l_policy_data.l_extent.start;
3139 /* no lock covering this page */
3140 if (index < cl_index(osc2cl(osc), start)) {
3141 /* no lock at @index, first lock at @start */
3142 info->oti_ng_index = cl_index(osc2cl(osc),
3146 /* Cache the first-non-overlapped index so as to
3147 * skip all pages within [index, oti_fn_index).
3148 * This is safe because if tmp lock is canceled,
3149 * it will discard these pages.
3151 info->oti_fn_index = cl_index(osc2cl(osc),
3153 if (end == OBD_OBJECT_EOF)
3154 info->oti_fn_index = CL_PAGE_EOF;
3158 info->oti_ng_index = CL_PAGE_EOF;
3164 if (cl_page_own(env, io, page) == 0) {
3165 cl_page_discard(env, io, page);
3166 cl_page_disown(env, io, page);
3168 LASSERT(page->cp_state == CPS_FREEING);
3172 info->oti_next_index = index + 1;
3176 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3177 struct osc_page *ops, void *cbdata)
3179 struct osc_thread_info *info = osc_env_info(env);
3180 struct cl_page *page = ops->ops_cl.cpl_page;
3182 /* page is top page. */
3183 info->oti_next_index = osc_index(ops) + 1;
3184 if (cl_page_own(env, io, page) == 0) {
3185 if (!ergo(page->cp_type == CPT_CACHEABLE,
3186 !PageDirty(cl_page_vmpage(page))))
3187 CL_PAGE_DEBUG(D_ERROR, env, page,
3188 "discard dirty page?\n");
3190 /* discard the page */
3191 cl_page_discard(env, io, page);
3192 cl_page_disown(env, io, page);
3194 LASSERT(page->cp_state == CPS_FREEING);
3199 EXPORT_SYMBOL(osc_discard_cb);
3202 * Discard pages protected by the given lock. This function traverses radix
3203 * tree to find all covering pages and discard them. If a page is being covered
3204 * by other locks, it should remain in cache.
3206 * If error happens on any step, the process continues anyway (the reasoning
3207 * behind this being that lock cancellation cannot be delayed indefinitely).
3209 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3210 pgoff_t start, pgoff_t end, bool discard)
3212 struct osc_thread_info *info = osc_env_info(env);
3213 struct cl_io *io = osc_env_thread_io(env);
3214 osc_page_gang_cbt cb;
3219 io->ci_obj = cl_object_top(osc2cl(osc));
3220 io->ci_ignore_layout = 1;
3221 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3225 cb = discard ? osc_discard_cb : check_and_discard_cb;
3226 info->oti_fn_index = info->oti_next_index = start;
3227 info->oti_ng_index = 0;
3229 osc_page_gang_lookup(env, io, osc,
3230 info->oti_next_index, end, cb, osc);
3232 cl_io_fini(env, io);