4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
34 * This file is part of Lustre, http://www.lustre.org/
35 * Lustre is a trademark of Sun Microsystems, Inc.
37 * osc cache management.
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_OSC
44 #include "osc_cl_internal.h"
45 #include "osc_internal.h"
47 static int extent_debug; /* set it to be true for more debug */
49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
53 struct osc_async_page *oap, int sent, int rc);
54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
56 static int osc_refresh_count(const struct lu_env *env,
57 struct osc_async_page *oap, int cmd);
58 static int osc_io_unplug_async(const struct lu_env *env,
59 struct client_obd *cli, struct osc_object *osc);
60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
61 unsigned int lost_grant);
63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
64 const char *func, int line);
65 #define osc_extent_tree_dump(lvl, obj) \
66 osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
72 /* ------------------ osc extent ------------------ */
73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
76 *buf++ = ext->oe_rw ? 'r' : 'w';
89 if (ext->oe_trunc_pending)
91 if (ext->oe_fsync_wait)
97 static inline char list_empty_marker(struct list_head *list)
99 return list_empty(list) ? '-' : '+';
102 #define EXTSTR "[%lu -> %lu/%lu]"
103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
104 static const char *oes_strings[] = {
105 "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
107 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do { \
108 struct osc_extent *__ext = (extent); \
112 "extent %p@{" EXTSTR ", " \
113 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt, \
114 /* ----- extent part 0 ----- */ \
115 __ext, EXTPARA(__ext), \
116 /* ----- part 1 ----- */ \
117 atomic_read(&__ext->oe_refc), \
118 atomic_read(&__ext->oe_users), \
119 list_empty_marker(&__ext->oe_link), \
120 oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \
122 /* ----- part 2 ----- */ \
123 __ext->oe_grants, __ext->oe_nr_pages, \
124 list_empty_marker(&__ext->oe_pages), \
125 waitqueue_active(&__ext->oe_waitq) ? '+' : '-', \
126 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner, \
127 /* ----- part 4 ----- */ \
129 if (lvl == D_ERROR && __ext->oe_dlmlock != NULL) \
130 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p\n", __ext); \
132 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p\n", __ext); \
136 #define EASSERTF(expr, ext, fmt, args...) do { \
138 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args); \
139 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj); \
145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
147 static inline struct osc_extent *rb_extent(struct rb_node *n)
152 return container_of(n, struct osc_extent, oe_node);
155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
160 LASSERT(ext->oe_intree);
161 return rb_extent(rb_next(&ext->oe_node));
164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
169 LASSERT(ext->oe_intree);
170 return rb_extent(rb_prev(&ext->oe_node));
173 static inline struct osc_extent *first_extent(struct osc_object *obj)
175 return rb_extent(rb_first(&obj->oo_root));
178 /* object must be locked by caller. */
179 static int osc_extent_sanity_check0(struct osc_extent *ext,
180 const char *func, const int line)
182 struct osc_object *obj = ext->oe_obj;
183 struct osc_async_page *oap;
187 if (!osc_object_is_locked(obj))
190 if (ext->oe_state >= OES_STATE_MAX)
193 if (atomic_read(&ext->oe_refc) <= 0)
196 if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
199 switch (ext->oe_state) {
201 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
206 if (atomic_read(&ext->oe_users) == 0)
210 if (ext->oe_fsync_wait && !ext->oe_urgent)
214 if (ext->oe_grants == 0)
216 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
219 if (atomic_read(&ext->oe_users) > 0)
223 if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
226 if (ext->oe_sync && ext->oe_grants > 0)
229 if (ext->oe_dlmlock != NULL) {
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(D_ERROR, ext,
265 "%s:%d sanity check %p failed with rc = %d\n",
266 func, line, ext, rc);
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); \
283 * sanity check - to make sure there is no overlapped extent in the tree.
285 static int osc_extent_is_overlapped(struct osc_object *obj,
286 struct osc_extent *ext)
288 struct osc_extent *tmp;
290 LASSERT(osc_object_is_locked(obj));
295 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
298 if (tmp->oe_end >= ext->oe_start &&
299 tmp->oe_start <= ext->oe_end)
305 static void osc_extent_state_set(struct osc_extent *ext, int state)
307 LASSERT(osc_object_is_locked(ext->oe_obj));
308 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
310 /* Never try to sanity check a state changing extent :-) */
311 /* LASSERT(sanity_check_nolock(ext) == 0); */
313 /* TODO: validate the state machine */
314 ext->oe_state = state;
315 wake_up_all(&ext->oe_waitq);
318 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
320 struct osc_extent *ext;
322 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
326 RB_CLEAR_NODE(&ext->oe_node);
328 atomic_set(&ext->oe_refc, 1);
329 atomic_set(&ext->oe_users, 0);
330 INIT_LIST_HEAD(&ext->oe_link);
331 ext->oe_state = OES_INV;
332 INIT_LIST_HEAD(&ext->oe_pages);
333 init_waitqueue_head(&ext->oe_waitq);
334 ext->oe_dlmlock = NULL;
339 static void osc_extent_free(struct osc_extent *ext)
341 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
344 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
346 LASSERT(atomic_read(&ext->oe_refc) >= 0);
347 atomic_inc(&ext->oe_refc);
351 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
353 LASSERT(atomic_read(&ext->oe_refc) > 0);
354 if (atomic_dec_and_test(&ext->oe_refc)) {
355 LASSERT(list_empty(&ext->oe_link));
356 LASSERT(atomic_read(&ext->oe_users) == 0);
357 LASSERT(ext->oe_state == OES_INV);
358 LASSERT(!ext->oe_intree);
360 if (ext->oe_dlmlock != NULL) {
361 lu_ref_add(&ext->oe_dlmlock->l_reference,
363 LDLM_LOCK_PUT(ext->oe_dlmlock);
364 ext->oe_dlmlock = NULL;
366 osc_extent_free(ext);
371 * osc_extent_put_trust() is a special version of osc_extent_put() when
372 * it's known that the caller is not the last user. This is to address the
373 * problem of lacking of lu_env ;-).
375 static void osc_extent_put_trust(struct osc_extent *ext)
377 LASSERT(atomic_read(&ext->oe_refc) > 1);
378 LASSERT(osc_object_is_locked(ext->oe_obj));
379 atomic_dec(&ext->oe_refc);
383 * Return the extent which includes pgoff @index, or return the greatest
384 * previous extent in the tree.
386 static struct osc_extent *osc_extent_search(struct osc_object *obj,
389 struct rb_node *n = obj->oo_root.rb_node;
390 struct osc_extent *tmp, *p = NULL;
392 LASSERT(osc_object_is_locked(obj));
395 if (index < tmp->oe_start) {
397 } else if (index > tmp->oe_end) {
408 * Return the extent covering @index, otherwise return NULL.
409 * caller must have held object lock.
411 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
414 struct osc_extent *ext;
416 ext = osc_extent_search(obj, index);
417 if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
418 return osc_extent_get(ext);
422 /* caller must have held object lock. */
423 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
425 struct rb_node **n = &obj->oo_root.rb_node;
426 struct rb_node *parent = NULL;
427 struct osc_extent *tmp;
429 LASSERT(ext->oe_intree == 0);
430 LASSERT(ext->oe_obj == obj);
431 LASSERT(osc_object_is_locked(obj));
436 if (ext->oe_end < tmp->oe_start)
438 else if (ext->oe_start > tmp->oe_end)
441 EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
443 rb_link_node(&ext->oe_node, parent, n);
444 rb_insert_color(&ext->oe_node, &obj->oo_root);
449 /* caller must have held object lock. */
450 static void osc_extent_erase(struct osc_extent *ext)
452 struct osc_object *obj = ext->oe_obj;
453 LASSERT(osc_object_is_locked(obj));
454 if (ext->oe_intree) {
455 rb_erase(&ext->oe_node, &obj->oo_root);
457 /* rbtree held a refcount */
458 osc_extent_put_trust(ext);
462 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
464 struct osc_object *obj = ext->oe_obj;
466 LASSERT(osc_object_is_locked(obj));
467 LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
468 if (ext->oe_state == OES_CACHE) {
469 osc_extent_state_set(ext, OES_ACTIVE);
470 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
472 atomic_inc(&ext->oe_users);
473 list_del_init(&ext->oe_link);
474 return osc_extent_get(ext);
477 static void __osc_extent_remove(struct osc_extent *ext)
479 LASSERT(osc_object_is_locked(ext->oe_obj));
480 LASSERT(list_empty(&ext->oe_pages));
481 osc_extent_erase(ext);
482 list_del_init(&ext->oe_link);
483 osc_extent_state_set(ext, OES_INV);
484 OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
487 static void osc_extent_remove(struct osc_extent *ext)
489 struct osc_object *obj = ext->oe_obj;
491 osc_object_lock(obj);
492 __osc_extent_remove(ext);
493 osc_object_unlock(obj);
497 * This function is used to merge extents to get better performance. It checks
498 * if @cur and @victim are contiguous at chunk level.
500 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
501 struct osc_extent *victim)
503 struct osc_object *obj = cur->oe_obj;
508 LASSERT(cur->oe_state == OES_CACHE);
509 LASSERT(osc_object_is_locked(obj));
513 if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
516 if (cur->oe_max_end != victim->oe_max_end)
519 LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
520 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
521 chunk_start = cur->oe_start >> ppc_bits;
522 chunk_end = cur->oe_end >> ppc_bits;
523 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
524 chunk_end + 1 != victim->oe_start >> ppc_bits)
527 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
529 cur->oe_start = min(cur->oe_start, victim->oe_start);
530 cur->oe_end = max(cur->oe_end, victim->oe_end);
531 cur->oe_grants += victim->oe_grants;
532 cur->oe_nr_pages += victim->oe_nr_pages;
533 /* only the following bits are needed to merge */
534 cur->oe_urgent |= victim->oe_urgent;
535 cur->oe_memalloc |= victim->oe_memalloc;
536 list_splice_init(&victim->oe_pages, &cur->oe_pages);
537 list_del_init(&victim->oe_link);
538 victim->oe_nr_pages = 0;
540 osc_extent_get(victim);
541 __osc_extent_remove(victim);
542 osc_extent_put(env, victim);
544 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
549 * Drop user count of osc_extent, and unplug IO asynchronously.
551 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
553 struct osc_object *obj = ext->oe_obj;
557 LASSERT(atomic_read(&ext->oe_users) > 0);
558 LASSERT(sanity_check(ext) == 0);
559 LASSERT(ext->oe_grants > 0);
561 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
562 LASSERT(ext->oe_state == OES_ACTIVE);
563 if (ext->oe_trunc_pending) {
564 /* a truncate process is waiting for this extent.
565 * This may happen due to a race, check
566 * osc_cache_truncate_start(). */
567 osc_extent_state_set(ext, OES_TRUNC);
568 ext->oe_trunc_pending = 0;
570 osc_extent_state_set(ext, OES_CACHE);
571 osc_update_pending(obj, OBD_BRW_WRITE,
574 /* try to merge the previous and next extent. */
575 osc_extent_merge(env, ext, prev_extent(ext));
576 osc_extent_merge(env, ext, next_extent(ext));
579 list_move_tail(&ext->oe_link,
580 &obj->oo_urgent_exts);
582 osc_object_unlock(obj);
584 osc_io_unplug_async(env, osc_cli(obj), obj);
586 osc_extent_put(env, ext);
590 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
592 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
596 * Find or create an extent which includes @index, core function to manage
599 static struct osc_extent *osc_extent_find(const struct lu_env *env,
600 struct osc_object *obj, pgoff_t index,
604 struct client_obd *cli = osc_cli(obj);
605 struct osc_lock *olck;
606 struct cl_lock_descr *descr;
607 struct osc_extent *cur;
608 struct osc_extent *ext;
609 struct osc_extent *conflict = NULL;
610 struct osc_extent *found = NULL;
613 int max_pages; /* max_pages_per_rpc */
615 int ppc_bits; /* pages per chunk bits */
620 cur = osc_extent_alloc(obj);
622 RETURN(ERR_PTR(-ENOMEM));
624 olck = osc_env_io(env)->oi_write_osclock;
625 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
626 LASSERT(olck->ols_state == OLS_GRANTED);
628 descr = &olck->ols_cl.cls_lock->cll_descr;
629 LASSERT(descr->cld_mode >= CLM_WRITE);
631 LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
632 ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
633 chunk_mask = ~((1 << ppc_bits) - 1);
634 chunksize = 1 << cli->cl_chunkbits;
635 chunk = index >> ppc_bits;
637 /* align end to rpc edge, rpc size may not be a power 2 integer. */
638 max_pages = cli->cl_max_pages_per_rpc;
639 LASSERT((max_pages & ~chunk_mask) == 0);
640 max_end = index - (index % max_pages) + max_pages - 1;
641 max_end = min_t(pgoff_t, max_end, descr->cld_end);
643 /* initialize new extent by parameters so far */
644 cur->oe_max_end = max_end;
645 cur->oe_start = index & chunk_mask;
646 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
647 if (cur->oe_start < descr->cld_start)
648 cur->oe_start = descr->cld_start;
649 if (cur->oe_end > max_end)
650 cur->oe_end = max_end;
652 cur->oe_mppr = max_pages;
653 if (olck->ols_dlmlock != NULL) {
654 LASSERT(olck->ols_hold);
655 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
656 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
659 /* grants has been allocated by caller */
660 LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
661 "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
662 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
666 osc_object_lock(obj);
667 ext = osc_extent_search(obj, cur->oe_start);
669 ext = first_extent(obj);
670 while (ext != NULL) {
671 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
672 loff_t ext_chk_end = ext->oe_end >> ppc_bits;
674 LASSERT(sanity_check_nolock(ext) == 0);
675 if (chunk > ext_chk_end + 1)
678 /* if covering by different locks, no chance to match */
679 if (olck->ols_dlmlock != ext->oe_dlmlock) {
680 EASSERTF(!overlapped(ext, cur), ext,
681 EXTSTR"\n", EXTPARA(cur));
683 ext = next_extent(ext);
687 /* discontiguous chunks? */
688 if (chunk + 1 < ext_chk_start) {
689 ext = next_extent(ext);
693 /* ok, from now on, ext and cur have these attrs:
694 * 1. covered by the same lock
695 * 2. contiguous at chunk level or overlapping. */
697 if (overlapped(ext, cur)) {
698 /* cur is the minimum unit, so overlapping means
700 EASSERTF((ext->oe_start <= cur->oe_start &&
701 ext->oe_end >= cur->oe_end),
702 ext, EXTSTR"\n", EXTPARA(cur));
704 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
705 /* for simplicity, we wait for this extent to
706 * finish before going forward. */
707 conflict = osc_extent_get(ext);
711 found = osc_extent_hold(ext);
715 /* non-overlapped extent */
716 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
717 /* we can't do anything for a non OES_CACHE extent, or
718 * if there is someone waiting for this extent to be
719 * flushed, try next one. */
720 ext = next_extent(ext);
724 /* check if they belong to the same rpc slot before trying to
725 * merge. the extents are not overlapped and contiguous at
726 * chunk level to get here. */
727 if (ext->oe_max_end != max_end) {
728 /* if they don't belong to the same RPC slot or
729 * max_pages_per_rpc has ever changed, do not merge. */
730 ext = next_extent(ext);
734 /* it's required that an extent must be contiguous at chunk
735 * level so that we know the whole extent is covered by grant
736 * (the pages in the extent are NOT required to be contiguous).
737 * Otherwise, it will be too much difficult to know which
738 * chunks have grants allocated. */
740 /* try to do front merge - extend ext's start */
741 if (chunk + 1 == ext_chk_start) {
742 /* ext must be chunk size aligned */
743 EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
745 /* pull ext's start back to cover cur */
746 ext->oe_start = cur->oe_start;
747 ext->oe_grants += chunksize;
748 *grants -= chunksize;
750 found = osc_extent_hold(ext);
751 } else if (chunk == ext_chk_end + 1) {
753 ext->oe_end = cur->oe_end;
754 ext->oe_grants += chunksize;
755 *grants -= chunksize;
757 /* try to merge with the next one because we just fill
759 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
760 /* we can save extent tax from next extent */
761 *grants += cli->cl_extent_tax;
763 found = osc_extent_hold(ext);
768 ext = next_extent(ext);
771 osc_extent_tree_dump(D_CACHE, obj);
773 LASSERT(conflict == NULL);
774 if (!IS_ERR(found)) {
775 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
776 OSC_EXTENT_DUMP(D_CACHE, found,
777 "found caching ext for %lu.\n", index);
779 } else if (conflict == NULL) {
780 /* create a new extent */
781 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
782 cur->oe_grants = chunksize + cli->cl_extent_tax;
783 *grants -= cur->oe_grants;
784 LASSERT(*grants >= 0);
786 cur->oe_state = OES_CACHE;
787 found = osc_extent_hold(cur);
788 osc_extent_insert(obj, cur);
789 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
790 index, descr->cld_end);
792 osc_object_unlock(obj);
794 if (conflict != NULL) {
795 LASSERT(found == NULL);
797 /* waiting for IO to finish. Please notice that it's impossible
798 * to be an OES_TRUNC extent. */
799 rc = osc_extent_wait(env, conflict, OES_INV);
800 osc_extent_put(env, conflict);
803 GOTO(out, found = ERR_PTR(rc));
810 osc_extent_put(env, cur);
811 LASSERT(*grants >= 0);
816 * Called when IO is finished to an extent.
818 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
821 struct client_obd *cli = osc_cli(ext->oe_obj);
822 struct osc_async_page *oap;
823 struct osc_async_page *tmp;
824 int nr_pages = ext->oe_nr_pages;
826 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
831 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
833 ext->oe_rc = rc ?: ext->oe_nr_pages;
834 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
836 osc_lru_add_batch(cli, &ext->oe_pages);
837 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
839 list_del_init(&oap->oap_rpc_item);
840 list_del_init(&oap->oap_pending_item);
841 if (last_off <= oap->oap_obj_off) {
842 last_off = oap->oap_obj_off;
843 last_count = oap->oap_count;
847 osc_ap_completion(env, cli, oap, sent, rc);
849 EASSERT(ext->oe_nr_pages == 0, ext);
852 lost_grant = ext->oe_grants;
853 } else if (blocksize < PAGE_CACHE_SIZE &&
854 last_count != PAGE_CACHE_SIZE) {
855 /* For short writes we shouldn't count parts of pages that
856 * span a whole chunk on the OST side, or our accounting goes
857 * wrong. Should match the code in filter_grant_check. */
858 int offset = last_off & ~CFS_PAGE_MASK;
859 int count = last_count + (offset & (blocksize - 1));
860 int end = (offset + last_count) & (blocksize - 1);
862 count += blocksize - end;
864 lost_grant = PAGE_CACHE_SIZE - count;
866 if (ext->oe_grants > 0)
867 osc_free_grant(cli, nr_pages, lost_grant);
869 osc_extent_remove(ext);
870 /* put the refcount for RPC */
871 osc_extent_put(env, ext);
875 static int extent_wait_cb(struct osc_extent *ext, int state)
879 osc_object_lock(ext->oe_obj);
880 ret = ext->oe_state == state;
881 osc_object_unlock(ext->oe_obj);
887 * Wait for the extent's state to become @state.
889 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
892 struct osc_object *obj = ext->oe_obj;
893 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
894 LWI_ON_SIGNAL_NOOP, NULL);
898 osc_object_lock(obj);
899 LASSERT(sanity_check_nolock(ext) == 0);
900 /* `Kick' this extent only if the caller is waiting for it to be
902 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
903 if (ext->oe_state == OES_ACTIVE) {
905 } else if (ext->oe_state == OES_CACHE) {
907 osc_extent_hold(ext);
911 osc_object_unlock(obj);
913 osc_extent_release(env, ext);
915 /* wait for the extent until its state becomes @state */
916 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
917 if (rc == -ETIMEDOUT) {
918 OSC_EXTENT_DUMP(D_ERROR, ext,
919 "%s: wait ext to %d timedout, recovery in progress?\n",
920 osc_export(obj)->exp_obd->obd_name, state);
922 lwi = LWI_INTR(NULL, NULL);
923 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
926 if (rc == 0 && ext->oe_rc < 0)
932 * Discard pages with index greater than @size. If @ext is overlapped with
933 * @size, then partial truncate happens.
935 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
938 struct cl_env_nest nest;
941 struct osc_object *obj = ext->oe_obj;
942 struct client_obd *cli = osc_cli(obj);
943 struct osc_async_page *oap;
944 struct osc_async_page *tmp;
945 int pages_in_chunk = 0;
946 int ppc_bits = cli->cl_chunkbits -
948 __u64 trunc_chunk = trunc_index >> ppc_bits;
954 LASSERT(sanity_check(ext) == 0);
955 LASSERT(ext->oe_state == OES_TRUNC);
956 LASSERT(!ext->oe_urgent);
958 /* Request new lu_env.
959 * We can't use that env from osc_cache_truncate_start() because
960 * it's from lov_io_sub and not fully initialized. */
961 env = cl_env_nested_get(&nest);
962 io = &osc_env_info(env)->oti_io;
963 io->ci_obj = cl_object_top(osc2cl(obj));
964 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
968 /* discard all pages with index greater then trunc_index */
969 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
971 pgoff_t index = osc_index(oap2osc(oap));
972 struct cl_page *page = oap2cl_page(oap);
974 LASSERT(list_empty(&oap->oap_rpc_item));
976 /* only discard the pages with their index greater than
977 * trunc_index, and ... */
978 if (index < trunc_index ||
979 (index == trunc_index && partial)) {
980 /* accounting how many pages remaining in the chunk
981 * so that we can calculate grants correctly. */
982 if (index >> ppc_bits == trunc_chunk)
987 list_del_init(&oap->oap_pending_item);
990 lu_ref_add(&page->cp_reference, "truncate", current);
992 if (cl_page_own(env, io, page) == 0) {
993 cl_page_discard(env, io, page);
994 cl_page_disown(env, io, page);
996 LASSERT(page->cp_state == CPS_FREEING);
1000 lu_ref_del(&page->cp_reference, "truncate", current);
1001 cl_page_put(env, page);
1006 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1007 ext->oe_nr_pages == 0),
1008 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1010 osc_object_lock(obj);
1011 if (ext->oe_nr_pages == 0) {
1012 LASSERT(pages_in_chunk == 0);
1013 grants = ext->oe_grants;
1015 } else { /* calculate how many grants we can free */
1016 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1020 /* if there is no pages in this chunk, we can also free grants
1021 * for the last chunk */
1022 if (pages_in_chunk == 0) {
1023 /* if this is the 1st chunk and no pages in this chunk,
1024 * ext->oe_nr_pages must be zero, so we should be in
1025 * the other if-clause. */
1026 LASSERT(trunc_chunk > 0);
1031 /* this is what we can free from this extent */
1032 grants = chunks << cli->cl_chunkbits;
1033 ext->oe_grants -= grants;
1034 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1035 ext->oe_end = min(last_index, ext->oe_max_end);
1036 LASSERT(ext->oe_end >= ext->oe_start);
1037 LASSERT(ext->oe_grants > 0);
1039 osc_object_unlock(obj);
1041 if (grants > 0 || nr_pages > 0)
1042 osc_free_grant(cli, nr_pages, grants);
1045 cl_io_fini(env, io);
1046 cl_env_nested_put(&nest, env);
1051 * This function is used to make the extent prepared for transfer.
1052 * A race with flusing page - ll_writepage() has to be handled cautiously.
1054 static int osc_extent_make_ready(const struct lu_env *env,
1055 struct osc_extent *ext)
1057 struct osc_async_page *oap;
1058 struct osc_async_page *last = NULL;
1059 struct osc_object *obj = ext->oe_obj;
1060 unsigned int page_count = 0;
1064 /* we're going to grab page lock, so object lock must not be taken. */
1065 LASSERT(sanity_check(ext) == 0);
1066 /* in locking state, any process should not touch this extent. */
1067 EASSERT(ext->oe_state == OES_LOCKING, ext);
1068 EASSERT(ext->oe_owner != NULL, ext);
1070 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1072 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1074 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1077 /* checking ASYNC_READY is race safe */
1078 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1081 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1084 spin_lock(&oap->oap_lock);
1085 oap->oap_async_flags |= ASYNC_READY;
1086 spin_unlock(&oap->oap_lock);
1089 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1092 LASSERTF(0, "unknown return code: %d\n", rc);
1096 LASSERT(page_count == ext->oe_nr_pages);
1097 LASSERT(last != NULL);
1098 /* the last page is the only one we need to refresh its count by
1099 * the size of file. */
1100 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1101 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1102 LASSERT(last_oap_count > 0);
1103 LASSERT(last->oap_page_off + last_oap_count <= PAGE_CACHE_SIZE);
1104 last->oap_count = last_oap_count;
1105 spin_lock(&last->oap_lock);
1106 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1107 spin_unlock(&last->oap_lock);
1110 /* for the rest of pages, we don't need to call osf_refresh_count()
1111 * because it's known they are not the last page */
1112 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1113 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1114 oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1115 spin_lock(&oap->oap_lock);
1116 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1117 spin_unlock(&oap->oap_lock);
1121 osc_object_lock(obj);
1122 osc_extent_state_set(ext, OES_RPC);
1123 osc_object_unlock(obj);
1124 /* get a refcount for RPC. */
1125 osc_extent_get(ext);
1131 * Quick and simple version of osc_extent_find(). This function is frequently
1132 * called to expand the extent for the same IO. To expand the extent, the
1133 * page index must be in the same or next chunk of ext->oe_end.
1135 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1137 struct osc_object *obj = ext->oe_obj;
1138 struct client_obd *cli = osc_cli(obj);
1139 struct osc_extent *next;
1140 int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1141 pgoff_t chunk = index >> ppc_bits;
1144 int chunksize = 1 << cli->cl_chunkbits;
1148 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1149 osc_object_lock(obj);
1150 LASSERT(sanity_check_nolock(ext) == 0);
1151 end_chunk = ext->oe_end >> ppc_bits;
1152 if (chunk > end_chunk + 1)
1153 GOTO(out, rc = -ERANGE);
1155 if (end_chunk >= chunk)
1158 LASSERT(end_chunk + 1 == chunk);
1159 /* try to expand this extent to cover @index */
1160 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1162 next = next_extent(ext);
1163 if (next != NULL && next->oe_start <= end_index)
1164 /* complex mode - overlapped with the next extent,
1165 * this case will be handled by osc_extent_find() */
1166 GOTO(out, rc = -EAGAIN);
1168 ext->oe_end = end_index;
1169 ext->oe_grants += chunksize;
1170 *grants -= chunksize;
1171 LASSERT(*grants >= 0);
1172 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1173 "overlapped after expanding for %lu.\n", index);
1177 osc_object_unlock(obj);
1181 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1182 const char *func, int line)
1184 struct osc_extent *ext;
1187 CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1188 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1190 /* osc_object_lock(obj); */
1192 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1193 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1196 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1197 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1200 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1201 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1204 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1205 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1206 /* osc_object_unlock(obj); */
1209 /* ------------------ osc extent end ------------------ */
1211 static inline int osc_is_ready(struct osc_object *osc)
1213 return !list_empty(&osc->oo_ready_item) ||
1214 !list_empty(&osc->oo_hp_ready_item);
1217 #define OSC_IO_DEBUG(OSC, STR, args...) \
1218 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1219 (OSC), osc_is_ready(OSC), \
1220 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1221 list_empty_marker(&(OSC)->oo_ready_item), \
1222 atomic_read(&(OSC)->oo_nr_writes), \
1223 list_empty_marker(&(OSC)->oo_hp_exts), \
1224 list_empty_marker(&(OSC)->oo_urgent_exts), \
1225 atomic_read(&(OSC)->oo_nr_reads), \
1226 list_empty_marker(&(OSC)->oo_reading_exts), \
1229 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1232 struct osc_page *opg = oap2osc_page(oap);
1233 struct cl_page *page = oap2cl_page(oap);
1236 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1239 result = cl_page_make_ready(env, page, CRT_WRITE);
1241 opg->ops_submit_time = cfs_time_current();
1245 static int osc_refresh_count(const struct lu_env *env,
1246 struct osc_async_page *oap, int cmd)
1248 struct osc_page *opg = oap2osc_page(oap);
1249 pgoff_t index = osc_index(oap2osc(oap));
1250 struct cl_object *obj;
1251 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1256 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1257 LASSERT(!(cmd & OBD_BRW_READ));
1258 LASSERT(opg != NULL);
1259 obj = opg->ops_cl.cpl_obj;
1261 cl_object_attr_lock(obj);
1262 result = cl_object_attr_get(env, obj, attr);
1263 cl_object_attr_unlock(obj);
1266 kms = attr->cat_kms;
1267 if (cl_offset(obj, index) >= kms)
1268 /* catch race with truncate */
1270 else if (cl_offset(obj, index + 1) > kms)
1271 /* catch sub-page write at end of file */
1272 return kms % PAGE_CACHE_SIZE;
1274 return PAGE_CACHE_SIZE;
1277 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1280 struct osc_page *opg = oap2osc_page(oap);
1281 struct cl_page *page = oap2cl_page(oap);
1282 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
1283 enum cl_req_type crt;
1288 cmd &= ~OBD_BRW_NOQUOTA;
1289 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1290 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1291 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1292 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1293 LASSERT(opg->ops_transfer_pinned);
1296 * page->cp_req can be NULL if io submission failed before
1297 * cl_req was allocated.
1299 if (page->cp_req != NULL)
1300 cl_req_page_done(env, page);
1301 LASSERT(page->cp_req == NULL);
1303 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1304 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1305 opg->ops_transfer_pinned = 0;
1307 spin_lock(&obj->oo_seatbelt);
1308 LASSERT(opg->ops_submitter != NULL);
1309 LASSERT(!list_empty(&opg->ops_inflight));
1310 list_del_init(&opg->ops_inflight);
1311 opg->ops_submitter = NULL;
1312 spin_unlock(&obj->oo_seatbelt);
1314 opg->ops_submit_time = 0;
1315 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1318 if (rc == 0 && srvlock) {
1319 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1320 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1321 int bytes = oap->oap_count;
1323 if (crt == CRT_READ)
1324 stats->os_lockless_reads += bytes;
1326 stats->os_lockless_writes += bytes;
1330 * This has to be the last operation with the page, as locks are
1331 * released in cl_page_completion() and nothing except for the
1332 * reference counter protects page from concurrent reclaim.
1334 lu_ref_del(&page->cp_reference, "transfer", page);
1336 cl_page_completion(env, page, crt, rc);
1341 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do { \
1342 struct client_obd *__tmp = (cli); \
1343 CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1344 "dropped: %ld avail: %ld, reserved: %ld, flight: %d }" \
1345 "lru {in list: %ld, left: %ld, waiters: %d }"fmt"\n", \
1346 __tmp->cl_import->imp_obd->obd_name, \
1347 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1348 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1349 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1350 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1351 atomic_long_read(&__tmp->cl_lru_in_list), \
1352 atomic_long_read(&__tmp->cl_lru_busy), \
1353 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1356 /* caller must hold loi_list_lock */
1357 static void osc_consume_write_grant(struct client_obd *cli,
1358 struct brw_page *pga)
1360 assert_spin_locked(&cli->cl_loi_list_lock);
1361 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1362 atomic_long_inc(&obd_dirty_pages);
1363 cli->cl_dirty_pages++;
1364 pga->flag |= OBD_BRW_FROM_GRANT;
1365 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1366 PAGE_CACHE_SIZE, pga, pga->pg);
1367 osc_update_next_shrink(cli);
1370 /* the companion to osc_consume_write_grant, called when a brw has completed.
1371 * must be called with the loi lock held. */
1372 static void osc_release_write_grant(struct client_obd *cli,
1373 struct brw_page *pga)
1377 assert_spin_locked(&cli->cl_loi_list_lock);
1378 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1383 pga->flag &= ~OBD_BRW_FROM_GRANT;
1384 atomic_long_dec(&obd_dirty_pages);
1385 cli->cl_dirty_pages--;
1386 if (pga->flag & OBD_BRW_NOCACHE) {
1387 pga->flag &= ~OBD_BRW_NOCACHE;
1388 atomic_long_dec(&obd_dirty_transit_pages);
1389 cli->cl_dirty_transit--;
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;
1424 cli->cl_avail_grant += unused;
1428 static void osc_unreserve_grant(struct client_obd *cli,
1429 unsigned int reserved, unsigned int unused)
1431 spin_lock(&cli->cl_loi_list_lock);
1432 __osc_unreserve_grant(cli, reserved, unused);
1434 osc_wake_cache_waiters(cli);
1435 spin_unlock(&cli->cl_loi_list_lock);
1439 * Free grant after IO is finished or canceled.
1441 * @lost_grant is used to remember how many grants we have allocated but not
1442 * used, we should return these grants to OST. There're two cases where grants
1445 * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1446 * written. In this case OST may use less chunks to serve this partial
1447 * write. OSTs don't actually know the page size on the client side. so
1448 * clients have to calculate lost grant by the blocksize on the OST.
1449 * See filter_grant_check() for details.
1451 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1452 unsigned int lost_grant)
1454 int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1456 spin_lock(&cli->cl_loi_list_lock);
1457 atomic_long_sub(nr_pages, &obd_dirty_pages);
1458 cli->cl_dirty_pages -= nr_pages;
1459 cli->cl_lost_grant += lost_grant;
1460 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1461 /* borrow some grant from truncate to avoid the case that
1462 * truncate uses up all avail grant */
1463 cli->cl_lost_grant -= grant;
1464 cli->cl_avail_grant += grant;
1466 osc_wake_cache_waiters(cli);
1467 spin_unlock(&cli->cl_loi_list_lock);
1468 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1469 lost_grant, cli->cl_lost_grant,
1470 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
1474 * The companion to osc_enter_cache(), called when @oap is no longer part of
1475 * the dirty accounting due to error.
1477 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1479 spin_lock(&cli->cl_loi_list_lock);
1480 osc_release_write_grant(cli, &oap->oap_brw_page);
1481 spin_unlock(&cli->cl_loi_list_lock);
1485 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1488 static int osc_enter_cache_try(struct client_obd *cli,
1489 struct osc_async_page *oap,
1490 int bytes, int transient)
1494 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1496 rc = osc_reserve_grant(cli, bytes);
1500 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1501 1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1502 osc_consume_write_grant(cli, &oap->oap_brw_page);
1504 cli->cl_dirty_transit++;
1505 atomic_long_inc(&obd_dirty_transit_pages);
1506 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1510 __osc_unreserve_grant(cli, bytes, bytes);
1516 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1519 spin_lock(&cli->cl_loi_list_lock);
1520 rc = list_empty(&ocw->ocw_entry);
1521 spin_unlock(&cli->cl_loi_list_lock);
1526 * The main entry to reserve dirty page accounting. Usually the grant reserved
1527 * in this function will be freed in bulk in osc_free_grant() unless it fails
1528 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1530 * The process will be put into sleep if it's already run out of grant.
1532 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1533 struct osc_async_page *oap, int bytes)
1535 struct osc_object *osc = oap->oap_obj;
1536 struct lov_oinfo *loi = osc->oo_oinfo;
1537 struct osc_cache_waiter ocw;
1538 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1539 LWI_ON_SIGNAL_NOOP, NULL);
1543 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1545 spin_lock(&cli->cl_loi_list_lock);
1547 /* force the caller to try sync io. this can jump the list
1548 * of queued writes and create a discontiguous rpc stream */
1549 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1550 cli->cl_dirty_max_pages == 0 ||
1551 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1552 GOTO(out, rc = -EDQUOT);
1554 /* Hopefully normal case - cache space and write credits available */
1555 if (osc_enter_cache_try(cli, oap, bytes, 0))
1558 /* We can get here for two reasons: too many dirty pages in cache, or
1559 * run out of grants. In both cases we should write dirty pages out.
1560 * Adding a cache waiter will trigger urgent write-out no matter what
1562 * The exiting condition is no avail grants and no dirty pages caching,
1563 * that really means there is no space on the OST. */
1564 init_waitqueue_head(&ocw.ocw_waitq);
1566 ocw.ocw_grant = bytes;
1567 while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1568 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1570 spin_unlock(&cli->cl_loi_list_lock);
1572 osc_io_unplug_async(env, cli, NULL);
1574 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1575 cli->cl_import->imp_obd->obd_name, &ocw, oap);
1577 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1579 spin_lock(&cli->cl_loi_list_lock);
1581 /* l_wait_event is interrupted by signal, or timed out */
1585 OSC_DUMP_GRANT(D_ERROR, cli,
1586 "try to reserve %d.\n", bytes);
1587 osc_extent_tree_dump(D_ERROR, osc);
1591 /* Ensures restartability - LU-3581 */
1595 CDEBUG(D_CACHE, "%s: event for cache space @"
1596 " %p never arrived due to %d\n",
1597 cli->cl_import->imp_obd->obd_name,
1601 list_del_init(&ocw.ocw_entry);
1605 LASSERT(list_empty(&ocw.ocw_entry));
1610 if (osc_enter_cache_try(cli, oap, bytes, 0))
1615 spin_unlock(&cli->cl_loi_list_lock);
1616 OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1620 /* caller must hold loi_list_lock */
1621 void osc_wake_cache_waiters(struct client_obd *cli)
1623 struct list_head *l, *tmp;
1624 struct osc_cache_waiter *ocw;
1627 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1628 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1629 list_del_init(&ocw->ocw_entry);
1631 ocw->ocw_rc = -EDQUOT;
1632 /* we can't dirty more */
1633 if ((cli->cl_dirty_pages >= cli->cl_dirty_max_pages) ||
1634 (1 + atomic_long_read(&obd_dirty_pages) >
1635 obd_max_dirty_pages)) {
1636 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1637 "osc max %ld, sys max %ld\n",
1638 cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1639 obd_max_dirty_pages);
1644 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1645 ocw->ocw_rc = -EDQUOT;
1648 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1649 ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1651 wake_up(&ocw->ocw_waitq);
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.. */
1697 if (!list_empty(&cli->cl_cache_waiters)) {
1698 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1701 if (atomic_read(&osc->oo_nr_writes) >=
1702 cli->cl_max_pages_per_rpc)
1705 if (atomic_read(&osc->oo_nr_reads) == 0)
1707 if (invalid_import) {
1708 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1711 /* all read are urgent. */
1712 if (!list_empty(&osc->oo_reading_exts))
1719 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1721 struct client_obd *cli = osc_cli(obj);
1722 if (cmd & OBD_BRW_WRITE) {
1723 atomic_add(delta, &obj->oo_nr_writes);
1724 atomic_add(delta, &cli->cl_pending_w_pages);
1725 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1727 atomic_add(delta, &obj->oo_nr_reads);
1728 atomic_add(delta, &cli->cl_pending_r_pages);
1729 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1731 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1734 static int osc_makes_hprpc(struct osc_object *obj)
1736 return !list_empty(&obj->oo_hp_exts);
1739 static void on_list(struct list_head *item, struct list_head *list,
1742 if (list_empty(item) && should_be_on)
1743 list_add_tail(item, list);
1744 else if (!list_empty(item) && !should_be_on)
1745 list_del_init(item);
1748 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1749 * can find pages to build into rpcs quickly */
1750 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1752 if (osc_makes_hprpc(osc)) {
1754 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1755 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1757 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1758 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1759 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1760 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1763 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1764 atomic_read(&osc->oo_nr_writes) > 0);
1766 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1767 atomic_read(&osc->oo_nr_reads) > 0);
1769 return osc_is_ready(osc);
1772 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1776 spin_lock(&cli->cl_loi_list_lock);
1777 is_ready = __osc_list_maint(cli, osc);
1778 spin_unlock(&cli->cl_loi_list_lock);
1783 /* this is trying to propogate async writeback errors back up to the
1784 * application. As an async write fails we record the error code for later if
1785 * the app does an fsync. As long as errors persist we force future rpcs to be
1786 * sync so that the app can get a sync error and break the cycle of queueing
1787 * pages for which writeback will fail. */
1788 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1795 ar->ar_force_sync = 1;
1796 ar->ar_min_xid = ptlrpc_sample_next_xid();
1801 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1802 ar->ar_force_sync = 0;
1805 /* this must be called holding the loi list lock to give coverage to exit_cache,
1806 * async_flag maintenance, and oap_request */
1807 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1808 struct osc_async_page *oap, int sent, int rc)
1810 struct osc_object *osc = oap->oap_obj;
1811 struct lov_oinfo *loi = osc->oo_oinfo;
1815 if (oap->oap_request != NULL) {
1816 xid = ptlrpc_req_xid(oap->oap_request);
1817 ptlrpc_req_finished(oap->oap_request);
1818 oap->oap_request = NULL;
1821 /* As the transfer for this page is being done, clear the flags */
1822 spin_lock(&oap->oap_lock);
1823 oap->oap_async_flags = 0;
1824 spin_unlock(&oap->oap_lock);
1825 oap->oap_interrupted = 0;
1827 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1828 spin_lock(&cli->cl_loi_list_lock);
1829 osc_process_ar(&cli->cl_ar, xid, rc);
1830 osc_process_ar(&loi->loi_ar, xid, rc);
1831 spin_unlock(&cli->cl_loi_list_lock);
1834 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1836 CERROR("completion on oap %p obj %p returns %d.\n",
1843 * Try to add extent to one RPC. We need to think about the following things:
1844 * - # of pages must not be over max_pages_per_rpc
1845 * - extent must be compatible with previous ones
1847 static int try_to_add_extent_for_io(struct client_obd *cli,
1848 struct osc_extent *ext,
1849 struct list_head *rpclist,
1850 int *pc, unsigned int *max_pages)
1852 struct osc_extent *tmp;
1853 struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1854 struct osc_async_page,
1858 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1861 *max_pages = max(ext->oe_mppr, *max_pages);
1862 if (*pc + ext->oe_nr_pages > *max_pages)
1865 list_for_each_entry(tmp, rpclist, oe_link) {
1866 struct osc_async_page *oap2;
1867 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1869 EASSERT(tmp->oe_owner == current, tmp);
1871 if (overlapped(tmp, ext)) {
1872 OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1876 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1877 CDEBUG(D_CACHE, "Do not permit different type of IO"
1878 " for a same RPC\n");
1882 if (tmp->oe_srvlock != ext->oe_srvlock ||
1883 !tmp->oe_grants != !ext->oe_grants)
1886 /* remove break for strict check */
1890 *pc += ext->oe_nr_pages;
1891 list_move_tail(&ext->oe_link, rpclist);
1892 ext->oe_owner = current;
1897 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1898 * get_write_extent() takes all appropriate extents in atomic.
1900 * The following policy is used to collect extents for IO:
1901 * 1. Add as many HP extents as possible;
1902 * 2. Add the first urgent extent in urgent extent list and take it out of
1904 * 3. Add subsequent extents of this urgent extent;
1905 * 4. If urgent list is not empty, goto 2;
1906 * 5. Traverse the extent tree from the 1st extent;
1907 * 6. Above steps exit if there is no space in this RPC.
1909 static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1911 struct client_obd *cli = osc_cli(obj);
1912 struct osc_extent *ext;
1914 unsigned int max_pages = cli->cl_max_pages_per_rpc;
1916 LASSERT(osc_object_is_locked(obj));
1917 while (!list_empty(&obj->oo_hp_exts)) {
1918 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1920 LASSERT(ext->oe_state == OES_CACHE);
1921 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1924 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1926 if (page_count == max_pages)
1929 while (!list_empty(&obj->oo_urgent_exts)) {
1930 ext = list_entry(obj->oo_urgent_exts.next,
1931 struct osc_extent, oe_link);
1932 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1936 if (!ext->oe_intree)
1939 while ((ext = next_extent(ext)) != NULL) {
1940 if ((ext->oe_state != OES_CACHE) ||
1941 (!list_empty(&ext->oe_link) &&
1942 ext->oe_owner != NULL))
1945 if (!try_to_add_extent_for_io(cli, ext, rpclist,
1946 &page_count, &max_pages))
1950 if (page_count == max_pages)
1953 ext = first_extent(obj);
1954 while (ext != NULL) {
1955 if ((ext->oe_state != OES_CACHE) ||
1956 /* this extent may be already in current rpclist */
1957 (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1958 ext = next_extent(ext);
1962 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1966 ext = next_extent(ext);
1972 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1973 struct osc_object *osc, pdl_policy_t pol)
1976 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
1977 struct osc_extent *ext;
1978 struct osc_extent *tmp;
1979 struct osc_extent *first = NULL;
1980 obd_count page_count = 0;
1985 LASSERT(osc_object_is_locked(osc));
1987 page_count = get_write_extents(osc, &rpclist);
1988 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1990 if (list_empty(&rpclist))
1993 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1995 list_for_each_entry(ext, &rpclist, oe_link) {
1996 LASSERT(ext->oe_state == OES_CACHE ||
1997 ext->oe_state == OES_LOCK_DONE);
1998 if (ext->oe_state == OES_CACHE)
1999 osc_extent_state_set(ext, OES_LOCKING);
2001 osc_extent_state_set(ext, OES_RPC);
2004 /* we're going to grab page lock, so release object lock because
2005 * lock order is page lock -> object lock. */
2006 osc_object_unlock(osc);
2008 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2009 if (ext->oe_state == OES_LOCKING) {
2010 rc = osc_extent_make_ready(env, ext);
2011 if (unlikely(rc < 0)) {
2012 list_del_init(&ext->oe_link);
2013 osc_extent_finish(env, ext, 0, rc);
2017 if (first == NULL) {
2019 srvlock = ext->oe_srvlock;
2021 LASSERT(srvlock == ext->oe_srvlock);
2025 if (!list_empty(&rpclist)) {
2026 LASSERT(page_count > 0);
2027 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2028 LASSERT(list_empty(&rpclist));
2031 osc_object_lock(osc);
2036 * prepare pages for ASYNC io and put pages in send queue.
2038 * \param cmd OBD_BRW_* macroses
2039 * \param lop pending pages
2041 * \return zero if no page added to send queue.
2042 * \return 1 if pages successfully added to send queue.
2043 * \return negative on errors.
2046 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2047 struct osc_object *osc, pdl_policy_t pol)
2050 struct osc_extent *ext;
2051 struct osc_extent *next;
2052 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2054 unsigned int max_pages = cli->cl_max_pages_per_rpc;
2058 LASSERT(osc_object_is_locked(osc));
2059 list_for_each_entry_safe(ext, next,
2060 &osc->oo_reading_exts, oe_link) {
2061 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2062 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2065 osc_extent_state_set(ext, OES_RPC);
2066 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2068 LASSERT(page_count <= max_pages);
2070 osc_update_pending(osc, OBD_BRW_READ, -page_count);
2072 if (!list_empty(&rpclist)) {
2073 osc_object_unlock(osc);
2075 LASSERT(page_count > 0);
2076 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2077 LASSERT(list_empty(&rpclist));
2079 osc_object_lock(osc);
2084 #define list_to_obj(list, item) ({ \
2085 struct list_head *__tmp = (list)->next; \
2086 list_del_init(__tmp); \
2087 list_entry(__tmp, struct osc_object, oo_##item); \
2090 /* This is called by osc_check_rpcs() to find which objects have pages that
2091 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2092 static struct osc_object *osc_next_obj(struct client_obd *cli)
2096 /* First return objects that have blocked locks so that they
2097 * will be flushed quickly and other clients can get the lock,
2098 * then objects which have pages ready to be stuffed into RPCs */
2099 if (!list_empty(&cli->cl_loi_hp_ready_list))
2100 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2101 if (!list_empty(&cli->cl_loi_ready_list))
2102 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2104 /* then if we have cache waiters, return all objects with queued
2105 * writes. This is especially important when many small files
2106 * have filled up the cache and not been fired into rpcs because
2107 * they don't pass the nr_pending/object threshhold */
2108 if (!list_empty(&cli->cl_cache_waiters) &&
2109 !list_empty(&cli->cl_loi_write_list))
2110 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2112 /* then return all queued objects when we have an invalid import
2113 * so that they get flushed */
2114 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2115 if (!list_empty(&cli->cl_loi_write_list))
2116 RETURN(list_to_obj(&cli->cl_loi_write_list,
2118 if (!list_empty(&cli->cl_loi_read_list))
2119 RETURN(list_to_obj(&cli->cl_loi_read_list,
2125 /* called with the loi list lock held */
2126 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2128 __must_hold(&cli->cl_loi_list_lock)
2130 struct osc_object *osc;
2134 while ((osc = osc_next_obj(cli)) != NULL) {
2135 struct cl_object *obj = osc2cl(osc);
2136 struct lu_ref_link link;
2138 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2140 if (osc_max_rpc_in_flight(cli, osc)) {
2141 __osc_list_maint(cli, osc);
2146 spin_unlock(&cli->cl_loi_list_lock);
2147 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2149 /* attempt some read/write balancing by alternating between
2150 * reads and writes in an object. The makes_rpc checks here
2151 * would be redundant if we were getting read/write work items
2152 * instead of objects. we don't want send_oap_rpc to drain a
2153 * partial read pending queue when we're given this object to
2154 * do io on writes while there are cache waiters */
2155 osc_object_lock(osc);
2156 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2157 rc = osc_send_write_rpc(env, cli, osc, pol);
2159 CERROR("Write request failed with %d\n", rc);
2161 /* osc_send_write_rpc failed, mostly because of
2164 * It can't break here, because if:
2165 * - a page was submitted by osc_io_submit, so
2167 * - no request in flight
2168 * - no subsequent request
2169 * The system will be in live-lock state,
2170 * because there is no chance to call
2171 * osc_io_unplug() and osc_check_rpcs() any
2172 * more. pdflush can't help in this case,
2173 * because it might be blocked at grabbing
2174 * the page lock as we mentioned.
2176 * Anyway, continue to drain pages. */
2180 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2181 rc = osc_send_read_rpc(env, cli, osc, pol);
2183 CERROR("Read request failed with %d\n", rc);
2185 osc_object_unlock(osc);
2187 osc_list_maint(cli, osc);
2188 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2189 cl_object_put(env, obj);
2191 spin_lock(&cli->cl_loi_list_lock);
2195 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2196 struct osc_object *osc, pdl_policy_t pol, int async)
2200 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2204 /* disable osc_lru_shrink() temporarily to avoid
2205 * potential stack overrun problem. LU-2859 */
2206 atomic_inc(&cli->cl_lru_shrinkers);
2207 spin_lock(&cli->cl_loi_list_lock);
2208 osc_check_rpcs(env, cli, pol);
2209 spin_unlock(&cli->cl_loi_list_lock);
2210 atomic_dec(&cli->cl_lru_shrinkers);
2212 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2213 LASSERT(cli->cl_writeback_work != NULL);
2214 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2219 static int osc_io_unplug_async(const struct lu_env *env,
2220 struct client_obd *cli, struct osc_object *osc)
2222 /* XXX: policy is no use actually. */
2223 return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2226 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2227 struct osc_object *osc, pdl_policy_t pol)
2229 (void)osc_io_unplug0(env, cli, osc, pol, 0);
2232 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2233 struct page *page, loff_t offset)
2235 struct obd_export *exp = osc_export(osc);
2236 struct osc_async_page *oap = &ops->ops_oap;
2240 return cfs_size_round(sizeof(*oap));
2242 oap->oap_magic = OAP_MAGIC;
2243 oap->oap_cli = &exp->exp_obd->u.cli;
2246 oap->oap_page = page;
2247 oap->oap_obj_off = offset;
2248 LASSERT(!(offset & ~CFS_PAGE_MASK));
2250 if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2251 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2253 INIT_LIST_HEAD(&oap->oap_pending_item);
2254 INIT_LIST_HEAD(&oap->oap_rpc_item);
2256 spin_lock_init(&oap->oap_lock);
2257 CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2258 oap, page, oap->oap_obj_off);
2262 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2263 struct osc_page *ops)
2265 struct osc_io *oio = osc_env_io(env);
2266 struct osc_extent *ext = NULL;
2267 struct osc_async_page *oap = &ops->ops_oap;
2268 struct client_obd *cli = oap->oap_cli;
2269 struct osc_object *osc = oap->oap_obj;
2272 int brw_flags = OBD_BRW_ASYNC;
2273 int cmd = OBD_BRW_WRITE;
2274 int need_release = 0;
2278 if (oap->oap_magic != OAP_MAGIC)
2281 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2284 if (!list_empty(&oap->oap_pending_item) ||
2285 !list_empty(&oap->oap_rpc_item))
2288 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2289 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2290 if (!client_is_remote(osc_export(osc)) &&
2291 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2292 brw_flags |= OBD_BRW_NOQUOTA;
2293 cmd |= OBD_BRW_NOQUOTA;
2296 /* check if the file's owner/group is over quota */
2297 if (!(cmd & OBD_BRW_NOQUOTA)) {
2298 struct cl_object *obj;
2299 struct cl_attr *attr;
2300 unsigned int qid[MAXQUOTAS];
2302 obj = cl_object_top(&osc->oo_cl);
2303 attr = &osc_env_info(env)->oti_attr;
2305 cl_object_attr_lock(obj);
2306 rc = cl_object_attr_get(env, obj, attr);
2307 cl_object_attr_unlock(obj);
2309 qid[USRQUOTA] = attr->cat_uid;
2310 qid[GRPQUOTA] = attr->cat_gid;
2311 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2318 oap->oap_page_off = ops->ops_from;
2319 oap->oap_count = ops->ops_to - ops->ops_from;
2320 /* No need to hold a lock here,
2321 * since this page is not in any list yet. */
2322 oap->oap_async_flags = 0;
2323 oap->oap_brw_flags = brw_flags;
2325 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2326 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2328 index = osc_index(oap2osc(oap));
2330 /* Add this page into extent by the following steps:
2331 * 1. if there exists an active extent for this IO, mostly this page
2332 * can be added to the active extent and sometimes we need to
2333 * expand extent to accomodate this page;
2334 * 2. otherwise, a new extent will be allocated. */
2336 ext = oio->oi_active;
2337 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2338 /* one chunk plus extent overhead must be enough to write this
2340 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2341 if (ext->oe_end >= index)
2344 /* it doesn't need any grant to dirty this page */
2345 spin_lock(&cli->cl_loi_list_lock);
2346 rc = osc_enter_cache_try(cli, oap, grants, 0);
2347 spin_unlock(&cli->cl_loi_list_lock);
2348 if (rc == 0) { /* try failed */
2351 } else if (ext->oe_end < index) {
2353 /* try to expand this extent */
2354 rc = osc_extent_expand(ext, index, &tmp);
2357 /* don't free reserved grant */
2359 OSC_EXTENT_DUMP(D_CACHE, ext,
2360 "expanded for %lu.\n", index);
2361 osc_unreserve_grant(cli, grants, tmp);
2366 } else if (ext != NULL) {
2367 /* index is located outside of active extent */
2371 osc_extent_release(env, ext);
2372 oio->oi_active = NULL;
2377 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2379 /* try to find new extent to cover this page */
2380 LASSERT(oio->oi_active == NULL);
2381 /* we may have allocated grant for this page if we failed
2382 * to expand the previous active extent. */
2383 LASSERT(ergo(grants > 0, grants >= tmp));
2387 /* we haven't allocated grant for this page. */
2388 rc = osc_enter_cache(env, cli, oap, tmp);
2395 ext = osc_extent_find(env, osc, index, &tmp);
2397 LASSERT(tmp == grants);
2398 osc_exit_cache(cli, oap);
2402 oio->oi_active = ext;
2406 osc_unreserve_grant(cli, grants, tmp);
2409 LASSERT(ergo(rc == 0, ext != NULL));
2411 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2412 ext, "index = %lu.\n", index);
2413 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2415 osc_object_lock(osc);
2416 if (ext->oe_nr_pages == 0)
2417 ext->oe_srvlock = ops->ops_srvlock;
2419 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2421 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2422 osc_object_unlock(osc);
2427 int osc_teardown_async_page(const struct lu_env *env,
2428 struct osc_object *obj, struct osc_page *ops)
2430 struct osc_async_page *oap = &ops->ops_oap;
2431 struct osc_extent *ext = NULL;
2435 LASSERT(oap->oap_magic == OAP_MAGIC);
2437 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2438 oap, ops, osc_index(oap2osc(oap)));
2440 osc_object_lock(obj);
2441 if (!list_empty(&oap->oap_rpc_item)) {
2442 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2444 } else if (!list_empty(&oap->oap_pending_item)) {
2445 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2446 /* only truncated pages are allowed to be taken out.
2447 * See osc_extent_truncate() and osc_cache_truncate_start()
2449 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2450 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2451 osc_index(oap2osc(oap)));
2455 osc_object_unlock(obj);
2457 osc_extent_put(env, ext);
2462 * This is called when a page is picked up by kernel to write out.
2464 * We should find out the corresponding extent and add the whole extent
2465 * into urgent list. The extent may be being truncated or used, handle it
2468 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2469 struct osc_page *ops)
2471 struct osc_extent *ext = NULL;
2472 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2473 struct cl_page *cp = ops->ops_cl.cpl_page;
2474 pgoff_t index = osc_index(ops);
2475 struct osc_async_page *oap = &ops->ops_oap;
2476 bool unplug = false;
2480 osc_object_lock(obj);
2481 ext = osc_extent_lookup(obj, index);
2483 osc_extent_tree_dump(D_ERROR, obj);
2484 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2487 switch (ext->oe_state) {
2490 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2494 /* If we know this extent is being written out, we should abort
2495 * so that the writer can make this page ready. Otherwise, there
2496 * exists a deadlock problem because other process can wait for
2497 * page writeback bit holding page lock; and meanwhile in
2498 * vvp_page_make_ready(), we need to grab page lock before
2499 * really sending the RPC. */
2501 /* race with truncate, page will be redirtied */
2503 /* The extent is active so we need to abort and let the caller
2504 * re-dirty the page. If we continued on here, and we were the
2505 * one making the extent active, we could deadlock waiting for
2506 * the page writeback to clear but it won't because the extent
2507 * is active and won't be written out. */
2508 GOTO(out, rc = -EAGAIN);
2513 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2517 spin_lock(&oap->oap_lock);
2518 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2519 spin_unlock(&oap->oap_lock);
2521 if (memory_pressure_get())
2522 ext->oe_memalloc = 1;
2525 if (ext->oe_state == OES_CACHE) {
2526 OSC_EXTENT_DUMP(D_CACHE, ext,
2527 "flush page %p make it urgent.\n", oap);
2528 if (list_empty(&ext->oe_link))
2529 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2536 osc_object_unlock(obj);
2537 osc_extent_put(env, ext);
2539 osc_io_unplug_async(env, osc_cli(obj), obj);
2544 * this is called when a sync waiter receives an interruption. Its job is to
2545 * get the caller woken as soon as possible. If its page hasn't been put in an
2546 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2547 * desiring interruption which will forcefully complete the rpc once the rpc
2550 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2552 struct osc_async_page *oap = &ops->ops_oap;
2553 struct osc_object *obj = oap->oap_obj;
2554 struct client_obd *cli = osc_cli(obj);
2555 struct osc_extent *ext;
2556 struct osc_extent *found = NULL;
2557 struct list_head *plist;
2558 pgoff_t index = osc_index(ops);
2563 LASSERT(!oap->oap_interrupted);
2564 oap->oap_interrupted = 1;
2566 /* Find out the caching extent */
2567 osc_object_lock(obj);
2568 if (oap->oap_cmd & OBD_BRW_WRITE) {
2569 plist = &obj->oo_urgent_exts;
2570 cmd = OBD_BRW_WRITE;
2572 plist = &obj->oo_reading_exts;
2575 list_for_each_entry(ext, plist, oe_link) {
2576 if (ext->oe_start <= index && ext->oe_end >= index) {
2577 LASSERT(ext->oe_state == OES_LOCK_DONE);
2578 /* For OES_LOCK_DONE state extent, it has already held
2579 * a refcount for RPC. */
2580 found = osc_extent_get(ext);
2584 if (found != NULL) {
2585 list_del_init(&found->oe_link);
2586 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2587 osc_object_unlock(obj);
2589 osc_extent_finish(env, found, 0, -EINTR);
2590 osc_extent_put(env, found);
2593 osc_object_unlock(obj);
2594 /* ok, it's been put in an rpc. only one oap gets a request
2596 if (oap->oap_request != NULL) {
2597 ptlrpc_mark_interrupted(oap->oap_request);
2598 ptlrpcd_wake(oap->oap_request);
2599 ptlrpc_req_finished(oap->oap_request);
2600 oap->oap_request = NULL;
2604 osc_list_maint(cli, obj);
2608 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2609 struct list_head *list, int cmd, int brw_flags)
2611 struct client_obd *cli = osc_cli(obj);
2612 struct osc_extent *ext;
2613 struct osc_async_page *oap;
2615 int mppr = cli->cl_max_pages_per_rpc;
2616 pgoff_t start = CL_PAGE_EOF;
2620 list_for_each_entry(oap, list, oap_pending_item) {
2621 pgoff_t index = osc_index(oap2osc(oap));
2627 mppr <<= (page_count > mppr);
2630 ext = osc_extent_alloc(obj);
2632 list_for_each_entry(oap, list, oap_pending_item) {
2633 list_del_init(&oap->oap_pending_item);
2634 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2639 ext->oe_rw = !!(cmd & OBD_BRW_READ);
2642 ext->oe_start = start;
2643 ext->oe_end = ext->oe_max_end = end;
2645 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2646 ext->oe_nr_pages = page_count;
2647 ext->oe_mppr = mppr;
2648 list_splice_init(list, &ext->oe_pages);
2650 osc_object_lock(obj);
2651 /* Reuse the initial refcount for RPC, don't drop it */
2652 osc_extent_state_set(ext, OES_LOCK_DONE);
2653 if (cmd & OBD_BRW_WRITE) {
2654 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2655 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2657 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2658 osc_update_pending(obj, OBD_BRW_READ, page_count);
2660 osc_object_unlock(obj);
2662 osc_io_unplug_async(env, cli, obj);
2667 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2669 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2670 struct osc_object *obj, __u64 size)
2672 struct client_obd *cli = osc_cli(obj);
2673 struct osc_extent *ext;
2674 struct osc_extent *waiting = NULL;
2676 struct list_head list = LIST_HEAD_INIT(list);
2681 /* pages with index greater or equal to index will be truncated. */
2682 index = cl_index(osc2cl(obj), size);
2683 partial = size > cl_offset(osc2cl(obj), index);
2686 osc_object_lock(obj);
2687 ext = osc_extent_search(obj, index);
2689 ext = first_extent(obj);
2690 else if (ext->oe_end < index)
2691 ext = next_extent(ext);
2692 while (ext != NULL) {
2693 EASSERT(ext->oe_state != OES_TRUNC, ext);
2695 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2696 /* if ext is in urgent state, it means there must exist
2697 * a page already having been flushed by write_page().
2698 * We have to wait for this extent because we can't
2699 * truncate that page. */
2700 LASSERT(!ext->oe_hp);
2701 OSC_EXTENT_DUMP(D_CACHE, ext,
2702 "waiting for busy extent\n");
2703 waiting = osc_extent_get(ext);
2707 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2709 osc_extent_get(ext);
2710 if (ext->oe_state == OES_ACTIVE) {
2711 /* though we grab inode mutex for write path, but we
2712 * release it before releasing extent(in osc_io_end()),
2713 * so there is a race window that an extent is still
2714 * in OES_ACTIVE when truncate starts. */
2715 LASSERT(!ext->oe_trunc_pending);
2716 ext->oe_trunc_pending = 1;
2718 EASSERT(ext->oe_state == OES_CACHE, ext);
2719 osc_extent_state_set(ext, OES_TRUNC);
2720 osc_update_pending(obj, OBD_BRW_WRITE,
2723 EASSERT(list_empty(&ext->oe_link), ext);
2724 list_add_tail(&ext->oe_link, &list);
2726 ext = next_extent(ext);
2728 osc_object_unlock(obj);
2730 osc_list_maint(cli, obj);
2732 while (!list_empty(&list)) {
2735 ext = list_entry(list.next, struct osc_extent, oe_link);
2736 list_del_init(&ext->oe_link);
2738 /* extent may be in OES_ACTIVE state because inode mutex
2739 * is released before osc_io_end() in file write case */
2740 if (ext->oe_state != OES_TRUNC)
2741 osc_extent_wait(env, ext, OES_TRUNC);
2743 rc = osc_extent_truncate(ext, index, partial);
2748 OSC_EXTENT_DUMP(D_ERROR, ext,
2749 "truncate error %d\n", rc);
2750 } else if (ext->oe_nr_pages == 0) {
2751 osc_extent_remove(ext);
2753 /* this must be an overlapped extent which means only
2754 * part of pages in this extent have been truncated.
2756 EASSERTF(ext->oe_start <= index, ext,
2757 "trunc index = %lu/%d.\n", index, partial);
2758 /* fix index to skip this partially truncated extent */
2759 index = ext->oe_end + 1;
2762 /* we need to hold this extent in OES_TRUNC state so
2763 * that no writeback will happen. This is to avoid
2765 LASSERT(oio->oi_trunc == NULL);
2766 oio->oi_trunc = osc_extent_get(ext);
2767 OSC_EXTENT_DUMP(D_CACHE, ext,
2768 "trunc at "LPU64"\n", size);
2770 osc_extent_put(env, ext);
2772 if (waiting != NULL) {
2775 /* ignore the result of osc_extent_wait the write initiator
2776 * should take care of it. */
2777 rc = osc_extent_wait(env, waiting, OES_INV);
2779 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2781 osc_extent_put(env, waiting);
2789 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2791 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2792 struct osc_object *obj)
2794 struct osc_extent *ext = oio->oi_trunc;
2796 oio->oi_trunc = NULL;
2798 bool unplug = false;
2800 EASSERT(ext->oe_nr_pages > 0, ext);
2801 EASSERT(ext->oe_state == OES_TRUNC, ext);
2802 EASSERT(!ext->oe_urgent, ext);
2804 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2805 osc_object_lock(obj);
2806 osc_extent_state_set(ext, OES_CACHE);
2807 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2809 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2812 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2813 osc_object_unlock(obj);
2814 osc_extent_put(env, ext);
2817 osc_io_unplug_async(env, osc_cli(obj), obj);
2822 * Wait for extents in a specific range to be written out.
2823 * The caller must have called osc_cache_writeback_range() to issue IO
2824 * otherwise it will take a long time for this function to finish.
2826 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2827 * nobody else can dirty this range of file while we're waiting for
2828 * extents to be written.
2830 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2831 pgoff_t start, pgoff_t end)
2833 struct osc_extent *ext;
2834 pgoff_t index = start;
2839 osc_object_lock(obj);
2840 ext = osc_extent_search(obj, index);
2842 ext = first_extent(obj);
2843 else if (ext->oe_end < index)
2844 ext = next_extent(ext);
2845 while (ext != NULL) {
2848 if (ext->oe_start > end)
2851 if (!ext->oe_fsync_wait) {
2852 ext = next_extent(ext);
2856 EASSERT(ergo(ext->oe_state == OES_CACHE,
2857 ext->oe_hp || ext->oe_urgent), ext);
2858 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2859 !ext->oe_hp && ext->oe_urgent), ext);
2861 index = ext->oe_end + 1;
2862 osc_extent_get(ext);
2863 osc_object_unlock(obj);
2865 rc = osc_extent_wait(env, ext, OES_INV);
2868 osc_extent_put(env, ext);
2871 osc_object_unlock(obj);
2873 OSC_IO_DEBUG(obj, "sync file range.\n");
2878 * Called to write out a range of osc object.
2880 * @hp : should be set this is caused by lock cancel;
2881 * @discard: is set if dirty pages should be dropped - file will be deleted or
2882 * truncated, this implies there is no partially discarding extents.
2884 * Return how many pages will be issued, or error code if error occurred.
2886 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2887 pgoff_t start, pgoff_t end, int hp, int discard)
2889 struct osc_extent *ext;
2890 struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2891 bool unplug = false;
2895 osc_object_lock(obj);
2896 ext = osc_extent_search(obj, start);
2898 ext = first_extent(obj);
2899 else if (ext->oe_end < start)
2900 ext = next_extent(ext);
2901 while (ext != NULL) {
2902 if (ext->oe_start > end)
2905 ext->oe_fsync_wait = 1;
2906 switch (ext->oe_state) {
2908 result += ext->oe_nr_pages;
2910 struct list_head *list = NULL;
2912 EASSERT(!ext->oe_hp, ext);
2914 list = &obj->oo_hp_exts;
2915 } else if (!ext->oe_urgent) {
2917 list = &obj->oo_urgent_exts;
2920 list_move_tail(&ext->oe_link, list);
2923 /* the only discarder is lock cancelling, so
2924 * [start, end] must contain this extent */
2925 EASSERT(ext->oe_start >= start &&
2926 ext->oe_max_end <= end, ext);
2927 osc_extent_state_set(ext, OES_LOCKING);
2928 ext->oe_owner = current;
2929 list_move_tail(&ext->oe_link,
2931 osc_update_pending(obj, OBD_BRW_WRITE,
2936 /* It's pretty bad to wait for ACTIVE extents, because
2937 * we don't know how long we will wait for it to be
2938 * flushed since it may be blocked at awaiting more
2939 * grants. We do this for the correctness of fsync. */
2940 LASSERT(hp == 0 && discard == 0);
2944 /* this extent is being truncated, can't do anything
2945 * for it now. it will be set to urgent after truncate
2946 * is finished in osc_cache_truncate_end(). */
2950 ext = next_extent(ext);
2952 osc_object_unlock(obj);
2954 LASSERT(ergo(!discard, list_empty(&discard_list)));
2955 if (!list_empty(&discard_list)) {
2956 struct osc_extent *tmp;
2959 osc_list_maint(osc_cli(obj), obj);
2960 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2961 list_del_init(&ext->oe_link);
2962 EASSERT(ext->oe_state == OES_LOCKING, ext);
2964 /* Discard caching pages. We don't actually write this
2965 * extent out but we complete it as if we did. */
2966 rc = osc_extent_make_ready(env, ext);
2967 if (unlikely(rc < 0)) {
2968 OSC_EXTENT_DUMP(D_ERROR, ext,
2969 "make_ready returned %d\n", rc);
2974 /* finish the extent as if the pages were sent */
2975 osc_extent_finish(env, ext, 0, 0);
2980 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2982 if (hp || discard) {
2984 rc = osc_cache_wait_range(env, obj, start, end);
2985 if (result >= 0 && rc < 0)
2989 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2994 * Returns a list of pages by a given [start, end] of \a obj.
2996 * \param resched If not NULL, then we give up before hogging CPU for too
2997 * long and set *resched = 1, in that case caller should implement a retry
3000 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3001 * crucial in the face of [offset, EOF] locks.
3003 * Return at least one page in @queue unless there is no covered page.
3005 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3006 struct osc_object *osc, pgoff_t start, pgoff_t end,
3007 osc_page_gang_cbt cb, void *cbdata)
3009 struct osc_page *ops;
3015 int res = CLP_GANG_OKAY;
3016 bool tree_lock = true;
3020 pvec = osc_env_info(env)->oti_pvec;
3021 spin_lock(&osc->oo_tree_lock);
3022 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3023 idx, OTI_PVEC_SIZE)) > 0) {
3024 struct cl_page *page;
3025 bool end_of_region = false;
3027 for (i = 0, j = 0; i < nr; ++i) {
3031 idx = osc_index(ops);
3033 end_of_region = true;
3037 page = ops->ops_cl.cpl_page;
3038 LASSERT(page->cp_type == CPT_CACHEABLE);
3039 if (page->cp_state == CPS_FREEING)
3043 lu_ref_add_atomic(&page->cp_reference,
3044 "gang_lookup", current);
3050 * Here a delicate locking dance is performed. Current thread
3051 * holds a reference to a page, but has to own it before it
3052 * can be placed into queue. Owning implies waiting, so
3053 * radix-tree lock is to be released. After a wait one has to
3054 * check that pages weren't truncated (cl_page_own() returns
3055 * error in the latter case).
3057 spin_unlock(&osc->oo_tree_lock);
3060 for (i = 0; i < j; ++i) {
3062 if (res == CLP_GANG_OKAY)
3063 res = (*cb)(env, io, ops, cbdata);
3065 page = ops->ops_cl.cpl_page;
3066 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3067 cl_page_put(env, page);
3069 if (nr < OTI_PVEC_SIZE || end_of_region)
3072 if (res == CLP_GANG_OKAY && need_resched())
3073 res = CLP_GANG_RESCHED;
3074 if (res != CLP_GANG_OKAY)
3077 spin_lock(&osc->oo_tree_lock);
3081 spin_unlock(&osc->oo_tree_lock);
3086 * Check if page @page is covered by an extra lock or discard it.
3088 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3089 struct osc_page *ops, void *cbdata)
3091 struct osc_thread_info *info = osc_env_info(env);
3092 struct osc_object *osc = cbdata;
3095 index = osc_index(ops);
3096 if (index >= info->oti_fn_index) {
3097 struct ldlm_lock *tmp;
3098 struct cl_page *page = ops->ops_cl.cpl_page;
3100 /* refresh non-overlapped index */
3101 tmp = osc_dlmlock_at_pgoff(env, osc, index, 0, 0);
3103 __u64 end = tmp->l_policy_data.l_extent.end;
3104 /* Cache the first-non-overlapped index so as to skip
3105 * all pages within [index, oti_fn_index). This is safe
3106 * because if tmp lock is canceled, it will discard
3108 info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3109 if (end == OBD_OBJECT_EOF)
3110 info->oti_fn_index = CL_PAGE_EOF;
3112 } else if (cl_page_own(env, io, page) == 0) {
3113 /* discard the page */
3114 cl_page_discard(env, io, page);
3115 cl_page_disown(env, io, page);
3117 LASSERT(page->cp_state == CPS_FREEING);
3121 info->oti_next_index = index + 1;
3122 return CLP_GANG_OKAY;
3125 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3126 struct osc_page *ops, void *cbdata)
3128 struct osc_thread_info *info = osc_env_info(env);
3129 struct cl_page *page = ops->ops_cl.cpl_page;
3131 /* page is top page. */
3132 info->oti_next_index = osc_index(ops) + 1;
3133 if (cl_page_own(env, io, page) == 0) {
3134 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3135 !PageDirty(cl_page_vmpage(page))));
3137 /* discard the page */
3138 cl_page_discard(env, io, page);
3139 cl_page_disown(env, io, page);
3141 LASSERT(page->cp_state == CPS_FREEING);
3144 return CLP_GANG_OKAY;
3148 * Discard pages protected by the given lock. This function traverses radix
3149 * tree to find all covering pages and discard them. If a page is being covered
3150 * by other locks, it should remain in cache.
3152 * If error happens on any step, the process continues anyway (the reasoning
3153 * behind this being that lock cancellation cannot be delayed indefinitely).
3155 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3156 pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3158 struct osc_thread_info *info = osc_env_info(env);
3159 struct cl_io *io = &info->oti_io;
3160 osc_page_gang_cbt cb;
3166 io->ci_obj = cl_object_top(osc2cl(osc));
3167 io->ci_ignore_layout = 1;
3168 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3172 cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3173 info->oti_fn_index = info->oti_next_index = start;
3175 res = osc_page_gang_lookup(env, io, osc,
3176 info->oti_next_index, end, cb, osc);
3177 if (info->oti_next_index > end)
3180 if (res == CLP_GANG_RESCHED)
3182 } while (res != CLP_GANG_OKAY);
3184 cl_io_fini(env, io);