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 struct osc_extent *osc_extent_find(const struct lu_env *env,
600 struct osc_object *obj, pgoff_t index,
603 struct client_obd *cli = osc_cli(obj);
604 struct osc_lock *olck;
605 struct cl_lock_descr *descr;
606 struct osc_extent *cur;
607 struct osc_extent *ext;
608 struct osc_extent *conflict = NULL;
609 struct osc_extent *found = NULL;
612 int max_pages; /* max_pages_per_rpc */
614 int ppc_bits; /* pages per chunk bits */
619 cur = osc_extent_alloc(obj);
621 RETURN(ERR_PTR(-ENOMEM));
623 olck = osc_env_io(env)->oi_write_osclock;
624 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
625 LASSERT(olck->ols_state == OLS_GRANTED);
627 descr = &olck->ols_cl.cls_lock->cll_descr;
628 LASSERT(descr->cld_mode >= CLM_WRITE);
630 LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
631 ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
632 chunk_mask = ~((1 << ppc_bits) - 1);
633 chunksize = 1 << cli->cl_chunkbits;
634 chunk = index >> ppc_bits;
636 /* align end to rpc edge, rpc size may not be a power 2 integer. */
637 max_pages = cli->cl_max_pages_per_rpc;
638 LASSERT((max_pages & ~chunk_mask) == 0);
639 max_end = index - (index % max_pages) + max_pages - 1;
640 max_end = min_t(pgoff_t, max_end, descr->cld_end);
642 /* initialize new extent by parameters so far */
643 cur->oe_max_end = max_end;
644 cur->oe_start = index & chunk_mask;
645 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
646 if (cur->oe_start < descr->cld_start)
647 cur->oe_start = descr->cld_start;
648 if (cur->oe_end > max_end)
649 cur->oe_end = max_end;
651 cur->oe_mppr = max_pages;
652 if (olck->ols_dlmlock != NULL) {
653 LASSERT(olck->ols_hold);
654 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
655 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
658 /* grants has been allocated by caller */
659 LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
660 "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
661 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
665 osc_object_lock(obj);
666 ext = osc_extent_search(obj, cur->oe_start);
668 ext = first_extent(obj);
669 while (ext != NULL) {
670 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
671 loff_t ext_chk_end = ext->oe_end >> ppc_bits;
673 LASSERT(sanity_check_nolock(ext) == 0);
674 if (chunk > ext_chk_end + 1)
677 /* if covering by different locks, no chance to match */
678 if (olck->ols_dlmlock != ext->oe_dlmlock) {
679 EASSERTF(!overlapped(ext, cur), ext,
680 EXTSTR"\n", EXTPARA(cur));
682 ext = next_extent(ext);
686 /* discontiguous chunks? */
687 if (chunk + 1 < ext_chk_start) {
688 ext = next_extent(ext);
692 /* ok, from now on, ext and cur have these attrs:
693 * 1. covered by the same lock
694 * 2. contiguous at chunk level or overlapping. */
696 if (overlapped(ext, cur)) {
697 /* cur is the minimum unit, so overlapping means
699 EASSERTF((ext->oe_start <= cur->oe_start &&
700 ext->oe_end >= cur->oe_end),
701 ext, EXTSTR"\n", EXTPARA(cur));
703 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
704 /* for simplicity, we wait for this extent to
705 * finish before going forward. */
706 conflict = osc_extent_get(ext);
710 found = osc_extent_hold(ext);
714 /* non-overlapped extent */
715 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
716 /* we can't do anything for a non OES_CACHE extent, or
717 * if there is someone waiting for this extent to be
718 * flushed, try next one. */
719 ext = next_extent(ext);
723 /* check if they belong to the same rpc slot before trying to
724 * merge. the extents are not overlapped and contiguous at
725 * chunk level to get here. */
726 if (ext->oe_max_end != max_end) {
727 /* if they don't belong to the same RPC slot or
728 * max_pages_per_rpc has ever changed, do not merge. */
729 ext = next_extent(ext);
733 /* it's required that an extent must be contiguous at chunk
734 * level so that we know the whole extent is covered by grant
735 * (the pages in the extent are NOT required to be contiguous).
736 * Otherwise, it will be too much difficult to know which
737 * chunks have grants allocated. */
739 /* try to do front merge - extend ext's start */
740 if (chunk + 1 == ext_chk_start) {
741 /* ext must be chunk size aligned */
742 EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
744 /* pull ext's start back to cover cur */
745 ext->oe_start = cur->oe_start;
746 ext->oe_grants += chunksize;
747 *grants -= chunksize;
749 found = osc_extent_hold(ext);
750 } else if (chunk == ext_chk_end + 1) {
752 ext->oe_end = cur->oe_end;
753 ext->oe_grants += chunksize;
754 *grants -= chunksize;
756 /* try to merge with the next one because we just fill
758 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
759 /* we can save extent tax from next extent */
760 *grants += cli->cl_extent_tax;
762 found = osc_extent_hold(ext);
767 ext = next_extent(ext);
770 osc_extent_tree_dump(D_CACHE, obj);
772 LASSERT(conflict == NULL);
773 if (!IS_ERR(found)) {
774 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
775 OSC_EXTENT_DUMP(D_CACHE, found,
776 "found caching ext for %lu.\n", index);
778 } else if (conflict == NULL) {
779 /* create a new extent */
780 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
781 cur->oe_grants = chunksize + cli->cl_extent_tax;
782 *grants -= cur->oe_grants;
783 LASSERT(*grants >= 0);
785 cur->oe_state = OES_CACHE;
786 found = osc_extent_hold(cur);
787 osc_extent_insert(obj, cur);
788 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
789 index, descr->cld_end);
791 osc_object_unlock(obj);
793 if (conflict != NULL) {
794 LASSERT(found == NULL);
796 /* waiting for IO to finish. Please notice that it's impossible
797 * to be an OES_TRUNC extent. */
798 rc = osc_extent_wait(env, conflict, OES_INV);
799 osc_extent_put(env, conflict);
802 GOTO(out, found = ERR_PTR(rc));
809 osc_extent_put(env, cur);
810 LASSERT(*grants >= 0);
815 * Called when IO is finished to an extent.
817 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
820 struct client_obd *cli = osc_cli(ext->oe_obj);
821 struct osc_async_page *oap;
822 struct osc_async_page *tmp;
823 int nr_pages = ext->oe_nr_pages;
825 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
830 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
832 ext->oe_rc = rc ?: ext->oe_nr_pages;
833 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
835 osc_lru_add_batch(cli, &ext->oe_pages);
836 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
838 list_del_init(&oap->oap_rpc_item);
839 list_del_init(&oap->oap_pending_item);
840 if (last_off <= oap->oap_obj_off) {
841 last_off = oap->oap_obj_off;
842 last_count = oap->oap_count;
846 osc_ap_completion(env, cli, oap, sent, rc);
848 EASSERT(ext->oe_nr_pages == 0, ext);
851 lost_grant = ext->oe_grants;
852 } else if (blocksize < PAGE_CACHE_SIZE &&
853 last_count != PAGE_CACHE_SIZE) {
854 /* For short writes we shouldn't count parts of pages that
855 * span a whole chunk on the OST side, or our accounting goes
856 * wrong. Should match the code in filter_grant_check. */
857 int offset = last_off & ~CFS_PAGE_MASK;
858 int count = last_count + (offset & (blocksize - 1));
859 int end = (offset + last_count) & (blocksize - 1);
861 count += blocksize - end;
863 lost_grant = PAGE_CACHE_SIZE - count;
865 if (ext->oe_grants > 0)
866 osc_free_grant(cli, nr_pages, lost_grant);
868 osc_extent_remove(ext);
869 /* put the refcount for RPC */
870 osc_extent_put(env, ext);
874 static int extent_wait_cb(struct osc_extent *ext, int state)
878 osc_object_lock(ext->oe_obj);
879 ret = ext->oe_state == state;
880 osc_object_unlock(ext->oe_obj);
886 * Wait for the extent's state to become @state.
888 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
891 struct osc_object *obj = ext->oe_obj;
892 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
893 LWI_ON_SIGNAL_NOOP, NULL);
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 = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
916 if (rc == -ETIMEDOUT) {
917 OSC_EXTENT_DUMP(D_ERROR, ext,
918 "%s: wait ext to %d timedout, recovery in progress?\n",
919 osc_export(obj)->exp_obd->obd_name, state);
921 lwi = LWI_INTR(NULL, NULL);
922 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
925 if (rc == 0 && ext->oe_rc < 0)
931 * Discard pages with index greater than @size. If @ext is overlapped with
932 * @size, then partial truncate happens.
934 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
937 struct cl_env_nest nest;
940 struct osc_object *obj = ext->oe_obj;
941 struct client_obd *cli = osc_cli(obj);
942 struct osc_async_page *oap;
943 struct osc_async_page *tmp;
944 int pages_in_chunk = 0;
945 int ppc_bits = cli->cl_chunkbits -
947 __u64 trunc_chunk = trunc_index >> ppc_bits;
953 LASSERT(sanity_check(ext) == 0);
954 LASSERT(ext->oe_state == OES_TRUNC);
955 LASSERT(!ext->oe_urgent);
957 /* Request new lu_env.
958 * We can't use that env from osc_cache_truncate_start() because
959 * it's from lov_io_sub and not fully initialized. */
960 env = cl_env_nested_get(&nest);
961 io = &osc_env_info(env)->oti_io;
962 io->ci_obj = cl_object_top(osc2cl(obj));
963 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
967 /* discard all pages with index greater then trunc_index */
968 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
970 pgoff_t index = osc_index(oap2osc(oap));
971 struct cl_page *page = oap2cl_page(oap);
973 LASSERT(list_empty(&oap->oap_rpc_item));
975 /* only discard the pages with their index greater than
976 * trunc_index, and ... */
977 if (index < trunc_index ||
978 (index == trunc_index && partial)) {
979 /* accounting how many pages remaining in the chunk
980 * so that we can calculate grants correctly. */
981 if (index >> ppc_bits == trunc_chunk)
986 list_del_init(&oap->oap_pending_item);
989 lu_ref_add(&page->cp_reference, "truncate", current);
991 if (cl_page_own(env, io, page) == 0) {
992 cl_page_discard(env, io, page);
993 cl_page_disown(env, io, page);
995 LASSERT(page->cp_state == CPS_FREEING);
999 lu_ref_del(&page->cp_reference, "truncate", current);
1000 cl_page_put(env, page);
1005 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1006 ext->oe_nr_pages == 0),
1007 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1009 osc_object_lock(obj);
1010 if (ext->oe_nr_pages == 0) {
1011 LASSERT(pages_in_chunk == 0);
1012 grants = ext->oe_grants;
1014 } else { /* calculate how many grants we can free */
1015 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1019 /* if there is no pages in this chunk, we can also free grants
1020 * for the last chunk */
1021 if (pages_in_chunk == 0) {
1022 /* if this is the 1st chunk and no pages in this chunk,
1023 * ext->oe_nr_pages must be zero, so we should be in
1024 * the other if-clause. */
1025 LASSERT(trunc_chunk > 0);
1030 /* this is what we can free from this extent */
1031 grants = chunks << cli->cl_chunkbits;
1032 ext->oe_grants -= grants;
1033 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1034 ext->oe_end = min(last_index, ext->oe_max_end);
1035 LASSERT(ext->oe_end >= ext->oe_start);
1036 LASSERT(ext->oe_grants > 0);
1038 osc_object_unlock(obj);
1040 if (grants > 0 || nr_pages > 0)
1041 osc_free_grant(cli, nr_pages, grants);
1044 cl_io_fini(env, io);
1045 cl_env_nested_put(&nest, env);
1050 * This function is used to make the extent prepared for transfer.
1051 * A race with flusing page - ll_writepage() has to be handled cautiously.
1053 static int osc_extent_make_ready(const struct lu_env *env,
1054 struct osc_extent *ext)
1056 struct osc_async_page *oap;
1057 struct osc_async_page *last = NULL;
1058 struct osc_object *obj = ext->oe_obj;
1063 /* we're going to grab page lock, so object lock must not be taken. */
1064 LASSERT(sanity_check(ext) == 0);
1065 /* in locking state, any process should not touch this extent. */
1066 EASSERT(ext->oe_state == OES_LOCKING, ext);
1067 EASSERT(ext->oe_owner != NULL, ext);
1069 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1071 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1073 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1076 /* checking ASYNC_READY is race safe */
1077 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1080 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1083 spin_lock(&oap->oap_lock);
1084 oap->oap_async_flags |= ASYNC_READY;
1085 spin_unlock(&oap->oap_lock);
1088 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1091 LASSERTF(0, "unknown return code: %d\n", rc);
1095 LASSERT(page_count == ext->oe_nr_pages);
1096 LASSERT(last != NULL);
1097 /* the last page is the only one we need to refresh its count by
1098 * the size of file. */
1099 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1100 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1101 LASSERT(last->oap_count > 0);
1102 LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
1103 spin_lock(&last->oap_lock);
1104 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1105 spin_unlock(&last->oap_lock);
1108 /* for the rest of pages, we don't need to call osf_refresh_count()
1109 * because it's known they are not the last page */
1110 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1111 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1112 oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1113 spin_lock(&oap->oap_lock);
1114 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1115 spin_unlock(&oap->oap_lock);
1119 osc_object_lock(obj);
1120 osc_extent_state_set(ext, OES_RPC);
1121 osc_object_unlock(obj);
1122 /* get a refcount for RPC. */
1123 osc_extent_get(ext);
1129 * Quick and simple version of osc_extent_find(). This function is frequently
1130 * called to expand the extent for the same IO. To expand the extent, the
1131 * page index must be in the same or next chunk of ext->oe_end.
1133 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1135 struct osc_object *obj = ext->oe_obj;
1136 struct client_obd *cli = osc_cli(obj);
1137 struct osc_extent *next;
1138 int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1139 pgoff_t chunk = index >> ppc_bits;
1142 int chunksize = 1 << cli->cl_chunkbits;
1146 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1147 osc_object_lock(obj);
1148 LASSERT(sanity_check_nolock(ext) == 0);
1149 end_chunk = ext->oe_end >> ppc_bits;
1150 if (chunk > end_chunk + 1)
1151 GOTO(out, rc = -ERANGE);
1153 if (end_chunk >= chunk)
1156 LASSERT(end_chunk + 1 == chunk);
1157 /* try to expand this extent to cover @index */
1158 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1160 next = next_extent(ext);
1161 if (next != NULL && next->oe_start <= end_index)
1162 /* complex mode - overlapped with the next extent,
1163 * this case will be handled by osc_extent_find() */
1164 GOTO(out, rc = -EAGAIN);
1166 ext->oe_end = end_index;
1167 ext->oe_grants += chunksize;
1168 *grants -= chunksize;
1169 LASSERT(*grants >= 0);
1170 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1171 "overlapped after expanding for %lu.\n", index);
1175 osc_object_unlock(obj);
1179 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1180 const char *func, int line)
1182 struct osc_extent *ext;
1185 CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1186 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1188 /* osc_object_lock(obj); */
1190 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1191 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1194 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1195 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1198 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1199 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1202 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1203 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1204 /* osc_object_unlock(obj); */
1207 /* ------------------ osc extent end ------------------ */
1209 static inline int osc_is_ready(struct osc_object *osc)
1211 return !list_empty(&osc->oo_ready_item) ||
1212 !list_empty(&osc->oo_hp_ready_item);
1215 #define OSC_IO_DEBUG(OSC, STR, args...) \
1216 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1217 (OSC), osc_is_ready(OSC), \
1218 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1219 list_empty_marker(&(OSC)->oo_ready_item), \
1220 atomic_read(&(OSC)->oo_nr_writes), \
1221 list_empty_marker(&(OSC)->oo_hp_exts), \
1222 list_empty_marker(&(OSC)->oo_urgent_exts), \
1223 atomic_read(&(OSC)->oo_nr_reads), \
1224 list_empty_marker(&(OSC)->oo_reading_exts), \
1227 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1230 struct osc_page *opg = oap2osc_page(oap);
1231 struct cl_page *page = oap2cl_page(oap);
1234 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1237 result = cl_page_make_ready(env, page, CRT_WRITE);
1239 opg->ops_submit_time = cfs_time_current();
1243 static int osc_refresh_count(const struct lu_env *env,
1244 struct osc_async_page *oap, int cmd)
1246 struct osc_page *opg = oap2osc_page(oap);
1247 pgoff_t index = osc_index(oap2osc(oap));
1248 struct cl_object *obj;
1249 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1254 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1255 LASSERT(!(cmd & OBD_BRW_READ));
1256 LASSERT(opg != NULL);
1257 obj = opg->ops_cl.cpl_obj;
1259 cl_object_attr_lock(obj);
1260 result = cl_object_attr_get(env, obj, attr);
1261 cl_object_attr_unlock(obj);
1264 kms = attr->cat_kms;
1265 if (cl_offset(obj, index) >= kms)
1266 /* catch race with truncate */
1268 else if (cl_offset(obj, index + 1) > kms)
1269 /* catch sub-page write at end of file */
1270 return kms % PAGE_CACHE_SIZE;
1272 return PAGE_CACHE_SIZE;
1275 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1278 struct osc_page *opg = oap2osc_page(oap);
1279 struct cl_page *page = oap2cl_page(oap);
1280 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
1281 enum cl_req_type crt;
1286 cmd &= ~OBD_BRW_NOQUOTA;
1287 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1288 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1289 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1290 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1291 LASSERT(opg->ops_transfer_pinned);
1294 * page->cp_req can be NULL if io submission failed before
1295 * cl_req was allocated.
1297 if (page->cp_req != NULL)
1298 cl_req_page_done(env, page);
1299 LASSERT(page->cp_req == NULL);
1301 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1302 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1303 opg->ops_transfer_pinned = 0;
1305 spin_lock(&obj->oo_seatbelt);
1306 LASSERT(opg->ops_submitter != NULL);
1307 LASSERT(!list_empty(&opg->ops_inflight));
1308 list_del_init(&opg->ops_inflight);
1309 opg->ops_submitter = NULL;
1310 spin_unlock(&obj->oo_seatbelt);
1312 opg->ops_submit_time = 0;
1313 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1316 if (rc == 0 && srvlock) {
1317 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1318 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1319 int bytes = oap->oap_count;
1321 if (crt == CRT_READ)
1322 stats->os_lockless_reads += bytes;
1324 stats->os_lockless_writes += bytes;
1328 * This has to be the last operation with the page, as locks are
1329 * released in cl_page_completion() and nothing except for the
1330 * reference counter protects page from concurrent reclaim.
1332 lu_ref_del(&page->cp_reference, "transfer", page);
1334 cl_page_completion(env, page, crt, rc);
1339 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do { \
1340 struct client_obd *__tmp = (cli); \
1341 CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1342 "dropped: %ld avail: %ld, reserved: %ld, flight: %d }" \
1343 "lru {in list: %ld, left: %ld, waiters: %d }"fmt"\n", \
1344 __tmp->cl_import->imp_obd->obd_name, \
1345 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1346 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1347 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1348 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1349 atomic_long_read(&__tmp->cl_lru_in_list), \
1350 atomic_long_read(&__tmp->cl_lru_busy), \
1351 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1354 /* caller must hold loi_list_lock */
1355 static void osc_consume_write_grant(struct client_obd *cli,
1356 struct brw_page *pga)
1358 assert_spin_locked(&cli->cl_loi_list_lock.lock);
1359 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1360 atomic_long_inc(&obd_dirty_pages);
1361 cli->cl_dirty_pages++;
1362 pga->flag |= OBD_BRW_FROM_GRANT;
1363 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1364 PAGE_CACHE_SIZE, pga, pga->pg);
1365 osc_update_next_shrink(cli);
1368 /* the companion to osc_consume_write_grant, called when a brw has completed.
1369 * must be called with the loi lock held. */
1370 static void osc_release_write_grant(struct client_obd *cli,
1371 struct brw_page *pga)
1375 assert_spin_locked(&cli->cl_loi_list_lock.lock);
1376 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1381 pga->flag &= ~OBD_BRW_FROM_GRANT;
1382 atomic_long_dec(&obd_dirty_pages);
1383 cli->cl_dirty_pages--;
1384 if (pga->flag & OBD_BRW_NOCACHE) {
1385 pga->flag &= ~OBD_BRW_NOCACHE;
1386 atomic_long_dec(&obd_dirty_transit_pages);
1387 cli->cl_dirty_transit--;
1393 * To avoid sleeping with object lock held, it's good for us allocate enough
1394 * grants before entering into critical section.
1396 * client_obd_list_lock held by caller
1398 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1402 if (cli->cl_avail_grant >= bytes) {
1403 cli->cl_avail_grant -= bytes;
1404 cli->cl_reserved_grant += bytes;
1410 static void __osc_unreserve_grant(struct client_obd *cli,
1411 unsigned int reserved, unsigned int unused)
1413 /* it's quite normal for us to get more grant than reserved.
1414 * Thinking about a case that two extents merged by adding a new
1415 * chunk, we can save one extent tax. If extent tax is greater than
1416 * one chunk, we can save more grant by adding a new chunk */
1417 cli->cl_reserved_grant -= reserved;
1418 if (unused > reserved) {
1419 cli->cl_avail_grant += reserved;
1420 cli->cl_lost_grant += unused - reserved;
1422 cli->cl_avail_grant += unused;
1426 void osc_unreserve_grant(struct client_obd *cli,
1427 unsigned int reserved, unsigned int unused)
1429 client_obd_list_lock(&cli->cl_loi_list_lock);
1430 __osc_unreserve_grant(cli, reserved, unused);
1432 osc_wake_cache_waiters(cli);
1433 client_obd_list_unlock(&cli->cl_loi_list_lock);
1437 * Free grant after IO is finished or canceled.
1439 * @lost_grant is used to remember how many grants we have allocated but not
1440 * used, we should return these grants to OST. There're two cases where grants
1443 * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1444 * written. In this case OST may use less chunks to serve this partial
1445 * write. OSTs don't actually know the page size on the client side. so
1446 * clients have to calculate lost grant by the blocksize on the OST.
1447 * See filter_grant_check() for details.
1449 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1450 unsigned int lost_grant)
1452 int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1454 client_obd_list_lock(&cli->cl_loi_list_lock);
1455 atomic_long_sub(nr_pages, &obd_dirty_pages);
1456 cli->cl_dirty_pages -= nr_pages;
1457 cli->cl_lost_grant += lost_grant;
1458 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1459 /* borrow some grant from truncate to avoid the case that
1460 * truncate uses up all avail grant */
1461 cli->cl_lost_grant -= grant;
1462 cli->cl_avail_grant += grant;
1464 osc_wake_cache_waiters(cli);
1465 client_obd_list_unlock(&cli->cl_loi_list_lock);
1466 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1467 lost_grant, cli->cl_lost_grant,
1468 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
1472 * The companion to osc_enter_cache(), called when @oap is no longer part of
1473 * the dirty accounting due to error.
1475 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1477 client_obd_list_lock(&cli->cl_loi_list_lock);
1478 osc_release_write_grant(cli, &oap->oap_brw_page);
1479 client_obd_list_unlock(&cli->cl_loi_list_lock);
1483 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1486 static int osc_enter_cache_try(struct client_obd *cli,
1487 struct osc_async_page *oap,
1488 int bytes, int transient)
1492 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1494 rc = osc_reserve_grant(cli, bytes);
1498 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1499 1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1500 osc_consume_write_grant(cli, &oap->oap_brw_page);
1502 cli->cl_dirty_transit++;
1503 atomic_long_inc(&obd_dirty_transit_pages);
1504 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1508 __osc_unreserve_grant(cli, bytes, bytes);
1514 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1517 client_obd_list_lock(&cli->cl_loi_list_lock);
1518 rc = list_empty(&ocw->ocw_entry);
1519 client_obd_list_unlock(&cli->cl_loi_list_lock);
1524 * The main entry to reserve dirty page accounting. Usually the grant reserved
1525 * in this function will be freed in bulk in osc_free_grant() unless it fails
1526 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1528 * The process will be put into sleep if it's already run out of grant.
1530 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1531 struct osc_async_page *oap, int bytes)
1533 struct osc_object *osc = oap->oap_obj;
1534 struct lov_oinfo *loi = osc->oo_oinfo;
1535 struct osc_cache_waiter ocw;
1536 struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1537 LWI_ON_SIGNAL_NOOP, NULL);
1541 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1543 client_obd_list_lock(&cli->cl_loi_list_lock);
1545 /* force the caller to try sync io. this can jump the list
1546 * of queued writes and create a discontiguous rpc stream */
1547 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1548 cli->cl_dirty_max_pages == 0 ||
1549 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1550 GOTO(out, rc = -EDQUOT);
1552 /* Hopefully normal case - cache space and write credits available */
1553 if (osc_enter_cache_try(cli, oap, bytes, 0))
1556 /* We can get here for two reasons: too many dirty pages in cache, or
1557 * run out of grants. In both cases we should write dirty pages out.
1558 * Adding a cache waiter will trigger urgent write-out no matter what
1560 * The exiting condition is no avail grants and no dirty pages caching,
1561 * that really means there is no space on the OST. */
1562 init_waitqueue_head(&ocw.ocw_waitq);
1564 ocw.ocw_grant = bytes;
1565 while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1566 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1568 client_obd_list_unlock(&cli->cl_loi_list_lock);
1570 osc_io_unplug_async(env, cli, NULL);
1572 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1573 cli->cl_import->imp_obd->obd_name, &ocw, oap);
1575 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1577 client_obd_list_lock(&cli->cl_loi_list_lock);
1579 /* l_wait_event is interrupted by signal, or timed out */
1583 OSC_DUMP_GRANT(D_ERROR, cli,
1584 "try to reserve %d.\n", bytes);
1585 osc_extent_tree_dump(D_ERROR, osc);
1589 /* Ensures restartability - LU-3581 */
1593 CDEBUG(D_CACHE, "%s: event for cache space @"
1594 " %p never arrived due to %d\n",
1595 cli->cl_import->imp_obd->obd_name,
1599 list_del_init(&ocw.ocw_entry);
1603 LASSERT(list_empty(&ocw.ocw_entry));
1608 if (osc_enter_cache_try(cli, oap, bytes, 0))
1613 client_obd_list_unlock(&cli->cl_loi_list_lock);
1614 OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1618 /* caller must hold loi_list_lock */
1619 void osc_wake_cache_waiters(struct client_obd *cli)
1621 struct list_head *l, *tmp;
1622 struct osc_cache_waiter *ocw;
1625 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1626 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1627 list_del_init(&ocw->ocw_entry);
1629 ocw->ocw_rc = -EDQUOT;
1630 /* we can't dirty more */
1631 if ((cli->cl_dirty_pages >= cli->cl_dirty_max_pages) ||
1632 (1 + atomic_long_read(&obd_dirty_pages) >
1633 obd_max_dirty_pages)) {
1634 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1635 "osc max %ld, sys max %ld\n",
1636 cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1637 obd_max_dirty_pages);
1642 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1643 ocw->ocw_rc = -EDQUOT;
1646 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1647 ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1649 wake_up(&ocw->ocw_waitq);
1655 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1657 int hprpc = !!list_empty(&osc->oo_hp_exts);
1658 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1661 /* This maintains the lists of pending pages to read/write for a given object
1662 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1663 * to quickly find objects that are ready to send an RPC. */
1664 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1667 int invalid_import = 0;
1670 /* if we have an invalid import we want to drain the queued pages
1671 * by forcing them through rpcs that immediately fail and complete
1672 * the pages. recovery relies on this to empty the queued pages
1673 * before canceling the locks and evicting down the llite pages */
1674 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1677 if (cmd & OBD_BRW_WRITE) {
1678 if (atomic_read(&osc->oo_nr_writes) == 0)
1680 if (invalid_import) {
1681 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1684 if (!list_empty(&osc->oo_hp_exts)) {
1685 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1688 if (!list_empty(&osc->oo_urgent_exts)) {
1689 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1692 /* trigger a write rpc stream as long as there are dirtiers
1693 * waiting for space. as they're waiting, they're not going to
1694 * create more pages to coalesce with what's waiting.. */
1695 if (!list_empty(&cli->cl_cache_waiters)) {
1696 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1699 if (atomic_read(&osc->oo_nr_writes) >=
1700 cli->cl_max_pages_per_rpc)
1703 if (atomic_read(&osc->oo_nr_reads) == 0)
1705 if (invalid_import) {
1706 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1709 /* all read are urgent. */
1710 if (!list_empty(&osc->oo_reading_exts))
1717 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1719 struct client_obd *cli = osc_cli(obj);
1720 if (cmd & OBD_BRW_WRITE) {
1721 atomic_add(delta, &obj->oo_nr_writes);
1722 atomic_add(delta, &cli->cl_pending_w_pages);
1723 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1725 atomic_add(delta, &obj->oo_nr_reads);
1726 atomic_add(delta, &cli->cl_pending_r_pages);
1727 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1729 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1732 static int osc_makes_hprpc(struct osc_object *obj)
1734 return !list_empty(&obj->oo_hp_exts);
1737 static void on_list(struct list_head *item, struct list_head *list,
1740 if (list_empty(item) && should_be_on)
1741 list_add_tail(item, list);
1742 else if (!list_empty(item) && !should_be_on)
1743 list_del_init(item);
1746 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1747 * can find pages to build into rpcs quickly */
1748 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1750 if (osc_makes_hprpc(osc)) {
1752 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1753 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1755 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1756 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1757 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1758 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1761 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1762 atomic_read(&osc->oo_nr_writes) > 0);
1764 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1765 atomic_read(&osc->oo_nr_reads) > 0);
1767 return osc_is_ready(osc);
1770 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1774 client_obd_list_lock(&cli->cl_loi_list_lock);
1775 is_ready = __osc_list_maint(cli, osc);
1776 client_obd_list_unlock(&cli->cl_loi_list_lock);
1781 /* this is trying to propogate async writeback errors back up to the
1782 * application. As an async write fails we record the error code for later if
1783 * the app does an fsync. As long as errors persist we force future rpcs to be
1784 * sync so that the app can get a sync error and break the cycle of queueing
1785 * pages for which writeback will fail. */
1786 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1793 ar->ar_force_sync = 1;
1794 ar->ar_min_xid = ptlrpc_sample_next_xid();
1799 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1800 ar->ar_force_sync = 0;
1803 /* this must be called holding the loi list lock to give coverage to exit_cache,
1804 * async_flag maintenance, and oap_request */
1805 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1806 struct osc_async_page *oap, int sent, int rc)
1808 struct osc_object *osc = oap->oap_obj;
1809 struct lov_oinfo *loi = osc->oo_oinfo;
1813 if (oap->oap_request != NULL) {
1814 xid = ptlrpc_req_xid(oap->oap_request);
1815 ptlrpc_req_finished(oap->oap_request);
1816 oap->oap_request = NULL;
1819 /* As the transfer for this page is being done, clear the flags */
1820 spin_lock(&oap->oap_lock);
1821 oap->oap_async_flags = 0;
1822 spin_unlock(&oap->oap_lock);
1823 oap->oap_interrupted = 0;
1825 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1826 client_obd_list_lock(&cli->cl_loi_list_lock);
1827 osc_process_ar(&cli->cl_ar, xid, rc);
1828 osc_process_ar(&loi->loi_ar, xid, rc);
1829 client_obd_list_unlock(&cli->cl_loi_list_lock);
1832 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1834 CERROR("completion on oap %p obj %p returns %d.\n",
1841 * Try to add extent to one RPC. We need to think about the following things:
1842 * - # of pages must not be over max_pages_per_rpc
1843 * - extent must be compatible with previous ones
1845 static int try_to_add_extent_for_io(struct client_obd *cli,
1846 struct osc_extent *ext,
1847 struct list_head *rpclist,
1848 int *pc, unsigned int *max_pages)
1850 struct osc_extent *tmp;
1851 struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1852 struct osc_async_page,
1856 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1859 *max_pages = max(ext->oe_mppr, *max_pages);
1860 if (*pc + ext->oe_nr_pages > *max_pages)
1863 list_for_each_entry(tmp, rpclist, oe_link) {
1864 struct osc_async_page *oap2;
1865 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1867 EASSERT(tmp->oe_owner == current, tmp);
1869 if (overlapped(tmp, ext)) {
1870 OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1874 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1875 CDEBUG(D_CACHE, "Do not permit different type of IO"
1876 " for a same RPC\n");
1880 if (tmp->oe_srvlock != ext->oe_srvlock ||
1881 !tmp->oe_grants != !ext->oe_grants)
1884 /* remove break for strict check */
1888 *pc += ext->oe_nr_pages;
1889 list_move_tail(&ext->oe_link, rpclist);
1890 ext->oe_owner = current;
1895 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1896 * get_write_extent() takes all appropriate extents in atomic.
1898 * The following policy is used to collect extents for IO:
1899 * 1. Add as many HP extents as possible;
1900 * 2. Add the first urgent extent in urgent extent list and take it out of
1902 * 3. Add subsequent extents of this urgent extent;
1903 * 4. If urgent list is not empty, goto 2;
1904 * 5. Traverse the extent tree from the 1st extent;
1905 * 6. Above steps exit if there is no space in this RPC.
1907 static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1909 struct client_obd *cli = osc_cli(obj);
1910 struct osc_extent *ext;
1912 unsigned int max_pages = cli->cl_max_pages_per_rpc;
1914 LASSERT(osc_object_is_locked(obj));
1915 while (!list_empty(&obj->oo_hp_exts)) {
1916 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1918 LASSERT(ext->oe_state == OES_CACHE);
1919 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1922 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1924 if (page_count == max_pages)
1927 while (!list_empty(&obj->oo_urgent_exts)) {
1928 ext = list_entry(obj->oo_urgent_exts.next,
1929 struct osc_extent, oe_link);
1930 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1934 if (!ext->oe_intree)
1937 while ((ext = next_extent(ext)) != NULL) {
1938 if ((ext->oe_state != OES_CACHE) ||
1939 (!list_empty(&ext->oe_link) &&
1940 ext->oe_owner != NULL))
1943 if (!try_to_add_extent_for_io(cli, ext, rpclist,
1944 &page_count, &max_pages))
1948 if (page_count == max_pages)
1951 ext = first_extent(obj);
1952 while (ext != NULL) {
1953 if ((ext->oe_state != OES_CACHE) ||
1954 /* this extent may be already in current rpclist */
1955 (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1956 ext = next_extent(ext);
1960 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1964 ext = next_extent(ext);
1970 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1971 struct osc_object *osc, pdl_policy_t pol)
1974 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
1975 struct osc_extent *ext;
1976 struct osc_extent *tmp;
1977 struct osc_extent *first = NULL;
1978 obd_count page_count = 0;
1983 LASSERT(osc_object_is_locked(osc));
1985 page_count = get_write_extents(osc, &rpclist);
1986 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1988 if (list_empty(&rpclist))
1991 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1993 list_for_each_entry(ext, &rpclist, oe_link) {
1994 LASSERT(ext->oe_state == OES_CACHE ||
1995 ext->oe_state == OES_LOCK_DONE);
1996 if (ext->oe_state == OES_CACHE)
1997 osc_extent_state_set(ext, OES_LOCKING);
1999 osc_extent_state_set(ext, OES_RPC);
2002 /* we're going to grab page lock, so release object lock because
2003 * lock order is page lock -> object lock. */
2004 osc_object_unlock(osc);
2006 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2007 if (ext->oe_state == OES_LOCKING) {
2008 rc = osc_extent_make_ready(env, ext);
2009 if (unlikely(rc < 0)) {
2010 list_del_init(&ext->oe_link);
2011 osc_extent_finish(env, ext, 0, rc);
2015 if (first == NULL) {
2017 srvlock = ext->oe_srvlock;
2019 LASSERT(srvlock == ext->oe_srvlock);
2023 if (!list_empty(&rpclist)) {
2024 LASSERT(page_count > 0);
2025 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2026 LASSERT(list_empty(&rpclist));
2029 osc_object_lock(osc);
2034 * prepare pages for ASYNC io and put pages in send queue.
2036 * \param cmd OBD_BRW_* macroses
2037 * \param lop pending pages
2039 * \return zero if no page added to send queue.
2040 * \return 1 if pages successfully added to send queue.
2041 * \return negative on errors.
2044 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2045 struct osc_object *osc, pdl_policy_t pol)
2048 struct osc_extent *ext;
2049 struct osc_extent *next;
2050 struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2052 unsigned int max_pages = cli->cl_max_pages_per_rpc;
2056 LASSERT(osc_object_is_locked(osc));
2057 list_for_each_entry_safe(ext, next,
2058 &osc->oo_reading_exts, oe_link) {
2059 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2060 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2063 osc_extent_state_set(ext, OES_RPC);
2064 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2066 LASSERT(page_count <= max_pages);
2068 osc_update_pending(osc, OBD_BRW_READ, -page_count);
2070 if (!list_empty(&rpclist)) {
2071 osc_object_unlock(osc);
2073 LASSERT(page_count > 0);
2074 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2075 LASSERT(list_empty(&rpclist));
2077 osc_object_lock(osc);
2082 #define list_to_obj(list, item) ({ \
2083 struct list_head *__tmp = (list)->next; \
2084 list_del_init(__tmp); \
2085 list_entry(__tmp, struct osc_object, oo_##item); \
2088 /* This is called by osc_check_rpcs() to find which objects have pages that
2089 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2090 static struct osc_object *osc_next_obj(struct client_obd *cli)
2094 /* First return objects that have blocked locks so that they
2095 * will be flushed quickly and other clients can get the lock,
2096 * then objects which have pages ready to be stuffed into RPCs */
2097 if (!list_empty(&cli->cl_loi_hp_ready_list))
2098 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2099 if (!list_empty(&cli->cl_loi_ready_list))
2100 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2102 /* then if we have cache waiters, return all objects with queued
2103 * writes. This is especially important when many small files
2104 * have filled up the cache and not been fired into rpcs because
2105 * they don't pass the nr_pending/object threshhold */
2106 if (!list_empty(&cli->cl_cache_waiters) &&
2107 !list_empty(&cli->cl_loi_write_list))
2108 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2110 /* then return all queued objects when we have an invalid import
2111 * so that they get flushed */
2112 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2113 if (!list_empty(&cli->cl_loi_write_list))
2114 RETURN(list_to_obj(&cli->cl_loi_write_list,
2116 if (!list_empty(&cli->cl_loi_read_list))
2117 RETURN(list_to_obj(&cli->cl_loi_read_list,
2123 /* called with the loi list lock held */
2124 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2126 __must_hold(&cli->cl_loi_list_lock)
2128 struct osc_object *osc;
2132 while ((osc = osc_next_obj(cli)) != NULL) {
2133 struct cl_object *obj = osc2cl(osc);
2134 struct lu_ref_link link;
2136 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2138 if (osc_max_rpc_in_flight(cli, osc)) {
2139 __osc_list_maint(cli, osc);
2144 client_obd_list_unlock(&cli->cl_loi_list_lock);
2145 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2147 /* attempt some read/write balancing by alternating between
2148 * reads and writes in an object. The makes_rpc checks here
2149 * would be redundant if we were getting read/write work items
2150 * instead of objects. we don't want send_oap_rpc to drain a
2151 * partial read pending queue when we're given this object to
2152 * do io on writes while there are cache waiters */
2153 osc_object_lock(osc);
2154 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2155 rc = osc_send_write_rpc(env, cli, osc, pol);
2157 CERROR("Write request failed with %d\n", rc);
2159 /* osc_send_write_rpc failed, mostly because of
2162 * It can't break here, because if:
2163 * - a page was submitted by osc_io_submit, so
2165 * - no request in flight
2166 * - no subsequent request
2167 * The system will be in live-lock state,
2168 * because there is no chance to call
2169 * osc_io_unplug() and osc_check_rpcs() any
2170 * more. pdflush can't help in this case,
2171 * because it might be blocked at grabbing
2172 * the page lock as we mentioned.
2174 * Anyway, continue to drain pages. */
2178 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2179 rc = osc_send_read_rpc(env, cli, osc, pol);
2181 CERROR("Read request failed with %d\n", rc);
2183 osc_object_unlock(osc);
2185 osc_list_maint(cli, osc);
2186 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2187 cl_object_put(env, obj);
2189 client_obd_list_lock(&cli->cl_loi_list_lock);
2193 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2194 struct osc_object *osc, pdl_policy_t pol, int async)
2198 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2202 /* disable osc_lru_shrink() temporarily to avoid
2203 * potential stack overrun problem. LU-2859 */
2204 atomic_inc(&cli->cl_lru_shrinkers);
2205 client_obd_list_lock(&cli->cl_loi_list_lock);
2206 osc_check_rpcs(env, cli, pol);
2207 client_obd_list_unlock(&cli->cl_loi_list_lock);
2208 atomic_dec(&cli->cl_lru_shrinkers);
2210 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2211 LASSERT(cli->cl_writeback_work != NULL);
2212 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2217 static int osc_io_unplug_async(const struct lu_env *env,
2218 struct client_obd *cli, struct osc_object *osc)
2220 /* XXX: policy is no use actually. */
2221 return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2224 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2225 struct osc_object *osc, pdl_policy_t pol)
2227 (void)osc_io_unplug0(env, cli, osc, pol, 0);
2230 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2231 struct page *page, loff_t offset)
2233 struct obd_export *exp = osc_export(osc);
2234 struct osc_async_page *oap = &ops->ops_oap;
2238 return cfs_size_round(sizeof(*oap));
2240 oap->oap_magic = OAP_MAGIC;
2241 oap->oap_cli = &exp->exp_obd->u.cli;
2244 oap->oap_page = page;
2245 oap->oap_obj_off = offset;
2246 LASSERT(!(offset & ~CFS_PAGE_MASK));
2248 if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2249 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2251 INIT_LIST_HEAD(&oap->oap_pending_item);
2252 INIT_LIST_HEAD(&oap->oap_rpc_item);
2254 spin_lock_init(&oap->oap_lock);
2255 CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2256 oap, page, oap->oap_obj_off);
2260 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2261 struct osc_page *ops)
2263 struct osc_io *oio = osc_env_io(env);
2264 struct osc_extent *ext = NULL;
2265 struct osc_async_page *oap = &ops->ops_oap;
2266 struct client_obd *cli = oap->oap_cli;
2267 struct osc_object *osc = oap->oap_obj;
2270 int brw_flags = OBD_BRW_ASYNC;
2271 int cmd = OBD_BRW_WRITE;
2272 int need_release = 0;
2276 if (oap->oap_magic != OAP_MAGIC)
2279 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2282 if (!list_empty(&oap->oap_pending_item) ||
2283 !list_empty(&oap->oap_rpc_item))
2286 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2287 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2288 if (!client_is_remote(osc_export(osc)) &&
2289 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2290 brw_flags |= OBD_BRW_NOQUOTA;
2291 cmd |= OBD_BRW_NOQUOTA;
2294 /* check if the file's owner/group is over quota */
2295 if (!(cmd & OBD_BRW_NOQUOTA)) {
2296 struct cl_object *obj;
2297 struct cl_attr *attr;
2298 unsigned int qid[MAXQUOTAS];
2300 obj = cl_object_top(&osc->oo_cl);
2301 attr = &osc_env_info(env)->oti_attr;
2303 cl_object_attr_lock(obj);
2304 rc = cl_object_attr_get(env, obj, attr);
2305 cl_object_attr_unlock(obj);
2307 qid[USRQUOTA] = attr->cat_uid;
2308 qid[GRPQUOTA] = attr->cat_gid;
2309 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2316 oap->oap_page_off = ops->ops_from;
2317 oap->oap_count = ops->ops_to - ops->ops_from;
2318 /* No need to hold a lock here,
2319 * since this page is not in any list yet. */
2320 oap->oap_async_flags = 0;
2321 oap->oap_brw_flags = brw_flags;
2323 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2324 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2326 index = osc_index(oap2osc(oap));
2328 /* Add this page into extent by the following steps:
2329 * 1. if there exists an active extent for this IO, mostly this page
2330 * can be added to the active extent and sometimes we need to
2331 * expand extent to accomodate this page;
2332 * 2. otherwise, a new extent will be allocated. */
2334 ext = oio->oi_active;
2335 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2336 /* one chunk plus extent overhead must be enough to write this
2338 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2339 if (ext->oe_end >= index)
2342 /* it doesn't need any grant to dirty this page */
2343 client_obd_list_lock(&cli->cl_loi_list_lock);
2344 rc = osc_enter_cache_try(cli, oap, grants, 0);
2345 client_obd_list_unlock(&cli->cl_loi_list_lock);
2346 if (rc == 0) { /* try failed */
2349 } else if (ext->oe_end < index) {
2351 /* try to expand this extent */
2352 rc = osc_extent_expand(ext, index, &tmp);
2355 /* don't free reserved grant */
2357 OSC_EXTENT_DUMP(D_CACHE, ext,
2358 "expanded for %lu.\n", index);
2359 osc_unreserve_grant(cli, grants, tmp);
2364 } else if (ext != NULL) {
2365 /* index is located outside of active extent */
2369 osc_extent_release(env, ext);
2370 oio->oi_active = NULL;
2375 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2377 /* try to find new extent to cover this page */
2378 LASSERT(oio->oi_active == NULL);
2379 /* we may have allocated grant for this page if we failed
2380 * to expand the previous active extent. */
2381 LASSERT(ergo(grants > 0, grants >= tmp));
2385 /* we haven't allocated grant for this page. */
2386 rc = osc_enter_cache(env, cli, oap, tmp);
2393 ext = osc_extent_find(env, osc, index, &tmp);
2395 LASSERT(tmp == grants);
2396 osc_exit_cache(cli, oap);
2400 oio->oi_active = ext;
2404 osc_unreserve_grant(cli, grants, tmp);
2407 LASSERT(ergo(rc == 0, ext != NULL));
2409 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2410 ext, "index = %lu.\n", index);
2411 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2413 osc_object_lock(osc);
2414 if (ext->oe_nr_pages == 0)
2415 ext->oe_srvlock = ops->ops_srvlock;
2417 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2419 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2420 osc_object_unlock(osc);
2425 int osc_teardown_async_page(const struct lu_env *env,
2426 struct osc_object *obj, struct osc_page *ops)
2428 struct osc_async_page *oap = &ops->ops_oap;
2429 struct osc_extent *ext = NULL;
2433 LASSERT(oap->oap_magic == OAP_MAGIC);
2435 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2436 oap, ops, osc_index(oap2osc(oap)));
2438 osc_object_lock(obj);
2439 if (!list_empty(&oap->oap_rpc_item)) {
2440 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2442 } else if (!list_empty(&oap->oap_pending_item)) {
2443 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2444 /* only truncated pages are allowed to be taken out.
2445 * See osc_extent_truncate() and osc_cache_truncate_start()
2447 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2448 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2449 osc_index(oap2osc(oap)));
2453 osc_object_unlock(obj);
2455 osc_extent_put(env, ext);
2460 * This is called when a page is picked up by kernel to write out.
2462 * We should find out the corresponding extent and add the whole extent
2463 * into urgent list. The extent may be being truncated or used, handle it
2466 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2467 struct osc_page *ops)
2469 struct osc_extent *ext = NULL;
2470 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2471 struct cl_page *cp = ops->ops_cl.cpl_page;
2472 pgoff_t index = osc_index(ops);
2473 struct osc_async_page *oap = &ops->ops_oap;
2474 bool unplug = false;
2478 osc_object_lock(obj);
2479 ext = osc_extent_lookup(obj, index);
2481 osc_extent_tree_dump(D_ERROR, obj);
2482 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2485 switch (ext->oe_state) {
2488 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2492 /* If we know this extent is being written out, we should abort
2493 * so that the writer can make this page ready. Otherwise, there
2494 * exists a deadlock problem because other process can wait for
2495 * page writeback bit holding page lock; and meanwhile in
2496 * vvp_page_make_ready(), we need to grab page lock before
2497 * really sending the RPC. */
2499 /* race with truncate, page will be redirtied */
2501 /* The extent is active so we need to abort and let the caller
2502 * re-dirty the page. If we continued on here, and we were the
2503 * one making the extent active, we could deadlock waiting for
2504 * the page writeback to clear but it won't because the extent
2505 * is active and won't be written out. */
2506 GOTO(out, rc = -EAGAIN);
2511 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2515 spin_lock(&oap->oap_lock);
2516 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2517 spin_unlock(&oap->oap_lock);
2519 if (memory_pressure_get())
2520 ext->oe_memalloc = 1;
2523 if (ext->oe_state == OES_CACHE) {
2524 OSC_EXTENT_DUMP(D_CACHE, ext,
2525 "flush page %p make it urgent.\n", oap);
2526 if (list_empty(&ext->oe_link))
2527 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2534 osc_object_unlock(obj);
2535 osc_extent_put(env, ext);
2537 osc_io_unplug_async(env, osc_cli(obj), obj);
2542 * this is called when a sync waiter receives an interruption. Its job is to
2543 * get the caller woken as soon as possible. If its page hasn't been put in an
2544 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2545 * desiring interruption which will forcefully complete the rpc once the rpc
2548 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2550 struct osc_async_page *oap = &ops->ops_oap;
2551 struct osc_object *obj = oap->oap_obj;
2552 struct client_obd *cli = osc_cli(obj);
2553 struct osc_extent *ext;
2554 struct osc_extent *found = NULL;
2555 struct list_head *plist;
2556 pgoff_t index = osc_index(ops);
2561 LASSERT(!oap->oap_interrupted);
2562 oap->oap_interrupted = 1;
2564 /* Find out the caching extent */
2565 osc_object_lock(obj);
2566 if (oap->oap_cmd & OBD_BRW_WRITE) {
2567 plist = &obj->oo_urgent_exts;
2568 cmd = OBD_BRW_WRITE;
2570 plist = &obj->oo_reading_exts;
2573 list_for_each_entry(ext, plist, oe_link) {
2574 if (ext->oe_start <= index && ext->oe_end >= index) {
2575 LASSERT(ext->oe_state == OES_LOCK_DONE);
2576 /* For OES_LOCK_DONE state extent, it has already held
2577 * a refcount for RPC. */
2578 found = osc_extent_get(ext);
2582 if (found != NULL) {
2583 list_del_init(&found->oe_link);
2584 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2585 osc_object_unlock(obj);
2587 osc_extent_finish(env, found, 0, -EINTR);
2588 osc_extent_put(env, found);
2591 osc_object_unlock(obj);
2592 /* ok, it's been put in an rpc. only one oap gets a request
2594 if (oap->oap_request != NULL) {
2595 ptlrpc_mark_interrupted(oap->oap_request);
2596 ptlrpcd_wake(oap->oap_request);
2597 ptlrpc_req_finished(oap->oap_request);
2598 oap->oap_request = NULL;
2602 osc_list_maint(cli, obj);
2606 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2607 struct list_head *list, int cmd, int brw_flags)
2609 struct client_obd *cli = osc_cli(obj);
2610 struct osc_extent *ext;
2611 struct osc_async_page *oap;
2613 int mppr = cli->cl_max_pages_per_rpc;
2614 pgoff_t start = CL_PAGE_EOF;
2618 list_for_each_entry(oap, list, oap_pending_item) {
2619 pgoff_t index = osc_index(oap2osc(oap));
2625 mppr <<= (page_count > mppr);
2628 ext = osc_extent_alloc(obj);
2630 list_for_each_entry(oap, list, oap_pending_item) {
2631 list_del_init(&oap->oap_pending_item);
2632 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2637 ext->oe_rw = !!(cmd & OBD_BRW_READ);
2640 ext->oe_start = start;
2641 ext->oe_end = ext->oe_max_end = end;
2643 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2644 ext->oe_nr_pages = page_count;
2645 ext->oe_mppr = mppr;
2646 list_splice_init(list, &ext->oe_pages);
2648 osc_object_lock(obj);
2649 /* Reuse the initial refcount for RPC, don't drop it */
2650 osc_extent_state_set(ext, OES_LOCK_DONE);
2651 if (cmd & OBD_BRW_WRITE) {
2652 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2653 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2655 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2656 osc_update_pending(obj, OBD_BRW_READ, page_count);
2658 osc_object_unlock(obj);
2660 osc_io_unplug_async(env, cli, obj);
2665 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2667 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2668 struct osc_object *obj, __u64 size)
2670 struct client_obd *cli = osc_cli(obj);
2671 struct osc_extent *ext;
2672 struct osc_extent *waiting = NULL;
2674 struct list_head list = LIST_HEAD_INIT(list);
2679 /* pages with index greater or equal to index will be truncated. */
2680 index = cl_index(osc2cl(obj), size);
2681 partial = size > cl_offset(osc2cl(obj), index);
2684 osc_object_lock(obj);
2685 ext = osc_extent_search(obj, index);
2687 ext = first_extent(obj);
2688 else if (ext->oe_end < index)
2689 ext = next_extent(ext);
2690 while (ext != NULL) {
2691 EASSERT(ext->oe_state != OES_TRUNC, ext);
2693 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2694 /* if ext is in urgent state, it means there must exist
2695 * a page already having been flushed by write_page().
2696 * We have to wait for this extent because we can't
2697 * truncate that page. */
2698 LASSERT(!ext->oe_hp);
2699 OSC_EXTENT_DUMP(D_CACHE, ext,
2700 "waiting for busy extent\n");
2701 waiting = osc_extent_get(ext);
2705 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2707 osc_extent_get(ext);
2708 if (ext->oe_state == OES_ACTIVE) {
2709 /* though we grab inode mutex for write path, but we
2710 * release it before releasing extent(in osc_io_end()),
2711 * so there is a race window that an extent is still
2712 * in OES_ACTIVE when truncate starts. */
2713 LASSERT(!ext->oe_trunc_pending);
2714 ext->oe_trunc_pending = 1;
2716 EASSERT(ext->oe_state == OES_CACHE, ext);
2717 osc_extent_state_set(ext, OES_TRUNC);
2718 osc_update_pending(obj, OBD_BRW_WRITE,
2721 EASSERT(list_empty(&ext->oe_link), ext);
2722 list_add_tail(&ext->oe_link, &list);
2724 ext = next_extent(ext);
2726 osc_object_unlock(obj);
2728 osc_list_maint(cli, obj);
2730 while (!list_empty(&list)) {
2733 ext = list_entry(list.next, struct osc_extent, oe_link);
2734 list_del_init(&ext->oe_link);
2736 /* extent may be in OES_ACTIVE state because inode mutex
2737 * is released before osc_io_end() in file write case */
2738 if (ext->oe_state != OES_TRUNC)
2739 osc_extent_wait(env, ext, OES_TRUNC);
2741 rc = osc_extent_truncate(ext, index, partial);
2746 OSC_EXTENT_DUMP(D_ERROR, ext,
2747 "truncate error %d\n", rc);
2748 } else if (ext->oe_nr_pages == 0) {
2749 osc_extent_remove(ext);
2751 /* this must be an overlapped extent which means only
2752 * part of pages in this extent have been truncated.
2754 EASSERTF(ext->oe_start <= index, ext,
2755 "trunc index = %lu/%d.\n", index, partial);
2756 /* fix index to skip this partially truncated extent */
2757 index = ext->oe_end + 1;
2760 /* we need to hold this extent in OES_TRUNC state so
2761 * that no writeback will happen. This is to avoid
2763 LASSERT(oio->oi_trunc == NULL);
2764 oio->oi_trunc = osc_extent_get(ext);
2765 OSC_EXTENT_DUMP(D_CACHE, ext,
2766 "trunc at "LPU64"\n", size);
2768 osc_extent_put(env, ext);
2770 if (waiting != NULL) {
2773 /* ignore the result of osc_extent_wait the write initiator
2774 * should take care of it. */
2775 rc = osc_extent_wait(env, waiting, OES_INV);
2777 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2779 osc_extent_put(env, waiting);
2787 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2789 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2790 struct osc_object *obj)
2792 struct osc_extent *ext = oio->oi_trunc;
2794 oio->oi_trunc = NULL;
2796 bool unplug = false;
2798 EASSERT(ext->oe_nr_pages > 0, ext);
2799 EASSERT(ext->oe_state == OES_TRUNC, ext);
2800 EASSERT(!ext->oe_urgent, ext);
2802 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2803 osc_object_lock(obj);
2804 osc_extent_state_set(ext, OES_CACHE);
2805 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2807 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2810 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2811 osc_object_unlock(obj);
2812 osc_extent_put(env, ext);
2815 osc_io_unplug_async(env, osc_cli(obj), obj);
2820 * Wait for extents in a specific range to be written out.
2821 * The caller must have called osc_cache_writeback_range() to issue IO
2822 * otherwise it will take a long time for this function to finish.
2824 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2825 * nobody else can dirty this range of file while we're waiting for
2826 * extents to be written.
2828 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2829 pgoff_t start, pgoff_t end)
2831 struct osc_extent *ext;
2832 pgoff_t index = start;
2837 osc_object_lock(obj);
2838 ext = osc_extent_search(obj, index);
2840 ext = first_extent(obj);
2841 else if (ext->oe_end < index)
2842 ext = next_extent(ext);
2843 while (ext != NULL) {
2846 if (ext->oe_start > end)
2849 if (!ext->oe_fsync_wait) {
2850 ext = next_extent(ext);
2854 EASSERT(ergo(ext->oe_state == OES_CACHE,
2855 ext->oe_hp || ext->oe_urgent), ext);
2856 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2857 !ext->oe_hp && ext->oe_urgent), ext);
2859 index = ext->oe_end + 1;
2860 osc_extent_get(ext);
2861 osc_object_unlock(obj);
2863 rc = osc_extent_wait(env, ext, OES_INV);
2866 osc_extent_put(env, ext);
2869 osc_object_unlock(obj);
2871 OSC_IO_DEBUG(obj, "sync file range.\n");
2876 * Called to write out a range of osc object.
2878 * @hp : should be set this is caused by lock cancel;
2879 * @discard: is set if dirty pages should be dropped - file will be deleted or
2880 * truncated, this implies there is no partially discarding extents.
2882 * Return how many pages will be issued, or error code if error occurred.
2884 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2885 pgoff_t start, pgoff_t end, int hp, int discard)
2887 struct osc_extent *ext;
2888 struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2889 bool unplug = false;
2893 osc_object_lock(obj);
2894 ext = osc_extent_search(obj, start);
2896 ext = first_extent(obj);
2897 else if (ext->oe_end < start)
2898 ext = next_extent(ext);
2899 while (ext != NULL) {
2900 if (ext->oe_start > end)
2903 ext->oe_fsync_wait = 1;
2904 switch (ext->oe_state) {
2906 result += ext->oe_nr_pages;
2908 struct list_head *list = NULL;
2910 EASSERT(!ext->oe_hp, ext);
2912 list = &obj->oo_hp_exts;
2913 } else if (!ext->oe_urgent) {
2915 list = &obj->oo_urgent_exts;
2918 list_move_tail(&ext->oe_link, list);
2921 /* the only discarder is lock cancelling, so
2922 * [start, end] must contain this extent */
2923 EASSERT(ext->oe_start >= start &&
2924 ext->oe_max_end <= end, ext);
2925 osc_extent_state_set(ext, OES_LOCKING);
2926 ext->oe_owner = current;
2927 list_move_tail(&ext->oe_link,
2929 osc_update_pending(obj, OBD_BRW_WRITE,
2934 /* It's pretty bad to wait for ACTIVE extents, because
2935 * we don't know how long we will wait for it to be
2936 * flushed since it may be blocked at awaiting more
2937 * grants. We do this for the correctness of fsync. */
2938 LASSERT(hp == 0 && discard == 0);
2942 /* this extent is being truncated, can't do anything
2943 * for it now. it will be set to urgent after truncate
2944 * is finished in osc_cache_truncate_end(). */
2948 ext = next_extent(ext);
2950 osc_object_unlock(obj);
2952 LASSERT(ergo(!discard, list_empty(&discard_list)));
2953 if (!list_empty(&discard_list)) {
2954 struct osc_extent *tmp;
2957 osc_list_maint(osc_cli(obj), obj);
2958 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2959 list_del_init(&ext->oe_link);
2960 EASSERT(ext->oe_state == OES_LOCKING, ext);
2962 /* Discard caching pages. We don't actually write this
2963 * extent out but we complete it as if we did. */
2964 rc = osc_extent_make_ready(env, ext);
2965 if (unlikely(rc < 0)) {
2966 OSC_EXTENT_DUMP(D_ERROR, ext,
2967 "make_ready returned %d\n", rc);
2972 /* finish the extent as if the pages were sent */
2973 osc_extent_finish(env, ext, 0, 0);
2978 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2980 if (hp || discard) {
2982 rc = osc_cache_wait_range(env, obj, start, end);
2983 if (result >= 0 && rc < 0)
2987 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2992 * Returns a list of pages by a given [start, end] of \a obj.
2994 * \param resched If not NULL, then we give up before hogging CPU for too
2995 * long and set *resched = 1, in that case caller should implement a retry
2998 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
2999 * crucial in the face of [offset, EOF] locks.
3001 * Return at least one page in @queue unless there is no covered page.
3003 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3004 struct osc_object *osc, pgoff_t start, pgoff_t end,
3005 osc_page_gang_cbt cb, void *cbdata)
3007 struct osc_page *ops;
3013 int res = CLP_GANG_OKAY;
3014 bool tree_lock = true;
3018 pvec = osc_env_info(env)->oti_pvec;
3019 spin_lock(&osc->oo_tree_lock);
3020 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3021 idx, OTI_PVEC_SIZE)) > 0) {
3022 struct cl_page *page;
3023 bool end_of_region = false;
3025 for (i = 0, j = 0; i < nr; ++i) {
3029 idx = osc_index(ops);
3031 end_of_region = true;
3035 page = ops->ops_cl.cpl_page;
3036 LASSERT(page->cp_type == CPT_CACHEABLE);
3037 if (page->cp_state == CPS_FREEING)
3041 lu_ref_add_atomic(&page->cp_reference,
3042 "gang_lookup", current);
3048 * Here a delicate locking dance is performed. Current thread
3049 * holds a reference to a page, but has to own it before it
3050 * can be placed into queue. Owning implies waiting, so
3051 * radix-tree lock is to be released. After a wait one has to
3052 * check that pages weren't truncated (cl_page_own() returns
3053 * error in the latter case).
3055 spin_unlock(&osc->oo_tree_lock);
3058 for (i = 0; i < j; ++i) {
3060 if (res == CLP_GANG_OKAY)
3061 res = (*cb)(env, io, ops, cbdata);
3063 page = ops->ops_cl.cpl_page;
3064 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3065 cl_page_put(env, page);
3067 if (nr < OTI_PVEC_SIZE || end_of_region)
3070 if (res == CLP_GANG_OKAY && need_resched())
3071 res = CLP_GANG_RESCHED;
3072 if (res != CLP_GANG_OKAY)
3075 spin_lock(&osc->oo_tree_lock);
3079 spin_unlock(&osc->oo_tree_lock);
3084 * Check if page @page is covered by an extra lock or discard it.
3086 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3087 struct osc_page *ops, void *cbdata)
3089 struct osc_thread_info *info = osc_env_info(env);
3090 struct osc_object *osc = cbdata;
3093 index = osc_index(ops);
3094 if (index >= info->oti_fn_index) {
3095 struct ldlm_lock *tmp;
3096 struct cl_page *page = ops->ops_cl.cpl_page;
3098 /* refresh non-overlapped index */
3099 tmp = osc_dlmlock_at_pgoff(env, osc, index, 0, 0);
3101 __u64 end = tmp->l_policy_data.l_extent.end;
3102 /* Cache the first-non-overlapped index so as to skip
3103 * all pages within [index, oti_fn_index). This is safe
3104 * because if tmp lock is canceled, it will discard
3106 info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3107 if (end == OBD_OBJECT_EOF)
3108 info->oti_fn_index = CL_PAGE_EOF;
3110 } else if (cl_page_own(env, io, page) == 0) {
3111 /* discard the page */
3112 cl_page_discard(env, io, page);
3113 cl_page_disown(env, io, page);
3115 LASSERT(page->cp_state == CPS_FREEING);
3119 info->oti_next_index = index + 1;
3120 return CLP_GANG_OKAY;
3123 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3124 struct osc_page *ops, void *cbdata)
3126 struct osc_thread_info *info = osc_env_info(env);
3127 struct cl_page *page = ops->ops_cl.cpl_page;
3129 /* page is top page. */
3130 info->oti_next_index = osc_index(ops) + 1;
3131 if (cl_page_own(env, io, page) == 0) {
3132 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3133 !PageDirty(cl_page_vmpage(page))));
3135 /* discard the page */
3136 cl_page_discard(env, io, page);
3137 cl_page_disown(env, io, page);
3139 LASSERT(page->cp_state == CPS_FREEING);
3142 return CLP_GANG_OKAY;
3146 * Discard pages protected by the given lock. This function traverses radix
3147 * tree to find all covering pages and discard them. If a page is being covered
3148 * by other locks, it should remain in cache.
3150 * If error happens on any step, the process continues anyway (the reasoning
3151 * behind this being that lock cancellation cannot be delayed indefinitely).
3153 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3154 pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3156 struct osc_thread_info *info = osc_env_info(env);
3157 struct cl_io *io = &info->oti_io;
3158 osc_page_gang_cbt cb;
3164 io->ci_obj = cl_object_top(osc2cl(osc));
3165 io->ci_ignore_layout = 1;
3166 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3170 cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3171 info->oti_fn_index = info->oti_next_index = start;
3173 res = osc_page_gang_lookup(env, io, osc,
3174 info->oti_next_index, end, cb, osc);
3175 if (info->oti_next_index > end)
3178 if (res == CLP_GANG_RESCHED)
3180 } while (res != CLP_GANG_OKAY);
3182 cl_io_fini(env, io);