4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
30 * This file is part of Lustre, http://www.lustre.org/
31 * Lustre is a trademark of Sun Microsystems, Inc.
33 * osc cache management.
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_OSC
40 #include <lustre_osc.h>
42 #include "osc_internal.h"
44 static int extent_debug; /* set it to be true for more debug */
46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
48 enum osc_extent_state state);
49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
50 struct osc_async_page *oap, int sent, int rc);
51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
53 static int osc_refresh_count(const struct lu_env *env,
54 struct osc_async_page *oap, int cmd);
55 static int osc_io_unplug_async(const struct lu_env *env,
56 struct client_obd *cli, struct osc_object *osc);
57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
58 unsigned int lost_grant, unsigned int dirty_grant);
60 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
61 const char *func, int line);
62 #define osc_extent_tree_dump(mask, obj) \
63 osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
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 #define EXTSTR "[%lu -> %lu/%lu]"
98 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
99 static const char *oes_strings[] = {
100 "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
102 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
103 static struct cfs_debug_limit_state cdls; \
104 struct osc_extent *__ext = (extent); \
107 __CDEBUG_WITH_LOC(file, func, line, mask, &cdls, \
108 "extent %p@{" EXTSTR ", " \
109 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt, \
110 /* ----- extent part 0 ----- */ \
111 __ext, EXTPARA(__ext), \
112 /* ----- part 1 ----- */ \
113 atomic_read(&__ext->oe_refc), \
114 atomic_read(&__ext->oe_users), \
115 list_empty_marker(&__ext->oe_link), \
116 oes_strings[__ext->oe_state], ext_flags(__ext, __buf), \
118 /* ----- part 2 ----- */ \
119 __ext->oe_grants, __ext->oe_nr_pages, \
120 list_empty_marker(&__ext->oe_pages), \
121 waitqueue_active(&__ext->oe_waitq) ? '+' : '-', \
122 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner, \
123 /* ----- part 4 ----- */ \
125 if (mask == D_ERROR && __ext->oe_dlmlock != NULL) \
126 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext); \
128 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext); \
131 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...) \
132 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__, \
133 mask, ext, fmt, ## __VA_ARGS__)
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 assert_osc_object_is_locked(obj);
189 if (ext->oe_state >= OES_STATE_MAX)
192 if (atomic_read(&ext->oe_refc) <= 0)
195 if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
198 switch (ext->oe_state) {
200 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
205 if (atomic_read(&ext->oe_users) == 0)
209 if (ext->oe_fsync_wait && !ext->oe_urgent)
213 if (ext->oe_grants == 0)
215 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 ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
231 !ldlm_is_failed(ext->oe_dlmlock)) {
232 struct ldlm_extent *extent;
234 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
235 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
236 extent->end >= cl_offset(osc2cl(obj), ext->oe_max_end)))
239 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
243 if (ext->oe_nr_pages > ext->oe_mppr)
246 /* Do not verify page list if extent is in RPC. This is because an
247 * in-RPC extent is supposed to be exclusively accessible w/o lock. */
248 if (ext->oe_state > OES_CACHE)
255 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
256 pgoff_t index = osc_index(oap2osc(oap));
258 if (index > ext->oe_end || index < ext->oe_start)
261 if (page_count != ext->oe_nr_pages)
266 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
267 "sanity check %p failed: rc = %d\n",
272 #define sanity_check_nolock(ext) \
273 osc_extent_sanity_check0(ext, __func__, __LINE__)
275 #define sanity_check(ext) ({ \
277 osc_object_lock((ext)->oe_obj); \
278 __res = sanity_check_nolock(ext); \
279 osc_object_unlock((ext)->oe_obj); \
285 * sanity check - to make sure there is no overlapped extent in the tree.
287 static int osc_extent_is_overlapped(struct osc_object *obj,
288 struct osc_extent *ext)
290 struct osc_extent *tmp;
292 assert_osc_object_is_locked(obj);
297 for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
300 if (tmp->oe_end >= ext->oe_start &&
301 tmp->oe_start <= ext->oe_end)
307 static void osc_extent_state_set(struct osc_extent *ext, int state)
309 assert_osc_object_is_locked(ext->oe_obj);
310 LASSERT(state >= OES_INV && state < OES_STATE_MAX);
312 /* Never try to sanity check a state changing extent :-) */
313 /* LASSERT(sanity_check_nolock(ext) == 0); */
315 /* TODO: validate the state machine */
316 ext->oe_state = state;
317 wake_up_all(&ext->oe_waitq);
320 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
322 struct osc_extent *ext;
324 OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
328 RB_CLEAR_NODE(&ext->oe_node);
330 cl_object_get(osc2cl(obj));
331 atomic_set(&ext->oe_refc, 1);
332 atomic_set(&ext->oe_users, 0);
333 INIT_LIST_HEAD(&ext->oe_link);
334 ext->oe_state = OES_INV;
335 INIT_LIST_HEAD(&ext->oe_pages);
336 init_waitqueue_head(&ext->oe_waitq);
337 ext->oe_dlmlock = NULL;
342 static void osc_extent_free(struct osc_extent *ext)
344 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
347 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
349 LASSERT(atomic_read(&ext->oe_refc) >= 0);
350 atomic_inc(&ext->oe_refc);
354 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
356 LASSERT(atomic_read(&ext->oe_refc) > 0);
357 if (atomic_dec_and_test(&ext->oe_refc)) {
358 LASSERT(list_empty(&ext->oe_link));
359 LASSERT(atomic_read(&ext->oe_users) == 0);
360 LASSERT(ext->oe_state == OES_INV);
361 LASSERT(!ext->oe_intree);
363 if (ext->oe_dlmlock != NULL) {
364 lu_ref_add(&ext->oe_dlmlock->l_reference,
366 LDLM_LOCK_PUT(ext->oe_dlmlock);
367 ext->oe_dlmlock = NULL;
369 cl_object_put(env, osc2cl(ext->oe_obj));
370 osc_extent_free(ext);
375 * osc_extent_put_trust() is a special version of osc_extent_put() when
376 * it's known that the caller is not the last user. This is to address the
377 * problem of lacking of lu_env ;-).
379 static void osc_extent_put_trust(struct osc_extent *ext)
381 LASSERT(atomic_read(&ext->oe_refc) > 1);
382 assert_osc_object_is_locked(ext->oe_obj);
383 atomic_dec(&ext->oe_refc);
387 * Return the extent which includes pgoff @index, or return the greatest
388 * previous extent in the tree.
390 static struct osc_extent *osc_extent_search(struct osc_object *obj,
393 struct rb_node *n = obj->oo_root.rb_node;
394 struct osc_extent *tmp, *p = NULL;
396 assert_osc_object_is_locked(obj);
399 if (index < tmp->oe_start) {
401 } else if (index > tmp->oe_end) {
412 * Return the extent covering @index, otherwise return NULL.
413 * caller must have held object lock.
415 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
418 struct osc_extent *ext;
420 ext = osc_extent_search(obj, index);
421 if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
422 return osc_extent_get(ext);
426 /* caller must have held object lock. */
427 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
429 struct rb_node **n = &obj->oo_root.rb_node;
430 struct rb_node *parent = NULL;
431 struct osc_extent *tmp;
433 LASSERT(ext->oe_intree == 0);
434 LASSERT(ext->oe_obj == obj);
435 assert_osc_object_is_locked(obj);
440 if (ext->oe_end < tmp->oe_start)
442 else if (ext->oe_start > tmp->oe_end)
445 EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
447 rb_link_node(&ext->oe_node, parent, n);
448 rb_insert_color(&ext->oe_node, &obj->oo_root);
453 /* caller must have held object lock. */
454 static void osc_extent_erase(struct osc_extent *ext)
456 struct osc_object *obj = ext->oe_obj;
457 assert_osc_object_is_locked(obj);
458 if (ext->oe_intree) {
459 rb_erase(&ext->oe_node, &obj->oo_root);
461 /* rbtree held a refcount */
462 osc_extent_put_trust(ext);
466 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
468 struct osc_object *obj = ext->oe_obj;
470 assert_osc_object_is_locked(obj);
471 LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
472 if (ext->oe_state == OES_CACHE) {
473 osc_extent_state_set(ext, OES_ACTIVE);
474 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
476 atomic_inc(&ext->oe_users);
477 list_del_init(&ext->oe_link);
478 return osc_extent_get(ext);
481 static void __osc_extent_remove(struct osc_extent *ext)
483 assert_osc_object_is_locked(ext->oe_obj);
484 LASSERT(list_empty(&ext->oe_pages));
485 osc_extent_erase(ext);
486 list_del_init(&ext->oe_link);
487 osc_extent_state_set(ext, OES_INV);
488 OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
491 static void osc_extent_remove(struct osc_extent *ext)
493 struct osc_object *obj = ext->oe_obj;
495 osc_object_lock(obj);
496 __osc_extent_remove(ext);
497 osc_object_unlock(obj);
501 * This function is used to merge extents to get better performance. It checks
502 * if @cur and @victim are contiguous at block level.
504 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
505 struct osc_extent *victim)
507 struct osc_object *obj = cur->oe_obj;
508 struct client_obd *cli = osc_cli(obj);
513 LASSERT(cur->oe_state == OES_CACHE);
514 assert_osc_object_is_locked(obj);
518 if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
521 if (cur->oe_max_end != victim->oe_max_end)
524 LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
525 ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
526 chunk_start = cur->oe_start >> ppc_bits;
527 chunk_end = cur->oe_end >> ppc_bits;
528 if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
529 chunk_end + 1 != victim->oe_start >> ppc_bits)
532 /* overall extent size should not exceed the max supported limit
533 * reported by the server */
534 if (cur->oe_end - cur->oe_start + 1 +
535 victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
538 OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
540 cur->oe_start = min(cur->oe_start, victim->oe_start);
541 cur->oe_end = max(cur->oe_end, victim->oe_end);
542 /* per-extent tax should be accounted only once for the whole extent */
543 cur->oe_grants += victim->oe_grants - cli->cl_grant_extent_tax;
544 cur->oe_nr_pages += victim->oe_nr_pages;
545 /* only the following bits are needed to merge */
546 cur->oe_urgent |= victim->oe_urgent;
547 cur->oe_memalloc |= victim->oe_memalloc;
548 list_splice_init(&victim->oe_pages, &cur->oe_pages);
549 list_del_init(&victim->oe_link);
550 victim->oe_nr_pages = 0;
552 osc_extent_get(victim);
553 __osc_extent_remove(victim);
554 osc_extent_put(env, victim);
556 OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
561 * Drop user count of osc_extent, and unplug IO asynchronously.
563 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
565 struct osc_object *obj = ext->oe_obj;
566 struct client_obd *cli = osc_cli(obj);
570 LASSERT(atomic_read(&ext->oe_users) > 0);
571 LASSERT(sanity_check(ext) == 0);
572 LASSERT(ext->oe_grants > 0);
574 if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
575 LASSERT(ext->oe_state == OES_ACTIVE);
576 if (ext->oe_trunc_pending) {
577 /* a truncate process is waiting for this extent.
578 * This may happen due to a race, check
579 * osc_cache_truncate_start(). */
580 osc_extent_state_set(ext, OES_TRUNC);
581 ext->oe_trunc_pending = 0;
582 osc_object_unlock(obj);
586 osc_extent_state_set(ext, OES_CACHE);
587 osc_update_pending(obj, OBD_BRW_WRITE,
590 /* try to merge the previous and next extent. */
591 if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
592 grant += cli->cl_grant_extent_tax;
593 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
594 grant += cli->cl_grant_extent_tax;
597 list_move_tail(&ext->oe_link,
598 &obj->oo_urgent_exts);
599 else if (ext->oe_nr_pages == ext->oe_mppr) {
600 list_move_tail(&ext->oe_link,
603 osc_object_unlock(obj);
605 osc_unreserve_grant(cli, 0, grant);
608 osc_io_unplug_async(env, cli, obj);
610 osc_extent_put(env, ext);
615 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
617 return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
621 * Find or create an extent which includes @index, core function to manage
624 static struct osc_extent *osc_extent_find(const struct lu_env *env,
625 struct osc_object *obj, pgoff_t index,
626 unsigned int *grants)
628 struct client_obd *cli = osc_cli(obj);
629 struct osc_lock *olck;
630 struct cl_lock_descr *descr;
631 struct osc_extent *cur;
632 struct osc_extent *ext;
633 struct osc_extent *conflict = NULL;
634 struct osc_extent *found = NULL;
637 unsigned int max_pages; /* max_pages_per_rpc */
638 unsigned int chunksize;
639 int ppc_bits; /* pages per chunk bits */
644 cur = osc_extent_alloc(obj);
646 RETURN(ERR_PTR(-ENOMEM));
648 olck = osc_env_io(env)->oi_write_osclock;
649 LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
650 LASSERT(olck->ols_state == OLS_GRANTED);
652 descr = &olck->ols_cl.cls_lock->cll_descr;
653 LASSERT(descr->cld_mode >= CLM_WRITE);
655 LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
656 "chunkbits: %u\n", cli->cl_chunkbits);
657 ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
658 chunk_mask = ~((1 << ppc_bits) - 1);
659 chunksize = 1 << cli->cl_chunkbits;
660 chunk = index >> ppc_bits;
662 /* align end to RPC edge. */
663 max_pages = cli->cl_max_pages_per_rpc;
664 if ((max_pages & ~chunk_mask) != 0) {
665 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
666 max_pages, cli->cl_chunkbits, chunk_mask);
667 RETURN(ERR_PTR(-EINVAL));
669 max_end = index - (index % max_pages) + max_pages - 1;
670 max_end = min_t(pgoff_t, max_end, descr->cld_end);
672 /* initialize new extent by parameters so far */
673 cur->oe_max_end = max_end;
674 cur->oe_start = index & chunk_mask;
675 cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
676 if (cur->oe_start < descr->cld_start)
677 cur->oe_start = descr->cld_start;
678 if (cur->oe_end > max_end)
679 cur->oe_end = max_end;
681 cur->oe_mppr = max_pages;
682 if (olck->ols_dlmlock != NULL) {
683 LASSERT(olck->ols_hold);
684 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
685 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
688 /* grants has been allocated by caller */
689 LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
690 "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
691 LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
695 osc_object_lock(obj);
696 ext = osc_extent_search(obj, cur->oe_start);
698 ext = first_extent(obj);
699 while (ext != NULL) {
700 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
701 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
703 LASSERT(sanity_check_nolock(ext) == 0);
704 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
707 /* if covering by different locks, no chance to match */
708 if (olck->ols_dlmlock != ext->oe_dlmlock) {
709 EASSERTF(!overlapped(ext, cur), ext,
710 EXTSTR"\n", EXTPARA(cur));
712 ext = next_extent(ext);
716 /* discontiguous chunks? */
717 if (chunk + 1 < ext_chk_start) {
718 ext = next_extent(ext);
722 /* ok, from now on, ext and cur have these attrs:
723 * 1. covered by the same lock
724 * 2. contiguous at chunk level or overlapping. */
726 if (overlapped(ext, cur)) {
727 /* cur is the minimum unit, so overlapping means
729 EASSERTF((ext->oe_start <= cur->oe_start &&
730 ext->oe_end >= cur->oe_end),
731 ext, EXTSTR"\n", EXTPARA(cur));
733 if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
734 /* for simplicity, we wait for this extent to
735 * finish before going forward. */
736 conflict = osc_extent_get(ext);
740 found = osc_extent_hold(ext);
744 /* non-overlapped extent */
745 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
746 /* we can't do anything for a non OES_CACHE extent, or
747 * if there is someone waiting for this extent to be
748 * flushed, try next one. */
749 ext = next_extent(ext);
753 /* check if they belong to the same rpc slot before trying to
754 * merge. the extents are not overlapped and contiguous at
755 * chunk level to get here. */
756 if (ext->oe_max_end != max_end) {
757 /* if they don't belong to the same RPC slot or
758 * max_pages_per_rpc has ever changed, do not merge. */
759 ext = next_extent(ext);
763 /* check whether maximum extent size will be hit */
764 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
765 cli->cl_max_extent_pages) {
766 ext = next_extent(ext);
770 /* it's required that an extent must be contiguous at chunk
771 * level so that we know the whole extent is covered by grant
772 * (the pages in the extent are NOT required to be contiguous).
773 * Otherwise, it will be too much difficult to know which
774 * chunks have grants allocated. */
776 /* try to do front merge - extend ext's start */
777 if (chunk + 1 == ext_chk_start) {
778 /* ext must be chunk size aligned */
779 EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
781 /* pull ext's start back to cover cur */
782 ext->oe_start = cur->oe_start;
783 ext->oe_grants += chunksize;
784 LASSERT(*grants >= chunksize);
785 *grants -= chunksize;
787 found = osc_extent_hold(ext);
788 } else if (chunk == ext_chk_end + 1) {
790 ext->oe_end = cur->oe_end;
791 ext->oe_grants += chunksize;
792 LASSERT(*grants >= chunksize);
793 *grants -= chunksize;
795 /* try to merge with the next one because we just fill
797 if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
798 /* we can save extent tax from next extent */
799 *grants += cli->cl_grant_extent_tax;
801 found = osc_extent_hold(ext);
806 ext = next_extent(ext);
809 osc_extent_tree_dump(D_CACHE, obj);
811 LASSERT(conflict == NULL);
812 if (!IS_ERR(found)) {
813 LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
814 OSC_EXTENT_DUMP(D_CACHE, found,
815 "found caching ext for %lu.\n", index);
817 } else if (conflict == NULL) {
818 /* create a new extent */
819 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
820 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
821 LASSERT(*grants >= cur->oe_grants);
822 *grants -= cur->oe_grants;
824 cur->oe_state = OES_CACHE;
825 found = osc_extent_hold(cur);
826 osc_extent_insert(obj, cur);
827 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
828 index, descr->cld_end);
830 osc_object_unlock(obj);
832 if (conflict != NULL) {
833 LASSERT(found == NULL);
835 /* waiting for IO to finish. Please notice that it's impossible
836 * to be an OES_TRUNC extent. */
837 rc = osc_extent_wait(env, conflict, OES_INV);
838 osc_extent_put(env, conflict);
841 GOTO(out, found = ERR_PTR(rc));
848 osc_extent_put(env, cur);
853 * Called when IO is finished to an extent.
855 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
858 struct client_obd *cli = osc_cli(ext->oe_obj);
859 struct osc_async_page *oap;
860 struct osc_async_page *tmp;
861 int nr_pages = ext->oe_nr_pages;
863 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
868 OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
870 ext->oe_rc = rc ?: ext->oe_nr_pages;
871 EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
873 osc_lru_add_batch(cli, &ext->oe_pages);
874 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
876 list_del_init(&oap->oap_rpc_item);
877 list_del_init(&oap->oap_pending_item);
878 if (last_off <= oap->oap_obj_off) {
879 last_off = oap->oap_obj_off;
880 last_count = oap->oap_count;
884 osc_ap_completion(env, cli, oap, sent, rc);
886 EASSERT(ext->oe_nr_pages == 0, ext);
889 lost_grant = ext->oe_grants;
890 } else if (blocksize < PAGE_SIZE &&
891 last_count != PAGE_SIZE) {
892 /* For short writes we shouldn't count parts of pages that
893 * span a whole chunk on the OST side, or our accounting goes
894 * wrong. Should match the code in filter_grant_check. */
895 int offset = last_off & ~PAGE_MASK;
896 int count = last_count + (offset & (blocksize - 1));
897 int end = (offset + last_count) & (blocksize - 1);
899 count += blocksize - end;
901 lost_grant = PAGE_SIZE - count;
903 if (ext->oe_grants > 0)
904 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
906 osc_extent_remove(ext);
907 /* put the refcount for RPC */
908 osc_extent_put(env, ext);
912 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
916 osc_object_lock(ext->oe_obj);
917 ret = ext->oe_state == state;
918 osc_object_unlock(ext->oe_obj);
924 * Wait for the extent's state to become @state.
926 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
927 enum osc_extent_state state)
929 struct osc_object *obj = ext->oe_obj;
933 osc_object_lock(obj);
934 LASSERT(sanity_check_nolock(ext) == 0);
935 /* `Kick' this extent only if the caller is waiting for it to be
937 if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
938 if (ext->oe_state == OES_ACTIVE) {
940 } else if (ext->oe_state == OES_CACHE) {
942 osc_extent_hold(ext);
946 osc_object_unlock(obj);
948 osc_extent_release(env, ext);
950 /* wait for the extent until its state becomes @state */
951 rc = wait_event_idle_timeout(ext->oe_waitq, extent_wait_cb(ext, state),
952 cfs_time_seconds(600));
954 OSC_EXTENT_DUMP(D_ERROR, ext,
955 "%s: wait ext to %u timedout, recovery in progress?\n",
956 cli_name(osc_cli(obj)), state);
958 wait_event_idle(ext->oe_waitq, extent_wait_cb(ext, state));
968 * Discard pages with index greater than @size. If @ext is overlapped with
969 * @size, then partial truncate happens.
971 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
976 struct osc_object *obj = ext->oe_obj;
977 struct client_obd *cli = osc_cli(obj);
978 struct osc_async_page *oap;
979 struct osc_async_page *tmp;
980 struct pagevec *pvec;
981 int pages_in_chunk = 0;
982 int ppc_bits = cli->cl_chunkbits -
984 __u64 trunc_chunk = trunc_index >> ppc_bits;
991 LASSERT(sanity_check(ext) == 0);
992 LASSERT(ext->oe_state == OES_TRUNC);
993 LASSERT(!ext->oe_urgent);
995 /* Request new lu_env.
996 * We can't use that env from osc_cache_truncate_start() because
997 * it's from lov_io_sub and not fully initialized. */
998 env = cl_env_get(&refcheck);
1000 RETURN(PTR_ERR(env));
1002 io = osc_env_thread_io(env);
1003 io->ci_obj = cl_object_top(osc2cl(obj));
1004 io->ci_ignore_layout = 1;
1005 pvec = &osc_env_info(env)->oti_pagevec;
1006 ll_pagevec_init(pvec, 0);
1007 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1011 /* discard all pages with index greater than trunc_index */
1012 list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1014 pgoff_t index = osc_index(oap2osc(oap));
1015 struct cl_page *page = oap2cl_page(oap);
1017 LASSERT(list_empty(&oap->oap_rpc_item));
1019 /* only discard the pages with their index greater than
1020 * trunc_index, and ... */
1021 if (index < trunc_index ||
1022 (index == trunc_index && partial)) {
1023 /* accounting how many pages remaining in the chunk
1024 * so that we can calculate grants correctly. */
1025 if (index >> ppc_bits == trunc_chunk)
1030 list_del_init(&oap->oap_pending_item);
1033 lu_ref_add(&page->cp_reference, "truncate", current);
1035 if (cl_page_own(env, io, page) == 0) {
1036 cl_page_discard(env, io, page);
1037 cl_page_disown(env, io, page);
1039 LASSERT(page->cp_state == CPS_FREEING);
1043 lu_ref_del(&page->cp_reference, "truncate", current);
1044 cl_pagevec_put(env, page, pvec);
1049 pagevec_release(pvec);
1051 EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1052 ext->oe_nr_pages == 0),
1053 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1055 osc_object_lock(obj);
1056 if (ext->oe_nr_pages == 0) {
1057 LASSERT(pages_in_chunk == 0);
1058 grants = ext->oe_grants;
1060 } else { /* calculate how many grants we can free */
1061 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1065 /* if there is no pages in this chunk, we can also free grants
1066 * for the last chunk */
1067 if (pages_in_chunk == 0) {
1068 /* if this is the 1st chunk and no pages in this chunk,
1069 * ext->oe_nr_pages must be zero, so we should be in
1070 * the other if-clause. */
1071 LASSERT(trunc_chunk > 0);
1076 /* this is what we can free from this extent */
1077 grants = chunks << cli->cl_chunkbits;
1078 ext->oe_grants -= grants;
1079 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1080 ext->oe_end = min(last_index, ext->oe_max_end);
1081 LASSERT(ext->oe_end >= ext->oe_start);
1082 LASSERT(ext->oe_grants > 0);
1084 osc_object_unlock(obj);
1086 if (grants > 0 || nr_pages > 0)
1087 osc_free_grant(cli, nr_pages, grants, grants);
1090 cl_io_fini(env, io);
1091 cl_env_put(env, &refcheck);
1096 * This function is used to make the extent prepared for transfer.
1097 * A race with flusing page - ll_writepage() has to be handled cautiously.
1099 static int osc_extent_make_ready(const struct lu_env *env,
1100 struct osc_extent *ext)
1102 struct osc_async_page *oap;
1103 struct osc_async_page *last = NULL;
1104 struct osc_object *obj = ext->oe_obj;
1105 unsigned int page_count = 0;
1109 /* we're going to grab page lock, so object lock must not be taken. */
1110 LASSERT(sanity_check(ext) == 0);
1111 /* in locking state, any process should not touch this extent. */
1112 EASSERT(ext->oe_state == OES_LOCKING, ext);
1113 EASSERT(ext->oe_owner != NULL, ext);
1115 OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1117 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1119 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1122 /* checking ASYNC_READY is race safe */
1123 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1126 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1129 spin_lock(&oap->oap_lock);
1130 oap->oap_async_flags |= ASYNC_READY;
1131 spin_unlock(&oap->oap_lock);
1134 LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1137 LASSERTF(0, "unknown return code: %d\n", rc);
1141 LASSERT(page_count == ext->oe_nr_pages);
1142 LASSERT(last != NULL);
1143 /* the last page is the only one we need to refresh its count by
1144 * the size of file. */
1145 if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1146 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1147 LASSERT(last_oap_count > 0);
1148 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1149 last->oap_count = last_oap_count;
1150 spin_lock(&last->oap_lock);
1151 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1152 spin_unlock(&last->oap_lock);
1155 /* for the rest of pages, we don't need to call osf_refresh_count()
1156 * because it's known they are not the last page */
1157 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1158 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1159 oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1160 spin_lock(&oap->oap_lock);
1161 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1162 spin_unlock(&oap->oap_lock);
1166 osc_object_lock(obj);
1167 osc_extent_state_set(ext, OES_RPC);
1168 osc_object_unlock(obj);
1169 /* get a refcount for RPC. */
1170 osc_extent_get(ext);
1176 * Quick and simple version of osc_extent_find(). This function is frequently
1177 * called to expand the extent for the same IO. To expand the extent, the
1178 * page index must be in the same or next chunk of ext->oe_end.
1180 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1181 unsigned int *grants)
1183 struct osc_object *obj = ext->oe_obj;
1184 struct client_obd *cli = osc_cli(obj);
1185 struct osc_extent *next;
1186 int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1187 pgoff_t chunk = index >> ppc_bits;
1190 unsigned int chunksize = 1 << cli->cl_chunkbits;
1194 LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1195 osc_object_lock(obj);
1196 LASSERT(sanity_check_nolock(ext) == 0);
1197 end_chunk = ext->oe_end >> ppc_bits;
1198 if (chunk > end_chunk + 1)
1199 GOTO(out, rc = -ERANGE);
1201 if (end_chunk >= chunk)
1204 LASSERT(end_chunk + 1 == chunk);
1206 /* try to expand this extent to cover @index */
1207 end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1209 /* don't go over the maximum extent size reported by server */
1210 if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1211 GOTO(out, rc = -ERANGE);
1213 next = next_extent(ext);
1214 if (next != NULL && next->oe_start <= end_index)
1215 /* complex mode - overlapped with the next extent,
1216 * this case will be handled by osc_extent_find() */
1217 GOTO(out, rc = -EAGAIN);
1219 ext->oe_end = end_index;
1220 ext->oe_grants += chunksize;
1221 LASSERT(*grants >= chunksize);
1222 *grants -= chunksize;
1223 EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1224 "overlapped after expanding for %lu.\n", index);
1228 osc_object_unlock(obj);
1232 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1233 const char *func, int line)
1235 struct osc_extent *ext;
1238 if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1241 CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1242 obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1244 /* osc_object_lock(obj); */
1246 for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1247 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1250 list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1251 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1254 list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1255 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1258 list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1259 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1260 /* osc_object_unlock(obj); */
1263 /* ------------------ osc extent end ------------------ */
1265 static inline int osc_is_ready(struct osc_object *osc)
1267 return !list_empty(&osc->oo_ready_item) ||
1268 !list_empty(&osc->oo_hp_ready_item);
1271 #define OSC_IO_DEBUG(OSC, STR, args...) \
1272 CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR, \
1273 (OSC), osc_is_ready(OSC), \
1274 list_empty_marker(&(OSC)->oo_hp_ready_item), \
1275 list_empty_marker(&(OSC)->oo_ready_item), \
1276 atomic_read(&(OSC)->oo_nr_writes), \
1277 list_empty_marker(&(OSC)->oo_hp_exts), \
1278 list_empty_marker(&(OSC)->oo_urgent_exts), \
1279 atomic_read(&(OSC)->oo_nr_reads), \
1280 list_empty_marker(&(OSC)->oo_reading_exts), \
1283 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1286 struct osc_page *opg = oap2osc_page(oap);
1287 struct cl_page *page = oap2cl_page(oap);
1290 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1293 result = cl_page_make_ready(env, page, CRT_WRITE);
1295 opg->ops_submit_time = ktime_get();
1299 static int osc_refresh_count(const struct lu_env *env,
1300 struct osc_async_page *oap, int cmd)
1302 struct osc_page *opg = oap2osc_page(oap);
1303 pgoff_t index = osc_index(oap2osc(oap));
1304 struct cl_object *obj;
1305 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1309 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1310 LASSERT(!(cmd & OBD_BRW_READ));
1311 LASSERT(opg != NULL);
1312 obj = opg->ops_cl.cpl_obj;
1314 cl_object_attr_lock(obj);
1315 result = cl_object_attr_get(env, obj, attr);
1316 cl_object_attr_unlock(obj);
1319 kms = attr->cat_kms;
1320 if (cl_offset(obj, index) >= kms)
1321 /* catch race with truncate */
1323 else if (cl_offset(obj, index + 1) > kms)
1324 /* catch sub-page write at end of file */
1325 return kms & ~PAGE_MASK;
1330 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1333 struct osc_page *opg = oap2osc_page(oap);
1334 struct cl_page *page = oap2cl_page(oap);
1335 enum cl_req_type crt;
1340 cmd &= ~OBD_BRW_NOQUOTA;
1341 LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1342 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1343 LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1344 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1345 LASSERT(opg->ops_transfer_pinned);
1347 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1348 /* Clear opg->ops_transfer_pinned before VM lock is released. */
1349 opg->ops_transfer_pinned = 0;
1351 opg->ops_submit_time = ktime_set(0, 0);
1352 srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1355 if (rc == 0 && srvlock) {
1356 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1357 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1358 size_t bytes = oap->oap_count;
1360 if (crt == CRT_READ)
1361 stats->os_lockless_reads += bytes;
1363 stats->os_lockless_writes += bytes;
1367 * This has to be the last operation with the page, as locks are
1368 * released in cl_page_completion() and nothing except for the
1369 * reference counter protects page from concurrent reclaim.
1371 lu_ref_del(&page->cp_reference, "transfer", page);
1373 cl_page_completion(env, page, crt, rc);
1374 cl_page_put(env, page);
1379 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do { \
1380 struct client_obd *__tmp = (cli); \
1381 CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1382 "dropped: %ld avail: %ld, dirty_grant: %ld, " \
1383 "reserved: %ld, flight: %d } lru {in list: %ld, " \
1384 "left: %ld, waiters: %d }" fmt "\n", \
1386 __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages, \
1387 atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1388 __tmp->cl_lost_grant, __tmp->cl_avail_grant, \
1389 __tmp->cl_dirty_grant, \
1390 __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, \
1391 atomic_long_read(&__tmp->cl_lru_in_list), \
1392 atomic_long_read(&__tmp->cl_lru_busy), \
1393 atomic_read(&__tmp->cl_lru_shrinkers), ##args); \
1396 /* caller must hold loi_list_lock */
1397 static void osc_consume_write_grant(struct client_obd *cli,
1398 struct brw_page *pga)
1400 assert_spin_locked(&cli->cl_loi_list_lock);
1401 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1402 cli->cl_dirty_pages++;
1403 pga->flag |= OBD_BRW_FROM_GRANT;
1404 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1405 PAGE_SIZE, pga, pga->pg);
1406 osc_update_next_shrink(cli);
1409 /* the companion to osc_consume_write_grant, called when a brw has completed.
1410 * must be called with the loi lock held. */
1411 static void osc_release_write_grant(struct client_obd *cli,
1412 struct brw_page *pga)
1416 assert_spin_locked(&cli->cl_loi_list_lock);
1417 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1422 pga->flag &= ~OBD_BRW_FROM_GRANT;
1423 atomic_long_dec(&obd_dirty_pages);
1424 cli->cl_dirty_pages--;
1429 * To avoid sleeping with object lock held, it's good for us allocate enough
1430 * grants before entering into critical section.
1432 * client_obd_list_lock held by caller
1434 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1438 if (cli->cl_avail_grant >= bytes) {
1439 cli->cl_avail_grant -= bytes;
1440 cli->cl_reserved_grant += bytes;
1446 static void __osc_unreserve_grant(struct client_obd *cli,
1447 unsigned int reserved, unsigned int unused)
1449 /* it's quite normal for us to get more grant than reserved.
1450 * Thinking about a case that two extents merged by adding a new
1451 * chunk, we can save one extent tax. If extent tax is greater than
1452 * one chunk, we can save more grant by adding a new chunk */
1453 cli->cl_reserved_grant -= reserved;
1454 if (unused > reserved) {
1455 cli->cl_avail_grant += reserved;
1456 cli->cl_lost_grant += unused - reserved;
1457 cli->cl_dirty_grant -= unused - reserved;
1459 cli->cl_avail_grant += unused;
1460 cli->cl_dirty_grant += reserved - unused;
1464 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1465 unsigned int reserved,
1466 unsigned int unused)
1468 __osc_unreserve_grant(cli, reserved, unused);
1470 osc_wake_cache_waiters(cli);
1473 static void osc_unreserve_grant(struct client_obd *cli,
1474 unsigned int reserved, unsigned int unused)
1476 spin_lock(&cli->cl_loi_list_lock);
1477 osc_unreserve_grant_nolock(cli, reserved, unused);
1478 spin_unlock(&cli->cl_loi_list_lock);
1482 * Free grant after IO is finished or canceled.
1484 * @lost_grant is used to remember how many grants we have allocated but not
1485 * used, we should return these grants to OST. There're two cases where grants
1488 * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1489 * written. In this case OST may use less chunks to serve this partial
1490 * write. OSTs don't actually know the page size on the client side. so
1491 * clients have to calculate lost grant by the blocksize on the OST.
1492 * See filter_grant_check() for details.
1494 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1495 unsigned int lost_grant, unsigned int dirty_grant)
1497 unsigned long grant;
1499 grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1501 spin_lock(&cli->cl_loi_list_lock);
1502 atomic_long_sub(nr_pages, &obd_dirty_pages);
1503 cli->cl_dirty_pages -= nr_pages;
1504 cli->cl_lost_grant += lost_grant;
1505 cli->cl_dirty_grant -= dirty_grant;
1506 if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1507 /* borrow some grant from truncate to avoid the case that
1508 * truncate uses up all avail grant */
1509 cli->cl_lost_grant -= grant;
1510 cli->cl_avail_grant += grant;
1512 osc_wake_cache_waiters(cli);
1513 spin_unlock(&cli->cl_loi_list_lock);
1514 CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1515 lost_grant, cli->cl_lost_grant,
1516 cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1517 cli->cl_dirty_grant);
1521 * The companion to osc_enter_cache(), called when @oap is no longer part of
1522 * the dirty accounting due to error.
1524 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1526 spin_lock(&cli->cl_loi_list_lock);
1527 osc_release_write_grant(cli, &oap->oap_brw_page);
1528 spin_unlock(&cli->cl_loi_list_lock);
1532 * Non-blocking version of osc_enter_cache() that consumes grant only when it
1535 static int osc_enter_cache_try(struct client_obd *cli,
1536 struct osc_async_page *oap,
1541 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1543 rc = osc_reserve_grant(cli, bytes);
1547 if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1548 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1549 obd_max_dirty_pages) {
1550 osc_consume_write_grant(cli, &oap->oap_brw_page);
1554 atomic_long_dec(&obd_dirty_pages);
1556 __osc_unreserve_grant(cli, bytes, bytes);
1562 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1565 spin_lock(&cli->cl_loi_list_lock);
1566 rc = list_empty(&ocw->ocw_entry);
1567 spin_unlock(&cli->cl_loi_list_lock);
1572 * The main entry to reserve dirty page accounting. Usually the grant reserved
1573 * in this function will be freed in bulk in osc_free_grant() unless it fails
1574 * to add osc cache, in that case, it will be freed in osc_exit_cache().
1576 * The process will be put into sleep if it's already run out of grant.
1578 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1579 struct osc_async_page *oap, int bytes)
1581 struct osc_object *osc = oap->oap_obj;
1582 struct lov_oinfo *loi = osc->oo_oinfo;
1583 struct osc_cache_waiter ocw;
1587 OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1589 spin_lock(&cli->cl_loi_list_lock);
1591 /* force the caller to try sync io. this can jump the list
1592 * of queued writes and create a discontiguous rpc stream */
1593 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1594 cli->cl_dirty_max_pages == 0 ||
1595 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1596 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1597 GOTO(out, rc = -EDQUOT);
1600 /* Hopefully normal case - cache space and write credits available */
1601 if (list_empty(&cli->cl_cache_waiters) &&
1602 osc_enter_cache_try(cli, oap, bytes)) {
1603 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1607 /* We can get here for two reasons: too many dirty pages in cache, or
1608 * run out of grants. In both cases we should write dirty pages out.
1609 * Adding a cache waiter will trigger urgent write-out no matter what
1611 * The exiting condition is no avail grants and no dirty pages caching,
1612 * that really means there is no space on the OST. */
1613 init_waitqueue_head(&ocw.ocw_waitq);
1615 ocw.ocw_grant = bytes;
1616 while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1617 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1619 spin_unlock(&cli->cl_loi_list_lock);
1621 osc_io_unplug_async(env, cli, NULL);
1623 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1624 cli_name(cli), &ocw, oap);
1626 rc = wait_event_idle_timeout(ocw.ocw_waitq,
1627 ocw_granted(cli, &ocw),
1628 cfs_time_seconds(AT_OFF ?
1632 spin_lock(&cli->cl_loi_list_lock);
1635 /* l_wait_event is interrupted by signal or timed out */
1636 list_del_init(&ocw.ocw_entry);
1641 LASSERT(list_empty(&ocw.ocw_entry));
1646 if (osc_enter_cache_try(cli, oap, bytes)) {
1654 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1657 OSC_DUMP_GRANT(D_CACHE, cli,
1658 "timeout, fall back to sync i/o\n");
1659 osc_extent_tree_dump(D_CACHE, osc);
1660 /* fall back to synchronous I/O */
1664 /* Ensures restartability - LU-3581 */
1665 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1669 OSC_DUMP_GRANT(D_CACHE, cli,
1670 "no grant space, fall back to sync i/o\n");
1673 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1674 "due to %d, fall back to sync i/o\n",
1675 cli_name(cli), &ocw, rc);
1680 spin_unlock(&cli->cl_loi_list_lock);
1684 /* caller must hold loi_list_lock */
1685 void osc_wake_cache_waiters(struct client_obd *cli)
1687 struct list_head *l, *tmp;
1688 struct osc_cache_waiter *ocw;
1691 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1692 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1694 ocw->ocw_rc = -EDQUOT;
1696 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant))
1699 if (ocw->ocw_rc == 0 ||
1700 !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1701 list_del_init(&ocw->ocw_entry);
1702 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1703 "%ld, %d\n", ocw, ocw->ocw_oap,
1704 cli->cl_avail_grant, ocw->ocw_rc);
1706 wake_up(&ocw->ocw_waitq);
1712 EXPORT_SYMBOL(osc_wake_cache_waiters);
1714 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1716 int hprpc = !!list_empty(&osc->oo_hp_exts);
1717 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1720 /* This maintains the lists of pending pages to read/write for a given object
1721 * (lop). This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1722 * to quickly find objects that are ready to send an RPC. */
1723 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1726 int invalid_import = 0;
1729 /* if we have an invalid import we want to drain the queued pages
1730 * by forcing them through rpcs that immediately fail and complete
1731 * the pages. recovery relies on this to empty the queued pages
1732 * before canceling the locks and evicting down the llite pages */
1733 if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1736 if (cmd & OBD_BRW_WRITE) {
1737 if (atomic_read(&osc->oo_nr_writes) == 0)
1739 if (invalid_import) {
1740 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1743 if (!list_empty(&osc->oo_hp_exts)) {
1744 CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1747 if (!list_empty(&osc->oo_urgent_exts)) {
1748 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1751 /* trigger a write rpc stream as long as there are dirtiers
1752 * waiting for space. as they're waiting, they're not going to
1753 * create more pages to coalesce with what's waiting.. */
1754 if (!list_empty(&cli->cl_cache_waiters)) {
1755 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1758 if (!list_empty(&osc->oo_full_exts)) {
1759 CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1763 if (atomic_read(&osc->oo_nr_reads) == 0)
1765 if (invalid_import) {
1766 CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1769 /* all read are urgent. */
1770 if (!list_empty(&osc->oo_reading_exts))
1777 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1779 struct client_obd *cli = osc_cli(obj);
1780 if (cmd & OBD_BRW_WRITE) {
1781 atomic_add(delta, &obj->oo_nr_writes);
1782 atomic_add(delta, &cli->cl_pending_w_pages);
1783 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1785 atomic_add(delta, &obj->oo_nr_reads);
1786 atomic_add(delta, &cli->cl_pending_r_pages);
1787 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1789 OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1792 static int osc_makes_hprpc(struct osc_object *obj)
1794 return !list_empty(&obj->oo_hp_exts);
1797 static void on_list(struct list_head *item, struct list_head *list,
1800 if (list_empty(item) && should_be_on)
1801 list_add_tail(item, list);
1802 else if (!list_empty(item) && !should_be_on)
1803 list_del_init(item);
1806 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1807 * can find pages to build into rpcs quickly */
1808 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1810 if (osc_makes_hprpc(osc)) {
1812 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1813 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1815 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1816 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1817 osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1818 osc_makes_rpc(cli, osc, OBD_BRW_READ));
1821 on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1822 atomic_read(&osc->oo_nr_writes) > 0);
1824 on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1825 atomic_read(&osc->oo_nr_reads) > 0);
1827 return osc_is_ready(osc);
1830 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1834 spin_lock(&cli->cl_loi_list_lock);
1835 is_ready = __osc_list_maint(cli, osc);
1836 spin_unlock(&cli->cl_loi_list_lock);
1841 /* this is trying to propogate async writeback errors back up to the
1842 * application. As an async write fails we record the error code for later if
1843 * the app does an fsync. As long as errors persist we force future rpcs to be
1844 * sync so that the app can get a sync error and break the cycle of queueing
1845 * pages for which writeback will fail. */
1846 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1853 ar->ar_force_sync = 1;
1854 ar->ar_min_xid = ptlrpc_sample_next_xid();
1859 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1860 ar->ar_force_sync = 0;
1863 /* this must be called holding the loi list lock to give coverage to exit_cache,
1864 * async_flag maintenance, and oap_request */
1865 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1866 struct osc_async_page *oap, int sent, int rc)
1868 struct osc_object *osc = oap->oap_obj;
1869 struct lov_oinfo *loi = osc->oo_oinfo;
1873 if (oap->oap_request != NULL) {
1874 xid = ptlrpc_req_xid(oap->oap_request);
1875 ptlrpc_req_finished(oap->oap_request);
1876 oap->oap_request = NULL;
1879 /* As the transfer for this page is being done, clear the flags */
1880 spin_lock(&oap->oap_lock);
1881 oap->oap_async_flags = 0;
1882 spin_unlock(&oap->oap_lock);
1883 oap->oap_interrupted = 0;
1885 if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1886 spin_lock(&cli->cl_loi_list_lock);
1887 osc_process_ar(&cli->cl_ar, xid, rc);
1888 osc_process_ar(&loi->loi_ar, xid, rc);
1889 spin_unlock(&cli->cl_loi_list_lock);
1892 rc = osc_completion(env, oap, oap->oap_cmd, rc);
1894 CERROR("completion on oap %p obj %p returns %d.\n",
1900 struct extent_rpc_data {
1901 struct list_head *erd_rpc_list;
1902 unsigned int erd_page_count;
1903 unsigned int erd_max_pages;
1904 unsigned int erd_max_chunks;
1905 unsigned int erd_max_extents;
1908 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1910 struct client_obd *cli = osc_cli(ext->oe_obj);
1911 unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1913 return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1917 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1919 if (ext->oe_no_merge || in_rpc->oe_no_merge)
1922 if (ext->oe_srvlock != in_rpc->oe_srvlock)
1925 if (ext->oe_ndelay != in_rpc->oe_ndelay)
1928 if (!ext->oe_grants != !in_rpc->oe_grants)
1931 if (ext->oe_dio != in_rpc->oe_dio)
1934 /* It's possible to have overlap on DIO */
1935 if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1942 * Try to add extent to one RPC. We need to think about the following things:
1943 * - # of pages must not be over max_pages_per_rpc
1944 * - extent must be compatible with previous ones
1946 static int try_to_add_extent_for_io(struct client_obd *cli,
1947 struct osc_extent *ext,
1948 struct extent_rpc_data *data)
1950 struct osc_extent *tmp;
1951 unsigned int chunk_count;
1954 EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1956 OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1958 if (data->erd_max_extents == 0)
1961 chunk_count = osc_extent_chunks(ext);
1962 EASSERTF(data->erd_page_count != 0 ||
1963 chunk_count <= data->erd_max_chunks, ext,
1964 "The first extent to be fit in a RPC contains %u chunks, "
1965 "which is over the limit %u.\n", chunk_count,
1966 data->erd_max_chunks);
1967 if (chunk_count > data->erd_max_chunks)
1970 data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1971 EASSERTF(data->erd_page_count != 0 ||
1972 ext->oe_nr_pages <= data->erd_max_pages, ext,
1973 "The first extent to be fit in a RPC contains %u pages, "
1974 "which is over the limit %u.\n", ext->oe_nr_pages,
1975 data->erd_max_pages);
1976 if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1979 list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1980 EASSERT(tmp->oe_owner == current, tmp);
1982 if (!can_merge(ext, tmp))
1986 data->erd_max_extents--;
1987 data->erd_max_chunks -= chunk_count;
1988 data->erd_page_count += ext->oe_nr_pages;
1989 list_move_tail(&ext->oe_link, data->erd_rpc_list);
1990 ext->oe_owner = current;
1995 * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1996 * get_write_extent() takes all appropriate extents in atomic.
1998 * The following policy is used to collect extents for IO:
1999 * 1. Add as many HP extents as possible;
2000 * 2. Add the first urgent extent in urgent extent list and take it out of
2002 * 3. Add subsequent extents of this urgent extent;
2003 * 4. If urgent list is not empty, goto 2;
2004 * 5. Traverse the extent tree from the 1st extent;
2005 * 6. Above steps exit if there is no space in this RPC.
2007 static unsigned int get_write_extents(struct osc_object *obj,
2008 struct list_head *rpclist)
2010 struct client_obd *cli = osc_cli(obj);
2011 struct osc_extent *ext;
2012 struct extent_rpc_data data = {
2013 .erd_rpc_list = rpclist,
2014 .erd_page_count = 0,
2015 .erd_max_pages = cli->cl_max_pages_per_rpc,
2016 .erd_max_chunks = osc_max_write_chunks(cli),
2017 .erd_max_extents = 256,
2020 assert_osc_object_is_locked(obj);
2021 while (!list_empty(&obj->oo_hp_exts)) {
2022 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2024 LASSERT(ext->oe_state == OES_CACHE);
2025 if (!try_to_add_extent_for_io(cli, ext, &data))
2026 return data.erd_page_count;
2027 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2029 if (data.erd_page_count == data.erd_max_pages)
2030 return data.erd_page_count;
2032 while (!list_empty(&obj->oo_urgent_exts)) {
2033 ext = list_entry(obj->oo_urgent_exts.next,
2034 struct osc_extent, oe_link);
2035 if (!try_to_add_extent_for_io(cli, ext, &data))
2036 return data.erd_page_count;
2038 if (data.erd_page_count == data.erd_max_pages)
2039 return data.erd_page_count;
2041 /* One key difference between full extents and other extents: full
2042 * extents can usually only be added if the rpclist was empty, so if we
2043 * can't add one, we continue on to trying to add normal extents. This
2044 * is so we don't miss adding extra extents to an RPC containing high
2045 * priority or urgent extents. */
2046 while (!list_empty(&obj->oo_full_exts)) {
2047 ext = list_entry(obj->oo_full_exts.next,
2048 struct osc_extent, oe_link);
2049 if (!try_to_add_extent_for_io(cli, ext, &data))
2052 if (data.erd_page_count == data.erd_max_pages)
2053 return data.erd_page_count;
2055 ext = first_extent(obj);
2056 while (ext != NULL) {
2057 if ((ext->oe_state != OES_CACHE) ||
2058 /* this extent may be already in current rpclist */
2059 (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2060 ext = next_extent(ext);
2064 if (!try_to_add_extent_for_io(cli, ext, &data))
2065 return data.erd_page_count;
2067 ext = next_extent(ext);
2069 return data.erd_page_count;
2073 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2074 struct osc_object *osc)
2078 struct osc_extent *ext;
2079 struct osc_extent *tmp;
2080 struct osc_extent *first = NULL;
2081 unsigned int page_count = 0;
2086 assert_osc_object_is_locked(osc);
2088 page_count = get_write_extents(osc, &rpclist);
2089 LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2091 if (list_empty(&rpclist))
2094 osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2096 list_for_each_entry(ext, &rpclist, oe_link) {
2097 LASSERT(ext->oe_state == OES_CACHE ||
2098 ext->oe_state == OES_LOCK_DONE);
2099 if (ext->oe_state == OES_CACHE)
2100 osc_extent_state_set(ext, OES_LOCKING);
2102 osc_extent_state_set(ext, OES_RPC);
2105 /* we're going to grab page lock, so release object lock because
2106 * lock order is page lock -> object lock. */
2107 osc_object_unlock(osc);
2109 list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2110 if (ext->oe_state == OES_LOCKING) {
2111 rc = osc_extent_make_ready(env, ext);
2112 if (unlikely(rc < 0)) {
2113 list_del_init(&ext->oe_link);
2114 osc_extent_finish(env, ext, 0, rc);
2118 if (first == NULL) {
2120 srvlock = ext->oe_srvlock;
2122 LASSERT(srvlock == ext->oe_srvlock);
2126 if (!list_empty(&rpclist)) {
2127 LASSERT(page_count > 0);
2128 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2129 LASSERT(list_empty(&rpclist));
2132 osc_object_lock(osc);
2137 * prepare pages for ASYNC io and put pages in send queue.
2139 * \param cmd OBD_BRW_* macroses
2140 * \param lop pending pages
2142 * \return zero if no page added to send queue.
2143 * \return 1 if pages successfully added to send queue.
2144 * \return negative on errors.
2147 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2148 struct osc_object *osc)
2151 struct osc_extent *ext;
2152 struct osc_extent *next;
2154 struct extent_rpc_data data = {
2155 .erd_rpc_list = &rpclist,
2156 .erd_page_count = 0,
2157 .erd_max_pages = cli->cl_max_pages_per_rpc,
2158 .erd_max_chunks = UINT_MAX,
2159 .erd_max_extents = UINT_MAX,
2164 assert_osc_object_is_locked(osc);
2165 list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2166 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2167 if (!try_to_add_extent_for_io(cli, ext, &data))
2169 osc_extent_state_set(ext, OES_RPC);
2170 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2172 LASSERT(data.erd_page_count <= data.erd_max_pages);
2174 osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2176 if (!list_empty(&rpclist)) {
2177 osc_object_unlock(osc);
2179 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2180 LASSERT(list_empty(&rpclist));
2182 osc_object_lock(osc);
2187 #define list_to_obj(list, item) ({ \
2188 struct list_head *__tmp = (list)->next; \
2189 list_del_init(__tmp); \
2190 list_entry(__tmp, struct osc_object, oo_##item); \
2193 /* This is called by osc_check_rpcs() to find which objects have pages that
2194 * we could be sending. These lists are maintained by osc_makes_rpc(). */
2195 static struct osc_object *osc_next_obj(struct client_obd *cli)
2199 /* First return objects that have blocked locks so that they
2200 * will be flushed quickly and other clients can get the lock,
2201 * then objects which have pages ready to be stuffed into RPCs */
2202 if (!list_empty(&cli->cl_loi_hp_ready_list))
2203 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2204 if (!list_empty(&cli->cl_loi_ready_list))
2205 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2207 /* then if we have cache waiters, return all objects with queued
2208 * writes. This is especially important when many small files
2209 * have filled up the cache and not been fired into rpcs because
2210 * they don't pass the nr_pending/object threshhold */
2211 if (!list_empty(&cli->cl_cache_waiters) &&
2212 !list_empty(&cli->cl_loi_write_list))
2213 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2215 /* then return all queued objects when we have an invalid import
2216 * so that they get flushed */
2217 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2218 if (!list_empty(&cli->cl_loi_write_list))
2219 RETURN(list_to_obj(&cli->cl_loi_write_list,
2221 if (!list_empty(&cli->cl_loi_read_list))
2222 RETURN(list_to_obj(&cli->cl_loi_read_list,
2228 /* called with the loi list lock held */
2229 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2230 __must_hold(&cli->cl_loi_list_lock)
2232 struct osc_object *osc;
2236 while ((osc = osc_next_obj(cli)) != NULL) {
2237 struct cl_object *obj = osc2cl(osc);
2238 struct lu_ref_link link;
2240 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2242 if (osc_max_rpc_in_flight(cli, osc)) {
2243 __osc_list_maint(cli, osc);
2248 spin_unlock(&cli->cl_loi_list_lock);
2249 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2251 /* attempt some read/write balancing by alternating between
2252 * reads and writes in an object. The makes_rpc checks here
2253 * would be redundant if we were getting read/write work items
2254 * instead of objects. we don't want send_oap_rpc to drain a
2255 * partial read pending queue when we're given this object to
2256 * do io on writes while there are cache waiters */
2257 osc_object_lock(osc);
2258 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2259 rc = osc_send_write_rpc(env, cli, osc);
2261 CERROR("Write request failed with %d\n", rc);
2263 /* osc_send_write_rpc failed, mostly because of
2266 * It can't break here, because if:
2267 * - a page was submitted by osc_io_submit, so
2269 * - no request in flight
2270 * - no subsequent request
2271 * The system will be in live-lock state,
2272 * because there is no chance to call
2273 * osc_io_unplug() and osc_check_rpcs() any
2274 * more. pdflush can't help in this case,
2275 * because it might be blocked at grabbing
2276 * the page lock as we mentioned.
2278 * Anyway, continue to drain pages. */
2282 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2283 rc = osc_send_read_rpc(env, cli, osc);
2285 CERROR("Read request failed with %d\n", rc);
2287 osc_object_unlock(osc);
2289 osc_list_maint(cli, osc);
2290 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2291 cl_object_put(env, obj);
2293 spin_lock(&cli->cl_loi_list_lock);
2297 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2298 struct osc_object *osc, int async)
2302 if (osc != NULL && osc_list_maint(cli, osc) == 0)
2306 spin_lock(&cli->cl_loi_list_lock);
2307 osc_check_rpcs(env, cli);
2308 spin_unlock(&cli->cl_loi_list_lock);
2310 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2311 LASSERT(cli->cl_writeback_work != NULL);
2312 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2316 EXPORT_SYMBOL(osc_io_unplug0);
2318 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2319 struct page *page, loff_t offset)
2321 struct obd_export *exp = osc_export(osc);
2322 struct osc_async_page *oap = &ops->ops_oap;
2326 return cfs_size_round(sizeof(*oap));
2328 oap->oap_magic = OAP_MAGIC;
2329 oap->oap_cli = &exp->exp_obd->u.cli;
2332 oap->oap_page = page;
2333 oap->oap_obj_off = offset;
2334 LASSERT(!(offset & ~PAGE_MASK));
2336 INIT_LIST_HEAD(&oap->oap_pending_item);
2337 INIT_LIST_HEAD(&oap->oap_rpc_item);
2339 spin_lock_init(&oap->oap_lock);
2340 CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2341 oap, page, oap->oap_obj_off);
2344 EXPORT_SYMBOL(osc_prep_async_page);
2346 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2347 struct osc_page *ops, cl_commit_cbt cb)
2349 struct osc_io *oio = osc_env_io(env);
2350 struct osc_extent *ext = NULL;
2351 struct osc_async_page *oap = &ops->ops_oap;
2352 struct client_obd *cli = oap->oap_cli;
2353 struct osc_object *osc = oap->oap_obj;
2354 struct pagevec *pvec = &osc_env_info(env)->oti_pagevec;
2357 unsigned int grants = 0;
2358 u32 brw_flags = OBD_BRW_ASYNC;
2359 int cmd = OBD_BRW_WRITE;
2360 int need_release = 0;
2364 if (oap->oap_magic != OAP_MAGIC)
2367 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2370 if (!list_empty(&oap->oap_pending_item) ||
2371 !list_empty(&oap->oap_rpc_item))
2374 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2375 brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2376 if (oio->oi_cap_sys_resource || io->ci_noquota) {
2377 brw_flags |= OBD_BRW_NOQUOTA;
2378 cmd |= OBD_BRW_NOQUOTA;
2381 /* check if the file's owner/group is over quota */
2382 if (!(cmd & OBD_BRW_NOQUOTA)) {
2383 struct cl_object *obj;
2384 struct cl_attr *attr;
2385 unsigned int qid[LL_MAXQUOTAS];
2387 obj = cl_object_top(&osc->oo_cl);
2388 attr = &osc_env_info(env)->oti_attr;
2390 cl_object_attr_lock(obj);
2391 rc = cl_object_attr_get(env, obj, attr);
2392 cl_object_attr_unlock(obj);
2394 qid[USRQUOTA] = attr->cat_uid;
2395 qid[GRPQUOTA] = attr->cat_gid;
2396 qid[PRJQUOTA] = attr->cat_projid;
2397 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2404 oap->oap_page_off = ops->ops_from;
2405 oap->oap_count = ops->ops_to - ops->ops_from;
2406 /* No need to hold a lock here,
2407 * since this page is not in any list yet. */
2408 oap->oap_async_flags = 0;
2409 oap->oap_brw_flags = brw_flags;
2411 OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2412 oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2414 index = osc_index(oap2osc(oap));
2416 /* Add this page into extent by the following steps:
2417 * 1. if there exists an active extent for this IO, mostly this page
2418 * can be added to the active extent and sometimes we need to
2419 * expand extent to accomodate this page;
2420 * 2. otherwise, a new extent will be allocated. */
2422 ext = oio->oi_active;
2423 if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2424 /* one chunk plus extent overhead must be enough to write this
2426 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2427 if (ext->oe_end >= index)
2430 /* it doesn't need any grant to dirty this page */
2431 spin_lock(&cli->cl_loi_list_lock);
2432 rc = osc_enter_cache_try(cli, oap, grants);
2433 if (rc == 0) { /* try failed */
2436 } else if (ext->oe_end < index) {
2438 /* try to expand this extent */
2439 rc = osc_extent_expand(ext, index, &tmp);
2442 /* don't free reserved grant */
2444 OSC_EXTENT_DUMP(D_CACHE, ext,
2445 "expanded for %lu.\n", index);
2446 osc_unreserve_grant_nolock(cli, grants, tmp);
2450 spin_unlock(&cli->cl_loi_list_lock);
2452 } else if (ext != NULL) {
2453 /* index is located outside of active extent */
2457 osc_extent_release(env, ext);
2458 oio->oi_active = NULL;
2463 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2465 /* try to find new extent to cover this page */
2466 LASSERT(oio->oi_active == NULL);
2467 /* we may have allocated grant for this page if we failed
2468 * to expand the previous active extent. */
2469 LASSERT(ergo(grants > 0, grants >= tmp));
2473 /* We haven't allocated grant for this page, and we
2474 * must not hold a page lock while we do enter_cache,
2475 * so we must mark dirty & unlock any pages in the
2476 * write commit pagevec. */
2477 if (pagevec_count(pvec)) {
2479 pagevec_reinit(pvec);
2481 rc = osc_enter_cache(env, cli, oap, tmp);
2488 ext = osc_extent_find(env, osc, index, &tmp);
2490 LASSERT(tmp == grants);
2491 osc_exit_cache(cli, oap);
2495 oio->oi_active = ext;
2499 osc_unreserve_grant(cli, grants, tmp);
2502 LASSERT(ergo(rc == 0, ext != NULL));
2504 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2505 ext, "index = %lu.\n", index);
2506 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2508 osc_object_lock(osc);
2509 if (ext->oe_nr_pages == 0)
2510 ext->oe_srvlock = ops->ops_srvlock;
2512 LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2514 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2515 osc_object_unlock(osc);
2517 if (!ext->oe_layout_version)
2518 ext->oe_layout_version = io->ci_layout_version;
2524 int osc_teardown_async_page(const struct lu_env *env,
2525 struct osc_object *obj, struct osc_page *ops)
2527 struct osc_async_page *oap = &ops->ops_oap;
2531 LASSERT(oap->oap_magic == OAP_MAGIC);
2533 CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2534 oap, ops, osc_index(oap2osc(oap)));
2536 if (!list_empty(&oap->oap_rpc_item)) {
2537 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2539 } else if (!list_empty(&oap->oap_pending_item)) {
2540 struct osc_extent *ext = NULL;
2542 osc_object_lock(obj);
2543 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2544 osc_object_unlock(obj);
2545 /* only truncated pages are allowed to be taken out.
2546 * See osc_extent_truncate() and osc_cache_truncate_start()
2548 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2549 OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2550 osc_index(oap2osc(oap)));
2554 osc_extent_put(env, ext);
2560 * This is called when a page is picked up by kernel to write out.
2562 * We should find out the corresponding extent and add the whole extent
2563 * into urgent list. The extent may be being truncated or used, handle it
2566 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2567 struct osc_page *ops)
2569 struct osc_extent *ext = NULL;
2570 struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2571 struct cl_page *cp = ops->ops_cl.cpl_page;
2572 pgoff_t index = osc_index(ops);
2573 struct osc_async_page *oap = &ops->ops_oap;
2574 bool unplug = false;
2578 osc_object_lock(obj);
2579 ext = osc_extent_lookup(obj, index);
2581 osc_extent_tree_dump(D_ERROR, obj);
2582 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2585 switch (ext->oe_state) {
2588 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2592 /* If we know this extent is being written out, we should abort
2593 * so that the writer can make this page ready. Otherwise, there
2594 * exists a deadlock problem because other process can wait for
2595 * page writeback bit holding page lock; and meanwhile in
2596 * vvp_page_make_ready(), we need to grab page lock before
2597 * really sending the RPC. */
2599 /* race with truncate, page will be redirtied */
2601 /* The extent is active so we need to abort and let the caller
2602 * re-dirty the page. If we continued on here, and we were the
2603 * one making the extent active, we could deadlock waiting for
2604 * the page writeback to clear but it won't because the extent
2605 * is active and won't be written out. */
2606 GOTO(out, rc = -EAGAIN);
2611 rc = cl_page_prep(env, io, cp, CRT_WRITE);
2615 spin_lock(&oap->oap_lock);
2616 oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2617 spin_unlock(&oap->oap_lock);
2619 if (memory_pressure_get())
2620 ext->oe_memalloc = 1;
2623 if (ext->oe_state == OES_CACHE) {
2624 OSC_EXTENT_DUMP(D_CACHE, ext,
2625 "flush page %p make it urgent.\n", oap);
2626 if (list_empty(&ext->oe_link))
2627 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2634 osc_object_unlock(obj);
2635 osc_extent_put(env, ext);
2637 osc_io_unplug_async(env, osc_cli(obj), obj);
2642 * this is called when a sync waiter receives an interruption. Its job is to
2643 * get the caller woken as soon as possible. If its page hasn't been put in an
2644 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2645 * desiring interruption which will forcefully complete the rpc once the rpc
2648 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2650 struct osc_async_page *oap = &ops->ops_oap;
2651 struct osc_object *obj = oap->oap_obj;
2652 struct client_obd *cli = osc_cli(obj);
2653 struct osc_extent *ext;
2654 struct osc_extent *found = NULL;
2655 struct list_head *plist;
2656 pgoff_t index = osc_index(ops);
2661 LASSERT(!oap->oap_interrupted);
2662 oap->oap_interrupted = 1;
2664 /* Find out the caching extent */
2665 osc_object_lock(obj);
2666 if (oap->oap_cmd & OBD_BRW_WRITE) {
2667 plist = &obj->oo_urgent_exts;
2668 cmd = OBD_BRW_WRITE;
2670 plist = &obj->oo_reading_exts;
2673 list_for_each_entry(ext, plist, oe_link) {
2674 if (ext->oe_start <= index && ext->oe_end >= index) {
2675 LASSERT(ext->oe_state == OES_LOCK_DONE);
2676 /* For OES_LOCK_DONE state extent, it has already held
2677 * a refcount for RPC. */
2678 found = osc_extent_get(ext);
2682 if (found != NULL) {
2683 list_del_init(&found->oe_link);
2684 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2685 osc_object_unlock(obj);
2687 osc_extent_finish(env, found, 0, -EINTR);
2688 osc_extent_put(env, found);
2691 osc_object_unlock(obj);
2692 /* ok, it's been put in an rpc. only one oap gets a request
2694 if (oap->oap_request != NULL) {
2695 ptlrpc_mark_interrupted(oap->oap_request);
2696 ptlrpcd_wake(oap->oap_request);
2697 ptlrpc_req_finished(oap->oap_request);
2698 oap->oap_request = NULL;
2702 osc_list_maint(cli, obj);
2706 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2707 struct osc_object *obj, struct list_head *list,
2710 struct client_obd *cli = osc_cli(obj);
2711 struct osc_extent *ext;
2712 struct osc_async_page *oap;
2714 int mppr = cli->cl_max_pages_per_rpc;
2715 bool can_merge = true;
2716 pgoff_t start = CL_PAGE_EOF;
2720 list_for_each_entry(oap, list, oap_pending_item) {
2721 struct osc_page *opg = oap2osc_page(oap);
2722 pgoff_t index = osc_index(opg);
2729 mppr <<= (page_count > mppr);
2731 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2735 ext = osc_extent_alloc(obj);
2737 struct osc_async_page *tmp;
2739 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2740 list_del_init(&oap->oap_pending_item);
2741 osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2746 ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2748 ext->oe_no_merge = !can_merge;
2750 ext->oe_start = start;
2751 ext->oe_end = ext->oe_max_end = end;
2753 ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2754 ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2755 ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2756 ext->oe_nr_pages = page_count;
2757 ext->oe_mppr = mppr;
2758 list_splice_init(list, &ext->oe_pages);
2759 ext->oe_layout_version = io->ci_layout_version;
2761 osc_object_lock(obj);
2762 /* Reuse the initial refcount for RPC, don't drop it */
2763 osc_extent_state_set(ext, OES_LOCK_DONE);
2764 if (!ext->oe_rw) { /* write */
2765 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2766 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2768 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2769 osc_update_pending(obj, OBD_BRW_READ, page_count);
2771 osc_object_unlock(obj);
2773 osc_io_unplug_async(env, cli, obj);
2778 * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2780 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2781 __u64 size, struct osc_extent **extp)
2783 struct client_obd *cli = osc_cli(obj);
2784 struct osc_extent *ext;
2785 struct osc_extent *waiting = NULL;
2792 /* pages with index greater or equal to index will be truncated. */
2793 index = cl_index(osc2cl(obj), size);
2794 partial = size > cl_offset(osc2cl(obj), index);
2797 osc_object_lock(obj);
2798 ext = osc_extent_search(obj, index);
2800 ext = first_extent(obj);
2801 else if (ext->oe_end < index)
2802 ext = next_extent(ext);
2803 while (ext != NULL) {
2804 EASSERT(ext->oe_state != OES_TRUNC, ext);
2806 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2807 /* if ext is in urgent state, it means there must exist
2808 * a page already having been flushed by write_page().
2809 * We have to wait for this extent because we can't
2810 * truncate that page. */
2811 OSC_EXTENT_DUMP(D_CACHE, ext,
2812 "waiting for busy extent\n");
2813 waiting = osc_extent_get(ext);
2817 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2819 osc_extent_get(ext);
2820 if (ext->oe_state == OES_ACTIVE) {
2821 /* though we grab inode mutex for write path, but we
2822 * release it before releasing extent(in osc_io_end()),
2823 * so there is a race window that an extent is still
2824 * in OES_ACTIVE when truncate starts. */
2825 LASSERT(!ext->oe_trunc_pending);
2826 ext->oe_trunc_pending = 1;
2828 EASSERT(ext->oe_state == OES_CACHE, ext);
2829 osc_extent_state_set(ext, OES_TRUNC);
2830 osc_update_pending(obj, OBD_BRW_WRITE,
2833 /* This extent could be on the full extents list, that's OK */
2834 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2835 if (!list_empty(&ext->oe_link))
2836 list_move_tail(&ext->oe_link, &list);
2838 list_add_tail(&ext->oe_link, &list);
2840 ext = next_extent(ext);
2842 osc_object_unlock(obj);
2844 osc_list_maint(cli, obj);
2846 while (!list_empty(&list)) {
2849 ext = list_entry(list.next, struct osc_extent, oe_link);
2850 list_del_init(&ext->oe_link);
2852 /* extent may be in OES_ACTIVE state because inode mutex
2853 * is released before osc_io_end() in file write case */
2854 if (ext->oe_state != OES_TRUNC)
2855 osc_extent_wait(env, ext, OES_TRUNC);
2857 rc = osc_extent_truncate(ext, index, partial);
2862 OSC_EXTENT_DUMP(D_ERROR, ext,
2863 "truncate error %d\n", rc);
2864 } else if (ext->oe_nr_pages == 0) {
2865 osc_extent_remove(ext);
2867 /* this must be an overlapped extent which means only
2868 * part of pages in this extent have been truncated.
2870 EASSERTF(ext->oe_start <= index, ext,
2871 "trunc index = %lu/%d.\n", index, partial);
2872 /* fix index to skip this partially truncated extent */
2873 index = ext->oe_end + 1;
2876 /* we need to hold this extent in OES_TRUNC state so
2877 * that no writeback will happen. This is to avoid
2879 * Only partial truncate can reach here, if @size is
2880 * not zero, the caller should provide a valid @extp. */
2881 LASSERT(*extp == NULL);
2882 *extp = osc_extent_get(ext);
2883 OSC_EXTENT_DUMP(D_CACHE, ext,
2884 "trunc at %llu\n", size);
2886 osc_extent_put(env, ext);
2888 if (waiting != NULL) {
2891 /* ignore the result of osc_extent_wait the write initiator
2892 * should take care of it. */
2893 rc = osc_extent_wait(env, waiting, OES_INV);
2895 OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2897 osc_extent_put(env, waiting);
2903 EXPORT_SYMBOL(osc_cache_truncate_start);
2906 * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2908 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2911 struct osc_object *obj = ext->oe_obj;
2912 bool unplug = false;
2914 EASSERT(ext->oe_nr_pages > 0, ext);
2915 EASSERT(ext->oe_state == OES_TRUNC, ext);
2916 EASSERT(!ext->oe_urgent, ext);
2918 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2919 osc_object_lock(obj);
2920 osc_extent_state_set(ext, OES_CACHE);
2921 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2923 list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2926 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2927 osc_object_unlock(obj);
2928 osc_extent_put(env, ext);
2931 osc_io_unplug_async(env, osc_cli(obj), obj);
2936 * Wait for extents in a specific range to be written out.
2937 * The caller must have called osc_cache_writeback_range() to issue IO
2938 * otherwise it will take a long time for this function to finish.
2940 * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2941 * nobody else can dirty this range of file while we're waiting for
2942 * extents to be written.
2944 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2945 pgoff_t start, pgoff_t end)
2947 struct osc_extent *ext;
2948 pgoff_t index = start;
2953 osc_object_lock(obj);
2954 ext = osc_extent_search(obj, index);
2956 ext = first_extent(obj);
2957 else if (ext->oe_end < index)
2958 ext = next_extent(ext);
2959 while (ext != NULL) {
2962 if (ext->oe_start > end)
2965 if (!ext->oe_fsync_wait) {
2966 ext = next_extent(ext);
2970 EASSERT(ergo(ext->oe_state == OES_CACHE,
2971 ext->oe_hp || ext->oe_urgent), ext);
2972 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2973 !ext->oe_hp && ext->oe_urgent), ext);
2975 index = ext->oe_end + 1;
2976 osc_extent_get(ext);
2977 osc_object_unlock(obj);
2979 rc = osc_extent_wait(env, ext, OES_INV);
2982 osc_extent_put(env, ext);
2985 osc_object_unlock(obj);
2987 OSC_IO_DEBUG(obj, "sync file range.\n");
2990 EXPORT_SYMBOL(osc_cache_wait_range);
2993 * Called to write out a range of osc object.
2995 * @hp : should be set this is caused by lock cancel;
2996 * @discard: is set if dirty pages should be dropped - file will be deleted or
2997 * truncated, this implies there is no partially discarding extents.
2999 * Return how many pages will be issued, or error code if error occurred.
3001 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3002 pgoff_t start, pgoff_t end, int hp, int discard)
3004 struct osc_extent *ext;
3005 LIST_HEAD(discard_list);
3006 bool unplug = false;
3010 osc_object_lock(obj);
3011 ext = osc_extent_search(obj, start);
3013 ext = first_extent(obj);
3014 else if (ext->oe_end < start)
3015 ext = next_extent(ext);
3016 while (ext != NULL) {
3017 if (ext->oe_start > end)
3020 ext->oe_fsync_wait = 1;
3021 switch (ext->oe_state) {
3023 result += ext->oe_nr_pages;
3025 struct list_head *list = NULL;
3027 EASSERT(!ext->oe_hp, ext);
3029 list = &obj->oo_hp_exts;
3030 } else if (!ext->oe_urgent) {
3032 list = &obj->oo_urgent_exts;
3035 list_move_tail(&ext->oe_link, list);
3038 struct client_obd *cli = osc_cli(obj);
3039 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
3040 pgoff_t align_by = (1 << pcc_bits);
3041 pgoff_t a_start = round_down(start, align_by);
3042 pgoff_t a_end = round_up(end, align_by);
3046 a_end = CL_PAGE_EOF;
3047 /* the only discarder is lock cancelling, so
3048 * [start, end], aligned by chunk size, must
3049 * contain this extent */
3050 LASSERTF(ext->oe_start >= a_start &&
3051 ext->oe_end <= a_end,
3052 "ext [%lu, %lu] reg [%lu, %lu] "
3053 "orig [%lu %lu] align %lu bits "
3054 "%d\n", ext->oe_start, ext->oe_end,
3055 a_start, a_end, start, end,
3056 align_by, pcc_bits);
3057 osc_extent_state_set(ext, OES_LOCKING);
3058 ext->oe_owner = current;
3059 list_move_tail(&ext->oe_link,
3061 osc_update_pending(obj, OBD_BRW_WRITE,
3066 /* It's pretty bad to wait for ACTIVE extents, because
3067 * we don't know how long we will wait for it to be
3068 * flushed since it may be blocked at awaiting more
3069 * grants. We do this for the correctness of fsync. */
3070 LASSERT(hp == 0 && discard == 0);
3074 /* this extent is being truncated, can't do anything
3075 * for it now. it will be set to urgent after truncate
3076 * is finished in osc_cache_truncate_end(). */
3080 ext = next_extent(ext);
3082 osc_object_unlock(obj);
3084 LASSERT(ergo(!discard, list_empty(&discard_list)));
3085 if (!list_empty(&discard_list)) {
3086 struct osc_extent *tmp;
3089 osc_list_maint(osc_cli(obj), obj);
3090 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3091 list_del_init(&ext->oe_link);
3092 EASSERT(ext->oe_state == OES_LOCKING, ext);
3094 /* Discard caching pages. We don't actually write this
3095 * extent out but we complete it as if we did. */
3096 rc = osc_extent_make_ready(env, ext);
3097 if (unlikely(rc < 0)) {
3098 OSC_EXTENT_DUMP(D_ERROR, ext,
3099 "make_ready returned %d\n", rc);
3104 /* finish the extent as if the pages were sent */
3105 osc_extent_finish(env, ext, 0, 0);
3110 osc_io_unplug(env, osc_cli(obj), obj);
3112 if (hp || discard) {
3114 rc = osc_cache_wait_range(env, obj, start, end);
3115 if (result >= 0 && rc < 0)
3119 OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3122 EXPORT_SYMBOL(osc_cache_writeback_range);
3125 * Returns a list of pages by a given [start, end] of \a obj.
3127 * \param resched If not NULL, then we give up before hogging CPU for too
3128 * long and set *resched = 1, in that case caller should implement a retry
3131 * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3132 * crucial in the face of [offset, EOF] locks.
3134 * Return at least one page in @queue unless there is no covered page.
3136 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3137 struct osc_object *osc, pgoff_t start, pgoff_t end,
3138 osc_page_gang_cbt cb, void *cbdata)
3140 struct osc_page *ops;
3141 struct pagevec *pagevec;
3147 int res = CLP_GANG_OKAY;
3148 bool tree_lock = true;
3152 pvec = osc_env_info(env)->oti_pvec;
3153 pagevec = &osc_env_info(env)->oti_pagevec;
3154 ll_pagevec_init(pagevec, 0);
3155 spin_lock(&osc->oo_tree_lock);
3156 while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3157 idx, OTI_PVEC_SIZE)) > 0) {
3158 struct cl_page *page;
3159 bool end_of_region = false;
3161 for (i = 0, j = 0; i < nr; ++i) {
3165 idx = osc_index(ops);
3167 end_of_region = true;
3171 page = ops->ops_cl.cpl_page;
3172 LASSERT(page->cp_type == CPT_CACHEABLE);
3173 if (page->cp_state == CPS_FREEING)
3177 lu_ref_add_atomic(&page->cp_reference,
3178 "gang_lookup", current);
3184 * Here a delicate locking dance is performed. Current thread
3185 * holds a reference to a page, but has to own it before it
3186 * can be placed into queue. Owning implies waiting, so
3187 * radix-tree lock is to be released. After a wait one has to
3188 * check that pages weren't truncated (cl_page_own() returns
3189 * error in the latter case).
3191 spin_unlock(&osc->oo_tree_lock);
3194 for (i = 0; i < j; ++i) {
3196 if (res == CLP_GANG_OKAY)
3197 res = (*cb)(env, io, ops, cbdata);
3199 page = ops->ops_cl.cpl_page;
3200 lu_ref_del(&page->cp_reference, "gang_lookup", current);
3201 cl_pagevec_put(env, page, pagevec);
3203 pagevec_release(pagevec);
3205 if (nr < OTI_PVEC_SIZE || end_of_region)
3208 if (res == CLP_GANG_OKAY && need_resched())
3209 res = CLP_GANG_RESCHED;
3210 if (res != CLP_GANG_OKAY)
3213 spin_lock(&osc->oo_tree_lock);
3217 spin_unlock(&osc->oo_tree_lock);
3220 EXPORT_SYMBOL(osc_page_gang_lookup);
3223 * Check if page @page is covered by an extra lock or discard it.
3225 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3226 struct osc_page *ops, void *cbdata)
3228 struct osc_thread_info *info = osc_env_info(env);
3229 struct osc_object *osc = cbdata;
3232 index = osc_index(ops);
3233 if (index >= info->oti_fn_index) {
3234 struct ldlm_lock *tmp;
3235 struct cl_page *page = ops->ops_cl.cpl_page;
3237 /* refresh non-overlapped index */
3238 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3239 OSC_DAP_FL_TEST_LOCK);
3241 __u64 end = tmp->l_policy_data.l_extent.end;
3242 /* Cache the first-non-overlapped index so as to skip
3243 * all pages within [index, oti_fn_index). This is safe
3244 * because if tmp lock is canceled, it will discard
3246 info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3247 if (end == OBD_OBJECT_EOF)
3248 info->oti_fn_index = CL_PAGE_EOF;
3250 } else if (cl_page_own(env, io, page) == 0) {
3251 /* discard the page */
3252 cl_page_discard(env, io, page);
3253 cl_page_disown(env, io, page);
3255 LASSERT(page->cp_state == CPS_FREEING);
3259 info->oti_next_index = index + 1;
3260 return CLP_GANG_OKAY;
3263 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3264 struct osc_page *ops, void *cbdata)
3266 struct osc_thread_info *info = osc_env_info(env);
3267 struct cl_page *page = ops->ops_cl.cpl_page;
3269 /* page is top page. */
3270 info->oti_next_index = osc_index(ops) + 1;
3271 if (cl_page_own(env, io, page) == 0) {
3272 if (!ergo(page->cp_type == CPT_CACHEABLE,
3273 !PageDirty(cl_page_vmpage(page))))
3274 CL_PAGE_DEBUG(D_ERROR, env, page,
3275 "discard dirty page?\n");
3277 /* discard the page */
3278 cl_page_discard(env, io, page);
3279 cl_page_disown(env, io, page);
3281 LASSERT(page->cp_state == CPS_FREEING);
3284 return CLP_GANG_OKAY;
3286 EXPORT_SYMBOL(osc_discard_cb);
3289 * Discard pages protected by the given lock. This function traverses radix
3290 * tree to find all covering pages and discard them. If a page is being covered
3291 * by other locks, it should remain in cache.
3293 * If error happens on any step, the process continues anyway (the reasoning
3294 * behind this being that lock cancellation cannot be delayed indefinitely).
3296 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3297 pgoff_t start, pgoff_t end, bool discard)
3299 struct osc_thread_info *info = osc_env_info(env);
3300 struct cl_io *io = osc_env_thread_io(env);
3301 osc_page_gang_cbt cb;
3307 io->ci_obj = cl_object_top(osc2cl(osc));
3308 io->ci_ignore_layout = 1;
3309 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3313 cb = discard ? osc_discard_cb : check_and_discard_cb;
3314 info->oti_fn_index = info->oti_next_index = start;
3316 res = osc_page_gang_lookup(env, io, osc,
3317 info->oti_next_index, end, cb, osc);
3318 if (info->oti_next_index > end)
3321 if (res == CLP_GANG_RESCHED)
3323 } while (res != CLP_GANG_OKAY);
3325 cl_io_fini(env, io);